Offline Model Loading When Previously Downloaded via llama-server fails

[1-ashraful-islam]

Issue Description:

When attempting to launch llama-server without an internet connection, it fails with a timeout error while making a GET request to Hugging Face, despite the model having been previously downloaded using the same command.

Steps to Reproduce:

1. Run the following command while connected to the internet to download and use the model:

llama-server \
    -hf ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF \
    --port 8012 -ngl 99 -fa -ub 1024 -b 1024 \
    --ctx-size 0 --cache-reuse 256

2. Disconnect from the internet.
3. Run the same command again.
4. The command fails, appearing to attempt a GET request to Hugging Face.
   Using the llama-server command without internet doesn't work even though the model was already downloaded previously using the same command. The command seems to timeout on a GET request to hugging face.

Expected Behavior:

If the model was previously downloaded, llama-server should detect and load it from cache without requiring an internet connection.

Actual Behavior:

The command times out while trying to reach Hugging Face, preventing offline usage.

Would appreciate any guidance on how to enforce offline usage or a workaround to bypass this issue! 🚀

If this is not a solved issue and you would like me to open the issue in the llama.cpp instead, let me know.

[ggerganov]

You can use the -m /path/to/model.gguf flag instead of -hf. This will skip the curl queries to the remote server.