| # q36-35b-a3b-fp8.recipe -- Qwen3.6-35B-A3B-FP8 with its MTP head, as the q36-35b-a3b-fp8.rad | |
| # container serves it. | |
| # | |
| # rad-convert <Qwen3.6-35B-A3B-FP8 checkpoint> --recipe data/recipes/q36-35b-a3b-fp8.recipe \ | |
| # -o q36-35b-a3b-fp8.rad | |
| # | |
| # The linears -- attention, delta net, shared and routed experts, the MTP layer's -- are | |
| # block-scaled fp8 in the checkpoint and kept as they are. What the recipe adds: | |
| # | |
| # - the lm_head, block fp8: the widest GEMM in the model at half the bytes | |
| # - the MTP head's 2-bit copy of the lm_head (u2 codes, an f16 scale and a u8 zero a group of | |
| # 128), which the draft rounds read in place of the full head | |
| output.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16 | |
| mtp.draft_head.weight rtn codes=u2 zero=u8 group=128 scale=f16 | |