| { | |
| "format": "mindpipe_hf_save_pretrained_debug", | |
| "algorithm": "smoothquant", | |
| "model_path": "/mnt/82_store/LLM-weights/Qwen3.5-4B", | |
| "result_dir": "/mnt/82_store/wxx/HWQuant/Mindpipe/results/Qwen3.5-4B/smoothquant/smoothquant_w8a8_seq512_alpha0p85", | |
| "weight_bits": 8, | |
| "activation_bits": 8, | |
| "sequence_length": 512, | |
| "group_size": 128, | |
| "output_dir": "/mnt/82_store/wxx/HWQuant/Mindpipe/tmp_hf_save/Qwen3.5-4B-SmoothQuant-w8a8/_pre_save_workflow", | |
| "pre_save_metrics": { | |
| "perplexity": 12.321548162109526, | |
| "evaluation_dataset": "wikitext2", | |
| "sequence_length": 512, | |
| "evaluated_chunks": 4, | |
| "batch_size": 1, | |
| "elapsed_seconds": 2.5450740084052086, | |
| "tokens_per_second": 803.1200637975982, | |
| "model_path": "/mnt/82_store/LLM-weights/Qwen3.5-4B", | |
| "quantization_algorithm": "smoothquant", | |
| "weight_bits": 8, | |
| "activation_bits": 8, | |
| "device": "cuda:0", | |
| "dtype": "float16", | |
| "artifacts_path": "artifacts.json" | |
| }, | |
| "load_note": "Use MindPipe to rebuild the compression structure before loading the HF state dict." | |
| } |