[Perf] Fix and reapply move apply w8a8 block fp8 linear to class (vllm-project#25696)

Signed-off-by: ElizaWszola <[email protected]>
Signed-off-by: ElizaWszola <[email protected]>
Signed-off-by: Luka Govedič <[email protected]>
Signed-off-by: Luka Govedič <[email protected]>
Co-authored-by: Luka Govedič <[email protected]>
Co-authored-by: Michael Goin <[email protected]>
Co-authored-by: Luka Govedič <[email protected]>

Author: 3 people

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -17,7 +17,7 @@
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    w8a8_block_fp8_matmul,
+    w8a8_triton_block_scaled_mm,
 )
 from vllm.utils import FlexibleArgumentParser, cdiv
 
@@ -158,7 +158,7 @@ def bench_fp8(
         "cutlass_fp8_fp8_fp16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm(
             a, b, scale_a, scale_b, torch.float16, bias.to(dtype=torch.float16)
         ),
-        "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_block_fp8_matmul(
+        "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_triton_block_scaled_mm(
             a_cont, b.t(), block_scale_a, block_scale_b.t(), (128, 128)
         ),
         "cutlass_fp8_fp8_fp16_scaled_mm_blockwise": lambda: ops.cutlass_scaled_mm(

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -9,7 +9,7 @@
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
-    w8a8_block_fp8_matmul,
+    w8a8_triton_block_scaled_mm,
 )
 from vllm.triton_utils import triton
 from vllm.utils.deep_gemm import (
@@ -63,7 +63,7 @@ def deepgemm_gemm():
 
     # === vLLM Triton Implementation ===
     def vllm_triton_gemm():
-        return w8a8_block_fp8_matmul(A_vllm,
+        return w8a8_triton_block_scaled_mm(A_vllm,
                                      B_vllm,
                                      A_scale_vllm,
                                      B_scale_vllm,

tests/kernels/quantization/test_block_fp8.py

@@ -11,7 +11,7 @@
                                        native_w8a8_block_matmul)
 from vllm.config import VllmConfig
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    cutlass_scaled_mm, per_token_group_quant_fp8, w8a8_block_fp8_matmul)
+    cutlass_scaled_mm, per_token_group_quant_fp8, w8a8_triton_block_scaled_mm)
 from vllm.platforms import current_platform
 from vllm.utils import has_deep_gemm
 from vllm.utils.deep_gemm import (fp8_gemm_nt,
@@ -91,7 +91,8 @@ def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed):
 
     ref_out = native_w8a8_block_matmul(A_fp8, B_fp8, As, Bs, block_size,
                                        out_dtype)
-    out = w8a8_block_fp8_matmul(A_fp8, B_fp8, As, Bs, block_size, out_dtype)
+    out = w8a8_triton_block_scaled_mm(A_fp8, B_fp8, As, Bs, block_size,
+                                      out_dtype)
 
     rel_diff = (torch.mean(
         torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /

tests/kernels/quantization/test_fp8_quant_group.py

@@ -20,9 +20,11 @@
         (8, 513, 64),  # Non-divisible (native only)
     ])
 @pytest.mark.parametrize("seed", [42])
+@pytest.mark.parametrize("use_ue8m0", [True, False])
 @torch.inference_mode()
 def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
-                                      group_size: int, seed: int) -> None:
+                                      group_size: int, seed: int,
+                                      use_ue8m0: bool) -> None:
     """Test QuantFP8 group quantization with various configurations.
 
     Tests both CUDA and native implementations, column-major scales,
@@ -38,7 +40,8 @@ def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
     group_shape = GroupShape(1, group_size)
     quant_op = QuantFP8(static=False,
                         group_shape=group_shape,
-                        column_major_scales=False)
+                        column_major_scales=False,
+                        use_ue8m0=use_ue8m0)
 
     # 1. Test native implementation (always available)
     x_quant_native, scales_native = quant_op.forward_native(x.clone())
@@ -48,9 +51,15 @@ def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
     # 2. Test column-major scales configuration
     quant_op_col = QuantFP8(static=False,
                             group_shape=group_shape,
-                            column_major_scales=True)
+                            column_major_scales=True,
+                            use_ue8m0=use_ue8m0)
     _, scales_col = quant_op_col.forward_native(x.clone())
-    assert scales_col.shape == (expected_num_groups, batch_size)
+    assert scales_col.shape == (batch_size, expected_num_groups)
+    assert scales_col.stride(0) == 1
+    assert scales_col.stride(1) == batch_size
+
+    # Test column-major scales consistency
+    assert torch.allclose(scales_col, scales_native, rtol=1e-9, atol=1e-8)
 
     # 3. Test CUDA implementation (only for divisible dimensions)
     if is_divisible:
@@ -68,21 +77,23 @@ def test_quantfp8_group_functionality(batch_size: int, hidden_dim: int,
 
