diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 7e975fdba..f13b8b6dd 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1882,6 +1882,7 @@ dsv4-fp4-b300-sglang:
     - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
     - { tp: 4, ep: 1, conc-start: 32, conc-end: 32 }
     - { tp: 4, ep: 4, dp-attn: true, conc-start: 512, conc-end: 512 }
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 8192, conc-end: 8192 }
   - isl: 8192
     osl: 1024
     search-space:
diff --git a/benchmarks/single_node/dsv4_fp4_b300_sglang.sh b/benchmarks/single_node/dsv4_fp4_b300_sglang.sh
index d50b57d72..8f43ea8a3 100755
--- a/benchmarks/single_node/dsv4_fp4_b300_sglang.sh
+++ b/benchmarks/single_node/dsv4_fp4_b300_sglang.sh
@@ -78,7 +78,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then
     # ep=8 in the yaml signals the mega_moe deepep backend; check high-conc
     # recipes first (they also have ep=8) so they aren't shadowed by the
     # medium-conc EP_SIZE=8 branch below.
-    if [ "$CONC" = "2048" ] || [ "$CONC" = "4096" ]; then
+    if [ "$CONC" = "2048" ] || [ "$CONC" = "4096" ] || [ "$CONC" = "8192" ]; then
         export NVSHMEM_DISABLE_IB=1
         export SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW=1
         export SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE=1
@@ -91,13 +91,21 @@ if [ "${DP_ATTENTION}" = "true" ]; then
             MEM_FRACTION_STATIC=0.87
             SWA_FULL_TOKENS_RATIO=0.06
             TOKENIZER_WORKER_NUM=4
-        else
+        elif [ "$CONC" = "4096" ]; then
             export SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8320
             CUDA_GRAPH_MAX_BS=544
             MAX_RUNNING_REQUESTS=4352
             MEM_FRACTION_STATIC=0.835
             SWA_FULL_TOKENS_RATIO=0.075
             TOKENIZER_WORKER_NUM=8
+        else
+            export SGLANG_OPT_USE_ONLINE_COMPRESS=1
+            export SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK=8256
+            CUDA_GRAPH_MAX_BS=1088
+            MAX_RUNNING_REQUESTS=8192
+            MEM_FRACTION_STATIC=0.80
+            SWA_FULL_TOKENS_RATIO=0.3
+            TOKENIZER_WORKER_NUM=16
         fi
         PARALLEL_ARGS=(
             --dp-size "$TP"
@@ -112,6 +120,9 @@ if [ "${DP_ATTENTION}" = "true" ]; then
         if [ "$CONC" = "4096" ]; then
             PARALLEL_ARGS+=(--decode-log-interval 5)
         fi
+        if [ "$CONC" = "8192" ]; then
+            PARALLEL_ARGS+=(--stream-interval 30)
+        fi
     elif [ "${EP_SIZE}" = "8" ]; then
         export NVSHMEM_DISABLE_IB=1
         export SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE=1
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2bd14e776..f74dbe770 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1985,3 +1985,10 @@
     - "Topology: 1 prefill DEP8 worker and 4 decode TP8 workers with dedicated NATS/etcd"
     - "Mirrors the historical 1P4D DEP8/TP8 offload point from srt-slurm aflowers/vllm-gb200-v0.20.0"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1218
+
+- config-keys:
+    - dsv4-fp4-b300-sglang
+  description:
+    - "Add conc=8192 recipe for 1k1k: deepep mega_moe backend with cuda-graph-max-bs 1088, max-running-requests 8192, mem-fraction-static 0.80, swa-full-tokens-ratio 0.3, tokenizer-worker-num 16"
+    - "conc=8192 enables SGLANG_OPT_USE_ONLINE_COMPRESS=1 and --stream-interval 30"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1209