ling-31-flash-fp4: fix KV OOM — weights are 86.43 GiB/rank (replicated embeddings/lm_head/dense/MLA on top of 291.7/4), GMU 0.78 left 0.29 GiB KV (65536 needs 0.77, engine init refused). Raise GMU 0.86, drop max-model-len 32768 + max-num-batched-tokens 4096 (bench scenarios short)

recipes/ling-31-flash-fp4/recipe.toml

@@ -20,9 +20,12 @@
 # hub (local-only requant), so force cache-only resolution.
 #
 # TP4 chosen (not TP6) to leave 2 Sparks for the GLM-5.3-Flash-FP4-Local 2x
-# lane on the same 6 free nodes. 78 GB weights/rank at GMU 0.78 leaves ~15 GB
-# KV (MLA KV is compact) — if the GB10 UMA boot gate (vLLM #48140) refuses at
-# 65536, drop --max-model-len to 32768.
+# lane on the same 6 free nodes. Weights are 86.43 GiB/rank (experts shard /4
+# but embeddings/lm_head/dense/MLA replicate — ~13.5 GiB/rank overhead on top
+# of 291.7/4). At GMU 0.78 that left only 0.29 GiB KV (65536 needs 0.77 → engine
+# init refused, est. max len 15360). So: GMU 0.86 + max-model-len 32768 +
+# max-num-batched-tokens 4096 (bench scenarios are short; ~15 GiB KV is ample).
+# If the GB10 UMA boot gate (vLLM #48140) bites at 0.86, drop to 0.82 / 16384.
 #
 # Quality A/B note: tool-eval-bench quality is TP-invariant (same weight math),
 # so TP4 fp4 vs the fp8 (TP8) isolates the quantization delta. A valid fp8
@@ -46,10 +49,10 @@ args = [
     "--enable-auto-tool-choice",
     "--tool-call-parser", "ling3",
     "--trust-remote-code",
-    "--max-model-len", "65536",
-    "--gpu-memory-utilization", "0.78",
+    "--max-model-len", "32768",
+    "--gpu-memory-utilization", "0.86",
     "--max-num-seqs", "32",
-    "--max-num-batched-tokens", "8192",
+    "--max-num-batched-tokens", "4096",
 ]
 
 [env]