[Perf] Tunings for SM100 FP8 CUTLASS kernel (#8818)

sgl-kernel/benchmark/bench_fp8_gemm.py

@@ -1,10 +1,12 @@
 import argparse
 import copy
 import itertools
+from typing import Optional, Tuple
 
 import torch
 import triton
 from sgl_kernel import fp8_scaled_mm as sgl_scaled_mm
+from sgl_kernel import sgl_per_tensor_quant_fp8
 from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
 from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant
 
@@ -69,6 +71,21 @@ WEIGHT_SHAPES = {
 }
 
 
+def sglang_scaled_fp8_quant(
+    input: torch.Tensor,
+    scale: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    fp8_type_: torch.dtype = torch.float8_e4m3fn
+    output = torch.empty_like(input, device=input.device, dtype=fp8_type_)
+    is_static = True
+    if scale is None:
+        scale = torch.zeros(1, device=input.device, dtype=torch.float32)
+        is_static = False
+    sgl_per_tensor_quant_fp8(input, output, scale, is_static)
+
+    return output, scale
+
+
 @triton.testing.perf_report(
     triton.testing.Benchmark(
         x_names=["batch_size"],
@@ -100,19 +117,22 @@ def benchmark(batch_size, provider, N, K):
     b = torch.ones((N, K), device="cuda") * 5.0
     scale_a = torch.randn((M,), device="cuda", dtype=torch.float32)
     scale_b = torch.randn((N,), device="cuda", dtype=torch.float32)
-    a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a)
-    b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
-    b_fp8 = b_fp8.t()
     quantiles = [0.5, 0.2, 0.8]
 
     dtype = torch.float16 if "fp16" in provider else torch.bfloat16
 
     if "vllm-fp8" in provider:
+        a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a)
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
+        b_fp8 = b_fp8.t()
         ms, min_ms, max_ms = triton.testing.do_bench(
             lambda: vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype),
             quantiles=quantiles,
         )
     elif "sglang-fp8" in provider:
+        a_fp8, scale_a_fp8 = sglang_scaled_fp8_quant(a, scale_a)
+        b_fp8, scale_b_fp8 = sglang_scaled_fp8_quant(b, scale_b)
+        b_fp8 = b_fp8.t()
         ms, min_ms, max_ms = triton.testing.do_bench(
             lambda: sgl_scaled_mm(
                 a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype, bias=None

sgl-kernel/csrc/gemm/fp8_gemm_kernel.cu

@@ -48,6 +48,7 @@ limitations under the License.
 #include <cutlass/gemm/kernel/gemm_universal.hpp>
 #include <cutlass/util/packed_stride.hpp>
 
+#include "math.hpp"
 #include "utils.h"
 
 using namespace cute;
@@ -1019,8 +1020,18 @@ void sm100_fp8_dispatch_bias(
     const torch::Tensor& scales_a,
     const torch::Tensor& scales_b,
     const c10::optional<torch::Tensor>& bias) {
-  using CTAShape = Shape<_256, _128, _64>;
-  using ClusterShape = Shape<_2, _2, _1>;
+  using CTAShapeDefault = Shape<_256, _128, _64>;
+  using ClusterShapeDefault = Shape<_2, _2, _1>;
+
+  using CTAShape256 = Shape<_128, _128, _128>;
+  using ClusterShape256 = Shape<_2, _1, _1>;
+
+  using CTAShape64 = Shape<_64, _64, _128>;
+  using ClusterShape64 = Shape<_1, _1, _1>;
+
+  using CTAShape16 = Shape<_64, _64, _128>;
+  using ClusterShape16 = Shape<_1, _4, _1>;
+
   using MainloopScheduleType = cutlass::gemm::collective::KernelScheduleAuto;
   using EpilogueScheduleType = cutlass::epilogue::collective::EpilogueScheduleAuto;
   using TileSchedulerType = void;
@@ -1029,30 +1040,121 @@ void sm100_fp8_dispatch_bias(
   using ElementOutput = OutType;
   using AccumElementType = float;
 
