From ae8efdb55640c50188456c537594b4ce939bf88d Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 16:54:10 -0700
Subject: [PATCH 01/23] new command

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 68 +++++++++++++++++++++++++++++++
 1 file changed, 68 insertions(+)
 create mode 100644 benchmarks/dsr1_fp4_b200_slurm.sh

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
new file mode 100644
index 0000000000..33c6c40514
--- /dev/null
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+
+echo "JOB $SLURM_JOB_ID running on NODE $SLURMD_NODENAME"
+
+huggingface-cli download $MODEL
+SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
+
+set -x
+PORT=$(( 8888 + $PORT_OFFSET ))
+
+TP = ''
+if [[ $CONC -eq 4 || $CONC -eq 8 || $CONC -eq 128 || $CONC -eq 256 ]]; then
+    TP = 8
+elif [[ $CONC -eq 16 || $CONC -eq 32 || $CONC -eq 64 ]]; then
+    TP = 4
+fi
+
+if [[ $TP == '' ]]; then
+    echo "Error: TP is not set"
+    exit 1
+fi
+
+python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
+--tensor-parallel-size=$TP --data-parallel-size=1 \
+--cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \
+--chunked-prefill-size 16384 --max-prefill-tokens 32768 \
+--enable-ep-moe --quantization modelopt_fp4  --enable-flashinfer-allreduce-fusion --scheduler-recv-interval 10 \
+--enable-symm-mem  --disable-radix-cache --attention-backend trtllm_mla --enable-flashinfer-trtllm-moe --stream-interval 10 \
+> $SERVER_LOG 2>&1 &
+
+set +x
+IGNORE_PAT="Ignore import error when loading sglang.srt.models.glm4v_moe: No module named 'transformers.models.glm4v_moe'"
+
+while IFS= read -r line; do
+  printf '%s\n' "$line"
+
+  # Skip the known benign "Ignore import error ..." line
+  if [[ "$line" == *"$IGNORE_PAT"* ]]; then
+    continue
+  fi
+
+  # Keep your original "error" trap for everything else
+  if [[ "$line" =~ [Ee][Rr][Rr][Oo][Rr] ]]; then
+    sleep 5
+    tail -n100 "$SERVER_LOG"
+    echo "JOB ${SLURM_JOB_ID:-NA} ran on NODE ${SLURMD_NODENAME:-unknown}"
+    exit 1
+  fi
+
+  # Break when server is ready
+  if [[ "$line" == *"The server is fired up and ready to roll"* ]]; then
+    break
+  fi
+# Start tail from the beginning so we don't miss early lines
+done < <(tail -n +1 -F "$SERVER_LOG")
+
+set -x
+git clone https://github.com/kimbochen/bench_serving.git
+python3 bench_serving/benchmark_serving.py \
+--model $MODEL --backend vllm \
+--base-url http://0.0.0.0:$PORT \
+--dataset-name random \
+--random-input-len $ISL --random-output-len $OSL --random-range-ratio $RANDOM_RANGE_RATIO \
+--num-prompts $(( $CONC * 10 )) --max-concurrency $CONC \
+--request-rate inf --ignore-eos \
+--save-result --percentile-metrics 'ttft,tpot,itl,e2el' \
+--result-dir /workspace/ \
+--result-filename $RESULT_FILENAME.json

From 580bd36a2e2be086edea7d6d2b1dba7363514e38 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 16:54:53 -0700
Subject: [PATCH 02/23] redundant stuff

---
 .github/workflows/dsr1-tmpl.yml | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/.github/workflows/dsr1-tmpl.yml b/.github/workflows/dsr1-tmpl.yml
index df05305d4b..75d4701740 100644
--- a/.github/workflows/dsr1-tmpl.yml
+++ b/.github/workflows/dsr1-tmpl.yml
@@ -153,8 +153,7 @@ jobs:
       osl: ${{ inputs.osl }}
       max-model-len: ${{ inputs.max-model-len }}
       random-range-ratio: ${{ inputs.random-range-ratio }}
-      tp-list: '[8]'
-
+      
   bmk-b200-trt-fp4:
     if: ${{ inputs.use_b200 }}
     uses: ./.github/workflows/benchmark-tmpl.yml

From 68cdf7a3e770a8d46123811542507602d20b4c07 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:10:35 -0700
Subject: [PATCH 03/23] new command

