Update DSv4 B300 vllm

.github/configs/nvidia-master.yaml

@@ -2516,7 +2516,7 @@ dsv4-fp8-h200-vllm:
 # field, so dp-attn=true is used as the existing vLLM script switch for DP4
 # layouts on 4 allocated GPUs.
 dsv4-fp4-b300-vllm:
-  image: vllm/vllm-openai:deepseekv4-cu130
+  image: vllm/vllm-openai:v0.20.0-cu130
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
@@ -2528,16 +2528,16 @@ dsv4-fp4-b300-vllm:
     osl: 1024
     search-space:
     - { tp: 4, conc-start: 1, conc-end: 128 }
-    - { tp: 8, conc-start: 1, conc-end: 128 }
+    - { tp: 8, conc-start: 1, conc-end: 4 }
     - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 512 }
     - { tp: 4, ep: 4, dp-attn: true, conc-start: 2048, conc-end: 2048 }
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 4096, conc-end: 8192 }
   - isl: 8192
     osl: 1024
     search-space:
     - { tp: 4, conc-start: 1, conc-end: 64 }
-    - { tp: 8, conc-start: 1, conc-end: 64 }
-    - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 1024 }
+    - { tp: 8, conc-start: 1, conc-end: 4 }
+    - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 512 }
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 2048, conc-end: 2048 }
 
 dsv4-fp4-b300-vllm-mtp:

benchmarks/single_node/dsv4_fp4_b300_vllm.sh

@@ -42,16 +42,18 @@ if [ "${EP_SIZE:-1}" -gt 1 ]; then
     EP_ARGS=(--enable-expert-parallel)
 fi
 
+MOE_ARGS=()
+if [ "${DP_ATTENTION}" = "true" ]; then
+    MOE_ARGS=(--moe-backend deep_gemm_mega_moe)
+fi
+
 if [ "${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
-fi
 
 if [ "${EVAL_ONLY}" = "true" ]; then
     EVAL_MAX_MODEL_LEN=$(compute_eval_context_length "$MODEL" "$BENCHMARK_MAX_MODEL_LEN")
@@ -73,6 +75,7 @@ vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
     --block-size 256 \
     --no-enable-prefix-caching \
     "${EP_ARGS[@]}" \
+    "${MOE_ARGS[@]}" \
     --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
     --attention_config.use_fp4_indexer_cache True \
     --tokenizer-mode deepseek_v4 \

perf-changelog.yaml

@@ -1999,3 +1999,10 @@
     - "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
+
+- config-keys:
+    - dsv4-fp4-b300-vllm
+  description:
+    - "Change image to vllm/vllm-openai:v0.20.0-cu130"
+    - "Use Mega MoE for DEP configs"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1221