compressed_config.json

{
  "format": "mindpipe_hf_save_pretrained_debug",
  "algorithm": "gptq",
  "model_path": "/mnt/82_store/LLM-weights/openbmb/MiniCPM-V",
  "result_dir": "/mnt/82_store/wxx/HWQuant/Mindpipe/results/MiniCPM-V/gptq/gptq_w4a16_seq512",
  "weight_bits": 4,
  "activation_bits": 16,
  "sequence_length": 512,
  "group_size": 128,
  "output_dir": "/mnt/82_store/wxx/HWQuant/Mindpipe/tmp_hf_save/MiniCPM-V-GPTQ-w4a16g128/_pre_save_workflow",
  "pre_save_metrics": {
    "perplexity": 9.280270004821208,
    "evaluation_dataset": "wikitext2",
    "sequence_length": 512,
    "evaluated_chunks": 4,
    "batch_size": 1,
    "elapsed_seconds": 0.21972699742764235,
    "tokens_per_second": 9302.452697798793,
    "model_path": "/mnt/82_store/LLM-weights/openbmb/MiniCPM-V",
    "quantization_algorithm": "gptq",
    "weight_bits": 4,
    "activation_bits": 16,
    "device": "cuda:0",
    "dtype": "float16",
    "artifacts_path": "artifacts.json"
  },
  "load_note": "Use MindPipe to rebuild the compression structure before loading the HF state dict."
}