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

fix: use BitsAndBytes INT4 for PEFT compatibility

Browse Source
This commit is contained in:
Christian Medina
2026-07-01 00:34:34 -04:00
parent d7301d37dd
commit be7b589c6b
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
training/scripts/train.py
View File

@@ -32,15 +32,17 @@ def train(config_path):
print(f"Loading model: {config['base_model']}") print(f"Loading model: {config['base_model']}")
# Load model - preserve NVFP4 quantization # Load model with BitsAndBytes INT4 (PEFT supports this)
print(f"Loading model: {config['base_model']}") print(f"Loading model: {config['base_model']}")
from transformers import BitsAndBytesConfig
model = AutoModelForCausalLM.from_pretrained( model = AutoModelForCausalLM.from_pretrained(
config["base_model"], config["base_model"],
torch_dtype=torch.float16, # Keep NVFP4 quantized quantization_config=BitsAndBytesConfig(load_in_4bit=True),
device_map="cpu", # Load to CPU first device_map="cpu", # Load to CPU first
trust_remote_code=True, trust_remote_code=True,
) )
print("Model loaded to CPU.") print("Model loaded with INT4 quantization.")
# Add LoRA # Add LoRA
lora_config = LoraConfig( lora_config = LoraConfig(