| 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 BatchSpec, create_common_attn_metadata |
| 11 | from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant |
| 12 | from vllm.compilation.passes.fusion.matcher_utils import QUANT_OPS |
| 13 | from vllm.compilation.passes.fusion.mla_attn_quant_fusion import ( |
| 14 | MLA_ATTN_OP, |
| 15 | MLAAttnQuantFusionPass, |
| 16 | ) |
| 17 | from vllm.compilation.passes.fx_utils import find_op_nodes |
| 18 | from vllm.compilation.passes.utility.noop_elimination import NoOpEliminationPass |
| 19 | from vllm.compilation.passes.utility.post_cleanup import PostCleanupPass |
| 20 | from vllm.config import ( |
| 21 | AttentionConfig, |
| 22 | CacheConfig, |
| 23 | CompilationConfig, |
| 24 | CompilationMode, |
| 25 | ModelConfig, |
| 26 | PassConfig, |
| 27 | SchedulerConfig, |
| 28 | VllmConfig, |
| 29 | set_current_vllm_config, |
| 30 | ) |
| 31 | from vllm.forward_context import get_forward_context, set_forward_context |
| 32 | from vllm.model_executor.kernels.linear.scaled_mm.cutlass import ( |
| 33 | CutlassFp8BlockScaledMMKernel, |
| 34 | ) |
| 35 | from vllm.model_executor.layers.attention import MLAAttention |
| 36 | from vllm.model_executor.layers.linear import ColumnParallelLinear |
| 37 | from vllm.model_executor.layers.quantization.fp8 import Fp8Config |
| 38 | from vllm.model_executor.layers.quantization.modelopt import ModelOptNvFp4Config |
| 39 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 40 | GroupShape, |
| 41 | QuantKey, |
| 42 | create_fp8_quant_key, |
| 43 | kFp8Dynamic128Sym, |
| 44 | kFp8StaticTensorSym, |
| 45 | kNvfp4Dynamic, |
| 46 | ) |
| 47 | from vllm.platforms import current_platform |
| 48 | from vllm.v1.attention.backend import AttentionMetadata |
| 49 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 50 | from vllm.v1.kv_cache_interface import MLAAttentionSpec |
| 51 | |
| 52 | FP8_DTYPE = current_platform.fp8_dtype() |
| 53 | FP4_DTYPE = torch.uint8 |
| 54 | DEVICE_TYPE = current_platform.device_type |
| 55 | |
| 56 | |
| 57 | class MLAAttentionQuantPatternModel(torch.nn.Module): |
| 58 | """Base model for MLA AttentionQuantPattern fusion.""" |
| 59 | |
| 60 | def __init__( |
| 61 | self, |
| 62 | num_heads: int, |
| 63 | qk_nope_head_dim: int, |
| 64 | qk_rope_head_dim: int, |
| 65 | v_head_dim: int, |
| 66 | kv_lora_rank: int, |
| 67 | kv_cache_dtype: torch.dtype, |
| 68 | device: torch.device, |
| 69 | vllm_config: VllmConfig, |
| 70 | **kwargs, |
| 71 | ): |
| 72 | super().__init__() |
| 73 | self.num_heads = num_heads |
| 74 | self.qk_nope_head_dim = qk_nope_head_dim |
| 75 | self.qk_rope_head_dim = qk_rope_head_dim |
| 76 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 77 | self.v_head_dim = v_head_dim |
| 78 | self.kv_lora_rank = kv_lora_rank |
| 79 | self.output_dim = num_heads * v_head_dim |
| 80 | self.head_size = kv_lora_rank + qk_rope_head_dim |
| 81 | self.kv_cache_dtype = kv_cache_dtype |
| 82 | self.device = device |
| 83 | self.vllm_config = vllm_config |
| 84 | self.dtype = vllm_config.model_config.dtype |
| 85 | |
| 86 | kv_b_proj = ColumnParallelLinear( |
| 87 | input_size=kv_lora_rank, |
| 88 | output_size=num_heads * (qk_nope_head_dim + v_head_dim), |
| 89 | bias=False, |
| 90 | prefix="model.layers.0.self_attn.kv_b_proj", |
