[AMD] Reorganize hip-related header files in sgl-kernel (#9320)

.github/workflows/pr-test-amd.yml

@@ -342,6 +342,7 @@ jobs:
           docker exec -w /sglang-checkout/sgl-kernel/tests ci_sglang python3 -m pytest test_moe_topk_softmax.py
           docker exec -w /sglang-checkout/sgl-kernel/tests/speculative ci_sglang python3 -m pytest test_eagle_utils.py
           docker exec -w /sglang-checkout/sgl-kernel/tests ci_sglang python3 -m pytest test_apply_token_bitmask_inplace.py
+          docker exec -w /sglang-checkout/sgl-kernel/tests ci_sglang python3 -m pytest test_activation.py
 
   pr-test-amd-finish:
     if: always()

sgl-kernel/csrc/allreduce/mscclpp_allreduce.cuh

 // Copyright (c) Microsoft Corporation.
 // Licensed under the MIT license.
 #pragma once
-#if defined(__HIP_PLATFORM_AMD__)
+#ifdef USE_ROCM
 #include <hip/hip_fp16.h>
 #else
 #include <cuda_bf16.h>

sgl-kernel/csrc/elementwise/activation.cu

@@ -25,7 +25,7 @@
 #include "utils.h"
 
 #else
-#include "hip_act_and_mul.cuh"
+#include "hip/hip_act_and_mul.cuh"
 #endif
 
 // Adapted from flashinfer activation

sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu

@@ -69,7 +69,7 @@ __global__ void per_tensor_quant_fp8_kernel(
 #pragma unroll
     for (uint32_t j = 0; j < VEC_SIZE; ++j) {
       float val = fmax(fmin(static_cast<float>(input_vec[j]) * scale_val, FP8_E4M3_MAX), -FP8_E4M3_MAX);
-#ifndef USE_ROCM
+#if !defined(USE_ROCM) || defined(HIP_FP8_TYPE_E4M3)
       output_arr[j] = static_cast<DST_DTYPE>(val);
 #else
       output_arr[j] = c10::Float8_e4m3fnuz(
@@ -83,7 +83,7 @@ __global__ void per_tensor_quant_fp8_kernel(
   const int32_t remaining_start = num_vec_elems * VEC_SIZE;
   for (int32_t idx = remaining_start + gid; idx < num_elements; idx += grid_size) {
     float val = fmax(-FP8_E4M3_MAX, fmin(static_cast<float>(input[idx]) * scale_val, FP8_E4M3_MAX));
-#ifndef USE_ROCM
+#if !defined(USE_ROCM) || defined(HIP_FP8_TYPE_E4M3)
     output[idx] = static_cast<DST_DTYPE>(val);
 #else
     output[idx] = c10::Float8_e4m3fnuz(

sgl-kernel/csrc/gemm/per_token_quant_fp8.cu

@@ -67,7 +67,7 @@ __global__ void per_token_quant_fp8_kernel(
     for (uint32_t j = 0; j < kVecSize; ++j) {
       float val = static_cast<float>(input_vec[j]) * scale_inv;
       val = fmaxf(fminf(val, FP8_E4M3_MAX), -FP8_E4M3_MAX);
-#ifndef USE_ROCM
+#if !defined(USE_ROCM) || defined(HIP_FP8_TYPE_E4M3)
       output_arr[j] = static_cast<DST_DTYPE>(val);
 #else
       output_arr[j] = c10::Float8_e4m3fnuz(
@@ -143,7 +143,7 @@ __global__ void per_token_quant_fp8_small_batch_kernel(
 #pragma unroll
     for (uint32_t j = 0; j < kVecSize; ++j) {
       float val = fmaxf(fminf(static_cast<float>(input_vec[j]) * scale_inv, FP8_E4M3_MAX), -FP8_E4M3_MAX);
-#ifndef USE_ROCM
+#if !defined(USE_ROCM) || defined(HIP_FP8_TYPE_E4M3)
       output_arr[j] = static_cast<DST_DTYPE>(val);
 #else
       output_arr[j] = c10::Float8_e4m3fnuz(

sgl-kernel/csrc/moe/moe_align_kernel.cu

@@ -21,8 +21,6 @@ limitations under the License.
 
 #include "utils.h"
 
-#define WARP_SIZE 32
-
 #define VEC_SIZE 4
 using Vec = int4;
 

sgl-kernel/include/hip_math_def.h → sgl-kernel/include/hip/hip_math_def.h

@@ -15,7 +15,7 @@ limitations under the License.
 
 #pragma once
 
-#if defined(__HIP_PLATFORM_AMD__)
+#ifdef USE_ROCM
 
 #include <hip/hip_bf16.h>
 #include <hip/hip_common.h>

sgl-kernel/include/utils.h

@@ -331,13 +331,15 @@ inline bool getEnvEnablePDL() {
 #ifndef USE_ROCM
 #define WARP_SIZE 32
 #else
-#define WARP_SIZE warpSize  // 64
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/macros/Macros.h>
+#define WARP_SIZE C10_WARP_SIZE
 #endif
 
-#if defined(__HIP_PLATFORM_AMD__)
+#ifdef USE_ROCM
 
-#include "hip_math_def.h"
-#include "hip_vec_dtypes.h"
+#include "hip/hip_math_def.h"
+#include "hip/hip_vec_dtypes.h"
 
 #else
 
@@ -354,14 +356,11 @@ __device__ __forceinline__ dstDtype castFromFloat(float val) {
 #endif
 
 // add FP8 support
-
 #ifndef USE_ROCM
 #include <c10/util/Float8_e4m3fn.h>
 using FP8_TYPE = c10::Float8_e4m3fn;
 C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
-
 #else  // USE_ROCM
-
 #if HIP_FP8_TYPE_FNUZ
 #include <c10/util/Float8_e4m3fnuz.h>
 using FP8_TYPE = c10::Float8_e4m3fnuz;

sgl-kernel/setup_rocm.py

@@ -72,6 +72,9 @@ if amdgpu_target not in ["gfx942", "gfx950"]:
     )
     sys.exit(1)
 
+fp8_macro = (
+    "-DHIP_FP8_TYPE_FNUZ" if amdgpu_target == "gfx942" else "-DHIP_FP8_TYPE_E4M3"
+)
 
 hipcc_flags = [
     "-DNDEBUG",
@@ -80,10 +83,10 @@ hipcc_flags = [
     "-Xcompiler",
     "-fPIC",
     "-std=c++17",
-    "-D__HIP_PLATFORM_AMD__=1",
     f"--amdgpu-target={amdgpu_target}",
     "-DENABLE_BF16",
     "-DENABLE_FP8",
+    fp8_macro,
 ]
 
 ext_modules = [