NameError: '_should_use_gpt_oss_bnb4bit' is not defined due to function definition ordering in gpt_oss.py

[HwangHanJae]

NameError: '_should_use_gpt_oss_bnb4bit' is not defined due to function definition ordering in gpt_oss.py

Environment

• unsloth==2026.3.11
• unsloth_zoo==2026.3.5
• Python 3.12
• PyTorch with CUDA

Bug Description

In unsloth_zoo/temporary_patches/gpt_oss.py, patch_gpt_oss_bnb4bit_auto is registered to TEMPORARY_PATCHES at line 1132, but the function _should_use_gpt_oss_bnb4bit() that it calls internally is defined much later at line 1351.

Between these two lines (~220 lines of module-level code), there are runtime statements that can fail:

# line 1132 - function reference registered
TEMPORARY_PATCHES.append(patch_gpt_oss_bnb4bit_auto)

# line 1141 - module-level CUDA call (can fail: OOM, no GPU context, etc.)
device_memory = torch.cuda.memory.mem_get_info(0)[-1]

# ... ~200 lines of module-level code ...

# line 1336 - module-level import
from .moe_utils import (...)

# line 1351 - never reached if anything above fails
def _should_use_gpt_oss_bnb4bit() -> bool:
    ...