diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
index f46ad956e..86328e283 100644
--- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh
@@ -61,8 +61,9 @@ export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
 # inter-turn idle gaps can reuse a socket exactly as the server closes it.
 # Outlast the client pool so the race cannot occur.
 export SGLANG_TIMEOUT_KEEP_ALIVE=900
-# PR #36684 & PR #36851 turned the v2 fused top-k on for GLM-5.x on ROCm.
-export SGLANG_OPT_USE_TOPK_V2=true
+# Use the precompiled HIP Top-K path; at c8 it improves P90 interactivity with
+# essentially unchanged throughput per GPU compared with fused Top-K v2.
+export SGLANG_OPT_USE_TOPK_V2=false
  
 # HiCache L2 (host DRAM), optionally extended with Mooncake L3.
 # KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=hicache or mooncake.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 381227752..172188f04 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -7753,3 +7753,12 @@
     - "Set a 64-token CUDA graph capture floor for B300 DSpark at c1/c2/c4; preserve c8+ sizing, full AgentX traces, AL 3.51, and Engram UVA"
     - "将 B300 DSpark 的 c1/c2/c4 CUDA graph 最小捕获范围设为 64 tokens；保持 c8+ 大小、完整 AgentX 轨迹、AL 3.51 和 Engram UVA"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3116
+
+- config-keys:
+    - glm5.2-fp4-mi355x-sglang-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Disable SGLANG_OPT_USE_TOPK_V2 so GLM-5.2 uses the precompiled HIP Top-K path on MI355X; at concurrency 8, P90 interactivity improves from 96.97 to 106.69 tok/s/user (+10.0%) while throughput per GPU changes from 9,553.2 to 9,539.0 tok/s (-0.15%)."
+    - "关闭 SGLANG_OPT_USE_TOPK_V2，使 GLM-5.2 在 MI355X 上使用预编译 HIP Top-K 路径；并发 8 时，P90 interactivity 从 96.97 提升到 106.69 tok/s/user（+10.0%），单卡吞吐从 9,553.2 变为 9,539.0 tok/s（-0.15%）。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3117