Fix: skip .cuda() when low_memory_mode; switch default to nvfp4

2026-05-07 03:06:33 +00:00
parent b8bdd00d19
commit 116933dcf6
1 changed files with 1 additions and 1 deletions
--- a/quantize_modelopt.py
+++ b/quantize_modelopt.py
@@ -110,7 +110,7 @@ def main():

    model = AutoModelForCausalLM.from_pretrained(args.model, **model_kwargs)

-    if not args.use_seq_device_map:
+    if not args.use_seq_device_map and not args.low_memory_mode:
        model = model.cuda()

    # Build calibration dataloader