mistral-large4-nvfp4-8x-sglang: fix 288K KV pool cap -> disable prefill graphs, mem-fraction 0.88 (boot-7: pool 288192/524288, ~17 GB eaten by prefill graphs)

recipes/mistral-large4-nvfp4-8x-sglang/recipe.toml

@@ -41,8 +41,13 @@ entrypoint = ["/opt/nvidia/nvidia_entrypoint.sh", "sglang", "serve"]
 args = [
     # Stable gateway alias (same as the vLLM lane).
     "--served-model-name", "mistral-large-4",
-    "--mem-fraction-static", "0.85",
+    # Boot-7 KV pool was 288K tokens (needs 524K): after weights 70.1 GiB only
+    # 15.9 GB reached the pool — ~17 GB eaten by prefill cuda graphs (bs list
+    # up to 2048), DSpark graphs and workspaces. Disable the prefill graphs
+    # (the KV/token is fine, ~51 KB) and raise the fraction to 0.88.
+    "--mem-fraction-static", "0.88",
     "--context-length", "524288",
+    "--disable-prefill-cuda-graph",
     "--chunked-prefill-size", "2048",
     # The embedded DSpark draft (do NOT pass --speculative-draft-model-path —
     # the checkpoint carries it).