recipes: nvfp4-lab 2x — 256K context (2.4 GiB KV budget at GMU 0.88)

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

@@ -50,10 +50,12 @@ args = [
     # GMU is ONLY a startup guardrail: 0.88 = 107.0 GiB, below any
     # realistic free-memory dip on GB10's 121.63 GiB (#48140).
     "--gpu-memory-utilization", "0.88",
-    # no KV pin: ~93 GiB weights/rank leaves ~10 GiB — let the engine
-    # auto-size the pool (a fixed pin overshoots the tighter budget)
+    # no KV pin: ~93 GiB weights/rank leaves only ~2.4 GiB KV at GMU 0.88
+    # (the re-cut is heavier than the old 186 GB cut) — the engine needs
+    # 3.73 GiB for one 524288-token request, so cap the context at 256K
+    # (needs ~1.9 GiB; bench contexts are far below that)
     "--block-size", "256",
-    "--max-model-len", "524288",
+    "--max-model-len", "262144",
     "--max-num-seqs", "16",
     "--max-num-batched-tokens", "4096",
     "--enable-chunked-prefill",