MedGemmaImpact/medgemma-1.5-4b-it-fp8

Quantized checkpoint of google/medgemma-1.5-4b-it using FP8 via llmcompressor.oneshot.

Calibration

• dataset: medqa
• num_samples: 512
• max_seq_length: 1024

Notes

• Vision modules are excluded from quantization to avoid breaking the vision encoder/projector.
• This repo contains the save_pretrained output produced by llmcompressor (plus the processor/config files).