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 33c6c40514..2801efbc7b 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -56,6 +56,17 @@ done < <(tail -n +1 -F "$SERVER_LOG")
 
 set -x
 git clone https://github.com/kimbochen/bench_serving.git
+
+python3 bench_serving/benchmark_serving.py \
+--model $MODEL --backend vllm \
+--base-url http://0.0.0.0:$PORT \
+--dataset-name random \
+--random-input-len $ISL --random-output-len $OSL --random-range-ratio $RANDOM_RANGE_RATIO \
+--num-prompts $CONC --max-concurrency $CONC \
+--request-rate inf --ignore-eos \
+--result-dir /workspace/ \
+--result-filename bak.json
+
 python3 bench_serving/benchmark_serving.py \
 --model $MODEL --backend vllm \
 --base-url http://0.0.0.0:$PORT \

From b9fd613417ff49d69e21794765061a500fe598d5 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:12:03 -0700
Subject: [PATCH 04/23] new command

---
 .github/workflows/runner-sweep-test.yml | 1 +
 benchmarks/dsr1_fp4_b200_slurm.sh       | 2 --
 2 files changed, 1 insertion(+), 2 deletions(-)

diff --git a/.github/workflows/runner-sweep-test.yml b/.github/workflows/runner-sweep-test.yml
index 6e6ca023de..63a2e04899 100644
--- a/.github/workflows/runner-sweep-test.yml
+++ b/.github/workflows/runner-sweep-test.yml
@@ -26,6 +26,7 @@ on:
           - 'lmsysorg/sglang:v0.4.9.post1-cu126'
           - 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
           - 'lmsysorg/sglang:v0.5.2rc2-cu126'
+          - 'lmsysorg/sglang:v0.5.2rc2-cu129'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'
diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 2801efbc7b..4929b9e27a 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -64,8 +64,6 @@ python3 bench_serving/benchmark_serving.py \
 --random-input-len $ISL --random-output-len $OSL --random-range-ratio $RANDOM_RANGE_RATIO \
 --num-prompts $CONC --max-concurrency $CONC \
 --request-rate inf --ignore-eos \
---result-dir /workspace/ \
---result-filename bak.json
 
 python3 bench_serving/benchmark_serving.py \
 --model $MODEL --backend vllm \

From 712856c421e9d8ae13ab650e0a30d059568047a4 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:13:07 -0700
Subject: [PATCH 05/23] new command

---
 .github/workflows/runner-sweep-test.yml | 2 +-
 .github/workflows/runner-test.yml       | 1 +
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/runner-sweep-test.yml b/.github/workflows/runner-sweep-test.yml
index 63a2e04899..ceae836343 100644
--- a/.github/workflows/runner-sweep-test.yml
+++ b/.github/workflows/runner-sweep-test.yml
@@ -26,7 +26,7 @@ on:
           - 'lmsysorg/sglang:v0.4.9.post1-cu126'
           - 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
           - 'lmsysorg/sglang:v0.5.2rc2-cu126'
-          - 'lmsysorg/sglang:v0.5.2rc2-cu129'
+          - 'lmsysorg/sglang:v0.5.3rc1-cu129'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'
diff --git a/.github/workflows/runner-test.yml b/.github/workflows/runner-test.yml
index 9db2644c02..3fd2edc2e7 100644
--- a/.github/workflows/runner-test.yml
+++ b/.github/workflows/runner-test.yml
@@ -51,6 +51,7 @@ on:
           - 'lmsysorg/sglang:v0.4.9.post1-cu126'
           - 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
           - 'lmsysorg/sglang:v0.5.2rc2-cu126'
+          - 'lmsysorg/sglang:v0.5.3rc1-cu129'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'

From d76f7a911db8c26b7ae4b4071419a05816820586 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:13:40 -0700
Subject: [PATCH 06/23] new command

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 1 +
 1 file changed, 1 insertion(+)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 4929b9e27a..7bba4b8a3a 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -57,6 +57,7 @@ done < <(tail -n +1 -F "$SERVER_LOG")
 set -x
 git clone https://github.com/kimbochen/bench_serving.git
 
+# warmup for JIT kernels
 python3 bench_serving/benchmark_serving.py \
 --model $MODEL --backend vllm \
 --base-url http://0.0.0.0:$PORT \

From e71bb8f6d05bd9a28e1217d442343e862fb6df2d Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:26:07 -0700
Subject: [PATCH 07/23] new yaml structure

