From ce4d3642c9ec9abcda3932455cba9f6a75dfb3de Mon Sep 17 00:00:00 2001 From: seungrokj Date: Sat, 25 Apr 2026 14:02:29 +0900 Subject: [PATCH 1/2] Day 0 Qwen3.5-397B-A17B BF16 MI355X Atom benchmarks Co-Authored-By: Claude Sonnet 4.6 --- .github/configs/amd-master.yaml | 20 +++++ .../single_node/qwen3.5_bf16_mi355x_atom.sh | 82 +++++++++++++++++++ perf-changelog.yaml | 7 ++ 3 files changed, 109 insertions(+) create mode 100644 benchmarks/single_node/qwen3.5_bf16_mi355x_atom.sh diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml index 7700edf09..3027f67ac 100644 --- a/.github/configs/amd-master.yaml +++ b/.github/configs/amd-master.yaml @@ -149,6 +149,26 @@ qwen3.5-bf16-mi355x-sglang-mtp: search-space: - { tp: 8, ep: 1, conc-start: 4, conc-end: 256, spec-decoding: mtp } +qwen3.5-bf16-mi355x-atom: + image: rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post + model: Qwen/Qwen3.5-397B-A17B + model-prefix: qwen3.5 + runner: mi355x + precision: bf16 + framework: atom + multinode: false + seq-len-configs: + - isl: 1024 + osl: 1024 + search-space: + - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 } + - { tp: 8, ep: 1, conc-start: 4, conc-end: 256 } + - isl: 8192 + osl: 1024 + search-space: + - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 } + - { tp: 8, ep: 1, conc-start: 4, conc-end: 256 } + qwen3.5-bf16-mi300x-sglang: image: lmsysorg/sglang:v0.5.10-rocm720-mi30x model: Qwen/Qwen3.5-397B-A17B diff --git a/benchmarks/single_node/qwen3.5_bf16_mi355x_atom.sh b/benchmarks/single_node/qwen3.5_bf16_mi355x_atom.sh new file mode 100644 index 000000000..2a8c67da0 --- /dev/null +++ b/benchmarks/single_node/qwen3.5_bf16_mi355x_atom.sh @@ -0,0 +1,82 @@ +#!/usr/bin/env bash + +source "$(dirname "$0")/../benchmark_lib.sh" + +check_env_vars \ + MODEL \ + TP \ + CONC \ + ISL \ + OSL \ + RANDOM_RANGE_RATIO \ + RESULT_FILENAME \ + EP_SIZE \ + DP_ATTENTION + +if [[ -n "$SLURM_JOB_ID" ]]; then + echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME" +fi + +echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION" + +SERVER_LOG=/workspace/server.log +PORT=${PORT:-8888} + +export OMP_NUM_THREADS=1 + +# Calculate max-model-len based on ISL and OSL +if [ "$ISL" = "1024" ] && [ "$OSL" = "1024" ]; then + CALCULATED_MAX_MODEL_LEN="" +else + CALCULATED_MAX_MODEL_LEN=" --max-model-len 10240 " +fi + +if [ "$EP_SIZE" -gt 1 ]; then + EP=" --enable-expert-parallel" +else + EP=" " +fi + +# Start GPU monitoring (power, temperature, clocks every second) +start_gpu_monitor +MEM_FRAC_STATIC=0.9 + +set -x + +python3 -m atom.entrypoints.openai_server \ + --model $MODEL \ + --server-port $PORT \ + -tp $TP \ + --kv_cache_dtype fp8 $CALCULATED_MAX_MODEL_LEN $EP \ + --gpu-memory-utilization $MEM_FRAC_STATIC \ + --trust-remote-code \ + > $SERVER_LOG 2>&1 & + +SERVER_PID=$! + +# Wait for server to be ready +wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID" + +export PYTHONDONTWRITEBYTECODE=1 +run_benchmark_serving \ + --model "$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" + append_lm_eval_summary +fi + +# Stop GPU monitoring +stop_gpu_monitor +set +x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index a6c811748..1f3cb61a2 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1812,3 +1812,10 @@ - "Topologies: low-conc 1p1d-dep8-tep8 (4 nodes, mirrored from NVIDIA srt-slurm PR #71 with offload kept and numa-bind dropped); mid 1p1d-dep8-dep16 (6 nodes) and high 3p1d-dep8-dep16 (10 nodes) hand-rolled, structurally derived from the kimi-k2.5 1k/1k pattern" - "Recipes stored under benchmarks/multi_node/srt-slurm-recipes/ and overlaid onto the upstream srt-slurm checkout at runtime" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1129 + +- config-keys: + - qwen3.5-bf16-mi355x-atom + description: + - "Day 0 Qwen3.5-397B-A17B BF16 MI355X benchmarks using Atom framework (rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post)" + - "Single-node sweep: TP4 and TP8, 1k/1k and 8k/1k ISL/OSL, FP8 KV cache, concurrency 4–256" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/ From 03ccd5750ecb83fc239027ca871f0ef04a56ef17 Mon Sep 17 00:00:00 2001 From: seungrokj Date: Sat, 25 Apr 2026 14:03:00 +0900 Subject: [PATCH 2/2] Backfill PR link in perf-changelog.yaml Co-Authored-By: Claude Sonnet 4.6 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 1f3cb61a2..4539699f8 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1818,4 +1818,4 @@ description: - "Day 0 Qwen3.5-397B-A17B BF16 MI355X benchmarks using Atom framework (rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post)" - "Single-node sweep: TP4 and TP8, 1k/1k and 8k/1k ISL/OSL, FP8 KV cache, concurrency 4–256" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/ + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1149