diff --git a/benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b300-fp4-mtp/agentic.yaml b/benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b300-fp4-mtp/agentic.yaml
index 34c77fe338..05bf3f6020 100644
--- a/benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b300-fp4-mtp/agentic.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/b300-fp4-mtp/agentic.yaml
@@ -6,7 +6,7 @@ base:
   name: dsv41flash-fp4-b300-vllm-agentic
   model:
     path: hf:deepseek-ai/DeepSeek-V4.1-Flash
-    container: vllm/vllm-openai:deepseekv41-flash-0909
+    container: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
     precision: fp4
   resources:
     gpu_type: b300
@@ -24,7 +24,7 @@ base:
   # The engine may take the full VLLM_ENGINE_READY_TIMEOUT_S to become ready.
   health_check:
     interval_seconds: 10
-    max_attempts: 360
+    max_attempts: 720
   roles:
     agg:
       nodes: 1
@@ -37,6 +37,10 @@ base:
         enable-auto-tool-choice: true
         reasoning-parser: deepseek_v41
         engram-config: '{"cpu_offload":true}'
+        # FlashInfer sparse attention with the merged Blackwell sparse indexer
+        # (MXFP4 indexer KV, sparse logits) at both TP sizes; fp8 KV cache.
+        attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV41","indexer_kv_dtype":"mxfp4","indexer_sparse_logits":true}'
+        kv-cache-dtype: fp8
         # Five-token DSpark with probabilistic drafting. Throughput runs replace
         # block rejection with the golden acceptance length.
         speculative-config: '{"method":"dspark","num_speculative_tokens":5,"draft_sample_method":"probabilistic","rejection_sample_method":"block","enable_adaptive_verification":false}'
@@ -46,7 +50,7 @@ base:
       env:
         VLLM_USE_RUST_FRONTEND: '1'
         PYTHONUNBUFFERED: '1'
-        VLLM_ENGINE_READY_TIMEOUT_S: '3600'
+        VLLM_ENGINE_READY_TIMEOUT_S: '7200'
   benchmark:
     type: custom
     command: bash /infmax-workspace/benchmarks/srt_agentic.sh
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index b0fc9ef451..feace66ce7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8328,7 +8328,7 @@ dsv41flash-fp4-gb200-sglang-agentic-dspark:
       - { tp: 2, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/gb200-fp4-mtp/agentic.yaml }
 
 dsv41flash-fp4-b300-vllm-agentic-dspark:
-  image: vllm/vllm-openai:deepseekv41-flash-0909
+  image: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:b300-dsxe
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 352c61fc0d..80e75c09b2 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8986,3 +8986,9 @@
     - "Hopper 使用 FA3，Blackwell FA4 fp8 descale 问题不适用，EAGLE3 draft 保持 attention_backend FLASH_ATTN。"
     - "This change does not alter the EAGLE3 draft model data type. The draft loads unmodified from the published Inferact/MiniMax-M3-EAGLE3-GQA checkpoint via --speculative-config (method=eagle3). kv-cache-dtype fp8 sets KV-cache storage precision, not the draft weights, and no flag overrides or re-quantizes the draft weights."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3503
+
+- config-keys:
+    - dsv41flash-fp4-b300-vllm-agentic-dspark
+  description:
+    - "Pin B300 DeepSeek-V4.1-Flash vLLM to nightly ddd6fbca with FlashInfer sparse attention, MXFP4 indexer KV, sparse logits, fp8 KV cache, and a 7200 s readiness timeout."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3458