metadata
license: mit
base_model:
- zai-org/GLM-4.5-Air
tags:
- fp8
- quantized
- quark
- fp8_e4m3
base_model_relation: quantized
This is an AMD Quark-quantized GLM-4.5 Air in fp8. Quantized on and for GFX1100 cards, in this case 2x W7900.
This is my first quantized model and I'm still evaluating. It was calibrated with wikitext; assuming success, a future iteration will be calibrated on other datasets.
Quantized perplexity on wikitext: 4.96421480178833