diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 3a7ba3df1..616aa6ff9 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2501,17 +2501,15 @@ dsv4-fp4-b300-vllm:
   - isl: 1024
     osl: 1024
     search-space:
-    - { tp: 8, conc-start: 4, conc-end: 4 }
     - { tp: 4, conc-start: 4, conc-end: 128 }
-    - { tp: 8, conc-start: 128, conc-end: 128 }
-    - { tp: 4, dp-attn: true, conc-start: 256, conc-end: 512 }
+    - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 4096 }
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 2048, conc-end: 8192 }
   - isl: 8192
     osl: 1024
     search-space:
-    - { tp: 8, conc-start: 4, conc-end: 4 }
-    - { tp: 4, conc-start: 4, conc-end: 128 }
-    - { tp: 8, conc-start: 128, conc-end: 128 }
-    - { tp: 4, dp-attn: true, conc-start: 256, conc-end: 512 }
+    - { tp: 4, conc-start: 4, conc-end: 64 }
+    - { tp: 4, ep: 4, dp-attn: true, conc-start: 128, conc-end: 1024 }
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 1024, conc-end: 4096 }
 
 qwen3.5-fp8-h200-sglang:
   image: lmsysorg/sglang:v0.5.9-cu129-amd64
diff --git a/benchmarks/single_node/dsv4_fp4_b300_vllm.sh b/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
index 52f38c4d9..6bb5b9049 100755
--- a/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
@@ -1,9 +1,8 @@
 #!/usr/bin/env bash
 
 # DeepSeek-V4-Pro B300 single-node aggregate recipe from the submitted B300
-# pareto sweep. The matrix uses dp-attn=true as the existing switch to flip a
-# 4-GPU run from TP4 to DP4. Expert parallel is always enabled to match the
-# provided vllm serve command exactly.
+# pareto sweep. TP mode (dp-attn=false) runs without expert parallel; DP mode
+# (dp-attn=true) enables expert parallel (EP_SIZE=TP value = DP size).
 
 source "$(dirname "$0")/../benchmark_lib.sh"
 
@@ -38,6 +37,17 @@ if [ "${DP_ATTENTION}" = "true" ]; then
     PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
 fi
 
+EP_ARGS=()
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+if [ "${DP_ATTENTION}" = "true" ]; then
+    MAX_NUM_BATCHED_TOKENS=2048
+else
+    MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
+fi
+
 BENCHMARK_MAX_MODEL_LEN="$MAX_MODEL_LEN"
 if [ "$ISL" -eq 1024 ] && [ "$OSL" -eq 1024 ]; then
     BENCHMARK_MAX_MODEL_LEN=4096
@@ -62,7 +72,7 @@ vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
     --trust-remote-code \
     --block-size 256 \
     --no-enable-prefix-caching \
-    --enable-expert-parallel \
+    "${EP_ARGS[@]}" \
     --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
     --attention_config.use_fp4_indexer_cache True \
     --tokenizer-mode deepseek_v4 \
@@ -71,7 +81,7 @@ vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
     --reasoning-parser deepseek_v4 \
     --max-cudagraph-capture-size 2048 \
     --max-model-len "$SERVE_MAX_MODEL_LEN" \
-    --max-num-batched-tokens 2048 > "$SERVER_LOG" 2>&1 &
+    --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" > "$SERVER_LOG" 2>&1 &
 
 SERVER_PID=$!
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e4c46268e..300d39c40 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1894,3 +1894,12 @@
     - "better performance for dp-attention"
     - "Recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1178
+
+- config-keys:
+   - dsv4-fp4-b300-vllm
+  description:
+    - "Update search space based on B300 pareto sweep results"
+    - "ISL=1024: TP4 conc 4-128; DP4 (dp-attn) conc 256-4096; DP8 (dp-attn) conc 2048-8192"
+    - "ISL=8192: TP4 conc 4-64; DP4 (dp-attn) conc 128-1024; DP8 (dp-attn) conc 1024-8192"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1155
+