diff --git a/quantize_streaming.py b/quantize_streaming.py index 84beac1..eac58e8 100644 --- a/quantize_streaming.py +++ b/quantize_streaming.py @@ -47,7 +47,7 @@ def streaming_quantize(model_path: str, output_path: str): if (output_path / shard_name).exists(): return - print(f"[{idx+1}/{len(shards)}] {Path(shard_file).name} (GPU {gpu_id})") + print(f"[{idx}/{len(shards)}] {Path(shard_file).name} (GPU {gpu_id})") # Load to assigned GPU state_dict = load_file(shard_file, device=f"cuda:{gpu_id}")