---
 .github/workflows/dsr1-tmpl.yml   | 4 +++-
 benchmarks/dsr1_fp4_b200_slurm.sh | 2 ++
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/dsr1-tmpl.yml b/.github/workflows/dsr1-tmpl.yml
index 75d4701740..326019edba 100644
--- a/.github/workflows/dsr1-tmpl.yml
+++ b/.github/workflows/dsr1-tmpl.yml
@@ -153,7 +153,9 @@ jobs:
       osl: ${{ inputs.osl }}
       max-model-len: ${{ inputs.max-model-len }}
       random-range-ratio: ${{ inputs.random-range-ratio }}
-      
+      tp-list: '[4, 8]'
+      conc-list: '[4, 8, 16, 32, 64, 128, 256]'  # Custom concurrency values for this job
+
   bmk-b200-trt-fp4:
     if: ${{ inputs.use_b200 }}
     uses: ./.github/workflows/benchmark-tmpl.yml
diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 7bba4b8a3a..6e13973b02 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -20,6 +20,8 @@ if [[ $TP == '' ]]; then
     exit 1
 fi
 
+echo "TP: $TP, CONC: $CONC"
+
 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \

From 00313310e11df64455b1cce57eadbeb183677411 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:28:12 -0700
Subject: [PATCH 08/23] new yaml structure

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 14 --------------
 1 file changed, 14 deletions(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 6e13973b02..f18d632c1e 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -8,20 +8,6 @@ SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
 set -x
 PORT=$(( 8888 + $PORT_OFFSET ))
 
-TP = ''
-if [[ $CONC -eq 4 || $CONC -eq 8 || $CONC -eq 128 || $CONC -eq 256 ]]; then
-    TP = 8
-elif [[ $CONC -eq 16 || $CONC -eq 32 || $CONC -eq 64 ]]; then
-    TP = 4
-fi
-
-if [[ $TP == '' ]]; then
-    echo "Error: TP is not set"
-    exit 1
-fi
-
-echo "TP: $TP, CONC: $CONC"
-
 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \

From 7ffd0e025490c210b5aa57ae51d21a2abf3d5dfa Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:28:44 -0700
Subject: [PATCH 09/23] redundant character

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index f18d632c1e..a76af3c863 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -52,7 +52,7 @@ python3 bench_serving/benchmark_serving.py \
 --dataset-name random \
 --random-input-len $ISL --random-output-len $OSL --random-range-ratio $RANDOM_RANGE_RATIO \
 --num-prompts $CONC --max-concurrency $CONC \
---request-rate inf --ignore-eos \
+--request-rate inf --ignore-eos
 
 python3 bench_serving/benchmark_serving.py \
 --model $MODEL --backend vllm \

From 682738499468c3e9afabeca38db15b154a58a95f Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 17:30:08 -0700
Subject: [PATCH 10/23] images

---
 .github/workflows/runner-sweep-test.yml | 2 +-
 .github/workflows/runner-test.yml       | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/runner-sweep-test.yml b/.github/workflows/runner-sweep-test.yml
index ceae836343..e3053b4109 100644
--- a/.github/workflows/runner-sweep-test.yml
+++ b/.github/workflows/runner-sweep-test.yml
@@ -26,7 +26,7 @@ on:
           - 'lmsysorg/sglang:v0.4.9.post1-cu126'
           - 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
           - 'lmsysorg/sglang:v0.5.2rc2-cu126'
-          - 'lmsysorg/sglang:v0.5.3rc1-cu129'
+          - 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'
diff --git a/.github/workflows/runner-test.yml b/.github/workflows/runner-test.yml
index 3fd2edc2e7..848d6665de 100644
--- a/.github/workflows/runner-test.yml
+++ b/.github/workflows/runner-test.yml
@@ -51,7 +51,7 @@ on:
           - 'lmsysorg/sglang:v0.4.9.post1-cu126'
           - 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
           - 'lmsysorg/sglang:v0.5.2rc2-cu126'
-          - 'lmsysorg/sglang:v0.5.3rc1-cu129'
+          - 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'

From 7e56d0c3e6d7a2fe8b119bb90890cd09723be4a1 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 19:17:58 -0700
Subject: [PATCH 11/23] added config

