# memebq

**Repository Path**: zhouyijava/memebq

## Basic Information

- **Project Name**: memebq
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-05-09
- **Last Updated**: 2026-08-28

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# MemeBQ:Memory Efficient Binary Quantization of LLMs

## Checkpoint usage

You can now enable per-layer checkpointing for long quantization jobs.

Basic usage:

```bash
python run.py <model> <dataset> <low_quant_method> <model_dtype> <drop_ratio> \
  --checkpoint --checkpoint_dir ./output/checkpoints
```

By default, checkpoints are saved every 2 layers. You can change it:

```bash
python run.py <model> <dataset> <low_quant_method> <model_dtype> <drop_ratio> \
  --checkpoint --checkpoint_interval 2
```

Resume behavior is automatic: when `--checkpoint` is enabled, the run will
auto-load the latest checkpoint for the current stage (if any).

Manual resume flag is no longer required.

```bash
python run.py <model> <dataset> <low_quant_method> <model_dtype> <drop_ratio> \
  --checkpoint --checkpoint_dir ./output/checkpoints
```

Optional:

- `--checkpoint_run_id <id>`: custom checkpoint identity for the run
- `--keep_checkpoint`: keep checkpoint files after successful completion
- `--checkpoint_interval <N>`: save every N layers (default `2`)

Checkpoint files are isolated by `run_id + stage`, so multi-stage quantization
flows can resume safely without file collisions.