recipes: dflash2-2x BLOCKED on r28 too — DFlash kernel BF16-KV-only vs fp8_ds_mla engine KV (fork-side)

recipes/glm53-flash-r28-dflash2-2x/recipe.toml

@@ -1,3 +1,12 @@
+# BLOCKED 2026-09-08 (r28 re-test): the r28 tree advanced the DCP+DFlash
+# boundary logic (fa4720ea) but the drafter still dies — now at
+# backend selection, not the old C4 assert: the DFlash attention kernel
+# is BF16-KV-only ("K and V caches must be bfloat16") while the GLM-5.3
+# engine runs fp8_ds_mla KV (fork default; no per-drafter dtype override
+# in this tree) -> "No valid attention backend found ... kv_cache_dtype=
+# fp8_ds_mla, use_mla=False, block_size=256". Fork-side work: teach
+# DFlashAttention the fp8_ds_mla KV, or add a drafter bf16-KV override.
+#
 # DFlash2 re-test on the r28 full stack (2026-09-08, operator: "может он
 # заработал"). The 2026-09-06 blocker (C4 parent-page stride assert + FP8-KV
 # drafter backend rejection) targeted the r27 tree; r28's vLLM carries