From fcdccef3059e1bb278146d6745e67896e3076be7 Mon Sep 17 00:00:00 2001 From: Kedar Potdar Date: Fri, 27 Feb 2026 11:21:33 -0800 Subject: [PATCH 1/5] add fp4 qwen35 configs --- .github/configs/nvidia-master.yaml | 23 +++++++ benchmarks/single_node/qwen3.5_fp4_b200.sh | 78 ++++++++++++++++++++++ perf-changelog.yaml | 9 +++ 3 files changed, 110 insertions(+) create mode 100755 benchmarks/single_node/qwen3.5_fp4_b200.sh diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index b4084e5e4..07c690890 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1804,6 +1804,29 @@ qwen3.5-bf16-b200-sglang: search-space: - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 } +qwen3.5-fp4-b200-sglang: + image: lmsysorg/sglang:v0.5.9-cu129-amd64 + model: nvidia/Qwen3.5-397B-A17B-NVFP4 + model-prefix: qwen3.5 + runner: b200 + precision: fp4 + framework: sglang + multinode: false + seq-len-configs: + - isl: 1024 + osl: 1024 + search-space: + - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 } + - { tp: 8, ep: 8, conc-start: 128, conc-end: 128 } + - isl: 1024 + osl: 8192 + search-space: + - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } + - isl: 8192 + osl: 1024 + search-space: + - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } + 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_fp4_b200.sh b/benchmarks/single_node/qwen3.5_fp4_b200.sh new file mode 100755 index 000000000..380dcacac --- /dev/null +++ b/benchmarks/single_node/qwen3.5_fp4_b200.sh @@ -0,0 +1,78 @@ +#!/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 $SLURMD_NODENAME" +fi + +nvidia-smi + +hf download "$MODEL" + +export NCCL_NVLS_ENABLE=1 +export SGL_ENABLE_JIT_DEEPGEMM=false +export SGLANG_ENABLE_FLASHINFER_GEMM=true +export PYTHONUNBUFFERED=1 + +SERVER_LOG=/workspace/server.log +PORT=${PORT:-8888} +CONTEXT_LENGTH=$((ISL + OSL + 20)) +CUDA_GRAPH_MAX_BS=$CONC +MAX_RUNNING_REQUESTS=$((CONC > 128 ? CONC : 128)) +EXTRA_ARGS="" + +if [[ $CONC -gt 4 ]]; then + MEM_FRAC_STATIC=0.85 + EXTRA_ARGS="--kv-cache-dtype fp8_e4m3 --enable-symm-mem" +else + MEM_FRAC_STATIC=0.82 +fi + +echo "Config: ISL=$ISL, OSL=$OSL, CONC=$CONC, EP=$EP_SIZE, MEM=$MEM_FRAC_STATIC, CUDA_BS=$CUDA_GRAPH_MAX_BS, MAX_RR=$MAX_RUNNING_REQUESTS" + +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 --ep-size $EP_SIZE \ +--cuda-graph-max-bs $CUDA_GRAPH_MAX_BS --max-running-requests $MAX_RUNNING_REQUESTS \ +--mem-fraction-static $MEM_FRAC_STATIC --chunked-prefill-size 32768 --max-prefill-tokens 32768 \ +--context-length $CONTEXT_LENGTH --disable-radix-cache \ +--attention-backend trtllm_mha --moe-runner-backend flashinfer_trtllm \ +--enable-flashinfer-allreduce-fusion --scheduler-recv-interval 30 \ +--stream-interval 30 --quantization modelopt_fp4 \ +$EXTRA_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" --concurrent-requests $CONC + append_lm_eval_summary +fi +set +x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 30782b124..89e041a3f 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -754,3 +754,12 @@ - "Update SGLang image from v0.5.8 to v0.5.9 for AMD single-node DeepSeek R1 configs" - "Key changes: AITER v0.1.10.post3 with FP8 Prefill/Decode/KV Cache, FP8 prefill attention kernel, MORI EP two-batch overlapping, OOM fix for DeepSeek weight loading" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/816 + +- config-keys: + - qwen3.5-fp4-b200-sglang + description: + - "Add Qwen3.5-397B-A17B NVFP4 B200 SGLang benchmark config and launch script" + - "Image: lmsysorg/sglang:v0.5.9-cu129-amd64" + - "Model: nvidia/Qwen3.5-397B-A17B-NVFP4" + - "Configs: 1k1k (conc 4-64 ep1, conc 128 ep8), 1k8k (conc 4-128 ep1), 8k1k (conc 4-128 ep1)" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX From 4e43260d134823793b94c3428ea45049b287e66f Mon Sep 17 00:00:00 2001 From: Kedar Potdar Date: Fri, 27 Feb 2026 11:23:20 -0800 Subject: [PATCH 2/5] add pr # --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 89e041a3f..f2a03d54f 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -762,4 +762,4 @@ - "Image: lmsysorg/sglang:v0.5.9-cu129-amd64" - "Model: nvidia/Qwen3.5-397B-A17B-NVFP4" - "Configs: 1k1k (conc 4-64 ep1, conc 128 ep8), 1k8k (conc 4-128 ep1), 8k1k (conc 4-128 ep1)" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/820 