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

Fix typos in comment (#3557)

Browse source
This commit is contained in:
mk0walsk 2025-11-06 05:29:36 +02:00 • committed by GitHub
commit 07f0dfe0dc
2 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

8
unsloth/kernels/fp8.py
View file

@ -323,8 +323,8 @@ class FP8BlockQuantLinear(torch.autograd.Function):
assert block_size is not None, "block_size is not set"
if triton.cdiv(m, block_size[0]) != p or triton.cdiv(n, block_size[1]) != q:
if triton.cdiv(m, block_size[0]) == q and triton.cdiv(n, block_size[1]) == p:
# weights are tranposed during backward pass for training :)
# We tranpose weight scale to counter that. Note that transposing weight would cause issues with matmul with input X
# weights are transposed during backward pass for training :)
# We transpose weight scale to counter that. Note that transposing weight would cause issues with matmul with input X
weight_scale = weight_scale.T
else:
raise ValueError(f"Weight shape {weight.shape} and scales shape {weight_scale.shape} is not compatible with block size {block_size}")
@ -437,8 +437,8 @@ class FP8_torch_linear(torch.autograd.Function):
if triton.cdiv(m, bs_n) != p or triton.cdiv(n, bs_k) != q:
if triton.cdiv(m, bs_n) == q and triton.cdiv(n, bs_k) == p:
# weights are tranposed during backward pass for training :)
# We tranpose weight scale to counter that. Note that transposing weight would cause issues with matmul with input X
# weights are transposed during backward pass for training :)
# We transpose weight scale to counter that. Note that transposing weight would cause issues with matmul with input X
weight_scale = weight_scale.T
else:
raise ValueError(f"Weight shape {weight.shape} and scales shape {weight_scale.shape} is not compatible with block size {block_size}")

2
unsloth/models/__init__.py
View file

@ -22,7 +22,7 @@ from .granite import FastGraniteModel
try:
from .falcon_h1 import FastFalconH1Model
except:
# transformers_version < 4.53.0 does not have falcon_h1 so silenty skip it for now
# transformers_version < 4.53.0 does not have falcon_h1 so silently skip it for now
pass
from .dpo import PatchDPOTrainer, PatchKTOTrainer
from ._utils import is_bfloat16_supported, is_vLLM_available, __version__