Update loader.py

This commit is contained in:
Daniel Han 2025-08-08 11:38:03 -07:00
commit 628bb6c97f

View file

@ -594,6 +594,9 @@ class FastModel(FastBaseModel):
elif "gpt-oss" in lowered_model_name:
os.environ["UNSLOTH_DISABLE_STATIC_GENERATION"] = "1"
if not load_in_4bit:
# CCE fails on Tesla T4
# OutOfResources: out of resource: shared memory, Required: 98304, Hardware limit: 65536. Reducing block sizes or `num_stages`
os.environ["UNSLOTH_ENABLE_CCE"] = "0"
# Only upcast MoE biases for MXFP4, not BnB
os.environ["UNSLOTH_FORCE_CUSTOM_DTYPE"] = \
"all;None;None;"\