Phi-4-mini-instruct support

[andrea-tomassi]

Phi-4-mini-instruct is out. Is there an ETA for supporting this model?

[ysjprojects]

Hi @andrea-tomassi I'm on it.

Will try to push it in a few hours.

[andrea-tomassi-sharelock]

@ysjprojects FYI: I did try your commit https://github.com/Lightning-AI/litgpt.git@646e372f9a47ace139910d03e7bafab630393d91 and it seems to be workig well on my env (VastAI PyTorch template 2.5 X86 and nvidia 4060TI).

However, at the end I get an error in conversion I don't get with other models, including phi 3.5-mini:

Step 4: Converting LoRA output to GGUF Format (streamlined)...
Converting LoRA model at 'out/microsoft_Phi-4-mini-instruct/emails/lora' to Hugging Face format in 'out/microsoft_Phi-4-mini-instruct/emails/hf_model'...
{'checkpoint_dir': PosixPath('out/microsoft_Phi-4-mini-instruct/emails/lora'),
 'output_dir': PosixPath('out/microsoft_Phi-4-mini-instruct/emails/hf_model')}
Traceback (most recent call last):
  File "/venv/main/bin/litgpt", line 8, in <module>
    sys.exit(main())
  File "/venv/main/lib/python3.10/site-packages/litgpt/__main__.py", line 71, in main
    CLI(parser_data)
  File "/venv/main/lib/python3.10/site-packages/jsonargparse/_cli.py", line 23, in CLI
    return auto_cli(*args, _stacklevel=3, **kwargs)
  File "/venv/main/lib/python3.10/site-packages/jsonargparse/_cli.py", line 125, in auto_cli
    return _run_component(component, init.get(subcommand))
  File "/venv/main/lib/python3.10/site-packages/jsonargparse/_cli.py", line 210, in _run_component
    return component(**cfg)
  File "/venv/main/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/venv/main/lib/python3.10/site-packages/litgpt/scripts/convert_lit_checkpoint.py", line 398, in convert_lit_checkpoint
    copy_fn(sd, lit_weights, saver=saver)
  File "/venv/main/lib/python3.10/site-packages/litgpt/scripts/convert_lit_checkpoint.py", line 277, in copy_weights_phi
    to_names = (weight_map[name_template].format(layer_idx),)
KeyError: 'transformer.h.{}.norm_2.weight'

Not sure this is related to your side, maybe it is something wrong in my fine-tuning pipeline? I'll get into this...

Ty for supporting this so quickly! Waiting for the final pull request!

[andrea-tomassi-sharelock]

Ok this commit: f6d2fd2 solved the issue in producing gguf, Ty!