Skip to content

LLM inference in C/C++

C++ ggml

convert : fix Nemotron-H LoRA GGUF conversion (#27356)

Kartik Guliabcb6084a4e20
diff --git a/conversion/nemotron.py b/conversion/nemotron.py
index cd0d48c8f0..07fbc65314 100644
--- a/conversion/nemotron.py
+++ b/conversion/nemotron.py
@@ -302,6 +302,10 @@ class NemotronHModel(GraniteHybridModel):
)
if not keep:
return None
+ # PEFT names adapter tensors using model.layers.*, while Nemotron-H checkpoints
+ # and the GGUF tensor map use backbone.layers.*
+ if name.startswith("model.layers.") and ".mixer." in name:
+ name = name.replace("model.layers.", "backbone.layers.", 1)
return super().filter_tensors((name, gen))
def prepare_metadata(self, vocab_only: bool):