glm53-fp4local-2x: cold_args GMU 0.72→0.75 — with correct TP2 the per-rank footprint is ~87 GiB (83 weights + overhead), 0.72 left 0.47 GiB KV vs 1.07 needed at 16384; 0.75 is the boot-safe floor (the cold tier is the ctx/seq/batched reduction)

recipes/glm53-flash-fp4-local-2x/recipe.toml

@@ -74,12 +74,12 @@ cold_args = [
     "--max-model-len", "16384",
     "--max-num-seqs", "1",
     "--max-num-batched-tokens", "2048",
-    # GMU 0.72, not lower: weights are ~83 GiB/rank (166 GB / 2) = 0.683 of
-    # the 121.63 GiB unified pool, so anything below ~0.70 cannot fit the
-    # weights at all (negative KV). 0.72 leaves ~4 GiB for KV + overhead —
-    # boot-safe for 16384 ctx / 1 seq, and colder than the 0.75 default
-    # attempt that stalled the nodes.
-    "--gpu-memory-utilization", "0.72",
+    # GMU 0.75: with the CORRECT TP2 the per-rank footprint is ~87 GiB
+    # (83 GiB weights + overhead), so 0.72 left only 0.47 GiB KV vs the
+    # 1.07 GiB needed at 16384 ctx — the boot-safe floor is the default GMU;
+    # the cold tier is the ctx/seq/batched reduction. (The earlier stalls were
+    # the missing distributed_backend — each node loaded the full model.)
+    "--gpu-memory-utilization", "0.75",
 ]
 
 [env]