vision 2x: parked note — B12X sparse o_proj DeepGEMM fp8_einsum assert at TP2 (his verified values); 4x DCP2 is the profile lane

recipes/dsv4-vision-aiden-2x/recipe.toml

@@ -17,6 +17,15 @@
 # DCP1 DSpark profile is 73.8-75.2 tok/s; 2x should sit below that.
 # Boot markers: "Model loading took ~82 GiB" → "GPU KV cache size" →
 # "Application startup complete". Smoke: 47*83 → 3901 + image_url ask.
+#
+# PARKED 2026-09-03 (2x aiden image): after clean weight load (82 GiB), the
+# warmup dies in the B12X sparse attention O-proj —
+# DeepGEMM fp8_einsum "layout.hpp:39: t.dim() == N" (shared by the
+# FlashMLA/FlashInfer sparse paths). Reproduced on his verified TP2 values
+# (GMU 0.85, KV pin off) and with the KV pin at 12 GiB. The 4x production
+# shape (DCP2) is the lane that carries the profile; 2x vision stays on our
+# proven dsv4-flash-vision overlay image (88/100) until the TP2 o_proj path
+# is sorted.
 
 engine = "vllm"
 name = "DeepSeek V4 Flash Vision-Exp (aiden image, 2x Spark, DSpark-6)"