vision 4x: parked note — same o_proj einsum assert in his exact production shape; image-level DeepGEMM layout issue on our GB10s

recipes/dsv4-vision-aiden-4x/recipe.toml

@@ -1,4 +1,11 @@
-# DeepSeek-V4-Flash-Vision-Exp on 4× GB10 — aiden production profile. DEFERRED.
+# DeepSeek-V4-Flash-Vision-Exp on 4× GB10 — aiden production profile. PARKED.
+#
+# 2026-09-03: parked — this image's FP8 O-proj einsum path (B12X/FlashMLA/
+# FlashInfer sparse attention shared kernel, DeepGEMM fp8_einsum) hard-asserts
+# "layout.hpp:39: t.dim() == N" at warmup on our GB10s in EVERY config tried
+# (2x TP2 GMU 0.85 no-pin; 4x TP4+DCP2 — his exact production shape). It is a
+# kernel/JIT layout issue in the image on our hardware, not a flag difference.
+# Vision lanes stay on our proven dsv4-flash-vision overlay image (2x: 88/100).
 #
 # Source: https://aidenle.com/recipes/deepseek-v4-flash-vision-4x-dgx-spark/
 # TP4 across 4 nodes with **DCP2 inside it** and DSpark-6 running together —