ArabicFP4 — FP8 with Arabic calibration
Collection
21 items • Updated • 1
Post-training W8A8 quantization with NVIDIA ModelOpt (FP8_DEFAULT_CFG).
| base model | inceptionai/jais-family-6p7b-chat |
| base dtype | bfloat16 |
| precision | FP8 (W8A8) |
| calibration variety | 64 MSA + 64 Gulf |
| calibration file | calib3_mixed.txt |
| calibration samples | 128 dialogues, truncated at 128 tokens |
| calibration source | Almheiri/ArabCulture-Dialogue rev 9acd60cbbb4f, seed 1448 |
| weight MSE | 7.106e-07 |
| activation quantizers calibrated | 225/225 |
| environment | torch 2.8.0+cu128 · transformers 4.57.6 · modelopt 0.45.0 |
FP8 weight scales are data-free. What the calibration variety changes is the static
per-tensor activation scales, exported as input_scale. Sibling checkpoints
-fp8-msa, -fp8-gulf and -fp8-mixed differ in exactly that and nothing else.
Not loadable by plain transformers — config.json declares quantization type modelopt.
vllm serve NouraAlqasim/jais-6.7b-fp8-mixed --quantization modelopt
Base model
inception42/jais-family-6p7b