diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm31flash/sglang/gb200-fp4-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm31flash/sglang/gb200-fp4-mtp/agentic.yaml
index 8f4fe1d22f..0c63ca368d 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm31flash/sglang/gb200-fp4-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm31flash/sglang/gb200-fp4-mtp/agentic.yaml
@@ -33,6 +33,8 @@ base:
         speculative-algorithm: DSPARK
         speculative-dspark-block-size: 7
         speculative-use-block-verification: true
+        # FP8 draft KV, as vLLM's DSpark drafter uses; bf16 costs ~10 GB of HBM per rank.
+        speculative-draft-kv-cache-dtype: fp8_e4m3
         cuda-graph-backend-prefill: breakable
         tokenizer-worker-num: 8
         watchdog-timeout: 3600
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 920eaaaf82..86938c1492 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -11499,3 +11499,11 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4509
   scenario-type: [agentic-coding]
   append-only: true
+
+- config-keys:
+    - minimaxm31flash-fp4-gb200-sglang-agentic-dspark
+  description:
+    - "Store the DSpark draft KV cache in FP8 (speculative-draft-kv-cache-dtype fp8_e4m3) instead of the bf16 default, matching the vLLM DSpark drafter; this frees HBM and HiCache host pool space for target KV."
+    - "Rerun the whole curve so every point uses the FP8 draft KV cache."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4599
+  scenario-type: [agentic-coding]