perf(agentic): probe GB200 frontier tails

Author: cquil11

.github/configs/nvidia-master.yaml

@@ -12827,6 +12827,21 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-3p2d-tep8-tp8:
     agentic-coding:
     - duration: 1800
       search-space:
+      # Ultra-high-interactivity probes below the historical c16 endpoint.
+      - spec-decoding: none
+        conc-list: [4, 8]
+        prefill:
+          num-worker: 3
+          tp: 8
+          ep: 8
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p2d-tep8-tp8-agentic.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 1
+          dp-attn: false
       - spec-decoding: none
         conc-list: [16, 24, 32, 40]
         prefill:
@@ -12901,6 +12916,21 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-2p1d-dep8-dep8:
           tp: 8
           ep: 8
           dp-attn: true
+      # Exploratory tail beyond the measured c160 normalized-throughput peak.
+      - spec-decoding: none
+        conc-list: [192, 224, 256]
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-agentic.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
 
 # Split from dsr1-fp4-b200-dynamo-trt: agentic-coding scenario only.
 dsr1-fp4-b200-dynamo-trt-agentic:

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/GB200_VLLM_AGENTIC_SWEEP_NOTES.md

@@ -1348,3 +1348,12 @@ The AIPerf submodule also merges `ajcasagrande/aiperf:ajc/agentx` at
 resulting merge commit retains both histories and supplies the latest AgentX
 warmup, timing, Weka-loader, metrics, Dynamo-session, and process-lifecycle
 fixes.
+
+## Exploratory Frontier Tails
+
+Five boundary probes extend the validated core sweep. The 3P/2D TEP8/TP8
+topology adds c4/c8 below its historical c16 endpoint to measure the
+ultra-high-interactivity limit. The 2P/1D DEP8/DEP8 topology adds
+c192/c224/c256 beyond the current c160 normalized-throughput leader to locate
+the actual one-decode saturation knee. These are isolated batches so they do
+not force reruns of already selected points when a boundary probe overloads.