Skip to main content ホーム クリエイター ascend-ai-coding awesome-ascend-skills ais-bench
ais-bench AISBench Benchmark - AI model evaluation tool for Ascend NPU. Supports accuracy evaluation (service/local models on text, multimodal datasets), performance evaluation (latency, throughput, stress testing, steady-state, real traffic simulation), vLLM/Triton inference services, 15+ benchmarks (MMLU, GSM8K, MMMU, docvqa, ocrbench_v2, etc.), multi-turn dialogue, Function Call (BFCL), and custom datasets.
インストールへ移動 Skills Marketplace コミュニティが作成したAIスキルを発見・探索
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ascend-ai-coding/awesome-ascend-skills --skill ais-benchコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Zipをダウンロード ダウンロード中... name ais-bench description AISBench Benchmark - AI model evaluation tool for Ascend NPU. Supports accuracy evaluation (service/local models on text, multimodal datasets), performance evaluation (latency, throughput, stress testing, steady-state, real traffic simulation), vLLM/Triton inference services, 15+ benchmarks (MMLU, GSM8K, MMMU, docvqa, ocrbench_v2, etc.), multi-turn dialogue, Function Call (BFCL), and custom datasets. keywords ["ais-bench","aisbench","model evaluation","benchmark","accuracy evaluation","performance evaluation","vllm","multimodal","llm evaluation"] github_url https://github.com/AISBench/benchmark github_hash 03f0f43383c3efdb6ecdcc845eb24bc7a1537575 version 1.0.0 created_at 2026-02-25T00:00:00.000Z entry_point ais_bench
AISBench Benchmark Tool
AISBench Benchmark is a model evaluation tool built based on OpenCompass. It supports evaluation scenarios for both accuracy and performance testing of AI models on Ascend NPU.
Overview
Accuracy Evaluation : Accuracy verification of service-deployed models and local models on various QA and reasoning benchmark datasets, covering text, multimodal, and other scenarios.Performance Evaluation : Latency and throughput evaluation of service-deployed models, extreme performance testing under stress test scenarios, steady-state performance evaluation, and real business traffic simulation.
Supported Scenarios
Scenario Description Accuracy Evaluation Model accuracy on text/multimodal datasets Performance Evaluation Latency, throughput, stress testing Steady-State Performance Obtain true optimal system performance Real Traffic Simulation Simulate real business traffic patterns Multi-turn Dialogue Evaluate multi-turn conversation models Function Call (BFCL) Function calling capability evaluation
Supported Benchmarks
Text : GSM8K, MMLU, Ceval, FewCLUE series, dapo_math, levalMultimodal : docvqa, infovqa, ocrbench_v2, omnidocbench, mmmu, mmmu_pro, mmstar, videomme, textvqa, videobench, vocalsoundMulti-turn Dialogue : sharegpt, mtbenchFunction Call : BFCL (Berkeley Function Calling Leaderboard)
Installation
Environment Requirements
Python Version : Only Python 3.10 , 3.11 , or 3.12 is supported.
