From efd5210f59cbadfbf8957277b156512013966ae8 Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Wed, 1 Apr 2026 18:30:21 -0500
Subject: [PATCH 1/6] Add Qwen3.5 MXFP4 single-node MI355X SGLang benchmark
 (TP8)

Uses amd/Qwen3.5-397B-A17B-MXFP4 with nightly ROCm SGLang image
(rocm/sgl-dev:v0.5.10rc0-rocm720-mi35x-20260330) which includes
MXFP4 support from sgl-project/sglang#21234.
---
 .github/configs/amd-master.yaml              | 18 ++++++
 benchmarks/single_node/qwen3.5_fp4_mi355x.sh | 68 ++++++++++++++++++++
 perf-changelog.yaml                          |  8 +++
 3 files changed, 94 insertions(+)
 create mode 100755 benchmarks/single_node/qwen3.5_fp4_mi355x.sh

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 72ef89c09..25ee52d0b 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -203,6 +203,24 @@ qwen3.5-fp8-mi355x-sglang:
     search-space:
     - { tp: 8, conc-start: 4, conc-end: 64 }
 
+qwen3.5-fp4-mi355x-sglang:
+  image: rocm/sgl-dev:v0.5.10rc0-rocm720-mi35x-20260330
+  model: amd/Qwen3.5-397B-A17B-MXFP4
+  model-prefix: qwen3.5
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: false
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - { tp: 8, conc-start: 4, conc-end: 64 }
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - { tp: 8, conc-start: 4, conc-end: 64 }
+
 qwen3.5-fp8-mi300x-sglang:
   image: lmsysorg/sglang:v0.5.9-rocm720-mi30x
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/benchmarks/single_node/qwen3.5_fp4_mi355x.sh b/benchmarks/single_node/qwen3.5_fp4_mi355x.sh
new file mode 100755
index 000000000..06b9a2ae8
--- /dev/null
+++ b/benchmarks/single_node/qwen3.5_fp4_mi355x.sh
@@ -0,0 +1,68 @@
+#!/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 $SLURMD_NODENAME"
+fi
+
+hf download "$MODEL"
+
+export SGLANG_USE_AITER=1
+export ROCM_QUICK_REDUCE_QUANTIZATION=INT4
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+set -x
+python3 -m sglang.launch_server --model-path=$MODEL --trust-remote-code \
+--host=0.0.0.0 --port=$PORT \
+--tensor-parallel-size=$TP \
+--attention-backend aiter \
+--mem-fraction-static=0.9 \
+--model-loader-extra-config '{"enable_multithread_load": true}' \
+--watchdog-timeout 1200 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+run_benchmark_serving \
+    --model "$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"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a82882f61..22f052fa2 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1,3 +1,11 @@
+- config-keys:
+    - qwen3.5-fp4-mi355x-sglang
+  description:
+    - "Add Qwen3.5 397B MXFP4 single-node MI355X SGLang benchmark (TP8)"
+    - "Uses nightly SGLang ROCm image with AMD MXFP4 support (sgl-project/sglang#21234)"
+    - "Model: amd/Qwen3.5-397B-A17B-MXFP4"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
 - config-keys:
     - kimik2.5-int4-mi300x-vllm
   description:

From bf5ce494818397952dcf8f50b647fd27b317b2e5 Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Wed, 1 Apr 2026 18:38:10 -0500
Subject: [PATCH 2/6] Move Qwen3.5 MXFP4 changelog entry to bottom of
 perf-changelog.yaml

---
 perf-changelog.yaml | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 22f052fa2..80e1f62a4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1,11 +1,3 @@
-- config-keys:
-    - qwen3.5-fp4-mi355x-sglang
-  description:
-    - "Add Qwen3.5 397B MXFP4 single-node MI355X SGLang benchmark (TP8)"
-    - "Uses nightly SGLang ROCm image with AMD MXFP4 support (sgl-project/sglang#21234)"
-    - "Model: amd/Qwen3.5-397B-A17B-MXFP4"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
-
 - config-keys:
     - kimik2.5-int4-mi300x-vllm
   description:
@@ -1243,3 +1235,11 @@
     - "New model support on ATOM framework"
     - "Kimi-K2.5 FP4, and MiniMax-M2.5 FP8 configs added for MI355X ATOM"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/963
+
+- config-keys:
+    - qwen3.5-fp4-mi355x-sglang
+  description:
+    - "Add Qwen3.5 397B MXFP4 single-node MI355X SGLang benchmark (TP8)"
+    - "Uses nightly SGLang ROCm image with AMD MXFP4 support (sgl-project/sglang#21234)"
+    - "Model: amd/Qwen3.5-397B-A17B-MXFP4"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From 002fd662a5d92307ea5c752b8b4509603f80a23f Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Wed, 1 Apr 2026 18:40:52 -0500
Subject: [PATCH 3/6] Remove ROCM_QUICK_REDUCE_QUANTIZATION=INT4 from Qwen3.5
 MXFP4 script