---
 .github/workflows/dsr1-tmpl.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/dsr1-tmpl.yml b/.github/workflows/dsr1-tmpl.yml
index 326019edba..29a6ebb496 100644
--- a/.github/workflows/dsr1-tmpl.yml
+++ b/.github/workflows/dsr1-tmpl.yml
@@ -144,7 +144,7 @@ jobs:
     secrets: inherit
     with:
       runner: b200
-      image: 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
+      image: 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
       model: 'nvidia/DeepSeek-R1-0528-FP4'
       framework: 'sglang'
       precision: 'fp4'
@@ -153,7 +153,7 @@ jobs:
       osl: ${{ inputs.osl }}
       max-model-len: ${{ inputs.max-model-len }}
       random-range-ratio: ${{ inputs.random-range-ratio }}
-      tp-list: '[4, 8]'
+      tp-list: '[8]'
       conc-list: '[4, 8, 16, 32, 64, 128, 256]'  # Custom concurrency values for this job
 
   bmk-b200-trt-fp4:

From 2e45e2d6ab1daefba7089a9b1bdb512ae13c4bdd Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 19:27:50 -0700
Subject: [PATCH 12/23] moved yaml

---
 .github/workflows/runner-sweep-test.yml | 2 +-
 .github/workflows/runner-test.yml       | 3 +--
 2 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/runner-sweep-test.yml b/.github/workflows/runner-sweep-test.yml
index e3053b4109..1eab7f9fd9 100644
--- a/.github/workflows/runner-sweep-test.yml
+++ b/.github/workflows/runner-sweep-test.yml
@@ -26,7 +26,6 @@ on:
           - 'lmsysorg/sglang:v0.4.9.post1-cu126'
           - 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
           - 'lmsysorg/sglang:v0.5.2rc2-cu126'
-          - 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'
@@ -36,6 +35,7 @@ on:
           - 'rocm/vllm-dev:open-mi300-08052025'
           - 'rocm/vllm-dev:open-mi355-08052025'
           - 'vllm/vllm-openai:v0.10.2'
+          - 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
 
       model:
         description: 'Model'
diff --git a/.github/workflows/runner-test.yml b/.github/workflows/runner-test.yml
index 848d6665de..21afc5e778 100644
--- a/.github/workflows/runner-test.yml
+++ b/.github/workflows/runner-test.yml
@@ -51,7 +51,6 @@ on:
           - 'lmsysorg/sglang:v0.4.9.post1-cu126'
           - 'lmsysorg/sglang:v0.5.0rc1-cu128-b200'
           - 'lmsysorg/sglang:v0.5.2rc2-cu126'
-          - 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
           - 'rocm/7.0-preview:rocm7.0_preview_ubuntu_22.04_sgl-dev-v0.5.2rc2-mi30x_rc1'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2'
           - 'nvcr.io#nvidia/tensorrt-llm/release:1.1.0rc2.post1'
@@ -61,7 +60,7 @@ on:
           - 'rocm/vllm-dev:open-mi300-08052025'
           - 'rocm/vllm-dev:open-mi355-08052025'
           - 'vllm/vllm-openai:v0.10.2'
-
+          - 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
       model:
         description: 'Model'
         required: true

From 5ee23ee2dac338cbb0a56cfd90302f6dfeb1285a Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 20:05:51 -0700
Subject: [PATCH 13/23] slurm script

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index a76af3c863..c3418ca42d 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -8,10 +8,35 @@ SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
 set -x
 PORT=$(( 8888 + $PORT_OFFSET ))
 
+SCHEDULER_RECV_INTERVAL=''
+
+if [[ "$ISL" == "1024" && "$OSL" == "1024" || "$ISL" == "1024" && "$OSL" == "8192" ]]; then
+    SCHEDULER_RECV_INTERVAL=10
+
+    if [[ "$CONC" -gte 16 ]]; then
+        SCHEDULER_RECV_INTERVAL=30
+    fi
+
+elif [[ "$ISL" == "1024" && "$OSL" == "8192" ]]; then
+    SCHEDULER_RECV_INTERVAL=10
+
+    if [[ "$CONC" -gte 32 ]]; then
+        SCHEDULER_RECV_INTERVAL=30
+    fi
+
+fi
+
+echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+if [[ "$SCHEDULER_RECV_INTERVAL" == '' ]]; then
+    echo "Error: SCHEDULER_RECV_INTERVAL is not set"
+    exit 1
+fi
+
 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \
---chunked-prefill-size 16384 --max-prefill-tokens 32768 \
+--chunked-prefill-size 16384 \
 --enable-ep-moe --quantization modelopt_fp4  --enable-flashinfer-allreduce-fusion --scheduler-recv-interval 10 \
 --enable-symm-mem  --disable-radix-cache --attention-backend trtllm_mla --enable-flashinfer-trtllm-moe --stream-interval 10 \
 > $SERVER_LOG 2>&1 &

From 989f10de50e4dc2ae332df86a3d5e4577146650f Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 20:09:53 -0700
Subject: [PATCH 14/23] slurm config

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 21 +++------------------
 1 file changed, 3 insertions(+), 18 deletions(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index c3418ca42d..705491982a 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -10,29 +10,14 @@ PORT=$(( 8888 + $PORT_OFFSET ))
 
 SCHEDULER_RECV_INTERVAL=''
 
-if [[ "$ISL" == "1024" && "$OSL" == "1024" || "$ISL" == "1024" && "$OSL" == "8192" ]]; then
+if [[ "$CONC" -lt 16 ]]; then
     SCHEDULER_RECV_INTERVAL=10
-
-    if [[ "$CONC" -gte 16 ]]; then
-        SCHEDULER_RECV_INTERVAL=30
-    fi
-
-elif [[ "$ISL" == "1024" && "$OSL" == "8192" ]]; then
-    SCHEDULER_RECV_INTERVAL=10
-
-    if [[ "$CONC" -gte 32 ]]; then
-        SCHEDULER_RECV_INTERVAL=30
-    fi
-
+elif [[ "$CONC" -gte 16 ]]; then
+    SCHEDULER_RECV_INTERVAL=30
 fi
 
 echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
 
-if [[ "$SCHEDULER_RECV_INTERVAL" == '' ]]; then
-    echo "Error: SCHEDULER_RECV_INTERVAL is not set"
-    exit 1
-fi
-
 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \

From 11092bdb9d2a181cfed038928665c9bb5f59c702 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 20:11:12 -0700
Subject: [PATCH 15/23] slurm config

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 705491982a..9e47baba5b 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -8,11 +8,9 @@ SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
 set -x
 PORT=$(( 8888 + $PORT_OFFSET ))
 
-SCHEDULER_RECV_INTERVAL=''
+SCHEDULER_RECV_INTERVAL=10
 
-if [[ "$CONC" -lt 16 ]]; then
-    SCHEDULER_RECV_INTERVAL=10
-elif [[ "$CONC" -gte 16 ]]; then
+if [[ "$CONC" -gte 16 ]]; then
     SCHEDULER_RECV_INTERVAL=30
 fi
 

From 6f18531831e58372ef402d8627f8a9dfc47cdcae Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 20:13:41 -0700
Subject: [PATCH 16/23] slurm config

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 9e47baba5b..13386331d3 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -20,7 +20,7 @@ python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \
 --chunked-prefill-size 16384 \
---enable-ep-moe --quantization modelopt_fp4  --enable-flashinfer-allreduce-fusion --scheduler-recv-interval 10 \
+--enable-ep-moe --quantization modelopt_fp4  --enable-flashinfer-allreduce-fusion --scheduler-recv-interval $SCHEDULER_RECV_INTERVAL \
 --enable-symm-mem  --disable-radix-cache --attention-backend trtllm_mla --enable-flashinfer-trtllm-moe --stream-interval 10 \
 > $SERVER_LOG 2>&1 &
 

From 07c41fcafc14d496fc153b60f29058b4d04c4ca1 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 20:15:08 -0700
Subject: [PATCH 17/23] slurm config

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 13386331d3..de93a855f3 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -10,7 +10,7 @@ PORT=$(( 8888 + $PORT_OFFSET ))
 
 SCHEDULER_RECV_INTERVAL=10
 
-if [[ "$CONC" -gte 16 ]]; then
+if [[ "$CONC" -ge 16 ]]; then
     SCHEDULER_RECV_INTERVAL=30
 fi
 

From 53d1e9c671b9374a726380d509e88c237689890b Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 20:33:39 -0700
Subject: [PATCH 18/23] slurm config

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index de93a855f3..db17f4eb7c 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -10,7 +10,7 @@ PORT=$(( 8888 + $PORT_OFFSET ))
 
 SCHEDULER_RECV_INTERVAL=10
 
-if [[ "$CONC" -ge 16 ]]; then
+if [[ $CONC -ge 16 ]]; then
     SCHEDULER_RECV_INTERVAL=30
 fi
 

