diff --git a/quantize_streaming.py b/quantize_streaming.py index 67c1a05..aaf4882 100644 --- a/quantize_streaming.py +++ b/quantize_streaming.py @@ -71,7 +71,7 @@ def streaming_quantize(model_path: str, output_path: str): print(f" Saved {shard_name}") - del state_dict, quant_states + del state_dict gc.collect() torch.cuda.empty_cache()