use flashinfer.sampling (#18696)

This commit is contained in:
pansicheng
2026-02-26 10:02:38 +08:00
committed by GitHub
parent 2739d7df62
commit 2ad475b4ed
8 changed files with 14 additions and 431 deletions
-15
View File
@@ -417,27 +417,12 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
{at::Tag::needs_fixed_stride_order});
m.impl("bmm_fp8", torch::kCUDA, &bmm_fp8);
m.def(
"min_p_sampling_from_probs(Tensor probs, Tensor output, Tensor? maybe_indices, Tensor? maybe_min_p_arr, float "
"min_p_val, bool deterministic, Generator? gen) -> ()");
m.impl("min_p_sampling_from_probs", torch::kCUDA, &min_p_sampling_from_probs);
m.def("top_k_renorm_probs(Tensor probs, Tensor! renorm_probs, Tensor? maybe_top_k_arr, int top_k_val) -> ()");
m.impl("top_k_renorm_probs", torch::kCUDA, &top_k_renorm_probs);
m.def("top_p_renorm_probs(Tensor probs, Tensor! renorm_probs, Tensor? maybe_top_p_arr, float top_p_val) -> ()");
m.impl("top_p_renorm_probs", torch::kCUDA, &top_p_renorm_probs);
m.def(
"top_p_sampling_from_probs(Tensor probs, Tensor output, Tensor? maybe_indices, Tensor? "
"maybe_top_p_arr, float top_p_val, bool deterministic, Generator? gen) -> ()");
m.impl("top_p_sampling_from_probs", torch::kCUDA, &top_p_sampling_from_probs);
m.def(
"top_k_top_p_sampling_from_probs(Tensor probs, Tensor output, Tensor? maybe_indices, Tensor? maybe_top_k_arr, "
"float top_k_val, Tensor? maybe_top_p_arr, float top_p_val, bool deterministic, Generator? gen) -> ()");
m.impl("top_k_top_p_sampling_from_probs", torch::kCUDA, &top_k_top_p_sampling_from_probs);
m.def("top_k_mask_logits(Tensor logits, Tensor mask_logits, Tensor? maybe_top_k_arr, int top_k_val) -> ()");
m.impl("top_k_mask_logits", torch::kCUDA, &top_k_mask_logits);