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

support cuda 13.0 and trtllm kernel by Aug 25 2025 (#9495)

Browse Source
This commit is contained in:
Rain Jiang
2025-08-26 23:13:27 -07:00
committed by GitHub
parent 8f7b1c31e8
commit 79e6a8a6ac
13 changed files with 81 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
sgl-kernel/csrc/moe/marlin_moe_wna16/kernel_fp16_ku8b128.cu
View File

@@ -1,109 +0,0 @@
// auto generated by generate.py
// clang-format off
#include "kernel.h"
#include "marlin_template.h"
namespace MARLIN_NAMESPACE_NAME {
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, true, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, true, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 2, 16, 4, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 2, 8, 4, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 3, 16, 4, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 3, 8, 4, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 4, 16, 4, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 4, 8, 4, false, pipe_stages, true, false, 0, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, true, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, true, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 2, 16, 4, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 2, 8, 4, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 3, 16, 4, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 3, 8, 4, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 4, 16, 4, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 4, 8, 4, false, pipe_stages, false, false, -1, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, true, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, true, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 2, 16, 4, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 2, 8, 4, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 3, 16, 4, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 3, 8, 4, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 4, 16, 4, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 4, 8, 4, false, pipe_stages, false, false, 2, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, true, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, true, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 2, 16, 4, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 2, 8, 4, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 3, 16, 4, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 3, 8, 4, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 4, 16, 4, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 4, 8, 4, false, pipe_stages, false, false, 4, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, true, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, true, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 1, 8, 8, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 1, 8, 4, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 2, 16, 4, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 2, 8, 4, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 3, 16, 4, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 3, 8, 4, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 256, 4, 16, 4, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
template __global__ void Marlin<half, sglang::kU8B128.id(), 128, 4, 8, 4, false, pipe_stages, false, false, 8, false>( MARLIN_KERNEL_PARAMS );
}