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

Update vision.py

Browse source
This commit is contained in:
Daniel Han 2025-10-01 04:36:58 -07:00
commit f78abbc751
1 changed files with 22 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

22
unsloth/models/vision.py
View file

@ -557,8 +557,30 @@ class FastBaseModel:
nbytes = embed_tokens.weight.numel() * embed_tokens.weight.itemsize
ngb = round(nbytes / 1024 / 1024 / 1024, 2)
print(f"Unsloth: Offloading embeddings to RAM to save {ngb} GB.")
# model.device also will change to CPU so change back
m = model
while hasattr(m, "model"):
if hasattr(m, "device"): m._old_device_ = m.device
m = m.model
if hasattr(m, "device"): m._old_device_ = m.device
# Move embeddings to CPU
embed_tokens.to("cpu")
# model.device also will change to CPU so change back
m = model
while hasattr(m, "model"):
if hasattr(m, "device") and hasattr(m, "_old_device_"):
try: m.device = m._old_device_
except: pass
del m._old_device_
m = m.model
if hasattr(m, "device"):
try: m.device = m._old_device_
except: pass
del m._old_device_
# Add hooks to move inputs to CPU and back to CUDA
# [TODO] Doesn't seem to work!
# def pre_hook(module, args):