 
 @pytest.mark.parametrize("seed", [42])
+@pytest.mark.parametrize("use_ue8m0", [True, False])
 @torch.inference_mode()
-def test_quantfp8_group_multidimensional(seed: int) -> None:
+def test_quantfp8_group_multidimensional(seed: int, use_ue8m0: bool) -> None:
     current_platform.seed_everything(seed)
 
     group_size = 64
 
     # Test with 3D input
-    batch1, batch2, hidden_dim = 4, 8, 512
+    batch1, batch2, hidden_dim = 4, 8, 1024
     x_3d = torch.randn(
         (batch1, batch2, hidden_dim), dtype=torch.bfloat16, device="cuda") * 8
 
     group_shape = GroupShape(1, group_size)
     quant_op = QuantFP8(static=False,
                         group_shape=group_shape,
-                        column_major_scales=False)
+                        column_major_scales=False,
+                        use_ue8m0=use_ue8m0)
 
     x_quant, scales = quant_op.forward_native(x_3d.clone())
     assert x_quant.shape == x_3d.shape
@@ -91,9 +102,10 @@ def test_quantfp8_group_multidimensional(seed: int) -> None:
     # Test column_major_scales with multi-dim
     quant_op_col = QuantFP8(static=False,
                             group_shape=group_shape,
-                            column_major_scales=True)
+                            column_major_scales=True,
+                            use_ue8m0=use_ue8m0)
     _, scales_col = quant_op_col.forward_native(x_3d.clone())
-    assert scales_col.shape == (batch1, hidden_dim // group_size, batch2)
+    assert scales_col.shape == (batch1, batch2, hidden_dim // group_size)
 
     # Test with 4D input
     batch1, batch2, batch3, hidden_dim = 2, 3, 4, 256

tests/model_executor/test_enabled_custom_ops.py

@@ -17,8 +17,6 @@
 from vllm.model_executor.layers.layernorm import (RMSNorm,
                                                   dispatch_rocm_rmsnorm_func,
                                                   fused_add_rms_norm, rms_norm)
-from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    cutlass_scaled_mm, dispatch_w8a8_blockscale_func, w8a8_block_fp8_matmul)
 from vllm.platforms import current_platform
 
 RMS_NORM_SUPPORTED_DTYPES = [torch.float16, torch.bfloat16]
@@ -111,34 +109,6 @@ def test_enabled_ops_invalid(env: str):
             RMSNorm(1024).enabled()
 
 
-@pytest.mark.skipif(
-    not current_platform.is_rocm() or not current_platform.is_fp8_fnuz(),
-    reason="AITER is a feature exclusive for ROCm and FP8_FNUZ")
-@pytest.mark.parametrize("use_cutlass", [True, False])
-@pytest.mark.parametrize("use_rocm_aiter", ["0", "1"])
-@pytest.mark.parametrize("use_rocm_aiter_gemm_w8a8_blockscale", ["0", "1"])
-def test_w8a8_blockscale_dispatch(use_cutlass: bool, use_rocm_aiter: str,
-                                  use_rocm_aiter_gemm_w8a8_blockscale: str,
-                                  monkeypatch):
-
-    monkeypatch.setenv("VLLM_ROCM_USE_AITER", use_rocm_aiter)
-    monkeypatch.setenv("VLLM_ROCM_USE_AITER_LINEAR",
-                       use_rocm_aiter_gemm_w8a8_blockscale)
-
-    use_aiter_and_is_supported = (bool(int(use_rocm_aiter)) and bool(
-        int(use_rocm_aiter_gemm_w8a8_blockscale)))
-    block_scale_func = dispatch_w8a8_blockscale_func(
-        use_cutlass, use_aiter_and_is_supported=use_aiter_and_is_supported)
-    if use_cutlass:
-        assert block_scale_func == cutlass_scaled_mm
-    elif current_platform.is_rocm() and int(use_rocm_aiter) and int(
-            use_rocm_aiter_gemm_w8a8_blockscale):
-        assert block_scale_func == (
-            torch.ops.vllm.rocm_aiter_gemm_w8a8_blockscale)
-    else:
-        assert block_scale_func == w8a8_block_fp8_matmul
-
-
 @pytest.mark.parametrize("use_rocm_aiter", ["0", "1"])
 def test_topk_dispatch(use_rocm_aiter: str, monkeypatch):
     monkeypatch.setenv("VLLM_ROCM_USE_AITER", use_rocm_aiter)

