Skip to content

fix sgl-kernel cu118 build #4872

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Mar 29, 2025
Merged

fix sgl-kernel cu118 build #4872

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ac56b50
upd
zhyncs Mar 28, 2025
946c59c
upd
zhyncs Mar 28, 2025
ba888ff
upd
zhyncs Mar 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion sgl-kernel/build.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,5 +25,6 @@ docker run --rm \
ln -s /usr/local/cuda-${CUDA_VERSION}/targets/x86_64-linux/lib/stubs/libcuda.so /usr/lib/x86_64-linux-gnu/libcuda.so && \
cd /sgl-kernel && \
ls -la ${PYTHON_ROOT_PATH}/lib/python${PYTHON_VERSION}/site-packages/wheel/ && \
PYTHONPATH=${PYTHON_ROOT_PATH}/lib/python${PYTHON_VERSION}/site-packages ${PYTHON_ROOT_PATH}/bin/python -m uv build --wheel -Cbuild-dir=build . --color=always
PYTHONPATH=${PYTHON_ROOT_PATH}/lib/python${PYTHON_VERSION}/site-packages ${PYTHON_ROOT_PATH}/bin/python -m uv build --wheel -Cbuild-dir=build . --color=always && \
./rename_wheels.sh
"
5 changes: 5 additions & 0 deletions sgl-kernel/csrc/gemm/awq_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
// Adapted from
// https://github.com/vllm-project/vllm/blob/eb59b5a6cba6727d3727c0372258db9002f687c1/csrc/quantization/awq/gemm_kernels.cu#L350
#include <c10/cuda/CUDAGuard.h>
#include <cuda.h>
#include <cuda_fp16.h>
#include <torch/all.h>
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
Expand Down Expand Up @@ -79,6 +80,7 @@ __device__ uint4 dequantize_s4_to_fp16x2(uint32_t const& source) {
}

__device__ uint4 dequantize_s4_to_bf16x2(uint32_t const& source) {
#if CUDA_VERSION >= 12000
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
uint4 result;
uint32_t* h = reinterpret_cast<uint32_t*>(&result);
Expand Down Expand Up @@ -118,6 +120,7 @@ __device__ uint4 dequantize_s4_to_bf16x2(uint32_t const& source) {
assert(false);
return {};
#endif
#endif
}

template <typename OutputT>
Expand All @@ -128,6 +131,7 @@ __global__ void __launch_bounds__(256) dequantize_weights(
OutputT* __restrict__ output,
int group_size,
int qweight_cols) {
#if CUDA_VERSION >= 12000
int col = blockIdx.x * blockDim.x + threadIdx.x;
int row = blockIdx.y * blockDim.y + threadIdx.y;

Expand Down Expand Up @@ -174,6 +178,7 @@ __global__ void __launch_bounds__(256) dequantize_weights(
static_assert(sizeof(uint4) == 8 * sizeof(OutputT), "Memory layout mismatch");
*reinterpret_cast<uint4*>(output_ptr) = weight_raw;
}
#endif
}

torch::Tensor awq_dequantize(torch::Tensor qweight, torch::Tensor scales, torch::Tensor qzeros) {
Expand Down
5 changes: 5 additions & 0 deletions sgl-kernel/csrc/gemm/nvfp4_quant_kernels.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@ limitations under the License.

#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda.h>
#include <cuda_fp8.h>
#include <cuda_runtime.h>
#include <cuda_runtime_api.h>
Expand Down Expand Up @@ -56,6 +57,7 @@ constexpr int CVT_FP4_SF_VEC_SIZE = 16;
// Convert 8 float32 values into 8 e2m1 values (represented as one uint32_t).
inline __device__ uint32_t fp32_vec_to_e2m1(float (&array)[8]) {
// PTX instructions used here requires sm100a.
#if CUDA_VERSION >= 12080
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000) && __CUDA_ARCH_HAS_FEATURE__(SM100_ALL)
uint32_t val;
asm volatile(
Expand Down Expand Up @@ -83,11 +85,13 @@ inline __device__ uint32_t fp32_vec_to_e2m1(float (&array)[8]) {
#else
return 0;
#endif
#endif
}

// Convert 4 float2 values into 8 e2m1 values (represented as one uint32_t).
inline __device__ uint32_t fp32_vec_to_e2m1(float2 (&array)[4]) {
// PTX instructions used here requires sm100a.
#if CUDA_VERSION >= 12080
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000) && __CUDA_ARCH_HAS_FEATURE__(SM100_ALL)
uint32_t val;
asm volatile(
Expand Down Expand Up @@ -115,6 +119,7 @@ inline __device__ uint32_t fp32_vec_to_e2m1(float2 (&array)[4]) {
#else
return 0;
#endif
#endif
}

// Fast reciprocal.
Expand Down
Loading