Commit History
Yesterday
-
Update to validated v2 power-of-two FP8 weights and activation scales
Verified signatureReplace 48 expert-weight shards with the v2 power-of-two row conversion. Select power-of-two activation scales by default and update fidelity, accuracy, performance, and multimodal validation documentation. Vision, dense, and MTP weights retain source precision. vLLM implementation is supplied by the deployment image.6eaa9b7 -
Add MI300X row-scaled FP8 Inkling 2 checkpoint with vision and DSpark
Verified signature78a4761