+  // Gemm type with bias
+  using BiasGemmDefault = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShapeDefault,
+      ClusterShapeDefault,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      true>;
+  using BiasGemm256 = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShape256,
+      ClusterShape256,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      true>;
+  using BiasGemm64 = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShape64,
+      ClusterShape64,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      true>;
+  using BiasGemm16 = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShape16,
+      ClusterShape16,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      true>;
+
+  // Gemm type without bias
+  using GemmDefault = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShapeDefault,
+      ClusterShapeDefault,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      false>;
+  using Gemm256 = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShape256,
+      ClusterShape256,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      false>;
+  using Gemm64 = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShape64,
+      ClusterShape64,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      false>;
+  using Gemm16 = DeviceGemmFp8RowwiseSm100<
+      ElementInput,
+      ElementOutput,
+      AccumElementType,
+      CTAShape16,
+      ClusterShape16,
+      MainloopScheduleType,
+      EpilogueScheduleType,
+      TileSchedulerType,
+      false>;
+
+  // next power of 2 (minimum 16)
+  uint32_t const m = a.size(0);
+  uint32_t const mp2 = std::max(static_cast<uint32_t>(16), next_pow_2(m));
+
   if (bias) {
-    using Gemm = DeviceGemmFp8RowwiseSm100<
-        ElementInput,
-        ElementOutput,
-        AccumElementType,
-        CTAShape,
-        ClusterShape,
-        MainloopScheduleType,
-        EpilogueScheduleType,
-        TileSchedulerType,
-        true>;
-    return launch_sm100_fp8_scaled_mm<Gemm, true>(out, a, b, scales_a, scales_b, bias);
+    if (mp2 <= 16) {
+      // m in [1, 16]
+      return launch_sm100_fp8_scaled_mm<BiasGemm16, true>(out, a, b, scales_a, scales_b, bias);
+    } else if (mp2 <= 64) {
+      // m in (16, 64]
+      return launch_sm100_fp8_scaled_mm<BiasGemm64, true>(out, a, b, scales_a, scales_b, bias);
+    } else if (mp2 <= 256) {
+      // m in (64, 256]
+      return launch_sm100_fp8_scaled_mm<BiasGemm256, true>(out, a, b, scales_a, scales_b, bias);
+    } else {
+      // m in (256, inf]
+      return launch_sm100_fp8_scaled_mm<BiasGemmDefault, true>(out, a, b, scales_a, scales_b, bias);
+    }
   } else {
-    using Gemm = DeviceGemmFp8RowwiseSm100<
-        ElementInput,
-        ElementOutput,
-        AccumElementType,
-        CTAShape,
-        ClusterShape,
-        MainloopScheduleType,
-        EpilogueScheduleType,
-        TileSchedulerType,
-        false>;
-    return launch_sm100_fp8_scaled_mm<Gemm, false>(out, a, b, scales_a, scales_b, bias);
+    if (mp2 <= 16) {
+      // m in [1, 16]
+      return launch_sm100_fp8_scaled_mm<Gemm16, false>(out, a, b, scales_a, scales_b, bias);
+    } else if (mp2 <= 64) {
+      // m in (16, 64]
+      return launch_sm100_fp8_scaled_mm<Gemm64, false>(out, a, b, scales_a, scales_b, bias);
+    } else if (mp2 <= 256) {
+      // m in (64, 256]
+      return launch_sm100_fp8_scaled_mm<Gemm256, false>(out, a, b, scales_a, scales_b, bias);
+    } else {
+      return launch_sm100_fp8_scaled_mm<GemmDefault, false>(out, a, b, scales_a, scales_b, bias);
+    }
   }
 }
 

sgl-kernel/csrc/gemm/math.hpp

@@ -0,0 +1,28 @@
+#pragma once
+
+#include <climits>
+#include <iostream>
+
+inline constexpr uint32_t next_pow_2(uint32_t const num) {
+  if (num <= 1) return num;
+  return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
+}
+
+template <typename A, typename B>
+static inline constexpr auto div_ceil(A a, B b) {
+  return (a + b - 1) / b;
+}
+
+// Round a down to the next multiple of b. The caller is responsible for making
+// sure that b is non-zero
+template <typename T>
+inline constexpr T round_to_previous_multiple_of(T a, T b) {
+  return a % b == 0 ? a : (a / b) * b;
+}
+
+// Round a up to the next multiple of b. The caller is responsible for making
+// sure that b is non-zero
+template <typename T>
+inline constexpr T round_to_next_multiple_of(T a, T b) {
+  return a % b == 0 ? a : ((a / b) + 1) * b;
+}