| 91 | ).to(device) |
| 92 | kv_b_proj_weight = kwargs.get("kv_b_proj_weight") |
| 93 | if kv_b_proj_weight is not None: |
| 94 | kv_b_proj.weight.data.copy_(kv_b_proj_weight) |
| 95 | else: |
| 96 | kv_b_proj.weight.data.normal_() |
| 97 | |
| 98 | # Create MLAAttention |
| 99 | self.mla_attn = MLAAttention( |
| 100 | num_heads=num_heads, |
| 101 | scale=1.0 / (self.qk_head_dim**0.5), |
| 102 | qk_nope_head_dim=qk_nope_head_dim, |
| 103 | qk_rope_head_dim=qk_rope_head_dim, |
| 104 | v_head_dim=v_head_dim, |
| 105 | q_lora_rank=None, |
| 106 | kv_lora_rank=kv_lora_rank, |
| 107 | kv_b_proj=kv_b_proj, |
| 108 | cache_config=vllm_config.cache_config, |
| 109 | quant_config=self.quant_config, |
| 110 | prefix="model.layers.0.self_attn.attn", |
| 111 | ) |
| 112 | self.mla_attn._k_scale = self.mla_attn._k_scale.to(device) |
| 113 | self.mla_attn._v_scale = self.mla_attn._v_scale.to(device) |
| 114 | |
| 115 | # Initialize W_UK_T and W_UV from kv_b_proj weights |
| 116 | self.mla_attn.process_weights_after_loading(torch.get_default_dtype()) |
| 117 | self.kv_b_proj_weight = kv_b_proj.weight.data.clone() |
| 118 | |
| 119 | self.block_size = 16 |
| 120 | |
| 121 | # Initialize MLA MetadataBuilder |
| 122 | self.builder = self.mla_attn.attn_backend.get_builder_cls()( |
| 123 | kv_cache_spec=MLAAttentionSpec( |
| 124 | block_size=self.block_size, |
| 125 | num_kv_heads=1, |
| 126 | head_size=self.head_size, |
| 127 | dtype=self.kv_cache_dtype, |
| 128 | ), |
| 129 | layer_names=[self.mla_attn.layer_name], |
| 130 | vllm_config=self.vllm_config, |
| 131 | device=self.device, |
| 132 | ) |
| 133 | |
| 134 | def build_attn_metadata(self, batch_size: int) -> AttentionMetadata: |
| 135 | """Initialize MLA attention metadata. |
| 136 | |
| 137 | NOTE: Uses decode-only batch (query_len=1 per request). The prefill |
| 138 | (forward_mha) path is not separately tested here because it requires |
| 139 | FlashAttention availability and different input tensor shapes. The |
| 140 | quant logic in forward_impl is identical for both paths — it quantizes |
| 141 | the full output[:num_actual_toks] buffer after both forward_mha and |
| 142 | forward_mqa have written their results. |
| 143 | """ |
| 144 | batch_spec = BatchSpec(seq_lens=[1] * batch_size, query_lens=[1] * batch_size) |
| 145 | common_attn_metadata = create_common_attn_metadata( |
| 146 | batch_spec, self.block_size, self.device, arange_block_indices=True |
| 147 | ) |
| 148 | |
| 149 | max_blocks = (max(batch_spec.seq_lens) + self.block_size - 1) // self.block_size |
| 150 | num_blocks = batch_size * max_blocks |
| 151 | |
| 152 | # MLA KV cache is 4D: (num_blocks, num_heads=1, block_size, head_size) |
| 153 | kv_cache = torch.zeros( |
| 154 | (num_blocks, 1, self.block_size, self.head_size), |
| 155 | dtype=self.kv_cache_dtype, |
| 156 | device=self.device, |
| 157 | ) |
| 158 | |
| 159 | self.mla_attn.bind_kv_cache(kv_cache) |
| 160 | |
| 161 | self.attn_metadata = self.builder.build( |
| 162 | common_prefix_len=0, common_attn_metadata=common_attn_metadata |
| 163 | ) |
| 164 | |
| 165 | return self.attn_metadata |
| 166 | |
| 167 | |
| 168 | class TestMLAAttentionFp8StaticQuantPatternModel(MLAAttentionQuantPatternModel): |
