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

Merge pull request #3224 from DefiWimar7/typos

Browse source 
chore: Fix Typos

Thank you @DefiWimar7
This commit is contained in:
Michael Han 2025-08-28 02:46:27 -07:00 • committed by GitHub
commit 20e6469959
2 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

2
unsloth/models/falcon_h1.py
View file

@ -583,7 +583,7 @@ def _fast_prepare_inputs_for_generation(
position_ids=None,
use_cache=True,
**kwargs,):
# Overwitten -- has a unique cache type, `FalconHybridMambaAttentionDynamicCache`
# Overwritten -- has a unique cache type, `FalconHybridMambaAttentionDynamicCache`
empty_past_kv = past_key_values is None
# If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens

4
unsloth/models/gemma.py
View file

@ -248,7 +248,7 @@ class GemmaFixedRotaryEmbedding(torch.nn.Module):
# in FP32. They are applied (multiplied) in FP32 as well.
self.current_rope_size = seq_len
# The difference is we do division explicity instead of t * (1/x) ie we do t/x.
# The difference is we do division explicitly instead of t * (1/x) ie we do t/x.
freq_exponents = (2.0 / self.dim) * (
torch.arange(self.dim // 2, dtype = torch.int64, device = "cpu").float()
)
@ -312,7 +312,7 @@ class GemmaFixedLinearScalingRotaryEmbedding(GemmaFixedRotaryEmbedding):
# in FP32. They are applied (multiplied) in FP32 as well.
self.current_rope_size = seq_len
# The difference is we do division explicity instead of t * (1/x) ie we do t/x.
# The difference is we do division explicitly instead of t * (1/x) ie we do t/x.
freq_exponents = (2.0 / self.dim) * (
torch.arange(self.dim // 2, dtype = torch.int64, device = "cpu").float()
)