EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
Hubert Lu
2025-08-18 16:53:44 -07:00
committed by GitHub
parent c2fbf60f39
commit c6c379ab31
14 changed files with 18 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
sgl-kernel/csrc/allreduce/mscclpp_allreduce.cuh
View File

@@ -1,7 +1,7 @@
// 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>

2
sgl-kernel/csrc/elementwise/activation.cu
View File

@@ -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

4
sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu
View File

@@ -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(

4
sgl-kernel/csrc/gemm/per_token_quant_fp8.cu
View File

@@ -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(

2
sgl-kernel/csrc/moe/moe_align_kernel.cu
View File

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