| 169 | """Test model for MLA Attention + FP8 static quant fusion.""" |
| 170 | |
| 171 | quant_key = kFp8StaticTensorSym |
| 172 | quant_config = Fp8Config() |
| 173 | |
| 174 | def __init__(self, *args, **kwargs): |
| 175 | super().__init__(*args, **kwargs) |
| 176 | |
| 177 | self.fp8_linear = TestFP8Layer( |
| 178 | weight_shape=(self.output_dim, self.output_dim), |
| 179 | activation_quant_key=self.quant_key, |
| 180 | weight_quant_key=self.quant_key, |
| 181 | device=self.device, |
| 182 | input_dtype=self.dtype, |
| 183 | ) |
| 184 | |
| 185 | w = kwargs.get("w") |
| 186 | if w is not None: |
| 187 | self.fp8_linear.weight = w["weight"] |
| 188 | self.fp8_linear.weight_scale = w["wscale"] |
| 189 | self.fp8_linear.input_scale = w["scale"] |
| 190 | |
| 191 | self.w = { |
| 192 | "weight": self.fp8_linear.weight, |
| 193 | "wscale": self.fp8_linear.weight_scale, |
| 194 | "scale": self.fp8_linear.input_scale, |
| 195 | } |
| 196 | |
| 197 | def forward( |
| 198 | self, |
| 199 | q: torch.Tensor, |
| 200 | kv_c_normed: torch.Tensor, |
| 201 | k_pe: torch.Tensor, |
| 202 | ): |
| 203 | """Forward pass that creates the MLA attention + FP8 quant pattern.""" |
| 204 | attn_output = self.mla_attn( |
| 205 | q, |
| 206 | kv_c_normed, |
| 207 | k_pe, |
| 208 | output_shape=(q.shape[0], self.output_dim), |
| 209 | ) |
| 210 | return self.fp8_linear(attn_output) |
| 211 | |
| 212 | |
| 213 | class TestMLAAttentionNvfp4QuantPatternModel(MLAAttentionQuantPatternModel): |
| 214 | """Test model for MLA Attention + NVFP4 quant fusion.""" |
| 215 | |
| 216 | quant_key = kNvfp4Dynamic |
| 217 | quant_config = ModelOptNvFp4Config( |
| 218 | is_checkpoint_nvfp4_serialized=False, |
| 219 | kv_cache_quant_algo=None, |
| 220 | exclude_modules=[], |
| 221 | ) |
| 222 | |
| 223 | def __init__(self, *args, **kwargs): |
| 224 | super().__init__(*args, **kwargs) |
| 225 | |
| 226 | self.w = kwargs.get( |
| 227 | "w", |
| 228 | { |
| 229 | "weight": torch.randint( |
| 230 | 256, |
| 231 | (self.output_dim, self.output_dim // 2), |
| 232 | dtype=FP4_DTYPE, |
| 233 | device=self.device, |
| 234 | ), |
| 235 | "wscale_swizzled": torch.randn( |
| 236 | self.output_dim, self.output_dim // 16 |
| 237 | ).to(dtype=FP8_DTYPE, device=self.device), |
| 238 | "wscale": torch.tensor([500], dtype=torch.float32, device=self.device), |
| 239 | "scale": torch.tensor([0.002], dtype=torch.float32, device=self.device), |
| 240 | }, |
| 241 | ) |
| 242 | |
| 243 | def forward( |
| 244 | self, |
| 245 | q: torch.Tensor, |
| 246 | kv_c_normed: torch.Tensor, |
| 247 | k_pe: torch.Tensor, |
| 248 | ): |
| 249 | """Forward pass that creates the MLA attention + NVFP4 quant pattern.""" |
| 250 | attn_output = self.mla_attn( |
| 251 | q, |
| 252 | kv_c_normed, |
| 253 | k_pe, |
| 254 | output_shape=(q.shape[0], self.output_dim), |
| 255 | ) |
| 256 | quant_output, output_block_scale = scaled_fp4_quant( |
| 257 | attn_output, 1 / self.w["scale"] |
| 258 | ) |
| 259 | return cutlass_scaled_fp4_mm( |
| 260 | a=quant_output, |
| 261 | b=self.w["weight"], |
| 262 | block_scale_a=output_block_scale, |
| 263 | block_scale_b=self.w["wscale_swizzled"], |
| 264 | alpha=self.w["scale"] * self.w["wscale"], |
| 265 | out_dtype=attn_output.dtype, |
| 266 | ) |
| 267 | |
| 268 | |
