Files
xas-test-library/catalog/test-definitions.json
T

639 lines
19 KiB
JSON
Raw Normal View History

{
"schema_version": 1,
"reviewed_at": "2026-10-20",
"basis": "Reviewed README entry points, script purpose comments and call references; registration does not authorize service/GPU actions.",
"example_script_ids": [
"499b615a7dc7435eb30789b7a7dc090e",
"fe5a2b28c8b4449591ef55bd88299a6d"
],
"tests": [
{
"key": "gpu-suite",
"source_group": "company-gpu-suite",
"entry": "xas-gpu-test.sh",
"name": "GPU 综合验收套件",
"description": "公司 XAS GPU Test v3.5.0 的统一入口,包含硬件压力、多卡通信、模型性能及报告。内部模块和套件自身测试统一收在详情中。",
"kind": "suite",
"members": [
{
"source_group": "company-gpu-suite",
"pattern": "**/*"
}
]
},
{
"key": "deepseek-suite",
"source_group": "dsv41-trial-20261010",
"entry": "run-b-suite.sh",
"name": "DeepSeek 综合性能测试",
"description": "在第二套 DeepSeek 服务上完成单路曲线、并发吞吐、自然语言生成、多轮缓存和工具调用的一整轮测试。",
"kind": "suite",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "bench-serve.sh"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "curves.py"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "multiturn_cache.py"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "spec_bench.py"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "spec_prompts.py"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "toolcall_stress.py"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-report-phase*.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-d2.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-d3.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-d456.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-unc*.sh",
"role": "variant"
}
]
},
{
"key": "model-suite",
"source_group": "model-trials-20261011",
"entry": "run-suite.sh",
"name": "模型综合测试套件",
"description": "面向被测模型的一整轮速度、长上下文、功能与稳定性测试。不同模型的历史调度脚本收在相关文件中。",
"kind": "suite",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "bench-serve.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "curves.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "functional_check.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "multiturn_cache.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "needle.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "spec_bench.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "spec_prompts.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "toolcall_stress.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "run-glm-chain.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "run-mimo-chain.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "run-qwen-chain.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "serve-model.sh",
"role": "variant"
}
]
},
{
"key": "functional",
"source_group": "model-trials-20261011",
"entry": "functional_check.py",
"name": "模型功能检查",
"description": "检查接口功能、结构化输出、多轮交互等可用性,不等同于答案准确率评测。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "functional_check.py",
"role": "variant"
}
]
},
{
"key": "curves",
"source_group": "model-trials-20261011",
"entry": "curves.py",
"name": "单路性能曲线",
"description": "单请求逐项测量生成长度和上下文长度变化时的性能曲线。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "curves.py",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-curves.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "context_probe.py",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "run-mimo-recheck.sh",
"role": "variant"
}
]
},
{
"key": "language",
"source_group": "model-trials-20261011",
"entry": "spec_bench.py",
"name": "自然语言生成速度",
"description": "用固定自然语言题库测量生成速度、并发和不同解码设置的表现。",
"kind": "test",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "spec_prompts.py"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "spec_bench.py",
"role": "variant"
}
]
},
{
"key": "conversation-cache",
"source_group": "model-trials-20261011",
"entry": "multiturn_cache.py",
"name": "多轮对话缓存效果",
"description": "模拟多轮对话,观察前缀缓存对每轮首字时间的影响。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "multiturn_cache.py",
"role": "variant"
}
]
},
{
"key": "needle",
"source_group": "model-trials-20261011",
"entry": "needle.py",
"name": "长上下文检索",
"description": "在指定上下文长度和位置中检索目标信息,检查长上下文能力。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "needle.py",
"role": "variant"
}
]
},
{
"key": "concurrent-input",
"source_group": "model-trials-20261011",
"entry": "concurrent_prefill.py",
"name": "并发长输入读入",
"description": "检查多个长输入同时到达时的读入表现。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "concurrent_prefill.py",
"role": "variant"
}
]
},
{
"key": "cache-revisit",
"source_group": "model-trials-20261011",
"entry": "kv_revisit.py",
"name": "长会话缓存回访",
"description": "检查长会话缓存被挤出后再次访问时的等待时间。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "kv_revisit.py",
"role": "variant"
}
]
},
{
"key": "interference",
"source_group": "model-trials-20261011",
"entry": "prefill_interference.py",
"name": "长输入对聊天延迟的影响",
"description": "测量超长输入进入时,对其他交互请求延迟造成的影响。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "prefill_interference.py",
"role": "variant"
}
]
},
{
"key": "tool-calls",
"source_group": "model-trials-20261011",
"entry": "toolcall_stress.py",
"name": "工具调用并发稳定性",
"description": "检查并发工具调用、参数 JSON 和必填字段的稳定性。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "toolcall_stress.py",
"role": "variant"
}
]
},
{
"key": "throughput",
"source_group": "model-trials-20261011",
"entry": "bench-serve.sh",
"name": "单服务并发吞吐",
"description": "对一个已部署的模型服务测量不同并发下的吞吐和延迟。",
"kind": "test",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "bench-one.sh"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "bench-serve.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "bench.py",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-fresh-seed-dsv41.sh",
