chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[CPU] layernorm & fused add-layernorm kernels (#14074)

Browse Source
This commit is contained in:
Zaili Wang
2025-12-11 16:58:23 -08:00
committed by GitHub
parent 9975acf50f
commit d6bd2d1126
5 changed files with 258 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
sgl-kernel/csrc/cpu/torch_extension_cpu.cpp
+10
View File
@@ -33,12 +33,18 @@ at::Tensor l2norm_cpu(at::Tensor& input, double eps);
// rmsnorm
at::Tensor rmsnorm_cpu(at::Tensor& input, at::Tensor& weight, double eps);
// layernorm
void layernorm_cpu(at::Tensor& input, at::Tensor& weight, double eps);
// qwen3_next_rmsnorm_gated
at::Tensor fused_rmsnorm_gated_cpu(at::Tensor& input, at::Tensor& weight, at::Tensor& gate, double eps);
// fused_add_rmsnorm
void fused_add_rmsnorm_cpu(at::Tensor& input, at::Tensor& residual, at::Tensor& weight, double eps);
// fused_add_layernorm
void fused_add_layernorm_cpu(at::Tensor& input, at::Tensor& residual, at::Tensor& weight, double eps);
// topk
std::tuple<at::Tensor, at::Tensor>
topk_sigmoid_cpu(at::Tensor& hidden_states, at::Tensor& gating_output, int64_t topk, bool renormalize);
@@ -324,12 +330,16 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
// norm
m.def("rmsnorm_cpu(Tensor input, Tensor weight, float eps) -> Tensor");
m.impl("rmsnorm_cpu", torch::kCPU, &rmsnorm_cpu);
m.def("layernorm_cpu(Tensor(a!) input, Tensor weight, float eps) -> ()");
m.impl("layernorm_cpu", torch::kCPU, &layernorm_cpu);
m.def("l2norm_cpu(Tensor input, float eps) -> Tensor");
m.impl("l2norm_cpu", torch::kCPU, &l2norm_cpu);
m.def("fused_rmsnorm_gated_cpu(Tensor input, Tensor weight, Tensor gate, float eps) -> Tensor");
m.impl("fused_rmsnorm_gated_cpu", torch::kCPU, &fused_rmsnorm_gated_cpu);
m.def("fused_add_rmsnorm_cpu(Tensor(a!) input, Tensor residual, Tensor weight, float eps) -> ()");
m.impl("fused_add_rmsnorm_cpu", torch::kCPU, &fused_add_rmsnorm_cpu);
m.def("fused_add_layernorm_cpu(Tensor(a!) input, Tensor residual, Tensor weight, float eps) -> ()");
m.impl("fused_add_layernorm_cpu", torch::kCPU, &fused_add_layernorm_cpu);
// topk
m.def("topk_sigmoid_cpu(Tensor hidden_states, Tensor gating_output, int topk, bool renormalize) -> (Tensor, Tensor)");