| 269 | class TestMLAAttentionFp8GroupQuantPatternModel(MLAAttentionQuantPatternModel): |
| 270 | """Test model for MLA Attention + per-group FP8 (block quant) fusion.""" |
| 271 | |
| 272 | quant_key = kFp8Dynamic128Sym |
| 273 | quant_config = Fp8Config( |
| 274 | weight_block_size=[128, 128], |
| 275 | ) |
| 276 | |
| 277 | def __init__(self, *args, **kwargs): |
| 278 | super().__init__(*args, **kwargs) |
| 279 | |
| 280 | weight_quant_key = create_fp8_quant_key( |
| 281 | static=True, group_shape=GroupShape(128, 128) |
| 282 | ) |
| 283 | device = kwargs.get("device", torch.device("cuda:0")) |
| 284 | |
| 285 | # Subclass to set weight_block_size before process_weights_after_loading |
| 286 | class _BlockFP8Layer(TestFP8Layer): |
| 287 | def __init__(self, *a, **kw): |
| 288 | self.weight_block_size = [128, 128] |
| 289 | super().__init__(*a, **kw) |
| 290 | |
| 291 | # Force CutlassFp8BlockScaledMMKernel to ensure the graph uses |
| 292 | # per_token_group_fp8_quant (not the deepgemm packed variant). |
| 293 | self.block_fp8_linear = _BlockFP8Layer( |
| 294 | weight_shape=(self.output_dim, self.output_dim), |
| 295 | activation_quant_key=self.quant_key, |
| 296 | weight_quant_key=weight_quant_key, |
| 297 | input_dtype=self.dtype, |
| 298 | device=device, |
| 299 | force_kernel=CutlassFp8BlockScaledMMKernel, |
| 300 | ) |
| 301 | |
| 302 | w = kwargs.get("w") |
| 303 | if w is not None: |
| 304 | self.block_fp8_linear.weight = w["weight"] |
| 305 | # Block-wise uses weight_scale_inv, not weight_scale |
| 306 | self.block_fp8_linear.weight_scale_inv = w["wscale"] |
| 307 | |
| 308 | self.w = { |
| 309 | "weight": self.block_fp8_linear.weight, |
| 310 | "wscale": self.block_fp8_linear.weight_scale_inv, |
| 311 | } |
| 312 | |
| 313 | def forward( |
| 314 | self, |
| 315 | q: torch.Tensor, |
| 316 | kv_c_normed: torch.Tensor, |
| 317 | k_pe: torch.Tensor, |
| 318 | ): |
| 319 | """Forward pass: MLA attention -> block FP8 linear (group quant).""" |
| 320 | attn_output = self.mla_attn( |
| 321 | q, |
| 322 | kv_c_normed, |
| 323 | k_pe, |
| 324 | output_shape=(q.shape[0], self.output_dim), |
| 325 | ) |
| 326 | return self.block_fp8_linear(attn_output) |
| 327 | |
| 328 | |
| 329 | def is_nvfp4_supported(): |
| 330 | return current_platform.has_device_capability(100) |
| 331 | |
| 332 | |
| 333 | # MLA test configuration |
| 334 | MLA_DIMS: list[tuple[int, int, int, int, int]] = [] |
| 335 | PATTERN_TEST_MODELS_MLA_FP8: list[tuple[str, type]] = [] |
| 336 | PATTERN_TEST_MODELS_MLA_GROUP_FP8: list[tuple[str, type]] = [] |
| 337 | PATTERN_TEST_MODELS_MLA_FP4: list[tuple[str, type]] = [] |
| 338 | BACKENDS_MLA_FP8: list[AttentionBackendEnum] = [] |
| 339 | BACKENDS_MLA_FP4: list[AttentionBackendEnum] = [] |
| 340 | |
| 341 | if current_platform.is_cuda(): |
| 342 | # (num_heads, qk_nope_head_dim, qk_rope_head_dim, v_head_dim, kv_lora_rank) |
| 343 | MLA_DIMS = [(16, 128, 64, 128, 512)] |
| 344 | PATTERN_TEST_MODELS_MLA_FP8 = [ |
| 345 | ( |
| 346 | "deepseek-ai/DeepSeek-V2-Lite", |
| 347 | TestMLAAttentionFp8StaticQuantPatternModel, |
| 348 | ) |
| 349 | ] |
| 350 | PATTERN_TEST_MODELS_MLA_GROUP_FP8 = [ |
| 351 | ( |
| 352 | "deepseek-ai/DeepSeek-V3", |
| 353 | TestMLAAttentionFp8GroupQuantPatternModel, |
| 354 | ) |
| 355 | ] |
