common, server : preserve HF file for cached models

[mrexodia]

Overview

Today I downloaded a model from: https://huggingface.co/mudler/Step-3.5-Flash-APEX-GGUF, which contains the following files:

Step-3.7-Flash-APEX-Balanced.gguf
Step-3.7-Flash-APEX-Compact.gguf
Step-3.7-Flash-APEX-I-Balanced.gguf
Step-3.7-Flash-APEX-I-Compact.gguf <-- I downloaded this one to my huggingface cache
Step-3.7-Flash-APEX-I-Mini.gguf
Step-3.7-Flash-APEX-I-Quality.gguf
Step-3.7-Flash-APEX-Quality.gguf

Using llama serve discovers the model with the following identifier: mudler/Step-3.7-Flash-APEX-GGUF:COMPACT. The issue is that the tag :COMPACT is a ambiguous and llama will (silently) start downloading Step-3.7-Flash-APEX-Compact.gguf from huggingface when you start using the model.

The fix is to pass the GGUF filename using LLAMA_ARG_HF_FILE so we avoid the ambiguity.

Additional information

I think there is a larger discussion to be had about the way the tag is determined, because a better tag name here would arguably be I-COMPACT. Unsloth's UD-xxx quants have a similar issue: https://https://huggingface.co/unsloth/Step-3.7-Flash-GGUF. This is not something I would feel comfortable submitting a PR for though.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, pi:gpt-5.5 was used to triage and fix the issue. I reviewed the code thoroughly by hand and fully understand the changes.

Duncan