From fa4f926278ec8bcd422ca6ddd233bc9c07bcabb4 Mon Sep 17 00:00:00 2001
From: Kaunil Dhruv <kdhruv@nvidia.com>
Date: Fri, 19 Sep 2025 20:47:54 -0700
Subject: [PATCH 19/23] slurm config

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index db17f4eb7c..7f32558add 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -16,7 +16,7 @@ fi
 
 echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
 
-python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
+PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \
 --chunked-prefill-size 16384 \

From c3f8da874bdea732fb70bf8c19edb37b73c443ab Mon Sep 17 00:00:00 2001
From: Kedar Potdar <115327600+kedarpotdar-nv@users.noreply.github.com>
Date: Fri, 19 Sep 2025 21:02:32 -0700
Subject: [PATCH 20/23] Update dsr1-tmpl.yml

update runner
---
 .github/workflows/dsr1-tmpl.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/dsr1-tmpl.yml b/.github/workflows/dsr1-tmpl.yml
index 29a6ebb496..375e53291e 100644
--- a/.github/workflows/dsr1-tmpl.yml
+++ b/.github/workflows/dsr1-tmpl.yml
@@ -143,7 +143,7 @@ jobs:
     uses: ./.github/workflows/benchmark-tmpl.yml
     secrets: inherit
     with:
-      runner: b200
+      runner: b200-trt
       image: 'lmsysorg/sglang:v0.5.3rc1-cu129-b200'
       model: 'nvidia/DeepSeek-R1-0528-FP4'
       framework: 'sglang'

From 306c378ae88bccfcf7acd0591cc27dd6b250c65c Mon Sep 17 00:00:00 2001
From: Kedar Potdar <115327600+kedarpotdar-nv@users.noreply.github.com>
Date: Fri, 19 Sep 2025 21:03:42 -0700
Subject: [PATCH 21/23] Update dsr1_fp4_b200_slurm.sh

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index 7f32558add..c7cef22360 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -16,6 +16,8 @@ fi
 
 echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
 
+export FLASHINFER_LOGGING_LEVEL=0
+
 PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
 --tensor-parallel-size=$TP --data-parallel-size=1 \
 --cuda-graph-max-bs 256 --max-running-requests 256 --mem-fraction-static 0.85 --kv-cache-dtype fp8_e4m3 \

From 8e61c53c861ed65a1674d87d86d2b9dbff80aff0 Mon Sep 17 00:00:00 2001
From: Kedar Potdar <115327600+kedarpotdar-nv@users.noreply.github.com>
Date: Fri, 19 Sep 2025 21:09:15 -0700
Subject: [PATCH 22/23] Update dsr1_fp4_b200_slurm.sh

---
 benchmarks/dsr1_fp4_b200_slurm.sh | 1 -
 1 file changed, 1 deletion(-)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index c7cef22360..a816fea7b5 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -16,7 +16,6 @@ fi
 
 echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
 
-export FLASHINFER_LOGGING_LEVEL=0
 
 PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path $MODEL --host 0.0.0.0 --port $PORT --trust-remote-code \
 --tensor-parallel-size=$TP --data-parallel-size=1 \

From 8ed75862dc1d241f3e0448421ef24cafe1e6bb0e Mon Sep 17 00:00:00 2001
From: Kedar Potdar <115327600+kedarpotdar-nv@users.noreply.github.com>
Date: Fri, 19 Sep 2025 21:40:20 -0700
Subject: [PATCH 23/23] Update dsr1_fp4_b200_slurm.sh

add comment on scheduler recv logic
---
 benchmarks/dsr1_fp4_b200_slurm.sh | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/benchmarks/dsr1_fp4_b200_slurm.sh b/benchmarks/dsr1_fp4_b200_slurm.sh
index a816fea7b5..d3a6b22225 100644
--- a/benchmarks/dsr1_fp4_b200_slurm.sh
+++ b/benchmarks/dsr1_fp4_b200_slurm.sh
@@ -8,12 +8,17 @@ SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
 set -x
 PORT=$(( 8888 + $PORT_OFFSET ))
 
+
+# Default: recv every ~10 requests; if CONC ≥ 16, relax to ~30 requests between scheduler recv polls.
+
 SCHEDULER_RECV_INTERVAL=10
 
 if [[ $CONC -ge 16 ]]; then
     SCHEDULER_RECV_INTERVAL=30
 fi
 
+
+
 echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"