Expand description
Scalar Quantization (SQ) and Binary Quantization (BQ) for extreme compression.
§Compression Methods
- Scalar Quantization (SQ8) - 8-bit quantization (4x compression)
- Scalar Quantization (SQ4) - 4-bit quantization (8x compression)
- Binary Quantization (BQ) - 1-bit quantization (32x compression)
§Trade-offs
| Method | Compression | Recall | Speed |
|---|---|---|---|
| Float32 | 1x | 100% | 1x |
| SQ8 | 4x | 98-99% | 2-3x |
| SQ4 | 8x | 95-97% | 3-4x |
| BQ | 32x | 85-95% | 4-8x |
Structs§
- Binary
Quantizer - Binary quantizer
- Quantization
Benchmark - Benchmark quantization performance
- Scalar
Quantizer4 - 4-bit scalar quantizer
- Scalar
Quantizer8 - 8-bit scalar quantizer