dsv4-fp4-b300-vllm: bump to vllm v0.20.0, deep_gemm_mega_moe MoE

.github/configs/nvidia-master.yaml

@@ -2513,7 +2513,7 @@ 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

benchmarks/single_node/dsv4_fp4_b300_vllm.sh

@@ -39,7 +39,7 @@ fi
 
 EP_ARGS=()
 if [ "${EP_SIZE:-1}" -gt 1 ]; then
-    EP_ARGS=(--enable-expert-parallel)
+    EP_ARGS=(--enable-expert-parallel --moe-backend deep_gemm_mega_moe)
 fi
 
 if [ "${DP_ATTENTION}" = "true" ]; then
@@ -66,20 +66,18 @@ start_gpu_monitor
 
 set -x
 vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
-    "${PARALLEL_ARGS[@]}" \
-    --pipeline-parallel-size 1 \
-    --kv-cache-dtype fp8 \
     --trust-remote-code \
+    --kv-cache-dtype fp8 \
     --block-size 256 \
     --no-enable-prefix-caching \
+    "${PARALLEL_ARGS[@]}" \
     "${EP_ARGS[@]}" \
-    --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
-    --attention_config.use_fp4_indexer_cache True \
+    --compilation-config '{"mode": 0, "cudagraph_mode": "FULL_DECODE_ONLY"}' \
+    --attention_config.use_fp4_indexer_cache=True \
     --tokenizer-mode deepseek_v4 \
     --tool-call-parser deepseek_v4 \
     --enable-auto-tool-choice \
     --reasoning-parser deepseek_v4 \
-    --max-cudagraph-capture-size 2048 \
     --max-model-len "$SERVE_MAX_MODEL_LEN" \
     --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" > "$SERVER_LOG" 2>&1 &
 

perf-changelog.yaml

@@ -1918,3 +1918,12 @@
     - "Three CONC bands: A=TP8 (1-8), B=TP4 (16-128), C=DP4 dp-attn (64-512); B/C overlap at conc 64,128"
     - "Configs: 1k1k and 8k1k, no validation.py / launcher / yaml-field changes (knob-free)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1180
+
+- config-keys:
+    - dsv4-fp4-b300-vllm
+  description:
+    - "Pin image to vllm/vllm-openai:v0.20.0-cu130 (was floating deepseekv4-cu130 tag); DeepGEMM is preinstalled in this image"
+    - "Switch compilation-config to {\"mode\": 0, \"cudagraph_mode\": \"FULL_DECODE_ONLY\"} and use --attention_config.use_fp4_indexer_cache=True"
+    - "Bundle --moe-backend deep_gemm_mega_moe into EP_ARGS so it only applies when EP is enabled (EP_SIZE>1)"
+    - "Drop --pipeline-parallel-size 1 and --max-cudagraph-capture-size 2048 (no longer in v0.20.0 recipe); keep --no-enable-prefix-caching"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1206