ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fix circular import in trainer.py after _gpu_init refactor

Browse source 
The MLX-aware unsloth/__init__.py loads the GPU surface via
"from ._gpu_init import *" at line 127. _gpu_init.py:332 then runs
"from .trainer import *", and trainer.py:26 was doing
"from . import is_bfloat16_supported".

That last import asks for the symbol on the `unsloth` package namespace,
but `unsloth/__init__.py` is paused at line 127 and the star export
from _gpu_init has not yet propagated. Result on a fresh editable
install (./install.sh --local):

  ImportError: cannot import name 'is_bfloat16_supported' from
  partially initialized module 'unsloth' (most likely due to a circular
  import) (.../unsloth/__init__.py)

PyPI 2026.5.2 is unaffected because that wheel ships the older inline
__init__.py which defines is_bfloat16_supported before
"from .trainer import *" runs.

Fix: import directly from the concrete _utils module, matching the
pattern already in models/loader.py and models/llama.py. This bypasses
the partially-initialized unsloth namespace and resolves cleanly via
unsloth.models._utils, which has already been fully loaded by the
preceding "from .models import *" in _gpu_init.py:327.

Verified end-to-end:
  ./install.sh --local on Linux+CUDA imports unsloth cleanly,
  FastLanguageModel.from_pretrained loads, 10-step LoRA training on
  Llama-3.2-1B runs (loss 1.76, 5.85 samples/s on B200), post-train
  generation works.
This commit is contained in:
Daniel Han 2026-05-06 07:11:27 +00:00
commit e594c38f3f
1 changed files with 6 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

7
unsloth/trainer.py
View file

@ -23,7 +23,12 @@ from functools import wraps
import trl
import inspect
from trl import SFTTrainer
from . import is_bfloat16_supported
# Import via the concrete _utils module instead of the partially-initialized
# `unsloth` namespace. The MLX-aware `unsloth/__init__.py` loads the GPU
# surface via `from ._gpu_init import *`, which transitively triggers
# `from .trainer import *` BEFORE the star export has propagated back to
# `unsloth`, so `from . import is_bfloat16_supported` raises at import time.
from .models._utils import is_bfloat16_supported
from unsloth.utils import (
configure_padding_free,
configure_sample_packing,