happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

read MoE params

This commit is contained in:
Yee Man Chan 2025-12-02 20:16:24 +08:00
parent 83d328d0d3
commit 772ca88070
1 changed files with 11 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
convert_hf_to_gguf.py
View File

@ -5046,6 +5046,17 @@ class KimiLinearModel(TextModel):
head_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
self.gguf_writer.add_rope_dimension_count(head_dim)
self.gguf_writer.add_rope_freq_base(self.hparams.get("rope_theta", 10000.0))
# MoE params
n_experts = self.hparams.get("num_local_experts", self.hparams.get("num_experts"))
if n_experts is not None:
self.gguf_writer.add_expert_count(n_experts)
# Support both num_experts_per_tok and num_experts_per_token
n_experts_used = self.hparams.get("num_experts_per_tok", self.hparams.get("num_experts_per_token"))
if n_experts_used is not None:
self.gguf_writer.add_expert_used_count(n_experts_used)
# moe_intermediate_size (1024 for Kimi)
moe_intermediate_size = self.hparams.get("moe_intermediate_size")
if moe_intermediate_size is not None: