From 4a70e241edfa21abd8c94d9ec8d41377cc65cd48 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Sat, 25 Apr 2026 08:29:40 -0700
Subject: [PATCH 01/16] Add DSv4 B200 configs

---
 .github/configs/nvidia-master.yaml           |  20 ++++
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 110 +++++++++++++++++++
 2 files changed, 130 insertions(+)
 create mode 100755 benchmarks/single_node/dsv4_fp4_b200_vllm.sh

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 42c720a63..769acb51c 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1705,6 +1705,26 @@ dsv4-fp4-b200-sglang:
     # max-throughput
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 512 }
 
+dsv4-fp4-b200-vllm:
+  image: vllm/vllm-openai:deepseekv4-cu130
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b200-dsv4
+  precision: fp4
+  framework: vllm
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, conc-start: 4, conc-end: 128 }
+    - { tp: 8, dp-attn: true, conc-start: 256, conc-end: 4096 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, conc-start: 4, conc-end: 32 }
+    - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 1024 }
+
 # NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/DeepSeek/DeepSeek-R1
 # does not have a B300-specific recipe, so this config reuses the existing DSR1 FP4
 # B200 SGLang recipe as-is until B300-specific tuning is available.
diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
new file mode 100755
index 000000000..88093aa62
--- /dev/null
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -0,0 +1,110 @@
+#!/usr/bin/env bash
+
+# DeepSeek-V4-Pro B200 single-node vLLM recipe derived from the B200 pareto
+# sweep. Uses TP8 for low concurrency and DP8 (dp-attn=true) for high
+# concurrency. Expert parallel is always enabled.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    DP_ATTENTION \
+    CONC \
+    ISL \
+    OSL \
+    MAX_MODEL_LEN \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+hf download "$MODEL"
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+# DeepSeek-V4-Pro weights are large; engine startup can exceed the default
+# 600s. Give it an hour to load.
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
+if [ "${DP_ATTENTION}" = "true" ]; then
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+fi
+
+# DP mode: mbt=ISL; TP mode: mbt=2*ISL; floor at 2048
+if [ "${DP_ATTENTION}" = "true" ]; then
+    MAX_NUM_BATCHED_TOKENS=$(( ISL < 2048 ? 2048 : ISL ))
+else
+    MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 < 2048 ? 2048 : ISL * 2 ))
+fi
+
+BENCHMARK_MAX_MODEL_LEN="$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")
+    export EVAL_MAX_MODEL_LEN
+    SERVE_MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+else
+    SERVE_MAX_MODEL_LEN="$BENCHMARK_MAX_MODEL_LEN"
+fi
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    --pipeline-parallel-size 1 \
+    --kv-cache-dtype fp8 \
+    --trust-remote-code \
+    --block-size 256 \
+    --no-enable-prefix-caching \
+    --enable-expert-parallel \
+    --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
+    --attention_config.use_fp4_indexer_cache True \
+    --tokenizer-mode deepseek_v4 \
+    --tool-call-parser deepseek_v4 \
+    --enable-auto-tool-choice \
+    --reasoning-parser deepseek_v4 \
+    --max-cudagraph-capture-size 2048 \
+    --max-model-len "$SERVE_MAX_MODEL_LEN" \
+    --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+pip install -q datasets pandas
+
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

From 6ee148f9b943fb806e8808a4b69365ec4667fce5 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Sat, 25 Apr 2026 08:45:26 -0700
Subject: [PATCH 02/16] Add changelog

---
 perf-changelog.yaml | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 397da6591..4309fe73a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1819,3 +1819,12 @@
     - "Restore the recipe-per-CONC split (low-latency / balanced / max-throughput) on top of the low-latency-only fallback from #1143; the DeepEP FP8 weight-postprocess path is fixed, so the high-throughput scenario runs again"
     - "Recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1132
+
+- config-keys:
+    - dsv4-fp4-b200-vllm
+  description:
+    - "Add DeepSeek-V4-Pro single-node B200 vLLM benchmark derived from B200 pareto sweep"
+    - "ISL=1024: TP8 conc 4-128; DP8 (dp-attn) conc 256-4096"
+    - "ISL=8192: TP8 conc 4-32; DP8 (dp-attn) conc 64-1024"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1156
+