Upstream PR (sgl-project/sglang#21234) does not use this flag.
---
 benchmarks/single_node/qwen3.5_fp4_mi355x.sh | 1 -
 1 file changed, 1 deletion(-)

diff --git a/benchmarks/single_node/qwen3.5_fp4_mi355x.sh b/benchmarks/single_node/qwen3.5_fp4_mi355x.sh
index 06b9a2ae8..e4734e66a 100755
--- a/benchmarks/single_node/qwen3.5_fp4_mi355x.sh
+++ b/benchmarks/single_node/qwen3.5_fp4_mi355x.sh
@@ -18,7 +18,6 @@ fi
 hf download "$MODEL"
 
 export SGLANG_USE_AITER=1
-export ROCM_QUICK_REDUCE_QUANTIZATION=INT4
 
 SERVER_LOG=/workspace/server.log
 PORT=${PORT:-8888}

From 0a42a1bc86b67767fe957e92b5f74b8f62d0eed9 Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Wed, 1 Apr 2026 19:51:29 -0500
Subject: [PATCH 4/6] Switch Qwen3.5 MXFP4 MI355X from TP8 to TP4

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

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 25ee52d0b..5b316bf2e 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -215,11 +215,11 @@ qwen3.5-fp4-mi355x-sglang:
   - isl: 1024
     osl: 1024
     search-space:
-    - { tp: 8, conc-start: 4, conc-end: 64 }
+    - { tp: 4, conc-start: 4, conc-end: 64 }
   - isl: 8192
     osl: 1024
     search-space:
-    - { tp: 8, conc-start: 4, conc-end: 64 }
+    - { tp: 4, conc-start: 4, conc-end: 64 }
 
 qwen3.5-fp8-mi300x-sglang:
   image: lmsysorg/sglang:v0.5.9-rocm720-mi30x

From 49ca9070c26e79d9632db67dca377901ac9b607b Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Wed, 1 Apr 2026 20:05:15 -0500
Subject: [PATCH 5/6] Update changelog: TP8 -> TP4, set pr-link to #994

---
 perf-changelog.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 80e1f62a4..2d85bd2ba 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1239,7 +1239,7 @@
 - config-keys:
     - qwen3.5-fp4-mi355x-sglang
   description:
-    - "Add Qwen3.5 397B MXFP4 single-node MI355X SGLang benchmark (TP8)"
+    - "Add Qwen3.5 397B MXFP4 single-node MI355X SGLang benchmark (TP4)"
     - "Uses nightly SGLang ROCm image with AMD MXFP4 support (sgl-project/sglang#21234)"
     - "Model: amd/Qwen3.5-397B-A17B-MXFP4"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/994

From e973588f2a1caf361bc81955369fa7eb4ff0cf27 Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Thu, 2 Apr 2026 10:06:39 -0500
Subject: [PATCH 6/6] Switch to upstream lmsysorg/sglang-rocm nightly image
 (20260401)

---
 .github/configs/amd-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 5b316bf2e..61257c4bc 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -204,7 +204,7 @@ qwen3.5-fp8-mi355x-sglang:
     - { tp: 8, conc-start: 4, conc-end: 64 }
 
 qwen3.5-fp4-mi355x-sglang:
-  image: rocm/sgl-dev:v0.5.10rc0-rocm720-mi35x-20260330
+  image: lmsysorg/sglang-rocm:v0.5.10rc0-rocm720-mi35x-20260401
   model: amd/Qwen3.5-397B-A17B-MXFP4
   model-prefix: qwen3.5
   runner: mi355x