search

dvlab-research / LISA

Project Page for "LISA: Reasoning Segmentation via Large Language Model"
Apache License 2.0
1.89k stars 132 forks source link

Generated several bin files instead of only one bin file using zero_to_fp32.py #155

Open skx6 opened 3 weeks ago

skx6 commented 3 weeks ago

Great work! I trained a 13B model, however, when I try to run the following codes, an error occurs:

python zero_to_fp32.py . ../pytorch_model.bin

The error is:

RuntimeError: Parent directory ../pytorch_model.bin does not exist.

Is the output of this step a file or directory? Is it a problem about deepspeed version or config?

Hishamew commented 3 weeks ago

That's because the version of deepspeed must be under or equal to 0.15.2. There is a change made to file zero_to_fp32.py in deepspeed0.15.3 which causes this problem.