From e1d3b91ed45668bd1730f3ea06563ceac38e719b Mon Sep 17 00:00:00 2001 From: ankursingh-nv Date: Tue, 3 Mar 2026 13:48:43 -0800 Subject: [PATCH 1/5] add kimi fp4 configs --- .github/configs/nvidia-master.yaml | 25 ++++++++ benchmarks/single_node/kimik2.5_fp4_b200.sh | 68 +++++++++++++++++++++ 2 files changed, 93 insertions(+) create mode 100644 benchmarks/single_node/kimik2.5_fp4_b200.sh diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 6885f36cb..5b107b621 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1873,6 +1873,31 @@ kimik2.5-int4-h200-vllm: search-space: - { tp: 8, conc-start: 4, conc-end: 64 } +kimik2.5-fp4-b200-vllm: + image: vllm/vllm-openai:v0.15.1 + model: nvidia/Kimi-K2.5-NVFP4 + model-prefix: kimik2.5 + runner: b200 + precision: fp4 + framework: vllm + multinode: false + seq-len-configs: + - isl: 1024 + osl: 1024 + search-space: + - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } + - { tp: 4, ep: 4, conc-start: 4, conc-end: 64 } + - isl: 1024 + osl: 8192 + search-space: + - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } + - { tp: 4, ep: 4, conc-start: 4, conc-end: 64 } + - isl: 8192 + osl: 1024 + search-space: + - { tp: 8, ep: 1, conc-start: 4, conc-end: 128 } + - { tp: 4, ep: 4, 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/single_node/kimik2.5_fp4_b200.sh b/benchmarks/single_node/kimik2.5_fp4_b200.sh new file mode 100644 index 000000000..eafb48aca --- /dev/null +++ b/benchmarks/single_node/kimik2.5_fp4_b200.sh @@ -0,0 +1,68 @@ +#!/usr/bin/env bash + +source "$(dirname "$0")/../benchmark_lib.sh" + +check_env_vars \ + MODEL \ + TP \ + EP \ + CONC \ + ISL \ + OSL \ + MAX_MODEL_LEN \ + RANDOM_RANGE_RATIO \ + RESULT_FILENAME + +if [[ -n "$SLURM_JOB_ID" ]]; then + echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME" +fi + +hf download "$MODEL" + +nvidia-smi + +export TORCH_CUDA_ARCH_LIST="10.0" +export PYTHONNOUSERSITE=1 + +SERVER_LOG=/workspace/server.log +PORT=${PORT:-8888} + +set -x +vllm serve $MODEL --host 0.0.0.0 --port $PORT \ +--tensor-parallel-size $TP \ +--expert-parallel-size $EP \ +--gpu-memory-utilization 0.90 \ +--max-model-len $MAX_MODEL_LEN \ +--max-num-seqs $CONC \ +--reasoning-parser kimi_k2 \ +--tool-call-parser kimi_k2 \ +--compilation_config.pass_config.fuse_allreduce_rms true \ +--trust-remote-code \ +--disable-log-requests > $SERVER_LOG 2>&1 & + +SERVER_PID=$! + +# 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 "$RESULT_FILENAME" \ + --result-dir /workspace/ \ + --trust-remote-code + +# After throughput, run evaluation only if RUN_EVAL is true +if [ "${RUN_EVAL}" = "true" ]; then + run_eval --framework lm-eval --port "$PORT" --concurrent-requests $CONC + append_lm_eval_summary +fi +set +x From 8140f32d9ff32a853da51a2072e87deaaf93756c Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 3 Mar 2026 21:54:03 +0000 Subject: [PATCH 2/5] Add perf-changelog entry for Kimi K2.5 FP4 B200 vLLM config Co-authored-by: Ankur Singh --- perf-changelog.yaml | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 15d00da6d..2fbedf2a8 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -875,3 +875,12 @@ - "Server: reasoning-parser qwen3, tool-call-parser qwen3_coder, enable-flashinfer-allreduce-fusion, mem-fraction-static 0.8" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/855 +- config-keys: + - kimik2.5-fp4-b200-vllm + description: + - "Add