From 1d765943dee930bedd5e2e7ca7cb9660b398e932 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 23 Jul 2026 15:15:14 -0700
Subject: [PATCH 1/3] feat: add GLM-5.1 FP8 B300 SGLang config
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add the B300 SGLang benchmark entry and script with writable model-cache fallback.

中文：新增 B300 SGLang 基准测试配置和脚本，并在模型未预置时使用可写缓存。
---
 .../fixed_seq_len/glm5_fp8_b300.sh            | 102 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  19 ++++
 perf-changelog.yaml                           |   9 ++
 3 files changed, 130 insertions(+)
 create mode 100644 benchmarks/single_node/fixed_seq_len/glm5_fp8_b300.sh

diff --git a/benchmarks/single_node/fixed_seq_len/glm5_fp8_b300.sh b/benchmarks/single_node/fixed_seq_len/glm5_fp8_b300.sh
new file mode 100644
index 000000000..018b666a0
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/glm5_fp8_b300.sh
@@ -0,0 +1,102 @@
+#!/usr/bin/env bash
+
+# NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/GLM/GLM-5.1
+# does not have a B300-specific recipe, so this script reuses the GLM5 FP8
+# B200 SGLang recipe until B300-specific guidance is available.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+# `hf download` creates the target directory if needed and is idempotent.
+# When MODEL_PATH is unset for a stand-alone run, fall back to the HF cache.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+    echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+export SGLANG_ENABLE_JIT_DEEPGEMM=1
+
+SERVER_LOG=/workspace/server.log
+
+echo "CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+EVAL_CONTEXT_ARGS=""
+if [[ "${EVAL_ONLY}" == "true" ]]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor
+
+set -x
+PYTHONNOUSERSITE=1 python3 -m sglang.launch_server \
+    --model-path "$MODEL_PATH" \
+    --served-model-name "$MODEL" \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --trust-remote-code \
+    --tensor-parallel-size "$TP" \
+    --data-parallel-size 1 \
+    --expert-parallel-size 1 \
+    --tool-call-parser glm47 \
+    --reasoning-parser glm45 \
+    --kv-cache-dtype fp8_e4m3 \
+    --quantization fp8 \
+    --attention-backend nsa \
+    --nsa-decode-backend trtllm \
+    --nsa-prefill-backend trtllm \
+    --moe-runner-backend flashinfer_trtllm \
+    --cuda-graph-max-bs "$CONC" \
+    --max-running-requests "$CONC" \
+    --mem-fraction-static 0.85 \
+    --chunked-prefill-size 32768 \
+    --max-prefill-tokens 32768 \
+    --enable-flashinfer-allreduce-fusion \
+    --disable-radix-cache \
+    --stream-interval 30 \
+    --model-loader-extra-config '{"enable_multithread_load": true}' \
+    $EVAL_CONTEXT_ARGS > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$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/
+
+if [[ "${RUN_EVAL}" == "true" ]]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index a167f4927..d823db9fe 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1281,6 +1281,25 @@ qwen3.5-fp4-b200-sglang-mtp:
       - { tp: 4, ep: 1, conc-start: 4, conc-end: 4, spec-decoding: mtp }
       - { tp: 2, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
 
+  # NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/GLM/GLM-5.1
+  # does not have a B300-specific recipe, so this config reuses the GLM5 FP8
+  # B200 SGLang recipe until B300-specific guidance is available.
+
+glm5-fp8-b300-sglang:
+  image: lmsysorg/sglang:v0.5.15.post1-cu130
+  model: zai-org/GLM-5.1-FP8
+  model-prefix: glm5
+  runner: b300
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, ep: 1, conc-start: 4, conc-end: 256 }
+
 qwen3.5-fp8-b200-sglang-mtp:
   image: lmsysorg/sglang:v0.5.14-cu130
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 335245a6c..1c0476f91 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5060,3 +5060,12 @@
     - "Re-pin VLLM_ROUTER_IMAGE to vllm/vllm-router:nightly-20260716-1fbcde7 (previous nightly-20260629-e667ebb was garbage-collected from Docker Hub)"
     - "Exclude known-bad nodes mia1-p01-g09,g14 from the disagg node pool"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2301
+
+- config-keys:
+    - glm5-fp8-b300-sglang
+  description:
+    - "Add the GLM-5.1 FP8 B300 SGLang configuration with the lmsysorg/sglang:v0.5.15.post1-cu130 image"
+    - "新增 GLM-5.1 FP8 B300 SGLang 配置，并使用 lmsysorg/sglang:v0.5.15.post1-cu130 镜像"
+    - "Use the writable /data/models/GLM-5.1-FP8 cache when the model is not pre-staged"
+    - "模型未预置时，使用可写的 /data/models/GLM-5.1-FP8 缓存"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From 4db6a5bacaaf21d07ee33f7499d8e79608f0b402 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 23 Jul 2026 16:08:12 -0700
Subject: [PATCH 2/3] chore: link perf changelog to #2320
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Record the pull request URL for the GLM-5.1 FP8 B300 SGLang configuration.

