@CISC High-jacking this thread to report a minor regression that I observe. I need the following patch to be able to convert https://huggingface.co/Qwen/Qwen3-30B-A3B-Base:

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 825080b58..1e5209cdc 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -4161,7 +4161,7 @@ class Qwen2MoeModel(TextModel):
             return
 
         if name.find("experts") != -1:
-            n_experts = self.hparams["num_experts"]
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
             if self._experts is None:

@CISC High-jacking this thread to report a minor regression that I observe. I need the following patch to be able to convert https://huggingface.co/Qwen/Qwen3-30B-A3B-Base:

Hmmm, ok, not sure why that would have regressed, I'll look into it.

Last I tested in mid-Jan, it was converting successfully with this commit: 2bbe4c2. Not sure if a problem in the convert script, or in something in the Python dependencies has changed.

Last I tested in mid-Jan, it was converting successfully with this commit: 2bbe4c2. Not sure if a problem in the convert script, or in something in the Python dependencies has changed.

Are you sure? I see the non-base model uses num_experts, I think this just changed in transformers by the time they released the base model.

Last I tested in mid-Jan, it was converting successfully with this commit: 2bbe4c2. Not sure if a problem in the convert script, or in something in the Python dependencies has changed.

Are you sure? I see the non-base model uses num_experts, I think this just changed in transformers by the time they released the base model.

Edit: Ah, I see the problem, it's transformers itself that changed, it's num_experts in the config, but AutoConfig returns num_local_experts.