vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix(conversion): Simplify tensor name mapping in conversion

Browse source 
Branch: GraniteMoE

Co-Authored-By: git@compilade.net
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart 2024-09-23 11:03:18 -06:00
parent 71bc4c1f93
commit 5eb28c4710
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
convert_hf_to_gguf.py
View file

@ -4129,8 +4129,8 @@ class GraniteMoeModel(GraniteModel):
if name.endswith("block_sparse_moe.input_linear.weight"):
gate, up = data_torch.chunk(2, dim=-2)
return [
(self.map_tensor_name(f"model.layers.{bid}.block_sparse_moe.input_linear.gate.weight"), gate),
(self.map_tensor_name(f"model.layers.{bid}.block_sparse_moe.input_linear.up.weight"), up),
(self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
(self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
]
return super().modify_tensors(data_torch, name, bid)