中文：记录 GLM-5.1 FP8 B300 SGLang 配置的拉取请求链接。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 1c0476f91..0756c1b9e 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5068,4 +5068,4 @@
     - "新增 GLM-5.1 FP8 B300 SGLang 配置，并使用 lmsysorg/sglang:v0.5.15.post1-cu130 镜像"
     - "Use the writable /data/models/GLM-5.1-FP8 cache when the model is not pre-staged"
     - "模型未预置时，使用可写的 /data/models/GLM-5.1-FP8 缓存"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2320

From 62b7be840e8788ced2da48c6c33c9731f8c82c7c Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 23 Jul 2026 16:15:22 -0700
Subject: [PATCH 3/3] feat: add GLM-5.1 FP8 B300 SGLang MTP config
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：新增 GLM-5.1 FP8 B300 SGLang MTP 配置，并将其变更日志关联到 #2320。
---
 .../fixed_seq_len/glm5_fp8_b300_mtp.sh        | 108 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  17 ++-
 perf-changelog.yaml                           |  11 ++
 3 files changed, 135 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/single_node/fixed_seq_len/glm5_fp8_b300_mtp.sh

diff --git a/benchmarks/single_node/fixed_seq_len/glm5_fp8_b300_mtp.sh b/benchmarks/single_node/fixed_seq_len/glm5_fp8_b300_mtp.sh
new file mode 100644
index 000000000..51cb4dfca
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/glm5_fp8_b300_mtp.sh
@@ -0,0 +1,108 @@
+#!/usr/bin/env bash
+
+# NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/GLM/GLM-5.1
+# does not have a B300-specific recipe, so this script reuses the GLM5 FP8
+# B200 SGLang recipe until B300-specific guidance is available.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME
+
+# `hf download` creates the target directory if needed and is idempotent.
+# When MODEL_PATH is unset for a stand-alone run, fall back to the HF cache.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+    echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+nvidia-smi
+
+export SGLANG_ENABLE_JIT_DEEPGEMM=1
+export SGLANG_ENABLE_SPEC_V2=1
+
+SERVER_LOG=/workspace/server.log
+
+echo "CONC: $CONC, ISL: $ISL, OSL: $OSL"
+
+EVAL_CONTEXT_ARGS=""
+if [[ "${EVAL_ONLY}" == "true" ]]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+
+start_gpu_monitor
+
+set -x
+PYTHONNOUSERSITE=1 python3 -m sglang.launch_server \
+    --model-path "$MODEL_PATH" \
+    --served-model-name "$MODEL" \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --trust-remote-code \
+    --tensor-parallel-size "$TP" \
+    --data-parallel-size 1 \
+    --expert-parallel-size 1 \
+    --tool-call-parser glm47 \
+    --reasoning-parser glm45 \
+    --kv-cache-dtype fp8_e4m3 \
+    --quantization fp8 \
+    --attention-backend nsa \
+    --nsa-decode-backend trtllm \
+    --nsa-prefill-backend trtllm \
+    --moe-runner-backend flashinfer_trtllm \
+    --cuda-graph-max-bs "$CONC" \
+    --max-running-requests "$CONC" \
+    --mem-fraction-static 0.85 \
+    --chunked-prefill-size 32768 \
+    --max-prefill-tokens 32768 \
+    --enable-flashinfer-allreduce-fusion \
+    --disable-radix-cache \
+    --stream-interval 30 \
+    --speculative-algorithm EAGLE \
+    --speculative-num-steps 3 \
+    --speculative-eagle-topk 1 \
+    --speculative-num-draft-tokens 4 \
+    --model-loader-extra-config '{"enable_multithread_load": true}' \
+    $EVAL_CONTEXT_ARGS > "$SERVER_LOG" 2>&1 &
+
+SERVER_PID=$!
+
+wait_for_server_ready --port "$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
+
+if [[ "${RUN_EVAL}" == "true" ]]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+stop_gpu_monitor
+set +x
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index d823db9fe..4c26e3061 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1282,7 +1282,7 @@ qwen3.5-fp4-b200-sglang-mtp:
       - { tp: 2, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp }
 
   # NOTE: At the time of submission, https://cookbook.sglang.io/autoregressive/GLM/GLM-5.1
-  # does not have a B300-specific recipe, so this config reuses the GLM5 FP8
+  # does not have a B300-specific recipe, so these configs reuse the GLM5 FP8
   # B200 SGLang recipe until B300-specific guidance is available.
 
 glm5-fp8-b300-sglang:
@@ -1300,6 +1300,21 @@ glm5-fp8-b300-sglang:
       search-space:
       - { tp: 8, ep: 1, conc-start: 4, conc-end: 256 }
 
+glm5-fp8-b300-sglang-mtp:
+  image: lmsysorg/sglang:v0.5.15.post1-cu130
+  model: zai-org/GLM-5.1-FP8
+  model-prefix: glm5
+  runner: b300
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, ep: 1, conc-start: 4, conc-end: 256, spec-decoding: mtp }
+
 qwen3.5-fp8-b200-sglang-mtp:
   image: lmsysorg/sglang:v0.5.14-cu130
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0756c1b9e..625136f43 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5069,3 +5069,14 @@
     - "Use the writable /data/models/GLM-5.1-FP8 cache when the model is not pre-staged"
     - "模型未预置时，使用可写的 /data/models/GLM-5.1-FP8 缓存"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2320
+
+- config-keys:
+    - glm5-fp8-b300-sglang-mtp
+  description:
+    - "Add the GLM-5.1 FP8 B300 SGLang MTP configuration with the lmsysorg/sglang:v0.5.15.post1-cu130 image"
+    - "新增 GLM-5.1 FP8 B300 SGLang MTP 配置，并使用 lmsysorg/sglang:v0.5.15.post1-cu130 镜像"
+    - "Enable EAGLE speculative decoding and use the chat template for benchmark requests"
+    - "启用 EAGLE 投机解码，并为基准测试请求使用聊天模板"
+    - "Use the writable /data/models/GLM-5.1-FP8 cache when the model is not pre-staged"
+    - "模型未预置时，使用可写的 /data/models/GLM-5.1-FP8 缓存"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2320