[Klaud Cold] Add conc 10/12/14 to kimik3-fp4-b300-vllm-agentic-dspark and move it to cluster:b300-dsxe / 为 kimik3-fp4-b300-vllm-agentic-dspark 新增并发 10/12/14 并迁移至 cluster:b300-dsxe

benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh

@@ -5,7 +5,7 @@ set -x
 # Agentic trace replay for Kimi-K3 (MXFP4) on B300: TP8 x DCP8, TokenspeedMLA,
 # Mooncake as the external KV tier. Concurrency selects the arm:
 #   conc <= 8   DSpark level 7, golden AL 3.84
-#   conc 16     DSpark level 3, golden AL 3.00
+#   conc 10-16  DSpark level 3, golden AL 3.00
 #   conc >  16  no drafting
 # Keep the arms disjoint in concurrency: exp-name carries conc and spec but not
 # the arm, so a shared concurrency would collide.

configs/nvidia-master.yaml

@@ -1489,7 +1489,7 @@ kimik3-fp4-b300-vllm-agentic-dspark:
   image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
-  runner: cluster:b300-nv
+  runner: cluster:b300-dsxe
   precision: fp4
   framework: vllm
   multinode: false
@@ -1502,9 +1502,9 @@ kimik3-fp4-b300-vllm-agentic-dspark:
       search-space:
       # TP8-only: a ~1.5 TB MXFP4 checkpoint does not fit below 8 GPUs.
       # One entry for every arm: the recipe drafts at DSpark 7 up to conc 8,
-      # DSpark 3 at conc 16, and not at all above. Keep the concurrencies
+      # DSpark 3 at conc 10-16, and not at all above. Keep the concurrencies
       # disjoint across arms so exp-names stay unique.
-      - { tp: 8, ep: 1, dcp-size: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 70] }
+      - { tp: 8, ep: 1, dcp-size: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 40, 48, 56, 70] }
 
 dsr1-fp8-b200-trt:
   image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc14

perf-changelog.yaml

@@ -6902,3 +6902,12 @@
     - "Lower mem-fraction-static from 0.89 to 0.86 on every arm: TP4, TP8 and TP8 with DP attention. swa-full-tokens-ratio becomes per-arm: 0.10 on the tensor-parallel arms as before, 0.15 under DP attention."
     - "Drop concurrency 2 and 10 from the TP4 arm, leaving [1, 4, 8], and drop concurrency 16 from the TP8 hicache arm, leaving [32, 48]. Concurrency 16 remains on the TP8 no-offload arm."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2800
+
+- config-keys:
+    - kimik3-fp4-b300-vllm-agentic-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add concurrency 10, 12 and 14 to the TP8 x DCP8 Mooncake sweep, so the conc-list becomes [1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 40, 48, 56, 70]. The new points fall in the recipe's existing conc 9-16 band and therefore draft with DSpark level 3 at the committed golden AL 3.00, the same arm as concurrency 16."
+    - "Repoint the runner from cluster:b300-nv, whose launcher and runner labels were retired in #2826, to cluster:b300-dsxe so the sweep can schedule."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2846