Skip to content

Ideogram loading and lowvram fix #902

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
PlagueKind wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Ideogram loading and lowvram fix #902

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
87e6bfb
Ideogram loading and lowvram fix
PlagueKind Jun 19, 2026
3d65f4d
fix quantization of text encoder
PlagueKind Jun 20, 2026
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
35 changes: 27 additions & 8 deletions extensions_built_in/diffusion_models/ideogram4/ideogram4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
import os
from typing import List, Optional

import torch
import yaml
from tqdm import tqdm
from safetensors.torch import load_file, save_file

from toolkit.config_modules import GenerateImageConfig, ModelConfig, NetworkConfig
Expand Down Expand Up @@ -360,15 +360,15 @@ def load_model(self):
self.print_and_status_update("Loading Ideogram4 model")
base = self.model_config.name_or_path

# ------------------------------------------------------------
# MODEL LOADING + QUANTIZE + OFFLOAD
# ------------------------------------------------------------
transformer = self._load_transformer(base)

if self.model_config.quantize:
self.print_and_status_update("Quantizing Transformer")
quantize_model(self, transformer)
flush()
else:
transformer.to(self.device_torch, dtype=dtype)
flush()

if (
self.model_config.layer_offloading
Expand All @@ -384,20 +384,39 @@ def load_model(self):
transformer.llm_cond_proj,
],
)

elif self.model_config.low_vram:
self.print_and_status_update("Moving transformer to CPU")
transformer.to("cpu")

else:
# quantize_model leaves the model on CPU; make sure it lands on device.
self.print_and_status_update("Moving transformer to device")
transformer.to(self.device_torch)
flush()

tokenizer, text_encoder = self._load_text_encoder(base)

# always start device-agnostic
text_encoder.to("cpu")

if self.model_config.quantize_te:
self.print_and_status_update("Quantizing Text Encoder")
text_encoder.to(self.device_torch)
quantize(text_encoder, weights=get_qtype(self.model_config.qtype_te))
freeze(text_encoder)
quantization_type = get_qtype(self.model_config.qtype_te)
te_blocks = getattr(getattr(text_encoder, 'language_model', None), 'layers', None)
if te_blocks is not None and self.model_config.low_vram:
for block in tqdm(te_blocks):
block.to(self.device_torch, dtype=self.torch_dtype)
quantize(block, weights=quantization_type)
freeze(block)
block.to("cpu")
text_encoder.to(self.device_torch, dtype=self.torch_dtype)
quantize(text_encoder, weights=quantization_type)
freeze(text_encoder)
text_encoder.to("cpu")
else:
text_encoder.to(self.device_torch)
quantize(text_encoder, weights=quantization_type)
freeze(text_encoder)
flush()
if (
self.model_config.layer_offloading
Expand Down