convert_hf_to_gguf.py incorrectly falls back to num_hidden_layers, causing KeyError on ChatGLM3 (num_layers present)

[shadowmmu]

Description

Some newer model configurations (such as ChatGLM3) no longer include the field "num_hidden_layers".
Instead, they provide "num_layers".

The current converter code contains the following line:

self.gguf_writer.add_block_count(
    self.hparams.get("num_layers", self.hparams["num_hidden_layers"])
)

This leads to a KeyError even if "num_layers" is present and "num_hidden_layers" is not present in the model config.

---

Questions

1. Is the absence of "num_hidden_layers" in newer models expected?
   It appears that recent architectures rely on "num_layers" instead.

2. Was the original line intended to assume "num_hidden_layers" always exists?
   If not, the KeyError suggests that the converter logic may be outdated for modern model configs.

3. Should the converter be updated to support both fields safely rather than assuming "num_hidden_layers" is always present?

---

Cause of the Error

The line:

self.hparams["num_hidden_layers"]

directly indexes into the dict.
If "num_hidden_layers" is missing, Python raises:

KeyError: 'num_hidden_layers'

This happens even when "num_layers" is present, because the dict lookup is only used when "num_layers" is missing from the get() call.

---

Suggested Fix

To avoid failing when "num_hidden_layers" is absent, the line could be updated to a safer fallback:

self.gguf_writer.add_block_count(
    self.hparams.get("num_layers", self.hparams.get("num_hidden_layers", 0)))

This ensures:

• "num_layers" is used when available
• "num_hidden_layers" is used if "num_layers" is missing
• a fallback value (e.g., 0) prevents a crash

[CISC]

It shouldn't read hparams at all as self.block_count contains the correct value already, made a PR.

Thanks for pointing out the issue. :)

[shadowmmu]

Awesome, thanks for the quick fix!

[shadowmmu]

@CISC BTW does llama.cpp support the ChatGLM3 architecture?

[CISC]

@CISC BTW does llama.cpp support the ChatGLM3 architecture?

Yes.