From 02487091201c0c5fce856c95160c60b5cab01479 Mon Sep 17 00:00:00 2001
From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com>
Date: Mon, 16 Feb 2026 19:01:36 +0000
Subject: [PATCH 1/2] Add Qwen3.5-397B-A17B FP8 B200 SGLang benchmark (no MTP)

Simple benchmark script without speculative decoding or MTP flags:
- No SGLANG_USE_CUDA_IPC_TRANSPORT
- No --reasoning-parser qwen3
- No --speculative-algo NEXTN or related flags
- TP=8, --mem-fraction-static 0.8
- Image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e

Co-authored-by: functionstackx <functionstackx@users.noreply.github.com>
---
 .github/configs/nvidia-master.yaml | 22 +++++++++++
 benchmarks/qwen3.5_fp8_b200.sh     | 62 ++++++++++++++++++++++++++++++
 perf-changelog.yaml                |  8 ++++
 3 files changed, 92 insertions(+)
 create mode 100755 benchmarks/qwen3.5_fp8_b200.sh

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 570e52d43..48a901353 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1782,6 +1782,28 @@ dsr1-fp8-b200-sglang:
     - { tp: 8, ep: 1, conc-start: 4, conc-end: 4 }
     - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 }
 
+qwen3.5-fp8-b200-sglang:
+  image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e
+  model: Qwen/Qwen3.5-397B-A17B
+  model-prefix: qwen3.5
+  runner: b200
+  precision: fp8
+  framework: sglang
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
+  - isl: 1024
+    osl: 8192
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 }
+
 dsr1-fp8-b200-sglang-mtp:
   image: lmsysorg/sglang:v0.5.8-cu130-amd64
   model: deepseek-ai/DeepSeek-R1-0528
diff --git a/benchmarks/qwen3.5_fp8_b200.sh b/benchmarks/qwen3.5_fp8_b200.sh
new file mode 100755
index 000000000..f6ded15e8
--- /dev/null
+++ b/benchmarks/qwen3.5_fp8_b200.sh
@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+if [[ -n "$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=${PORT:-8888}
+
+MEM_FRAC_STATIC=0.8
+
+ps aux
+
+set -x
+PYTHONNOUSERSITE=1 python3 -m sglang.launch_server \
+    --model-path=$MODEL \
+    --host=0.0.0.0 \
+    --port=$PORT \
+    --tensor-parallel-size=$TP \
+    --mem-fraction-static $MEM_FRAC_STATIC \
+    > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# 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 "$RESULT_FILENAME" \
+    --result-dir /workspace/
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT" --concurrent-requests $CONC
+    append_lm_eval_summary
+fi
+set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 5fafe2d64..fc247d5e4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -636,3 +636,11 @@
   description:
     - "Bump MI355X disagg FP8 recipe commit to fix perf regression on 8k1k DEP8"
   pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/701
+
+- config-keys:
+    - qwen3.5-fp8-b200-sglang
+  description:
+    - "Add Qwen3.5-397B-A17B FP8 B200 SGLang benchmark"
+    - "Image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e"
+    - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From ce477cfaa4d46252d7f54ee0434f2173224e9d70 Mon Sep 17 00:00:00 2001
From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com>
Date: Tue, 17 Feb 2026 03:26:33 +0000
Subject: [PATCH 2/2] Rename Qwen3.5-397B-A17B benchmark from FP8 to BF16
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The benchmark script runs without --quantization fp8, so the model
loads in BF16. Rename the script, config key, and changelog entry
to accurately reflect the precision.

- Rename benchmarks/qwen3.5_fp8_b200.sh → qwen3.5_bf16_b200.sh
- Update nvidia-master.yaml config key and precision field
- Update perf-changelog.yaml references and PR link

Co-authored-by: functionstackx <functionstackx@users.noreply.github.com>
---
 .github/configs/nvidia-master.yaml                       | 4 ++--
 benchmarks/{qwen3.5_fp8_b200.sh => qwen3.5_bf16_b200.sh} | 0
 perf-changelog.yaml                                      | 6 +++---
 3 files changed, 5 insertions(+), 5 deletions(-)
 rename benchmarks/{qwen3.5_fp8_b200.sh => qwen3.5_bf16_b200.sh} (100%)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 48a901353..46112ff50 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1782,12 +1782,12 @@ dsr1-fp8-b200-sglang:
     - { tp: 8, ep: 1, conc-start: 4, conc-end: 4 }
     - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 }
 
-qwen3.5-fp8-b200-sglang:
+qwen3.5-bf16-b200-sglang:
   image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e
   model: Qwen/Qwen3.5-397B-A17B
   model-prefix: qwen3.5
   runner: b200
-  precision: fp8
+  precision: bf16
   framework: sglang
   multinode: false
   seq-len-configs:
diff --git a/benchmarks/qwen3.5_fp8_b200.sh b/benchmarks/qwen3.5_bf16_b200.sh
similarity index 100%
rename from benchmarks/qwen3.5_fp8_b200.sh
rename to benchmarks/qwen3.5_bf16_b200.sh
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index fc247d5e4..f828d705c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -638,9 +638,9 @@
   pr-link: https://github.com/InferenceMAX/InferenceMAX/pull/701
 
 - config-keys:
-    - qwen3.5-fp8-b200-sglang
+    - qwen3.5-bf16-b200-sglang
   description:
-    - "Add Qwen3.5-397B-A17B FP8 B200 SGLang benchmark"
+    - "Add Qwen3.5-397B-A17B BF16 B200 SGLang benchmark"
     - "Image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e"
     - "TP=8, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/704