lab-dflash2: parked note — lab image GLM5Next lacks SupportsEagle3 (dflash requires aux hidden states); the A/B moves to the aiden image

recipes/glm53-flash-nvfp4-lab-dflash2/recipe.toml

@@ -1,3 +1,11 @@
+# PARKED 2026-09-04: the lab image's GLM5Next target model does not implement
+# the SupportsEagle3 interface, but its model_runner requires aux hidden-state
+# outputs for the dflash method (model_runner.py:240) → "Model does not
+# support EAGLE3 interface" at boot. The lab's own qualified DFlash2 runs use
+# a newer fork with the aux-capture plumbing; porting it here is not worth it
+# tonight. The competitive A/B lives on the aiden image (glm53-flash-aiden-4x-
+# labdrafter) which ships the DFlash2 overlay.
+#
 # NOTE: the lab's DFlash2 checkpoints are MXFP8-quantized (weight_scale
 # tensors) and their config names the arch DFlash2DraftModel — the lab image
 # registers the drafter under DFlashDraftModel and its loader has no MXFP8