From e76cbda60a784fbbd732bf1e4d4f9aa19e4f17dc Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Fri, 17 Apr 2026 01:49:51 -0400
Subject: [PATCH 1/4] Add B300 config: qwen3.5-fp8-sglang

Mirrors the existing qwen3.5-fp8-b300-sglang-mtp config but without
speculative decoding, for a non-MTP baseline on B300.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 .github/configs/nvidia-master.yaml         | 18 +++++
 benchmarks/single_node/qwen3.5_fp8_b300.sh | 83 ++++++++++++++++++++++
 perf-changelog.yaml                        |  8 +++
 3 files changed, 109 insertions(+)
 create mode 100644 benchmarks/single_node/qwen3.5_fp8_b300.sh

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 484be6899..5b550879c 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -1882,6 +1882,24 @@ qwen3.5-fp8-b300-sglang-mtp:
     search-space:
     - { tp: 4, ep: 1, conc-start: 4, conc-end: 256, spec-decoding: mtp }
 
+qwen3.5-fp8-b300-sglang:
+  image: lmsysorg/sglang:v0.5.10.post1-cu130
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: b300
+  precision: fp8
+  framework: sglang
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 }
+
 kimik2.5-int4-b200-vllm:
   image: vllm/vllm-openai:v0.15.1
   model: moonshotai/Kimi-K2.5
diff --git a/benchmarks/single_node/qwen3.5_fp8_b300.sh b/benchmarks/single_node/qwen3.5_fp8_b300.sh
new file mode 100644
index 000000000..b87d25e91
--- /dev/null
+++ b/benchmarks/single_node/qwen3.5_fp8_b300.sh
@@ -0,0 +1,83 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    EP_SIZE
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+CONTEXT_LENGTH=$((ISL + OSL + 20))
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    CONTEXT_LENGTH="$EVAL_MAX_MODEL_LEN"
+fi
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0.0.0.0 --port=$PORT \
+--trust-remote-code \
+--tensor-parallel-size=$TP --data-parallel-size=1 --expert-parallel-size=$EP_SIZE \
+--enable-symm-mem \
+--disable-radix-cache \
+--quantization fp8 \
+--kv-cache-dtype fp8_e4m3 \
+--mamba-ssm-dtype bfloat16 \
+--attention-backend trtllm_mha \
+--moe-runner-backend flashinfer_trtllm \
+--cuda-graph-max-bs $CONC \
+--max-running-requests $CONC \
+--max-prefill-tokens 16384 \
+--chunked-prefill-size 16384 \
+--mem-fraction-static 0.8 \
+--stream-interval 50 \
+--scheduler-recv-interval 10 \
+--tokenizer-worker-num 6 \
+--tokenizer-path $MODEL \
+--context-length $CONTEXT_LENGTH > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# 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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --use-chat-template
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7ca73d687..730a95afe 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1388,3 +1388,11 @@
     - "Image: lmsysorg/sglang:v0.5.10.post1-cu130"
     - "EAGLE speculative decoding with MTP, TP=4, concurrency 4-256 for 1k1k and 8k1k"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1035
+
+- config-keys:
+    - qwen3.5-fp8-b300-sglang
+  description:
+    - "Add Qwen3.5-397B-A17B-FP8 B300 SGLang benchmark (non-MTP)"
+    - "Image: lmsysorg/sglang:v0.5.10.post1-cu130"
+    - "TP=4, concurrency 4-256 for 1k1k and 8k1k"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX

From 8685590de4e464220b47977b909e23f60bf86e79 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Fri, 17 Apr 2026 01:50:15 -0400
Subject: [PATCH 2/4] Fill in PR link for qwen3.5-fp8-b300-sglang changelog
 entry

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 730a95afe..3f081c240 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1395,4 +1395,4 @@
     - "Add Qwen3.5-397B-A17B-FP8 B300 SGLang benchmark (non-MTP)"
     - "Image: lmsysorg/sglang:v0.5.10.post1-cu130"
     - "TP=4, concurrency 4-256 for 1k1k and 8k1k"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1048

From 90a3b460a44d085a2f8536dc6e69cf899a712fe1 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Fri, 17 Apr 2026 02:47:10 -0400
Subject: [PATCH 3/4] Pin B300 single-node salloc to known-good nodelist

B300 cluster has a subset of nodes with hardware/network issues that cause
benchmarks to hang or fail to start. Constrain salloc to the known-working
set (b300-[001-006,008-012,017-020]) until those nodes are remediated.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 runners/launch_b300-nv.sh | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 4bb623780..f4181e893 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -222,7 +222,9 @@ else
     FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
 
-    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$TP --exclusive --time=180 --no-shell --job-name="$RUNNER_NAME"
+    # Pin to the known-good B300 nodes; others have hardware/network issues that
+    # cause benchmarks to hang or fail to start.
+    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --nodelist=b300-[001-006,008-012,017-020] --gres=gpu:$TP --exclusive --time=180 --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID bash -c "enroot import -o $SQUASH_FILE docker://$IMAGE"

From e18c3087e0d5e6b1f29cce5a51af0a4b56ff99f9 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Fri, 17 Apr 2026 02:50:45 -0400
Subject: [PATCH 4/4] Request a single node from the B300 nodelist

Without -N 1, salloc would allocate every node matching --nodelist; the
single-node benchmark only needs one.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 runners/launch_b300-nv.sh | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index f4181e893..3d863b54c 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -222,9 +222,9 @@ else
     FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
 
-    # Pin to the known-good B300 nodes; others have hardware/network issues that
-    # cause benchmarks to hang or fail to start.
-    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --nodelist=b300-[001-006,008-012,017-020] --gres=gpu:$TP --exclusive --time=180 --no-shell --job-name="$RUNNER_NAME"
+    # Pin to one of the known-good B300 nodes; others have hardware/network
+    # issues that cause benchmarks to hang or fail to start.
+    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --nodelist=b300-[001-006,008-012,017-020] -N 1 --gres=gpu:$TP --exclusive --time=180 --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID bash -c "enroot import -o $SQUASH_FILE docker://$IMAGE"