From 707f2256d6f494f05ccdf89184fb283c5d0cca84 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Sat, 25 Apr 2026 09:36:41 -0700
Subject: [PATCH 03/16] fixup

---
 .github/configs/nvidia-master.yaml           |  4 ++--
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 11 ++++++++---
 2 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 769acb51c..d1b0813ba 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1718,12 +1718,12 @@ dsv4-fp4-b200-vllm:
     osl: 1024
     search-space:
     - { tp: 8, conc-start: 4, conc-end: 128 }
-    - { tp: 8, dp-attn: true, conc-start: 256, conc-end: 4096 }
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 4096 }
   - isl: 8192
     osl: 1024
     search-space:
     - { tp: 8, conc-start: 4, conc-end: 32 }
-    - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 1024 }
+    - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 1024 }
 
 # NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/DeepSeek/DeepSeek-R1
 # does not have a B300-specific recipe, so this config reuses the existing DSR1 FP4
diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index 88093aa62..be9f84f71 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -1,8 +1,8 @@
 #!/usr/bin/env bash
 
 # DeepSeek-V4-Pro B200 single-node vLLM recipe derived from the B200 pareto
-# sweep. Uses TP8 for low concurrency and DP8 (dp-attn=true) for high
-# concurrency. Expert parallel is always enabled.
+# 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"
 
@@ -37,6 +37,11 @@ if [ "${DP_ATTENTION}" = "true" ]; then
     PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
 fi
 
+EP_ARGS=()
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
 # DP mode: mbt=ISL; TP mode: mbt=2*ISL; floor at 2048
 if [ "${DP_ATTENTION}" = "true" ]; then
     MAX_NUM_BATCHED_TOKENS=$(( ISL < 2048 ? 2048 : ISL ))
@@ -68,7 +73,7 @@ vllm serve "$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 \

From 8ec131053eafbd98baf5011516d5f521e9c2e503 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Sat, 25 Apr 2026 09:46:20 -0700
Subject: [PATCH 04/16] fixup

---
 .github/configs/nvidia-master.yaml | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index d1b0813ba..4ec19cf27 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1717,7 +1717,8 @@ dsv4-fp4-b200-vllm:
   - isl: 1024
     osl: 1024
     search-space:
-    - { tp: 8, conc-start: 4, conc-end: 128 }
+    - { tp: 8, conc-start: 4, conc-end: 64 }
+    - { tp: 8, ep: 8, conc-start: 128, conc-end: 128 }
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 4096 }
   - isl: 8192
     osl: 1024

From 81594d748eccff1078727ff541d644336d0e64dc Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Sat, 25 Apr 2026 09:59:04 -0700
Subject: [PATCH 05/16] fixup

---
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index be9f84f71..338b80bd9 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -42,6 +42,12 @@ if [ "${EP_SIZE:-1}" -gt 1 ]; then
     EP_ARGS=(--enable-expert-parallel)
 fi
 
+# DP mode uses gpu-memory-utilization=0.85 (matched from pareto sweep)
+GMU_ARGS=()
+if [ "${DP_ATTENTION}" = "true" ]; then
+    GMU_ARGS=(--gpu-memory-utilization 0.85)
+fi
+
 # DP mode: mbt=ISL; TP mode: mbt=2*ISL; floor at 2048
 if [ "${DP_ATTENTION}" = "true" ]; then
     MAX_NUM_BATCHED_TOKENS=$(( ISL < 2048 ? 2048 : ISL ))
@@ -74,6 +80,7 @@ vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
     --block-size 256 \
     --no-enable-prefix-caching \
     "${EP_ARGS[@]}" \
+    "${GMU_ARGS[@]}" \
     --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
     --attention_config.use_fp4_indexer_cache True \
     --tokenizer-mode deepseek_v4 \

From f2fcfae6d0558bb85790182f71a15204e5287318 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Sat, 25 Apr 2026 10:18:59 -0700
Subject: [PATCH 06/16] fix runners

---
 runners/launch_b200-cw.sh         | 10 +++++++++-
 runners/launch_b200-dgxc-slurm.sh | 10 +++++++++-
 runners/launch_b200-nb.sh         | 10 +++++++++-
 3 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/runners/launch_b200-cw.sh b/runners/launch_b200-cw.sh
index ec7ba9a97..0b2dbf305 100644
--- a/runners/launch_b200-cw.sh
+++ b/runners/launch_b200-cw.sh
@@ -6,6 +6,14 @@ export PORT=8888
 MODEL_CODE="${EXP_NAME%%_*}"
 FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
 SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+# Prefer a framework-tagged script (e.g. dsv4_fp4_b200_vllm.sh) so models
+# with multiple inference engines can coexist; fall back to the historical
+# name without an engine suffix (`_trt` for trt, bare for everyone else).
+BENCH_BASE="benchmarks/single_node/${MODEL_CODE}_${PRECISION}_b200"
+BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+if [[ ! -f "$BENCH_SCRIPT" ]]; then
+    BENCH_SCRIPT="${BENCH_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
+fi
 
 PARTITION="b200"
 SQUASH_FILE="/tmp/gharunner/squash/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
@@ -58,6 +66,6 @@ srun --jobid=$JOB_ID \
 --container-mount-home \
 --container-workdir=$CONTAINER_MOUNT_DIR \
 --no-container-entrypoint --export=ALL \
-bash benchmarks/single_node/${MODEL_CODE}_${PRECISION}_b200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
+bash "$BENCH_SCRIPT"
 
 scancel $JOB_ID
diff --git a/runners/launch_b200-dgxc-slurm.sh b/runners/launch_b200-dgxc-slurm.sh
index c07037ff4..edf5db957 100644
--- a/runners/launch_b200-dgxc-slurm.sh
+++ b/runners/launch_b200-dgxc-slurm.sh
@@ -253,6 +253,14 @@ else
     SQUASH_FILE="/home/sa-shared/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
     FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+    # Prefer a framework-tagged script (e.g. dsv4_fp4_b200_vllm.sh) so models
+    # with multiple inference engines can coexist; fall back to the historical
+    # name without an engine suffix (`_trt` for trt, bare for everyone else).
+    BENCH_BASE="benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b200"
+    BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+    if [[ ! -f "$BENCH_SCRIPT" ]]; then
+        BENCH_SCRIPT="${BENCH_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
+    fi
     LOCK_FILE="${SQUASH_FILE}.lock"
 
     # TODO(Cam): lmsysorg/sglang:deepseek-v4-blackwell installs sglang editable at
@@ -290,5 +298,5 @@ else
         --no-container-mount-home \
         --container-workdir=$CONTAINER_MOUNT_DIR \
         --no-container-entrypoint --export=ALL,PORT=8888 \
-        bash benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
+        bash "$BENCH_SCRIPT"
 fi
diff --git a/runners/launch_b200-nb.sh b/runners/launch_b200-nb.sh
index 6b411fec2..e0c8d92fb 100644
--- a/runners/launch_b200-nb.sh
+++ b/runners/launch_b200-nb.sh
@@ -4,6 +4,14 @@ HF_HUB_CACHE_MOUNT="/mnt/data/gharunners/hf-hub-cache/"
 PARTITION="main"
 FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
 SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
+# Prefer a framework-tagged script (e.g. dsv4_fp4_b200_vllm.sh) so models
+# with multiple inference engines can coexist; fall back to the historical
+# name without an engine suffix (`_trt` for trt, bare for everyone else).
+BENCH_BASE="benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b200"
+BENCH_SCRIPT="${BENCH_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
+if [[ ! -f "$BENCH_SCRIPT" ]]; then
+    BENCH_SCRIPT="${BENCH_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
+fi
 
 UCX_NET_DEVICES=eth0
 
@@ -27,4 +35,4 @@ srun --partition=$PARTITION --gres=gpu:$TP --exclusive --job-name="$RUNNER_NAME"
 --container-writable \
 --container-workdir=$CONTAINER_MOUNT_DIR \
 --no-container-entrypoint --export=ALL,PORT=8888,UCX_NET_DEVICES=$UCX_NET_DEVICES \
-bash benchmarks/single_node/${EXP_NAME%%_*}_${PRECISION}_b200${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh
\ No newline at end of file
+bash "$BENCH_SCRIPT"
\ No newline at end of file

From 98d83ffab9db0cf9360c4e707e04307d05ecad91 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sat, 25 Apr 2026 20:12:12 -0400
Subject: [PATCH 07/16] Simplify MAX_NUM_BATCHED_TOKENS calculation

Set MAX_NUM_BATCHED_TOKENS to a fixed value of 2048.
---
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index 338b80bd9..7eeebddda 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -48,12 +48,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then
     GMU_ARGS=(--gpu-memory-utilization 0.85)
 fi
 
-# DP mode: mbt=ISL; TP mode: mbt=2*ISL; floor at 2048
-if [ "${DP_ATTENTION}" = "true" ]; then
-    MAX_NUM_BATCHED_TOKENS=$(( ISL < 2048 ? 2048 : ISL ))
-else
-    MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 < 2048 ? 2048 : ISL * 2 ))
-fi
+MAX_NUM_BATCHED_TOKENS=2048
 
 BENCHMARK_MAX_MODEL_LEN="$MAX_MODEL_LEN"
 if [ "$ISL" -eq 1024 ] && [ "$OSL" -eq 1024 ]; then

From c18f4132b6aa6b172c5c04809a979f08572e0a51 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sat, 25 Apr 2026 20:36:21 -0400
Subject: [PATCH 08/16] Update perf-changelog.yaml with benchmark changes

Removed and re-added DeepSeek-V4-Pro benchmark details in the changelog.
---
 perf-changelog.yaml | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6ae2c64a5..9ac4a1e9a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1820,14 +1820,6 @@
     - "Recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1132
 
-- config-keys:
-    - dsv4-fp4-b200-vllm
-  description:
-    - "Add DeepSeek-V4-Pro single-node B200 vLLM benchmark derived from B200 pareto sweep"
-    - "ISL=1024: TP8 conc 4-128; DP8 (dp-attn) conc 256-4096"
-    - "ISL=8192: TP8 conc 4-32; DP8 (dp-attn) conc 64-1024"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1156
-
 - config-keys:
     - dsv4-fp8-mi355x-sglang
   description:
@@ -1842,3 +1834,11 @@
     - "Bump --chunked-prefill-size from 4096 to 8192"
     - "Retrigger dsv4-fp8-mi355x-sglang"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1159
+
+- config-keys:
+    - dsv4-fp4-b200-vllm
+  description:
+    - "Add DeepSeek-V4-Pro single-node B200 vLLM benchmark derived from B200 pareto sweep"
+    - "ISL=1024: TP8 conc 4-128; DP8 (dp-attn) conc 256-4096"
+    - "ISL=8192: TP8 conc 4-32; DP8 (dp-attn) conc 64-1024"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1156

From 5bab835a1f38937351c71d5d5c60f778de136bbb Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sun, 26 Apr 2026 00:46:43 -0400
Subject: [PATCH 09/16] Modify MAX_NUM_BATCHED_TOKENS logic

Adjust MAX_NUM_BATCHED_TOKENS based on ISL and concurrency.
---
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index 7eeebddda..9901f3966 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -48,7 +48,11 @@ if [ "${DP_ATTENTION}" = "true" ]; then
     GMU_ARGS=(--gpu-memory-utilization 0.85)
 fi
 
-MAX_NUM_BATCHED_TOKENS=2048
+if [ "${ISL}" -eq 8192 ] && [ "${concurrency}" -le 128 ]; then
+    MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
+else
+    MAX_NUM_BATCHED_TOKENS=2048
+fi
 
 BENCHMARK_MAX_MODEL_LEN="$MAX_MODEL_LEN"
 if [ "$ISL" -eq 1024 ] && [ "$OSL" -eq 1024 ]; then

From e28c638b43c157c9b52c8acce2702a8599665e5e Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sun, 26 Apr 2026 10:26:23 -0400
Subject: [PATCH 10/16] Fix variable name from 'concurrency' to 'CONC'

---
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index 9901f3966..53b7e645f 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -48,7 +48,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then
     GMU_ARGS=(--gpu-memory-utilization 0.85)
 fi
 
-if [ "${ISL}" -eq 8192 ] && [ "${concurrency}" -le 128 ]; then
+if [ "${ISL}" -eq 8192 ] && [ "${CONC}" -le 128 ]; then
     MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
 else
     MAX_NUM_BATCHED_TOKENS=2048

From e86d6e9b070e6ba5480cfddbca31cb4820309b66 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sun, 26 Apr 2026 16:29:33 -0400
Subject: [PATCH 11/16] Revise perf-changelog.yaml with new benchmarks

Updated performance changelog with new benchmarks and configurations for DeepSeek-V4-Pro and vLLM. Removed obsolete entries and added links to relevant pull requests.
---
 perf-changelog.yaml | 36 ++++++++++++++----------------------
 1 file changed, 14 insertions(+), 22 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 705a3f019..ac01a5c28 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1857,32 +1857,24 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1170
 
 - config-keys:
-    - dsv4-fp4-b300-sglang
-  description:
-    - "Restore the recipe-per-CONC split (low-latency / balanced / max-throughput) on top of the low-latency-only fallback from #1143; the DeepEP FP8 weight-postprocess path is fixed, so the high-throughput scenario runs again"
-    - "Recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1158
-
-- config-keys:
-    - dsv4-fp4-b300-sglang
+    - dsv4-fp4-b300-sglang-mtp
   description:
-    - "Floor --max-running-requests at 8 in dsv4_fp4_b300_sglang.sh so low-CONC sweeps don't drop below the queue depth needed for stable benchmarking (CONC * 3 / 2 still applies above CONC=5)"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1173
-
-- config-keys:
-    - dsv4-fp4-b300-sglang
-  description:
-    - "better performance for dp-attention"
-    - "Recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1174
+    - "Add DeepSeek-V4-Pro FP4 B300 SGLang benchmark with EAGLE/MTP speculative decoding"
+    - "Image: lmsysorg/sglang:deepseek-v4-b300@sha256:26e116bd211e300dbb76924d56c5cbe6cc3ee5ee2fe314859cb8774f5bc070f3 (pinned for deep_gemm transform_weights_for_mega_moe support; same digest as PR #1158)"
+    - "Model: deepseek-ai/DeepSeek-V4-Pro"
+    - "EAGLE/MTP flags hardcoded in script: num-steps=3, eagle-topk=1, num-draft-tokens=4"
+    - "Recipe (MoE backend, chunked-prefill) selected in script by dp-attn: TP-only + flashinfer_mxfp4 (small batch) vs DP-attn + deepep mega_moe (large batch)"
+    - "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/1166
 
 - config-keys:
-    - dsv4-fp4-b200-vllm
+   - dsv4-fp4-b300-vllm
   description:
-    - "Add DeepSeek-V4-Pro single-node B200 vLLM benchmark derived from B200 pareto sweep"
-    - "ISL=1024: TP8 conc 4-128; DP8 (dp-attn) conc 256-4096"
-    - "ISL=8192: TP8 conc 4-32; DP8 (dp-attn) conc 64-1024"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1156
+    - "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
 
 - config-keys:
     - dsv4-fp4-b200-vllm

From a222193d444db56bcb5862d41bebc66656f828d6 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sun, 26 Apr 2026 16:30:51 -0400
Subject: [PATCH 12/16] Update perf-changelog.yaml

---
 perf-changelog.yaml | 7 -------
 1 file changed, 7 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ac01a5c28..eb61839df 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1779,13 +1779,6 @@
     - "Prefix caching and speculative decoding disabled for baseline numbers"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1131
 