tests/quantization/test_compressed_tensors.py

@@ -18,6 +18,9 @@
     CompressedTensorsW4A16Fp4, CompressedTensorsW4A16Sparse24,
     CompressedTensorsW8A8Fp8, CompressedTensorsW8A8Int8,
     CompressedTensorsW8A16Fp8, CompressedTensorsWNA16)
+from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    W8A8BlockFp8LinearOp)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     cutlass_fp4_supported)
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
@@ -742,3 +745,35 @@ def test_compressed_tensors_transforms_perplexity(vllm_runner, model, prompt,
         perplexity = llm.generate_prompt_perplexity([prompt])[0]
         print(perplexity)
         assert perplexity <= exp_perplexity
+
+
+def test_compressed_tensors_fp8_block_enabled(vllm_runner):
+    model_path = "RedHatAI/Qwen3-0.6B-FP8-BLOCK"
+    with vllm_runner(model_path) as llm:
+
+        fp8_dtype = current_platform.fp8_dtype()
+
+        def check_model(model):
+            layer = model.model.layers[0]
+
+            qkv_proj = layer.self_attn.qkv_proj
+            assert isinstance(qkv_proj.quant_method,
+                              CompressedTensorsLinearMethod)
+            assert isinstance(qkv_proj.scheme, CompressedTensorsW8A8Fp8)
+            assert isinstance(qkv_proj.scheme.w8a8_block_fp8_linear,
+                              W8A8BlockFp8LinearOp)
+
+            assert qkv_proj.weight.dtype is fp8_dtype
+            assert qkv_proj.weight_scale.dtype is torch.float32
+            assert len(qkv_proj.weight.shape) == 2
+            assert len(qkv_proj.weight_scale.shape) == 2
+
+            input_quant_op = \
+                qkv_proj.scheme.w8a8_block_fp8_linear.input_quant_op
+            assert isinstance(input_quant_op, QuantFP8)
+            assert input_quant_op._forward_method == input_quant_op.forward_cuda
+
+        llm.apply_model(check_model)
+
+        output = llm.generate_greedy("Hello my name is", max_tokens=20)
+        assert output

vllm/config/vllm.py

@@ -516,6 +516,23 @@ def __post_init__(self):
                     " by VLLM_DEBUG_DUMP_PATH to %s", env_path)
             self.compilation_config.debug_dump_path = env_path
 
+        def has_blocked_weights():
+            if self.quant_config is not None:
+                if hasattr(self.quant_config, "weight_block_size"):
+                    return self.quant_config.weight_block_size is not None
+                elif hasattr(self.quant_config, "has_blocked_weights"):
+                    return self.quant_config.has_blocked_weights()
+            return False
+
+        # Enable quant_fp8 CUDA ops (TODO disable in follow up)
+        # On H100 the CUDA kernel is faster than
+        # native implementation
+        # https://github.com/vllm-project/vllm/issues/25094
+        if has_blocked_weights():
+            custom_ops = self.compilation_config.custom_ops
+            if "none" not in custom_ops and "-quant_fp8" not in custom_ops:
+                custom_ops.append("+quant_fp8")
+
     def update_sizes_for_sequence_parallelism(self,
                                               possible_sizes: list) -> list:
         # remove the sizes that not multiple of tp_size when

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -644,6 +644,14 @@ def get_cache_scale(self, name: str) -> Optional[str]:
         # If no matches, return None
         return None
 
+    def has_blocked_weights(self) -> bool:
+        for scheme in self.target_scheme_map.values():
+            weight_quant = scheme.get("weights")
+            if (weight_quant is not None
+                    and weight_quant.strategy == QuantizationStrategy.BLOCK):
+                return True
+        return False
+
     @staticmethod
     def supports_cutlass_24(
             weight_quant: Optional[QuantizationArgs],

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py

@@ -11,7 +11,7 @@
 from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
     CompressedTensorsScheme)
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    apply_fp8_block_linear, check_aiter_fp8_linear_support,
+    W8A8BlockFp8LinearOp, check_aiter_fp8_linear_support,
     create_fp8_input_scale, create_fp8_scale_parameter,
     create_fp8_weight_parameter, maybe_post_process_fp8_weight_block,
     process_fp8_weight_block_strategy, process_fp8_weight_channel_strategy,
@@ -41,16 +41,30 @@ def __init__(self, weight_quant: QuantizationArgs,
         self.strategy = weight_quant.strategy
         self.out_dtype = torch.get_default_dtype()
         self.is_static_input_scheme = is_static_input_scheme
-        self.act_q_group_shape = GroupShape.PER_TENSOR \
-            if is_static_input_scheme else GroupShape.PER_TOKEN
-        self.fp8_linear = Fp8LinearOp(
-            act_quant_static=self.is_static_input_scheme,
-            act_quant_group_shape=self.act_q_group_shape)
 
         self.weight_block_size = self.weight_quant.block_structure
+        if self.weight_block_size is not None:
+            self.act_q_group_shape = GroupShape(1, self.weight_block_size[0])
+        else:
+            self.act_q_group_shape = GroupShape.PER_TENSOR \
+                if is_static_input_scheme else GroupShape.PER_TOKEN
+
         self.cutlass_block_fp8_supported = cutlass_block_fp8_supported()
         self.use_aiter_and_is_supported = check_aiter_fp8_linear_support()
 