conda create --name ais_bench python=3.10 -y
conda activate ais_bench
Install from Source
git clone https://github.com/AISBench/benchmark.git
cd benchmark/
pip3 install -e ./ --use-pep517
Optional Dependencies
pip3 install -r requirements/api.txt
pip3 install -r requirements/extra.txt
pip3 install -r requirements/hf_vl_dependency.txt
pip3 install -r requirements/datasets/bfcl_dependencies.txt --no-deps
Quick Start
Basic Command Structure ais_bench --models <model_task> --datasets <dataset_task> [--summarizer example]
--models: Specifies the model task configuration--datasets: Specifies the dataset task configuration--summarizer: Result presentation task (default: example)
Find Configuration Files
ais_bench --models vllm_api_general_chat --datasets demo_gsm8k_gen_4_shot_cot_chat_prompt --search
Example: Service Model Accuracy Evaluation
Start vLLM inference service (follow vLLM documentation)
Prepare dataset :
Download GSM8K from opencompass
Extract to ais_bench/datasets/gsm8k/
Modify model configuration (vllm_api_general_chat.py):
from ais_bench.benchmark.models import VLLMCustomAPIChat
models = [
dict (
attr="service" ,
type =VLLMCustomAPIChat,
abbr='vllm-api-general-chat' ,
path="" ,
model="" ,
stream=False ,
request_rate=0 ,
retry=2 ,
api_key="" ,
host_ip="localhost" ,
host_port=8080 ,
url="" ,
max_out_len=512 ,
batch_size=1 ,
trust_remote_code=False ,
generation_kwargs=dict (
temperature=0.01 ,
ignore_eos=False ,
)
)
]
Run evaluation :
ais_bench --models vllm_api_general_chat --datasets demo_gsm8k_gen_4_shot_cot_chat_prompt
Output Results dataset version metric mode vllm_api_general_chat
----------------------- -------- -------- ----- ----------------------
demo_gsm8k 401e4c accuracy gen 62.50
Model Task Types
Service-Deployed Models Model Type Description vllm_api_general_chatGeneral vLLM API chat model vllm_api_function_call_chatFunction calling model (BFCL) triton_api_*Triton inference service
Local Models Model Type Description hf_*HuggingFace models vllm_offline_*vLLM offline inference
Performance Evaluation
Key Metrics Metric Description TTFT Time to First Token TPOT Time Per Output Token Throughput Tokens per second Latency Request latency (P50, P90, P99)
Performance Test Example ais_bench --models vllm_api_general_chat --datasets custom_performance \
--mode performance --concurrency 100
Steady-State Performance For obtaining true optimal system performance:
ais_bench --models vllm_api_general_chat --datasets sharegpt \
--stable-stage --duration 300
Real Traffic Simulation ais_bench --models vllm_api_general_chat --datasets custom \
--rps-distribution rps_config.json
Multi-task Evaluation
Multiple Models ais_bench --models model1 model2 model3 --datasets dataset1
Multiple Datasets ais_bench --models model1 --datasets dataset1 dataset2 dataset3
Parallel Execution ais_bench --models model1 model2 --datasets dataset1 dataset2 --parallel 4
Custom Datasets
Performance Custom Dataset Create a JSONL file with custom requests:
{ "input" : "Your prompt here" , "max_output_length" : 512 }
Accuracy Custom Dataset
Output Structure outputs/default/20250628_151326/
├── configs/ # Combined configuration
├── logs/ # Execution logs
│ ├── eval/ # Evaluation logs
│ └── infer/ # Inference logs
├── predictions/ # Raw inference results
├── results/ # Calculated scores
└── summary/ # Final summaries
├── summary_*.csv
├── summary_*.md
└── summary_*.txt
Task Management Interface During execution, a real-time task management interface displays:
Task name and progress
Time cost and status
Log path
Extended parameters
P key: Pause/Resume screen refreshCtrl+C: Exit
Common CLI Options Option Description --modelsModel task name(s) --datasetsDataset task name(s) --summarizerResult summarizer --searchList config file paths --debugPrint detailed logs --modeEvaluation mode (accuracy/performance) --parallelNumber of parallel tasks --resumeResume from breakpoint --failed-onlyRe-run failed cases only
Advanced Features
Breakpoint Resume ais_bench --models model1 --datasets dataset1 --resume outputs/default/20250628_151326
Failed Case Re-run ais_bench --models model1 --datasets dataset1 --failed-only --resume outputs/default/20250628_151326
Multi-file Dataset Merge For datasets like MMLU with multiple files:
ais_bench --models model1 --datasets mmlu_merged
Repeated Inference for pass@k ais_bench --models model1 --datasets dataset1 --repeat-n 5
Troubleshooting
Installation Issues
Python version mismatch : Use Python 3.10/3.11/3.12Dependency conflicts : Use conda environmentbfcl_eval pathlib issue : Use --no-deps flag
Runtime Issues
Model connection failed : Check host_ip, host_port, and service statusDataset not found : Download dataset to ais_bench/datasets/Memory issues : Reduce batch_size or use smaller dataset
Helper Scripts Quick utility scripts for common operations:
bash scripts/check_env.sh
bash scripts/run_accuracy_test.sh vllm_api_general_chat demo_gsm8k --host-port 8080
bash scripts/run_performance_test.sh vllm_api_general_chat sharegpt --concurrency 100
python scripts/parse_results.py outputs/default/20250628_151326
References Detailed documentation for specific use cases:
Templates Ready-to-use templates for custom evaluation:
