gemma2-9b · fp8 · W8A16 · calibrated on mixed
Post-training quantization with NVIDIA ModelOpt (FP8_DEFAULT_CFG); activation quantizers were
disabled after calibration, so the exported model is weight-only.
| base model | google/gemma-2-9b-it |
| base dtype | float16 |
| mode | W8A16 — FP8 weights (E4M3, per-tensor scale), activations FP16 |
| calibration | calib3_mixed.txt, 512 samples, 512 tokens, variety mixed |
| calibration reaches weights | no — see below |
| weight MSE | 3.163557e-08 |
| weight quantizers | 294 |
| environment | torch 2.8.0+cu128 · transformers 4.57.6 · modelopt 0.46.0 |
Weight scales here are computed data-free from the weights themselves. With activations left in FP16 there is nothing for the calibration text to affect, so this checkpoint would be identical whatever text were used. It is a control the AWQ variants are measured against.
The checkpoint stores plain quantized weights; nothing inside it records the weight-only
mode. That lives in the repo name and calib_stats.json only.
Not loadable by plain transformers — config.json declares quantization type modelopt.
vllm serve Azzam15/gemma2-9b-fp8-w8a16-mixed --quantization modelopt
Not comparable to a W4A4 checkpoint: the gap between them is partly precision and partly which tensors were quantized at all.
- Downloads last month
- 9