+        if self.weight_block_size is not None:
+            assert not self.is_static_input_scheme
+            self.w8a8_block_fp8_linear = W8A8BlockFp8LinearOp(
+                weight_group_shape=GroupShape(*self.weight_block_size),
+                act_quant_group_shape=self.act_q_group_shape,
+                cutlass_block_fp8_supported=self.cutlass_block_fp8_supported,
+                use_aiter_and_is_supported=self.use_aiter_and_is_supported,
+            )
+        else:
+            self.fp8_linear = Fp8LinearOp(
+                act_quant_static=self.is_static_input_scheme,
+                act_quant_group_shape=self.act_q_group_shape)
+
     @classmethod
     def get_min_capability(cls) -> int:
         # lovelace and up
@@ -142,13 +156,14 @@ def apply_weights(self,
                       x: torch.Tensor,
                       bias: Optional[torch.Tensor] = None) -> torch.Tensor:
 
-        if layer.weight_block_size is not None:
-            return apply_fp8_block_linear(
-                layer,
+        if self.weight_block_size is not None:
+            return self.w8a8_block_fp8_linear.apply(
                 input=x,
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                input_scale=layer.input_scale,
                 bias=bias,
-                cutlass_block_fp8_supported=self.cutlass_block_fp8_supported,
-                use_aiter_and_is_supported=self.use_aiter_and_is_supported)
+            )
 
         return self.fp8_linear.apply(input=x,
                                      weight=layer.weight,

vllm/model_executor/layers/quantization/fp8.py

@@ -33,7 +33,7 @@
     register_moe_scaling_factors, rotate_flashinfer_fp8_moe_weights,
     select_cutlass_fp8_gemm_impl, swap_w13_to_w31)
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    apply_fp8_block_linear, check_aiter_fp8_linear_support,
+    W8A8BlockFp8LinearOp, check_aiter_fp8_linear_support,
     create_fp8_input_scale, create_fp8_scale_parameter,
     create_fp8_weight_parameter, expert_weight_is_col_major,
     maybe_post_process_fp8_weight_block, process_fp8_weight_block_strategy,
@@ -242,15 +242,28 @@ def __init__(self, quant_config: Fp8Config):
         self.weight_block_size = self.quant_config.weight_block_size
         self.block_quant = self.weight_block_size is not None
         self.act_q_static = self.quant_config.activation_scheme == "static"
-        # Use per-token quantization for better perf if dynamic and cutlass
-        if not self.act_q_static and cutlass_fp8_supported():
-            self.act_q_group_shape = GroupShape.PER_TOKEN
+        if self.weight_block_size:
+            self.act_q_group_shape = GroupShape(1, self.weight_block_size[0])
         else:
-            self.act_q_group_shape = GroupShape.PER_TENSOR
+            # Use per-token quantization for better perf if dynamic and cutlass
+            if not self.act_q_static and cutlass_fp8_supported():
+                self.act_q_group_shape = GroupShape.PER_TOKEN
+            else:
+                self.act_q_group_shape = GroupShape.PER_TENSOR
 
-        self.fp8_linear = Fp8LinearOp(
-            act_quant_static=self.act_q_static,
-            act_quant_group_shape=self.act_q_group_shape)
+        if self.block_quant:
+            assert not self.act_q_static
+            assert self.weight_block_size is not None
+            self.w8a8_block_fp8_linear = W8A8BlockFp8LinearOp(
+                weight_group_shape=GroupShape(*self.weight_block_size),
+                act_quant_group_shape=self.act_q_group_shape,
+                cutlass_block_fp8_supported=self.cutlass_block_fp8_supported,
+                use_aiter_and_is_supported=self.use_aiter_and_is_supported,
+            )
+        else:
+            self.fp8_linear = Fp8LinearOp(
+                act_quant_static=self.act_q_static,
+                act_quant_group_shape=self.act_q_group_shape)
 
     def create_weights(
         self,
@@ -399,12 +412,15 @@ def apply(self,
                 bias=bias)
 
         if self.block_quant:
-            return apply_fp8_block_linear(
-                layer,
+            assert self.weight_block_size is not None
+
+            return self.w8a8_block_fp8_linear.apply(
                 input=x,
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                input_scale=layer.input_scale,
                 bias=bias,
-                cutlass_block_fp8_supported=self.cutlass_block_fp8_supported,
-                use_aiter_and_is_supported=self.use_aiter_and_is_supported)
+            )
 
         return self.fp8_linear.apply(input=x,
                                      weight=layer.weight,