Conversion recipe

README.md

@@ -67,7 +67,9 @@ docker run --rm -v ~/models:/models --entrypoint rad-convert radiance \
     -o /models/qwen3.6-35b-a3b-fp8.rad
 ```
 
-The recipe (`rad-info --recipe qwen3.6-35b-a3b-fp8.rad`); everything it does not name is the checkpoint's own:
+The recipe file both commands name is in this repository as `q36-35b-a3b-fp8.recipe`, with its
+comments. Its rules, as the container records them (`rad-info --recipe qwen3.6-35b-a3b-fp8.rad`);
+everything they do not name is the checkpoint's own:
 
 ```
 output.weight rtn codes=fp8_e4m3 block=128x128 scale=bf16

q36-35b-a3b-fp8.recipe

@@ -0,0 +1,15 @@
+# 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