Kimi K2.5 NVFP4 vLLM benchmark configuration for B200" + - "Model: nvidia/Kimi-K2.5-NVFP4 with --reasoning-parser kimi_k2 and --trust-remote-code" + - "Image: vllm/vllm-openai:v0.15.1" + - "TP=8/EP=1 (conc 4-128) and TP=4/EP=4 (conc 4-64) for 1k1k, 1k8k, and 8k1k" + - "Enables fuse_allreduce_rms compilation pass" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/862 From b140111aac7cea50e0796262167fc13100b32227 Mon Sep 17 00:00:00 2001 From: ankursingh-nv Date: Tue, 3 Mar 2026 13:58:48 -0800 Subject: [PATCH 3/5] update vllm version --- .github/configs/nvidia-master.yaml | 2 +- perf-changelog.yaml | 4 +--- 2 files changed, 2 insertions(+), 4 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 5b107b621..b2ba206f6 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -1874,7 +1874,7 @@ kimik2.5-int4-h200-vllm: - { tp: 8, conc-start: 4, conc-end: 64 } kimik2.5-fp4-b200-vllm: - image: vllm/vllm-openai:v0.15.1 + image: vllm/vllm-openai:v0.16.0 model: nvidia/Kimi-K2.5-NVFP4 model-prefix: kimik2.5 runner: b200 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 2fbedf2a8..37b369bca 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -880,7 +880,5 @@ description: - "Add Kimi K2.5 NVFP4 vLLM benchmark configuration for B200" - "Model: nvidia/Kimi-K2.5-NVFP4 with --reasoning-parser kimi_k2 and --trust-remote-code" - - "Image: vllm/vllm-openai:v0.15.1" - - "TP=8/EP=1 (conc 4-128) and TP=4/EP=4 (conc 4-64) for 1k1k, 1k8k, and 8k1k" - - "Enables fuse_allreduce_rms compilation pass" + - "Image: vllm/vllm-openai:v0.16.0" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/862 From 065f123dadbf4545e63940abd22ae06ceb709285 Mon Sep 17 00:00:00 2001 From: ankursingh-nv Date: Wed, 4 Mar 2026 11:02:43 -0800 Subject: [PATCH 4/5] fix ep variable name --- benchmarks/single_node/kimik2.5_fp4_b200.sh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/benchmarks/single_node/kimik2.5_fp4_b200.sh b/benchmarks/single_node/kimik2.5_fp4_b200.sh index eafb48aca..bb426659b 100644 --- a/benchmarks/single_node/kimik2.5_fp4_b200.sh +++ b/benchmarks/single_node/kimik2.5_fp4_b200.sh @@ -5,7 +5,7 @@ source "$(dirname "$0")/../benchmark_lib.sh" check_env_vars \ MODEL \ TP \ - EP \ + EP_SIZE \ CONC \ ISL \ OSL \ @@ -30,7 +30,7 @@ PORT=${PORT:-8888} set -x vllm serve $MODEL --host 0.0.0.0 --port $PORT \ --tensor-parallel-size $TP \ ---expert-parallel-size $EP \ +--expert-parallel-size $EP_SIZE \ --gpu-memory-utilization 0.90 \ --max-model-len $MAX_MODEL_LEN \ --max-num-seqs $CONC \ From 04107b3d532554df33c29ba07d72c8f79565212f Mon Sep 17 00:00:00 2001 From: ankursingh-nv Date: Wed, 4 Mar 2026 11:53:24 -0800 Subject: [PATCH 5/5] fix ep --- benchmarks/single_node/kimik2.5_fp4_b200.sh | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/benchmarks/single_node/kimik2.5_fp4_b200.sh b/benchmarks/single_node/kimik2.5_fp4_b200.sh index bb426659b..7b49acecb 100644 --- a/benchmarks/single_node/kimik2.5_fp4_b200.sh +++ b/benchmarks/single_node/kimik2.5_fp4_b200.sh @@ -5,7 +5,6 @@ source "$(dirname "$0")/../benchmark_lib.sh" check_env_vars \ MODEL \ TP \ - EP_SIZE \ CONC \ ISL \ OSL \ @@ -29,8 +28,7 @@ PORT=${PORT:-8888} set -x vllm serve $MODEL --host 0.0.0.0 --port $PORT \ ---tensor-parallel-size $TP \ ---expert-parallel-size $EP_SIZE \ +--tensor-parallel-size=$TP \ --gpu-memory-utilization 0.90 \ --max-model-len $MAX_MODEL_LEN \ --max-num-seqs $CONC \