Quantization¶
FP8¶
FP8 checkpoints store two-dimensional model weights as float8_e4m3fn and
retain biases, normalization scales, and mask embeddings in their original
floating-point type. They load through the normal constructor and can remain
FP8 in the Triton CUDA backend.
relational-transformers-quantize \
RelativeDB/rt-j-fp16 RelativeDB/rt-j-fp8-local
Int8¶
Int8 checkpoints use one floating-point scale per output row. Portable PyTorch dequantizes them during loading. RelativeDB’s native engine can operate on the quantized representation directly.
Int4¶
Int4 checkpoints pack two weights per byte in groups of 32 and store a scale and minimum per group. They provide the smallest artifact and the largest accuracy tradeoff.
Validation¶
Always compare logits, ranking, and task metrics against FP16 on representative
relational contexts. The test suite validates every published format when
RUN_HUB_TESTS=1 and compares Triton with PyTorch when RUN_CUDA_TESTS=1.