-- config-keys:
-    - dsv4-fp4-b300-sglang
-  description:
-    - "Restore the recipe-per-CONC split (low-latency / balanced / max-throughput) on top of the low-latency-only fallback from #1143; the DeepEP FP8 weight-postprocess path is fixed, so the high-throughput scenario runs again"
-    - "Recipes from https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1158
-
 - config-keys:
     - dsv4-fp8-mi355x-sglang
   description:

From 3f038c42113ba870130de09bd9b575c32ab98076 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sun, 26 Apr 2026 19:46:57 -0400
Subject: [PATCH 13/16] lower batch size to fix OOM

---
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index 53b7e645f..2db6d899e 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -49,7 +49,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then
 fi
 
 if [ "${ISL}" -eq 8192 ] && [ "${CONC}" -le 128 ]; then
-    MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
+    MAX_NUM_BATCHED_TOKENS=${ISL}
 else
     MAX_NUM_BATCHED_TOKENS=2048
 fi

From fae14d9be14f70fdf7b7f845462bb86da18cb395 Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sun, 26 Apr 2026 19:50:21 -0400
Subject: [PATCH 14/16] Remove GPU memory utilization for DP mode

Removed GPU memory utilization settings for DP mode.
---
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 6 ------
 1 file changed, 6 deletions(-)

diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index 2db6d899e..c85637506 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -42,12 +42,6 @@ if [ "${EP_SIZE:-1}" -gt 1 ]; then
     EP_ARGS=(--enable-expert-parallel)
 fi
 
-# DP mode uses gpu-memory-utilization=0.85 (matched from pareto sweep)
-GMU_ARGS=()
-if [ "${DP_ATTENTION}" = "true" ]; then
-    GMU_ARGS=(--gpu-memory-utilization 0.85)
-fi
-
 if [ "${ISL}" -eq 8192 ] && [ "${CONC}" -le 128 ]; then
     MAX_NUM_BATCHED_TOKENS=${ISL}
 else

From acb851022464d992b1197844c0b9a98e1374da4c Mon Sep 17 00:00:00 2001
From: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Date: Sun, 26 Apr 2026 20:35:52 -0400
Subject: [PATCH 15/16] Introduce GMU_ARGS for GPU memory utilization

Add GPU memory utilization argument for DP attention.
---
 benchmarks/single_node/dsv4_fp4_b200_vllm.sh | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
index c85637506..d37dc5282 100755
--- a/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b200_vllm.sh
@@ -42,6 +42,11 @@ if [ "${EP_SIZE:-1}" -gt 1 ]; then
     EP_ARGS=(--enable-expert-parallel)
 fi
 
+GMU_ARGS=()
+if [ "${DP_ATTENTION}" = "true" ]; then
+    GMU_ARGS=(--gpu-memory-utilization 0.85)
+fi
+
 if [ "${ISL}" -eq 8192 ] && [ "${CONC}" -le 128 ]; then
     MAX_NUM_BATCHED_TOKENS=${ISL}
 else

From 984064aa17c7b86e7730c6020408d72ff3e9ab63 Mon Sep 17 00:00:00 2001
From: wzhao18 <wzhao18.sz@gmail.com>
Date: Sun, 26 Apr 2026 22:10:07 -0700
Subject: [PATCH 16/16] Add low-latency configs

---
 .github/configs/nvidia-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 57157ba40..17d9b2b1d 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1718,13 +1718,13 @@ dsv4-fp4-b200-vllm:
   - isl: 1024
     osl: 1024
     search-space:
-    - { tp: 8, conc-start: 4, conc-end: 64 }
+    - { tp: 8, conc-start: 1, conc-end: 64 }
     - { tp: 8, ep: 8, conc-start: 128, conc-end: 128 }
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 4096 }
   - isl: 8192
     osl: 1024
     search-space:
-    - { tp: 8, conc-start: 4, conc-end: 32 }
+    - { tp: 8, conc-start: 1, conc-end: 32 }
     - { tp: 8, ep: 8, dp-attn: true, conc-start: 64, conc-end: 1024 }
 
 # NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/DeepSeek/DeepSeek-R1