[Kernel] Switch fp8 layers to use the CUTLASS kernels (#5183)

Switching from torch._scaled_mm to vLLM's cutlass fp8 kernels when supported as we are seeing 5-15% improvement in e2e performance on neuralmagic/Meta-Llama-3-8B-Instruct-FP8

see https://docs.google.com/spreadsheets/d/1GiAnmzyGHgZ6zL_LDSTm35Bdrt4A8AaFEurDlISYYA4/ for some quick e2e benchmarks and #5144 for comparisons across different GEMM sizes.

vllm/_custom_ops.py

@@ -179,7 +179,7 @@ def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
 
 # cutlass
 def cutlass_scaled_mm_dq(a: torch.Tensor, b: torch.Tensor,
-                         a_scales: torch.Tensor, b_scales: torch.Tensor,
+                         scale_a: torch.Tensor, scale_b: torch.Tensor,
                          out_dtype: Type[torch.dtype]) -> torch.Tensor:
     assert (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
     assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
@@ -188,7 +188,7 @@ def cutlass_scaled_mm_dq(a: torch.Tensor, b: torch.Tensor,
     n = b.shape[1]
     out = torch.empty((m, n), dtype=out_dtype, device=a.device)
 
-    vllm_ops.cutlass_scaled_mm_dq(out, a, b, a_scales, b_scales)
+    vllm_ops.cutlass_scaled_mm_dq(out, a, b, scale_a, scale_b)
 
     return out
 

vllm/model_executor/layers/quantization/fp8.py

@@ -17,6 +17,24 @@ ACTIVATION_SCHEMES = ["static", "dynamic"]
 logger = init_logger(__name__)
 
 
+def cutlass_fp8_supported() -> bool:
+    capability = torch.cuda.get_device_capability()
+    capability = capability[0] * 10 + capability[1]
+    version = torch.version.cuda
+    version = version[0] * 10 + version[1]
+
+    # CUTLASS FP8 kernels need at least
+    #   CUDA 12.0 on SM90 systems (Hopper)
+    #   CUDA 12.4 on SM89 systems (Lovelace)
+    gpu_is_supported = False
+    if capability >= 900:
+        gpu_is_supported = version > 120
+    elif capability >= 890:
+        gpu_is_supported = version > 124
+
+    return gpu_is_supported
+
+
 class Fp8Config(QuantizationConfig):
     """Config class for FP8."""
 
@@ -92,6 +110,7 @@ class Fp8LinearMethod(LinearMethodBase):
 
     def __init__(self, quant_config: Fp8Config):
         self.quant_config = quant_config
+        self.cutlass_fp8_supported = cutlass_fp8_supported()
 
     def _create_scale_param(
         self,
@@ -233,25 +252,40 @@ class Fp8LinearMethod(LinearMethodBase):
               layer: torch.nn.Module,
               x: torch.Tensor,
               bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+
         # ops.scaled_fp8_quant supports both dynamic and static quant.
         #   If dynamic, layer.act_scale is None and x_scale computed from x.
         #   If static,  layer.act_scale is scalar and x_scale set to act_scale.
-        qinput, x_scale = ops.scaled_fp8_quant(x,
-                                               layer.act_scale,
-                                               batch_dim_padding=17)
 
-        # Fused GEMM_DQ -- note we padded the input above because
+        if bias is None and self.cutlass_fp8_supported:
-        # torch._scaled_mm is more performant for matrices with
+            qinput, x_scale = ops.scaled_fp8_quant(x, layer.act_scale)
-        # batch dimension > 16. Note that this could change
+
-        # in the future.
+            # Fused GEMM_DQ
-        output, _ = torch._scaled_mm(
+            output = ops.cutlass_scaled_mm_dq(
-            qinput,
+                qinput,
-            layer.weight,
+                layer.weight,
-            out_dtype=x.dtype,
+                out_dtype=x.dtype,
-            scale_a=x_scale,
+                scale_a=x_scale,
-            scale_b=layer.weight_scale,
+                scale_b=layer.weight_scale,
-            bias=bias,
+            )
-        )
+
+        else:
+            qinput, x_scale = ops.scaled_fp8_quant(x,
+                                                   layer.act_scale,
+                                                   batch_dim_padding=17)
+
+            # Fused GEMM_DQ -- note we padded the input above because
+            # torch._scaled_mm is more performant for matrices with
+            # batch dimension > 16. Note that this could change
+            # in the future.
+            output, _ = torch._scaled_mm(
+                qinput,
+                layer.weight,
+                out_dtype=x.dtype,
+                scale_a=x_scale,
+                scale_b=layer.weight_scale,
+                bias=bias,
+            )
 
         return torch.narrow(output, 0, 0, x.shape[0])