| 356 | PATTERN_TEST_MODELS_MLA_FP4 = [ |
| 357 | ( |
| 358 | "deepseek-ai/DeepSeek-V2-Lite", |
| 359 | TestMLAAttentionNvfp4QuantPatternModel, |
| 360 | ) |
| 361 | ] |
| 362 | BACKENDS_MLA_FP8 = [AttentionBackendEnum.TRITON_MLA] |
| 363 | BACKENDS_MLA_FP4 = [AttentionBackendEnum.TRITON_MLA] |
| 364 | elif current_platform.is_rocm(): |
| 365 | # ROCm supports the static-FP8 output-quant fusion through AITER MLA. |
| 366 | # Per-group FP8 still relies on the CUDA-only CUTLASS block-scaled |
| 367 | # kernel, and NVFP4 is NVIDIA-specific. |
| 368 | MLA_DIMS = [(16, 128, 64, 128, 512)] |
| 369 | PATTERN_TEST_MODELS_MLA_FP8 = [ |
| 370 | ( |
| 371 | "deepseek-ai/DeepSeek-V2-Lite", |
| 372 | TestMLAAttentionFp8StaticQuantPatternModel, |
| 373 | ) |
| 374 | ] |
| 375 | BACKENDS_MLA_FP8 = [AttentionBackendEnum.ROCM_AITER_MLA] |
| 376 | |
| 377 | |
| 378 | @pytest.mark.parametrize( |
| 379 | "num_heads, qk_nope_head_dim, qk_rope_head_dim, v_head_dim, kv_lora_rank", |
| 380 | MLA_DIMS, |
| 381 | ) |
| 382 | @pytest.mark.parametrize( |
| 383 | "batch_size", [7, 256] if current_platform.is_cuda_alike() else [8] |
| 384 | ) |
| 385 | @pytest.mark.parametrize("dtype", [torch.bfloat16]) |
| 386 | @pytest.mark.parametrize( |
| 387 | "backend, model_name, model_class, custom_ops", |
| 388 | list( |
| 389 | flat_product( |
| 390 | BACKENDS_MLA_FP8, |
| 391 | PATTERN_TEST_MODELS_MLA_FP8, |
| 392 | ["+quant_fp8", "-quant_fp8"], |
| 393 | ) |
| 394 | ) |
| 395 | + list( |
| 396 | flat_product( |
| 397 | BACKENDS_MLA_FP8, |
| 398 | PATTERN_TEST_MODELS_MLA_GROUP_FP8, |
| 399 | ["+quant_fp8"], |
| 400 | ) |
| 401 | ) |
| 402 | + list(flat_product(BACKENDS_MLA_FP4, PATTERN_TEST_MODELS_MLA_FP4, [""])), |
| 403 | ) |
| 404 | @pytest.mark.skipif( |
| 405 | not current_platform.is_cuda_alike(), reason="Only test ROCm or CUDA" |
| 406 | ) |
| 407 | @pytest.mark.skipif(not current_platform.supports_fp8(), reason="Need FP8") |
| 408 | def test_mla_attention_quant_pattern( |
| 409 | num_heads: int, |
| 410 | qk_nope_head_dim: int, |
| 411 | qk_rope_head_dim: int, |
| 412 | v_head_dim: int, |
| 413 | kv_lora_rank: int, |
| 414 | batch_size: int, |
| 415 | dtype: torch.dtype, |
| 416 | custom_ops: str, |
| 417 | model_name: str, |
| 418 | model_class: type[MLAAttentionQuantPatternModel], |
| 419 | backend: AttentionBackendEnum, |
| 420 | dist_init, |
| 421 | disable_vllm_compile_cache, |
| 422 | workspace_init, |
| 423 | ): |
| 424 | """Test MLA AttentionQuantPattern fusion pass.""" |
| 425 | if ( |
| 426 | model_class is TestMLAAttentionNvfp4QuantPatternModel |
| 427 | and not is_nvfp4_supported() |
| 428 | ): |
| 429 | pytest.skip("NVFP4 is not supported on this GPU (requires SM 100+).") |
| 430 | |
| 431 | custom_ops_list = custom_ops.split(",") if custom_ops else [] |
| 432 | |
| 433 | device = torch.device(f"{DEVICE_TYPE}:0") |
| 434 | torch.set_default_dtype(dtype) |
| 435 | torch.manual_seed(42) |
| 436 | |
| 437 | model_config = ModelConfig( |
| 438 | model=model_name, |
| 439 | max_model_len=2048, |
| 440 | dtype=dtype, |
| 441 | ) |
| 442 | vllm_config = VllmConfig( |
| 443 | model_config=model_config, |
| 444 | scheduler_config=SchedulerConfig( |
| 445 | max_num_seqs=1024, |
| 446 | max_model_len=model_config.max_model_len, |