"role": "variant"
}
]
},
{
"key": "math",
"source_group": "dsv41-trial-20261010",
"entry": "math_check.py",
"name": "数学应用题回归",
"description": "通过可自动判对错的算术应用题检查解码配置变更后是否出现明显退化。",
"kind": "test",
"members": []
},
{
"key": "vision",
"source_group": "dsv41-trial-20261010",
"entry": "vision_check.py",
"name": "图片理解功能检查",
"description": "检查模型图片输入功能是否可用,需要准备图片素材。",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-d7v-d8.sh",
"role": "variant"
}
]
},
{
"key": "robust",
"source_group": "model-trials-20261011",
"entry": "robust_check.py",
"name": "异常输入与断连检查",
"description": "检查异常输入、请求中断等情况下的模型服务稳定性。",
"kind": "test",
"members": []
},
{
"key": "agent",
"source_group": "model-trials-20261011",
"entry": "agent_sim.py",
"name": "Agent 多轮场景测试",
"description": "用多轮 Agent 场景请求测量运行表现和 token 消耗,汇总脚本和历史调度链作为配套文件。",
"kind": "test",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "agent_summary.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "agent-chain.sh"
}
]
},
{
"key": "replicas",
"source_group": "model-trials-20261011",
"entry": "run-replicas.sh",
"name": "多实例并行吞吐",
"description": "测量多个模型实例同时运行时的合计能力;实例数、GPU 和端口仍由原配置确定。",
"kind": "suite",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "bench-replicas.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "bench-multi.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "spec_bench.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "spec_prompts.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "serve-replicas.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "serve-sets.sh"
}
]
},
{
"key": "engine-comparison",
"source_group": "model-trials-20261011",
"entry": "run-compare.sh",
"name": "推理引擎与解码配置对比",
"description": "比较推理引擎、权重精度和解码配置的速度表现;同类历史对照脚本不再单独平铺。",
"kind": "suite",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "bench-serve.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "curves.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "spec_bench.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "spec_prompts.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "qwen38-compare-chain.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "qwen38-fast-chain.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "qwen38-sglang-track.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "run-glm-mtp-retry.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "run-qwen-pin-check.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "run-qwen-shm-check.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-spec-compare.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-dspark-tests.sh",
"role": "variant"
}
]
},
{
"key": "soak",
"source_group": "model-trials-20261011",
"entry": "soak.sh",
"name": "长时间稳定性测试",
"description": "持续施加并发和长输入负载,观察失败、吞吐和资源变化。",
"kind": "suite",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "bench-serve.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "concurrent_prefill.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "soak24-chain.sh",
"role": "variant"
},
{
"source_group": "dsv41-trial-20261010",
"pattern": "run-d8.sh",
"role": "variant"
}
]
},
{
"key": "capacity",
"source_group": "model-trials-20261011",
"entry": "cap-8x-chain.sh",
"name": "单卡与整机并发上限",
"description": "测量单卡或单套服务的并发上限,以及多实例整机满载表现。",
"kind": "suite",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "bench-multi.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "serve-replicas.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "serve-sets.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "qwen38-cap-chain.sh",
"role": "variant"
},
{
"source_group": "model-trials-20261011",
"pattern": "x16-chain.sh",
"role": "variant"
}
]
},
{
"key": "cost",
"source_group": "model-trials-20261011",
"entry": "value-chain.sh",
"name": "性能与成本评估",
"description": "按聊天、检索和代码生成场景测量性能,结合既定价格口径汇总成本。",
"kind": "suite",
"members": [
{
"source_group": "model-trials-20261011",
"pattern": "bench-multi.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "bench-one.sh"
},
{
"source_group": "model-trials-20261011",
"pattern": "value_ladder.py"
},
{
"source_group": "model-trials-20261011",
"pattern": "value2-chain.sh",
"role": "variant"
}
]
},
{
"key": "light-smoke",
"managed_profile": "legacy_smoke",
"kind": "test",
"members": [
{
"source_group": "dsv41-trial-20261010",
"pattern": "smoke.sh",
"role": "variant"
}
]
},
{
"key": "drama3-asr",
"source_group": "drama3-tests-20261015",
"entry": "test_asr.py",
"name": "语音识别文件与实时转写",
"description": "用现有 Drama3 合成三段录音,核验 Qwen3-ASR 文件转写、实时转写、字错率和响应延迟;会占用共享语音服务。",
"kind": "suite",
"members": []
},
{
"key": "drama3-api",
"source_group": "drama3-tests-20261015",
"entry": "test_api.py",
"name": "语音合成网关兼容性与质量回归",
"description": "核验本地语音网关的多格式、四种音色、流式、变速、SDK 兼容和八路并发;保留原 selftest 凭据读取方式。",
"kind": "suite",
"members": []
},
{
"key": "drama3-voice-chat",
"source_group": "drama3-tests-20261015",
"entry": "test_voice_chat.py",
"name": "实时语音对话端到端回归",
"description": "核验输入、回答、打断和四人对话的完整链路;当前语音服务预设仍指向旧 DeepSeek 模型名,待服务方确认后再开放。",
"kind": "suite",
"members": [
{
"source_group": "drama3-tests-20261015",
"pattern": "test_voice_logic.py"
}
]
},
{
"key": "drama3-engine",
"source_group": "drama3-tests-20261015",
"entry": "test_drama3.py",
"name": "Drama3 引擎合成与流式回归",
"description": "核验 Drama3 引擎的合成、克隆、流式和并发;依赖原管理凭据及外网参考音频下载,暂未接通执行。",
"kind": "suite",
"members": []
},
{
"key": "drama3-voice-capacity",
"source_group": "drama3-tests-20261015",
"entry": "load_voice.py",
"name": "语音链路并发容量测试",
"description": "用模拟大脑测语音链路的并发容量、首声延迟和播放卡顿;需补齐临时辅助容器生命周期与每次资源上限。",
"kind": "suite",
"members": [
{
"source_group": "drama3-tests-20261015",
"pattern": "mock_brain.py"
},
{
"source_group": "drama3-tests-20261015",
"pattern": "load_summary.py"
}
]
}
]
}