Updated DSv4 vllm B300 MTP (#1271)

* add DP to b300 mtp

* Update changelog

* Update Docker image version for dsv4-fp4-b300-vllm-mtp

* Update Docker image version to v0.20.2

* Modify search-space parameters in nvidia-master.yaml

* Merge duplicate DP_ATTENTION conditions in benchmark script

Co-authored-by: Cameron Quilici <cquil11@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: Cameron Quilici <cquil11@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cameron Quilici <cjquilici@gmail.com>

Author: 4 people

.github/configs/nvidia-master.yaml

@@ -2825,7 +2825,7 @@ dsv4-fp4-b300-trt-mtp:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 1024, spec-decoding: mtp }
 
 dsv4-fp4-b300-vllm-mtp:
-  image: vllm/vllm-openai:v0.20.0-cu130
+  image: vllm/vllm-openai:v0.20.2
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
@@ -2838,13 +2838,15 @@ dsv4-fp4-b300-vllm-mtp:
       osl: 1024
       search-space:
       - { tp: 4, conc-start: 1, conc-end: 256, spec-decoding: mtp }
-      - { tp: 8, conc-start: 1, conc-end: 64, spec-decoding: mtp }
+      - { tp: 8, conc-start: 1, conc-end: 8, spec-decoding: mtp }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 1024, spec-decoding: mtp }
     - isl: 8192
       osl: 1024
       search-space:
       - { tp: 4, conc-start: 1, conc-end: 64, spec-decoding: mtp }
-      - { tp: 8, conc-start: 1, conc-end: 64, spec-decoding: mtp }
+      - { tp: 8, conc-start: 1, conc-end: 8, spec-decoding: mtp }
       - { tp: 4, ep: 4, conc-start: 64, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 256, conc-end: 512, spec-decoding: mtp }
 
 qwen3.5-fp8-h200-sglang:
   image: lmsysorg/sglang:v0.5.9-cu129-amd64

benchmarks/single_node/dsv4_fp4_b300_vllm_mtp.sh

@@ -36,7 +36,14 @@ if [ "${EP_SIZE:-1}" -gt 1 ]; then
     EP_ARGS=(--enable-expert-parallel)
 fi
 
-MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
+MOE_ARGS=()
+if [ "${DP_ATTENTION}" = "true" ]; then
+    MOE_ARGS=(--moe-backend deep_gemm_mega_moe)
+    MAX_NUM_BATCHED_TOKENS=2048
+else
+    MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
+fi
+
 BENCHMARK_MAX_MODEL_LEN=$MAX_MODEL_LEN
 
 if [ "${EVAL_ONLY}" = "true" ]; then
@@ -61,6 +68,7 @@ vllm serve "$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

@@ -2486,3 +2486,10 @@
   description:
     - "Update SGLang image from v0.5.9-cu130 to v0.5.11-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1322
+  
+- config-keys:
+    - dsv4-fp4-b300-vllm-mtp
+  description:
+    - "Update image tag to vllm/vllm-openai:v0.20.2"
+    - "Add DEP configs for B300 vLLM MTP"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1271