kindling recipes: drop_caches before_serve — the 0.88 GMU startup gate reads MemFree on GB10 UMA and random nodes carried ~15 GiB of weight/NFS page cache (106.53 free vs 107.03 required on rank1, the '2 of 4 nodes' boot failure); reclaim cache per node before containers start

recipes/glm53-kindling-2x/recipe.toml

@@ -38,6 +38,12 @@ pipeline_parallel_size = 1
 distributed_backend = "mp"
 shm_size = "32gb"
 
+# GB10 UMA gate (vLLM #48140): reclaim the page cache before the engine startup check reads MemFree — leftover weight/NFS page cache (~10-15 GiB)
+# tipped random nodes below the 0.88 GMU threshold at boot.
+before_serve = [
+  { command = "sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'", on_failure = "warn" },
+]
+
 args = [
     # --- upstream entrypoint.sh `exec vllm serve …`, TP=2 case block -----
     "--served-model-name", "glm53",

recipes/glm53-kindling-4x/recipe.toml

@@ -35,6 +35,12 @@ pipeline_parallel_size = 1
 distributed_backend = "mp"
 shm_size = "32gb"
 
+# GB10 UMA gate (vLLM #48140): reclaim the page cache before the engine startup check reads MemFree — leftover weight/NFS page cache (~10-15 GiB)
+# tipped random nodes below the 0.88 GMU threshold at boot.
+before_serve = [
+  { command = "sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'", on_failure = "warn" },
+]
+
 args = [
     # --- upstream entrypoint.sh `exec vllm serve …`, TP=4 case block -----
     "--served-model-name", "glm53",

recipes/glm53-kindling-6x/recipe.toml

@@ -44,6 +44,12 @@ pipeline_parallel_size = 1
 distributed_backend = "mp"
 shm_size = "32gb"
 
+# GB10 UMA gate (vLLM #48140): reclaim the page cache before the engine startup check reads MemFree — leftover weight/NFS page cache (~10-15 GiB)
+# tipped random nodes below the 0.88 GMU threshold at boot.
+before_serve = [
+  { command = "sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'", on_failure = "warn" },
+]
+
 args = [
     # --- upstream entrypoint.sh `exec vllm serve …`, TP=6 case block -----
     "--served-model-name", "glm53",