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

GptAttention turn training off during inference (#3289)

Browse source
This commit is contained in:
DoubleMathew 2025-09-08 15:47:32 -05:00 • committed by GitHub
commit c9c068fa0b
2 changed files with 4 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

2
unsloth/models/llama.py
View file

@ -2986,6 +2986,7 @@ class FastLlamaModel:
_for_inference(m)
m = m.model
_for_inference(m)
model.eval() # to turn off training on modules deeper in
# Since transformers 4.53, must turn off explicitly
for module in model.modules():
@ -3030,6 +3031,7 @@ class FastLlamaModel:
_for_training(m)
m = m.model
_for_training(m)
model.train() # to turn on training on modules deeper in
# Since transformers 4.53, must turn on explicitly
for module in model.modules():

2
unsloth/models/vision.py
View file

@ -772,6 +772,7 @@ class FastBaseModel:
_for_inference(m)
m = m.model
_for_inference(m)
model.eval() # to turn off training on modules deeper in
# Since transformers 4.53, must turn off explicitly
for module in model.modules():
@ -823,6 +824,7 @@ class FastBaseModel:
_for_training(m)
m = m.model
_for_training(m)
model.train() # to turn on training on modules deeper in
# Since transformers 4.53, must turn on explicitly
for module in model.modules():