Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

CUDA: support for weight clamp in top-k norm (#16702)

Browse Source
This commit is contained in:
Aman Gupta
2025-10-27 09:06:16 +08:00
committed by GitHub
parent 3470a5c891
commit 75d33b9302
4 changed files with 60 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/topk-moe.cuh
+3 -2
View File
@@ -8,8 +8,9 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
ggml_tensor * weights,
ggml_tensor * ids,
const bool with_norm,
const bool delayed_softmax = false);
const bool delayed_softmax = false,
ggml_tensor * weight_clamp = nullptr);
bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights);
bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax, const ggml_tensor * weights, const ggml_tensor * clamp = nullptr);
std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool with_norm, bool delayed_softmax = false);