[0.39.0][Bug] bitsandbytes introduced scipy but not claimed as dependencies

[lanking520]

Using HuggingFace with bitsandbytes having the following issues. This seemed to be a bug in the setup. Using 0.38.1 all works fine

[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/tmp/.djl.ai/python/0.23.0-SNAPSHOT/djl_python/huggingface.py", line 183, in get_pipeline
[INFO ] 2023-05-24 15:17:10 PyProcess -     model = AutoModelForCausalLM.from_pretrained(
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/transformers/models/auto/auto_factory.py", line 471, in from_pretrained
[INFO ] 2023-05-24 15:17:10 PyProcess -     return model_class.from_pretrained(
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/transformers/modeling_utils.py", line 2508, in from_pretrained
[INFO ] 2023-05-24 15:17:10 PyProcess -     from .utils.bitsandbytes import get_keys_to_not_convert, replace_8bit_linear
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/transformers/utils/bitsandbytes.py", line 7, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     import bitsandbytes as bnb
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/__init__.py", line 6, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     from . import cuda_setup, utils, research
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/research/__init__.py", line 1, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     from . import nn
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/research/nn/__init__.py", line 1, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     from .modules import LinearFP8Mixed, LinearFP8Global
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/research/nn/modules.py", line 8, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     from bitsandbytes.optim import GlobalOptimManager
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/optim/__init__.py", line 8, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     from .adagrad import Adagrad, Adagrad8bit, Adagrad32bit
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/optim/adagrad.py", line 5, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     from bitsandbytes.optim.optimizer import Optimizer1State
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/optim/optimizer.py", line 12, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     import bitsandbytes.functional as F
[INFO ] 2023-05-24 15:17:10 PyProcess -   File "/usr/local/lib/python3.9/dist-packages/bitsandbytes/functional.py", line 12, in <module>
[INFO ] 2023-05-24 15:17:10 PyProcess -     from scipy.stats import norm
[INFO ] 2023-05-24 15:17:10 PyProcess - ModuleNotFoundError: No module named 'scipy'

[lanking520]

@TimDettmers should be an easy fix. Or we could find a way to make a workaround if scipy is not found to use numpy norm function

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.