computer-10-fp8

Per-row FP8 (e4m3) weight quantization of cosmicoptima/computer-10, made to fit a single 96 GB GPU.

Every *_proj.weight in the 80 decoder layers is stored as float8_e4m3fn with a sibling *_proj.weight_scale (float32, one per output row, amax / 448); dequantize as weight.to(bf16) * weight_scale[:, None]. Embeddings, LM head and norms are unchanged bf16. There is no quantization_config; load the layers yourself (see the Space that uses it).

Prompt format and sampling are unchanged from the original: see its card.