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

fix: use float16 dtype to preserve NVFP4 quantization

Browse Source
This commit is contained in:
Christian Medina
2026-07-01 00:08:42 -04:00
parent af09d47c6f
commit d7301d37dd
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
training/scripts/train.py
View File

@@ -32,14 +32,15 @@ def train(config_path):
print(f"Loading model: {config['base_model']}") print(f"Loading model: {config['base_model']}")
# Load model to CPU first, DeepSpeed will distribute # Load model - preserve NVFP4 quantization
print("Loading model to CPU...") print(f"Loading model: {config['base_model']}")
model = AutoModelForCausalLM.from_pretrained( model = AutoModelForCausalLM.from_pretrained(
config["base_model"], config["base_model"],
torch_dtype=torch.float16, # Keep NVFP4 quantized
device_map="cpu", # Load to CPU first device_map="cpu", # Load to CPU first
trust_remote_code=True, trust_remote_code=True,
) )
print(f"Model loaded to CPU. Moving to GPU with DeepSpeed...") print("Model loaded to CPU.")
# Add LoRA # Add LoRA
lora_config = LoraConfig( lora_config = LoraConfig(