cmedina/agenx-lora-training
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: remove PEFT prepare (causes OOM)

Browse Source
This commit is contained in:
Christian Medina
2026-07-02 20:17:23 -04:00
parent ac891fe25b
commit 3716051abc
1 changed files with 1 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
quantize_to_bnb.py
View File

@@ -25,14 +25,8 @@ def quantize_model(model_path, output_path):
total_params = sum(p.numel() for p in model.parameters()) total_params = sum(p.numel() for p in model.parameters())
print(f" Total parameters: {total_params / 1e9:.2f}B") print(f" Total parameters: {total_params / 1e9:.2f}B")
# Apply PEFT prepare for k-bit training
print("\nApplying PEFT prepare_model_for_kbit_training...")
from peft import prepare_model_for_kbit_training
model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=False)
print("✓ Model prepared for k-bit training")
# Quantize using BnB's actual API # Quantize using BnB's actual API
print("Quantizing with BnB 4-bit...") print("\nQuantizing with BnB 4-bit...")
from bitsandbytes.nn import Linear4bit from bitsandbytes.nn import Linear4bit
from torch import nn from torch import nn