| 447 | is_encoder_decoder=model_config.is_encoder_decoder, |
| 448 | ), |
| 449 | compilation_config=CompilationConfig( |
| 450 | mode=CompilationMode.VLLM_COMPILE, |
| 451 | custom_ops=custom_ops_list, |
| 452 | ), |
| 453 | cache_config=CacheConfig(cache_dtype="auto"), |
| 454 | attention_config=AttentionConfig(backend=backend), |
| 455 | ) |
| 456 | |
| 457 | # MLA inputs: q(B, N, qk_head_dim), kv_c_normed(B, L), k_pe(B, 1, R) |
| 458 | qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 459 | q = torch.randn(batch_size, num_heads, qk_head_dim, dtype=dtype, device=device) |
| 460 | kv_c_normed = torch.randn(batch_size, kv_lora_rank, dtype=dtype, device=device) |
| 461 | k_pe = torch.randn(batch_size, 1, qk_rope_head_dim, dtype=dtype, device=device) |
| 462 | |
| 463 | # Mark first dimension as dynamic |
| 464 | torch._dynamo.mark_dynamic(q, 0) |
| 465 | torch._dynamo.mark_dynamic(kv_c_normed, 0) |
| 466 | torch._dynamo.mark_dynamic(k_pe, 0) |
| 467 | |
| 468 | # Run model without fusion |
| 469 | vllm_config_unfused = copy.deepcopy(vllm_config) |
| 470 | with ( |
| 471 | set_current_vllm_config(vllm_config_unfused), |
| 472 | set_forward_context(attn_metadata=None, vllm_config=vllm_config_unfused), |
| 473 | ): |
| 474 | model_unfused = model_class( |
| 475 | num_heads=num_heads, |
| 476 | qk_nope_head_dim=qk_nope_head_dim, |
| 477 | qk_rope_head_dim=qk_rope_head_dim, |
| 478 | v_head_dim=v_head_dim, |
| 479 | kv_lora_rank=kv_lora_rank, |
| 480 | kv_cache_dtype=dtype, |
| 481 | device=device, |
| 482 | vllm_config=vllm_config_unfused, |
| 483 | ) |
| 484 | if ( |
| 485 | model_class is TestMLAAttentionFp8GroupQuantPatternModel |
| 486 | and type(model_unfused.block_fp8_linear.kernel) |
| 487 | is not CutlassFp8BlockScaledMMKernel |
| 488 | ): |
| 489 | pytest.skip("CUTLASS FP8 block kernel is not supported on this platform") |
| 490 | model_unfused = model_unfused.to(device) |
| 491 | # HACK: See #131044 |
| 492 | result_unfused_0 = model_unfused(q, kv_c_normed, k_pe) # noqa: F841 |
| 493 | |
| 494 | forward_ctx = get_forward_context() |
| 495 | forward_ctx.attn_metadata = model_unfused.build_attn_metadata(batch_size) |
| 496 | |
| 497 | compiled_unfused = torch.compile(model_unfused, fullgraph=True) |
| 498 | result_unfused = compiled_unfused(q, kv_c_normed, k_pe) |
| 499 | |
| 500 | # Run model with attn fusion enabled |
| 501 | vllm_config.compilation_config.pass_config = PassConfig( |
| 502 | fuse_attn_quant=True, eliminate_noops=True |
| 503 | ) |
| 504 | with ( |
| 505 | set_current_vllm_config(vllm_config), |
| 506 | set_forward_context(attn_metadata=None, vllm_config=vllm_config), |
| 507 | ): |
| 508 | model_fused = model_class( |
| 509 | num_heads=num_heads, |
| 510 | qk_nope_head_dim=qk_nope_head_dim, |
| 511 | qk_rope_head_dim=qk_rope_head_dim, |
| 512 | v_head_dim=v_head_dim, |
| 513 | kv_lora_rank=kv_lora_rank, |
| 514 | kv_cache_dtype=dtype, |
| 515 | device=device, |
| 516 | vllm_config=vllm_config, |
| 517 | w=model_unfused.w, |
| 518 | kv_b_proj_weight=model_unfused.kv_b_proj_weight, |
| 519 | ) |
| 520 | model_fused = model_fused.to(device) |
| 521 | |
| 522 | forward_ctx = get_forward_context() |
| 523 | forward_ctx.attn_metadata = model_fused.build_attn_metadata(batch_size) |
| 524 | |
| 525 | # Create test backend with fusion passes |
