Run the Dynamo frontend and the AgentX client on a dedicated node

On the head node the frontend and the client grew next to two DEP2 prefill instances (Engram CPU tables plus 70 GiB
host KV per rank) until the 8P1D c5120 job was OOM-killed during warmup. Both now share one CPU-only node
(frontend and benchmark placement dedicated, colocated), which adds one node per disaggregated point and leaves the
GPU count unchanged. The frontend is pinned to NUMA node 0 (frontend.numa_bind) now that worker steps get every CPU.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: Juntian777

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv41flash/vllm/gb300-fp4/agentx/disagg-variants.yaml

@@ -52,9 +52,15 @@ base:
       type: etcd
       placement:
         node: infra
+  # The Dynamo frontend and the AgentX client share one CPU-only node: next to two DEP2 prefill instances (Engram CPU
+  # tables plus 70 GiB host KV per rank) the client and frontend memory at conc 5120 exhausted the head node. The
+  # frontend is pinned to NUMA node 0.
   frontend:
     type: dynamo
     enable_multiple_frontends: false
+    numa_bind: true
+    placement:
+      node: dedicated
     args:
       router-mode: "kv"
       router-session-affinity-ttl-secs: 14400
@@ -186,6 +192,8 @@ base:
   benchmark:
     type: custom
     command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+    placement:
+      node: dedicated
     env:
       INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
       RESULT_DIR: "/logs/agentic"