Skip to content

Enable Chroma finetune within 32gb RAM & 24gb VRAM #488

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
+13 −4
Open

Enable Chroma finetune within 32gb RAM & 24gb VRAM #488

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
12ba0dd
fix FakeTextEncoder bug when loading
stefanasandei Oct 29, 2025
ef8e579
enable Chroma loading within 32gb RAM & 24gb VRAM
stefanasandei Oct 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 13 additions & 4 deletions extensions_built_in/diffusion_models/chroma/chroma_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,8 +167,13 @@ def load_model(self):

chroma_params.depth = double_blocks
chroma_params.depth_single_blocks = single_blocks

# load Chroma into RAM in bfloat16, go back to fp32 afterwards
def_dtype = torch.get_default_dtype()
torch.set_default_dtype(torch.bfloat16)
transformer = Chroma(chroma_params)

torch.set_default_dtype(def_dtype)

# add dtype, not sure why it doesnt have it
transformer.dtype = dtype
# load the state dict into the model
Expand Down Expand Up @@ -420,9 +425,13 @@ def get_model_has_grad(self):
return self.model.final_layer.linear.weight.requires_grad

def get_te_has_grad(self):
# return from a weight if it has grad
return self.text_encoder[1].encoder.block[0].layer[0].SelfAttention.q.weight.requires_grad

from toolkit.unloader import FakeTextEncoder

te = self.text_encoder[1]
if isinstance(te, FakeTextEncoder):
return False
return te.encoder.block[0].layer[0].SelfAttention.q.weight.requires_grad

def save_model(self, output_path, meta, save_dtype):
if not output_path.endswith(".safetensors"):
output_path = output_path + ".safetensors"
Expand Down