| 526 | noop_pass = NoOpEliminationPass(vllm_config) |
| 527 | attn_pass = LazyInitPass(MLAAttnQuantFusionPass, vllm_config) |
| 528 | cleanup_pass = PostCleanupPass(vllm_config) |
| 529 | |
| 530 | test_backend = TestBackend(noop_pass, attn_pass, cleanup_pass) |
| 531 | # HACK: See https://github.com/vllm-project/vllm/issues/31044 |
| 532 | result_fused_0 = model_fused(q, kv_c_normed, k_pe) # noqa: F841 |
| 533 | |
| 534 | compiled_fused = torch.compile( |
| 535 | model_fused, backend=test_backend, fullgraph=True |
| 536 | ) |
| 537 | |
| 538 | result_fused = compiled_fused(q, kv_c_normed, k_pe) |
| 539 | |
| 540 | # Check attn fusion support |
| 541 | quant_key: QuantKey = model_class.quant_key |
| 542 | attn_fusion_supported = [ |
| 543 | layer.impl.fused_output_quant_supported(quant_key) |
| 544 | for key, layer in vllm_config.compilation_config.static_forward_context.items() |
| 545 | if isinstance(layer, MLAAttention) |
| 546 | ] |
| 547 | assert sum(attn_fusion_supported) == len(attn_fusion_supported), ( |
| 548 | "All MLA layers should support attention fusion" |
| 549 | ) |
| 550 | |
| 551 | # Check quantization ops in the graph |
| 552 | is_per_group = quant_key.scale.group_shape.is_per_group() |
| 553 | quant_op = ( |
| 554 | torch.ops.aten.reciprocal |
| 555 | if "-quant_fp8" in custom_ops_list |
| 556 | else QUANT_OPS[quant_key] |
| 557 | ) |
| 558 | test_backend.check_before_ops([quant_op], fully_replaced=is_per_group) |
| 559 | |
| 560 | assert attn_pass.pass_.matched_count == sum(attn_fusion_supported) |
| 561 | |
| 562 | # Check MLA attention ops in the graph |
| 563 | attn_nodes_pre = list(find_op_nodes(MLA_ATTN_OP, test_backend.graph_pre_pass)) |
| 564 | attn_nodes_post = list(find_op_nodes(MLA_ATTN_OP, test_backend.graph_post_pass)) |
| 565 | |
| 566 | assert len(attn_nodes_pre) > 0, "Should have MLA attention nodes before fusion" |
| 567 | assert len(attn_nodes_pre) == len(attn_nodes_post), ( |
| 568 | "Should have same number of MLA attention nodes before and after fusion" |
| 569 | ) |
| 570 | |
| 571 | # Before fusion: neither scale should be set |
| 572 | assert attn_nodes_pre[0].kwargs.get("output_scale") is None |
| 573 | assert attn_nodes_pre[0].kwargs.get("output_block_scale") is None |
| 574 | |
| 575 | # After fusion: derive expected scale presence from quant_key properties. |
| 576 | # - output_scale: present for static quant or non-FP8 (NVFP4 carries input_scale) |
| 577 | # - output_block_scale: present when quant uses per-group/block scaling |
| 578 | has_output_scale = attn_nodes_post[0].kwargs.get("output_scale") is not None |
| 579 | has_block_scale = attn_nodes_post[0].kwargs.get("output_block_scale") is not None |
| 580 | |
| 581 | expects_output_scale = quant_key.scale.static or quant_key.dtype != FP8_DTYPE |
| 582 | assert has_output_scale == expects_output_scale, ( |
| 583 | f"output_scale: expected present={expects_output_scale}, got {has_output_scale}" |
| 584 | ) |
| 585 | assert has_block_scale == is_per_group, ( |
| 586 | f"output_block_scale: expected present={is_per_group}, got {has_block_scale}" |
| 587 | ) |
| 588 | |
| 589 | # Check numerical correctness |
| 590 | torch.testing.assert_close(result_unfused, result_fused, atol=1e-2, rtol=1e-2) |