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

[pre-commit.ci] auto fixes from pre-commit.com hooks

Browse source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot] 2026-03-17 08:24:17 +00:00
commit 7f5b7b368c
2 changed files with 12 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

5
unsloth/kernels/fp8.py
View file

@ -605,9 +605,8 @@ except:
pass
HAS_FBGEMM_FP8_OPS = (
hasattr(torch.ops, "fbgemm")
and hasattr(torch.ops.fbgemm, "quantize_fp8_per_row")
HAS_FBGEMM_FP8_OPS = hasattr(torch.ops, "fbgemm") and hasattr(
torch.ops.fbgemm, "quantize_fp8_per_row"
)

16
unsloth/models/loader.py
View file

@ -819,7 +819,9 @@ class FastLanguageModel(FastLlamaModel):
if load_in_fp8 != False:
_tag_model_with_fp8_torchao_config(model, fp8_mode)
if load_in_fp8 != False or _has_prequantized_fp8_config(model_name, token=token, trust_remote_code=trust_remote_code):
if load_in_fp8 != False or _has_prequantized_fp8_config(
model_name, token = token, trust_remote_code = trust_remote_code
):
maybe_patch_stacked_moe_expert_fp8_scales(
model,
model_name = model_name,
@ -1363,10 +1365,10 @@ class FastModel(FastBaseModel):
model_name = get_model_name(
model_name,
load_in_4bit,
load_in_fp8=load_in_fp8,
fast_inference=fast_inference,
token=token,
trust_remote_code=trust_remote_code,
load_in_fp8 = load_in_fp8,
fast_inference = fast_inference,
token = token,
trust_remote_code = trust_remote_code,
)
# Check if pre-quantized models are allowed
# AMD Instinct GPUs need blocksize = 128 on bitsandbytes < 0.49.2 (our pre-quants use blocksize = 64)
@ -1576,7 +1578,9 @@ class FastModel(FastBaseModel):
if load_in_fp8 != False:
_tag_model_with_fp8_torchao_config(model, fp8_mode)
if load_in_fp8 != False or _has_prequantized_fp8_config(model_name, token=token, trust_remote_code=trust_remote_code):
if load_in_fp8 != False or _has_prequantized_fp8_config(
model_name, token = token, trust_remote_code = trust_remote_code
):
maybe_patch_stacked_moe_expert_fp8_scales(
model,
model_name = model_name,