Incorrect Default Config File Paths for Llama 3.1 8B and Qwen 2.5 7B Models

[MaxHastings]

I've noticed an issue where the downloaded model directories for Llama 3.1 8B and Qwen 2.5 7B do not match the paths expected in their respective default config files.

Llama 3.1 8B Issue
The downloaded model is placed in:

/tmp/Llama-3.1-8B-Instruct
However, when running fine-tuning using the default config, it looks for:

/tmp/Meta-Llama-3.1-8B-Instruct/config.json
This results in the following error:

FileNotFoundError: [Errno 2] No such file or directory: '/tmp/Meta-Llama-3.1-8B-Instruct/config.json'
It seems like the configs at this location need to be updated to reflect the correct path.
Notably, I do not experience this issue with Llama 3.2 1B and 3B models—their default configs seem correct.

Qwen 2.5 7B Issue
A similar issue exists for Qwen 2.5 7B:

The default config expects the model at:

/tmp/Qwen2_5-7B-Instruct
But when downloading via:

tune download Qwen/Qwen2.5-7B-Instruct --ignore-patterns 'original/consolidated.00.pth'
It is saved to:

/tmp/Qwen2.5-7B-Instruct
Running fine-tuning results in an error due to the mismatch:

FileNotFoundError: [Errno 2] No such file or directory: '/tmp/Qwen2_5-7B-Instruct/vocab.json'
This can be traced to line 33 in the default config:
Qwen2.5 7B Config (line 33)

path: /tmp/Qwen2_5-7B-Instruct/vocab.json
Potential Workarounds
Manually specifying the output directory during the download seems to resolve the issue:

--output-dir /tmp/Qwen2_5-7B-Instruct
However, I wanted to flag this as it seems like the default configs should be updated to match the actual model download paths.

Would appreciate any insights or fixes for this! Thanks! 😊

[acisseJZhong]

Hi @MaxHastings when I tried to download llama3.1 using command tune download meta-llama/Meta-Llama-3.1-8B-Instruct --ignore-patterns "original/consolidated.00.pth", it is downloading to /tmp/Meta-Llama-3.1-8B-Instruct. Curious what command are you using?

For qwen2_5, I reproduced. Seems like our parsing logic here

torchtune/torchtune/_cli/download.py

Line 141 in a8aeb13

output_dir = Path("/tmp") / model_name

is just directly taking the model name as output dir.

And we are actually being inconsistent, e.g.

torchtune/recipes/configs/qwen2_5/0.5B_full.yaml

Line 32 in a8aeb13

checkpoint_dir: /tmp/Qwen2.5-0.5B-Instruct

this has checkpoint_dir Qwen2.5-0.5B-Instruct. We should probably fix the default config for qwen 7b. cc @felipemello1 in case I am missing anything.