compressed_config.json

{
  "format": "mindpipe_hf_save_pretrained_debug",
  "algorithm": "smoothquant",
  "model_path": "/mnt/82_store/LLM-weights/Meta-Llama-3.1-8B-Instruct",
  "result_dir": "/mnt/82_store/wxx/HWQuant/Mindpipe/results/Meta-Llama-3.1-8B-Instruct/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/Meta-Llama-3.1-8B-Instruct-SmoothQuant-w8a8/_pre_save_workflow",
  "pre_save_metrics": {
    "perplexity": 9.494966108833099,
    "evaluation_dataset": "wikitext2",
    "sequence_length": 512,
    "evaluated_chunks": 4,
    "batch_size": 1,
    "elapsed_seconds": 0.8245680080726743,
    "tokens_per_second": 2478.8737617623524,
    "model_path": "/mnt/82_store/LLM-weights/Meta-Llama-3.1-8B-Instruct",
    "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."
}