From c738532e43fb8df006e3bf709b9369825c3f2941 Mon Sep 17 00:00:00 2001 From: Kedar Potdar Date: Fri, 27 Feb 2026 11:54:49 -0800 Subject: [PATCH 3/5] add tp4 --- .github/configs/nvidia-master.yaml | 3 +++ benchmarks/single_node/qwen3.5_fp4_b200.sh | 11 ++--------- 2 files changed, 5 insertions(+), 9 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 07c690890..58a46d4a5 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1816,15 +1816,18 @@ qwen3.5-fp4-b200-sglang: - isl: 1024 osl: 1024 search-space: + - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 } - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 } - { tp: 8, ep: 8, conc-start: 128, conc-end: 128 } - isl: 1024 osl: 8192 search-space: + - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 } - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } - isl: 8192 osl: 1024 search-space: + - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 } - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } kimik2.5-int4-b200-vllm: diff --git a/benchmarks/single_node/qwen3.5_fp4_b200.sh b/benchmarks/single_node/qwen3.5_fp4_b200.sh index 380dcacac..9773e7f58 100755 --- a/benchmarks/single_node/qwen3.5_fp4_b200.sh +++ b/benchmarks/single_node/qwen3.5_fp4_b200.sh @@ -30,14 +30,7 @@ PORT=${PORT:-8888} CONTEXT_LENGTH=$((ISL + OSL + 20)) CUDA_GRAPH_MAX_BS=$CONC MAX_RUNNING_REQUESTS=$((CONC > 128 ? CONC : 128)) -EXTRA_ARGS="" - -if [[ $CONC -gt 4 ]]; then - MEM_FRAC_STATIC=0.85 - EXTRA_ARGS="--kv-cache-dtype fp8_e4m3 --enable-symm-mem" -else - MEM_FRAC_STATIC=0.82 -fi +MEM_FRAC_STATIC=0.85 echo "Config: ISL=$ISL, OSL=$OSL, CONC=$CONC, EP=$EP_SIZE, MEM=$MEM_FRAC_STATIC, CUDA_BS=$CUDA_GRAPH_MAX_BS, MAX_RR=$MAX_RUNNING_REQUESTS" @@ -51,7 +44,7 @@ PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0. --attention-backend trtllm_mha --moe-runner-backend flashinfer_trtllm \ --enable-flashinfer-allreduce-fusion --scheduler-recv-interval 30 \ --stream-interval 30 --quantization modelopt_fp4 \ -$EXTRA_ARGS > $SERVER_LOG 2>&1 & +--kv-cache-dtype fp8_e4m3 --fp4-gemm-backend flashinfer_cutlass > $SERVER_LOG 2>&1 & SERVER_PID=$! From 5c0a4e8280fa19d953d24cdf7972e5898eb75444 Mon Sep 17 00:00:00 2001 From: Kedar Potdar Date: Fri, 27 Feb 2026 15:47:57 -0800 Subject: [PATCH 4/5] limit search space --- .github/configs/nvidia-master.yaml | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 58a46d4a5..41d3952b1 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1816,19 +1816,21 @@ qwen3.5-fp4-b200-sglang: - isl: 1024 osl: 1024 search-space: - - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 } - - { tp: 8, ep: 1, conc-start: 4, conc-end: 64 } - - { tp: 8, ep: 8, conc-start: 128, conc-end: 128 } + - { tp: 4, ep: 1, conc-start: 4, conc-end: 64 } + - { tp: 8, ep: 1, conc-start: 4, conc-end: 4 } + - { tp: 8, ep: 8, conc-start: 64, conc-end: 128 } - isl: 1024 osl: 8192 search-space: - - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 } - - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } + - { tp: 4, ep: 1, conc-start: 4, conc-end: 64 } + - { tp: 8, ep: 1, conc-start: 4, conc-end: 4 } + - { tp: 8, ep: 1, conc-start: 64, conc-end: 128 } - isl: 8192 osl: 1024 search-space: - - { tp: 4, ep: 1, conc-start: 4, conc-end: 32 } - - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } + - { tp: 4, ep: 1, conc-start: 4, conc-end: 64 } + - { tp: 8, ep: 1, conc-start: 4, conc-end: 4 } + - { tp: 8, ep: 1, conc-start: 64, conc-end: 128 } kimik2.5-int4-b200-vllm: image: vllm/vllm-openai:v0.15.1 From cf443319080f88db648e99c6f151fe32e06424a3 Mon Sep 17 00:00:00 2001 From: Kedar Potdar <115327600+kedarpotdar-nv@users.noreply.github.com> Date: Mon, 2 Mar 2026 10:34:21 -0800 Subject: [PATCH 5/5] newline in perf-changelog.yaml --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 44bb2e136..0df5b5dad 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -832,4 +832,4 @@ - "Image: lmsysorg/sglang:v0.5.9-cu129-amd64" - "Model: nvidia/Qwen3.5-397B-A17B-NVFP4" - "Configs: 1k1k (conc 4-64 ep1, conc 128 ep8), 1k8k (conc 4-128 ep1), 8k1k (conc 4-128 ep1)" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/820 \ No newline at end of file + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/820