公开数据 API
InferenceX API 参考文档
从 InferenceX 使用的同一组数据源中，读取基准测试数据、溯源信息、数据集、CollectiveX 数据和诊断数据。
- 规范版本
- v1 · OpenAPI 3.1
- 身份验证
- 已发布的只读端点不需要身份验证。
- 响应格式
- 除非端点明确说明 CSV，否则响应均为 JSON。日期使用 YYYY-MM-DD，时间戳使用 UTC ISO 8601。
- 基础 URL
- https://inferencex.semianalysis.com
快速入门
只需几步，即可从查看契约到获得真实响应。
- 1
查询可用配置
先查询 availability，选择实际可用的模型、硬件、框架和序列取值。
查询可用配置curl "https://inferencex.semianalysis.com/api/v1/availability" - 2
获取基准行
然后为受支持的展示模型请求最新原始基准行。
获取基准行curl "https://inferencex.semianalysis.com/api/v1/benchmarks?model=DeepSeek-R1-0528"
技能版本 · 1.2.0
通过智能体使用 API
inferencex-api 技能帮助智能体查找和使用公开 API，涵盖基准测试、溯源、数据集、CollectiveX 和诊断接口。已验证的单轮请求 PowerX 导出是首个完整示例。
@semianalysisai/inferencex-skills
需要 Node 24 或更新版本、npm，以及 Codex 或 Claude Code。安装和 API 查询均需联网。
查询通过公开 API 完成，无需数据库凭据。独立的 MCP server 有自己的配置流程；本技能不依赖它，也不会改动它的配置。
安装到项目
在项目目录中执行对应智能体的安装命令，然后在该项目中启动智能体会话。
Codex
npm exec --yes --package @semianalysisai/inferencex-skills@1.2.0 -- inferencex-skills install --target codexClaude Code
npm exec --yes --package @semianalysisai/inferencex-skills@1.2.0 -- inferencex-skills install --target claude升级时，将安装命令中的版本号改为新的已发布版本，并加上 --force 重新执行。默认会跳过已有技能。请先保存本地修改：--force 会覆盖同名文件，但保留不再随包提供的旧文件。
检查已复制的技能版本时，将上方命令中的 install 改为 status，并保留相同的 --target。Installer version 表示安装器版本，Installed version 表示项目内已复制技能的版本。旧版安装可能显示 unknown。
使用示例（英文）首个示例：实测 PowerX 数据
使用 inferencex-api 导出 DeepSeek-V4-Pro 最新可用的实测 PowerX 数据：
- 仅选取输入恰好为 8192、输出恰好为 1024 个 token 的单轮请求，并要求 strictV2。
- 在 InferenceX 仓库之外，使用已安装 CLI 的 powerx export 命令，分别生成 CSV 和 JSON 证据包。将 result.csv 或 result.json 与对应的 manifest.json 和 responses/ 一起保留，核验每个证据包后再报告结果。
- 区分实测单 GPU 功率、整个部署的 GPU 能耗与预留功率估算。
- 保留原始模型键、来源标识和 URL、测量日期，以及独立的快照元数据。
- 记录请求 URL、提取时间、包版本、本地筛选条件，以及返回和选中的数据条数。
- 说明数据行被排除的原因，并列出所请求指标的缺失项。
- 缺失指标保持不可用，真实零值保持为零。说明空结果的含义，不据此推断所有基准测试数据都不存在。PowerX 指南与直接导出
打开下方安装路径中的指南，查看验证规则、单位、缺失数据处理和溯源说明。随包提供的 Node 24 命令行工具（CLI）可在项目中直接运行，无需检出 InferenceX 仓库或提供数据库凭据。
- Codex
- .agents/skills/inferencex-api/references/powerx.md
- Claude Code
- .claude/skills/inferencex-api/references/powerx.md
运行已安装的 CLI（Codex）
mkdir -p evidence
node .agents/skills/inferencex-api/scripts/inferencex.mjs powerx export \
--model DeepSeek-V4-Pro --isl 8192 --osl 1024 \
--format csv --output-dir evidence/powerx-csv
node .agents/skills/inferencex-api/scripts/inferencex.mjs verify evidence/powerx-csvClaude Code 使用 .claude/skills/inferencex-api 路径。导出 JSON 时改用 --format json --output-dir evidence/powerx-json，并核验该目录。每次导出都使用新的输出目录，其中包含 result.csv 或 result.json、manifest.json 和 responses/。请保留整个证据包：即使结果为空，也能保留请求和数据覆盖范围的元数据。省略 --date 表示查询最新可用观测值，也可添加 --date YYYY-MM-DD 指定截止日期。
avg_power_w 是实测单 GPU 平均功率，单位为 W。schema v2 中不带角色前缀的 joules 指标表示整个部署的 GPU 能耗；带 prefill/decode 前缀的能耗仅对应相应角色。这些数据是已有观测值，不是新运行的基准测试，也不是设施总能耗测量值。
约定
适用于受支持接口的通用请求、错误与缓存行为。
- 客户端请求统计
- 可选的 inferencex-cli/<version> 或 inferencex-skill/<version> User-Agent 标识用于归类公开 GET 请求，包括缓存命中及重试。X-InferenceX-Traffic 区分 normal、ci 和 validation 流量。仅记录接口路径模板、包版本、客户端类型（CLI 或 skill）、部署环境与流量类别，不记录查询参数值或用户标识。包内客户端可设置 INFERENCEX_TELEMETRY=0 关闭来源标记。这些请求头不影响身份验证、响应数据或缓存。
- 错误
- JSON 错误包含 error 字符串。400 表示参数缺失或无效，404 表示请求的记录不存在，500 表示服务器查询失败。
- 缓存
- 只读端点可能由共享缓存提供。若 Accept-Encoding 允许，JSON 以 brotli 或 gzip 压缩，否则不压缩，并附带 Vary: Accept-Encoding。/api/v1/benchmarks 例外：由共享缓存负责压缩，因此其 ETag 和 Vary 头由该缓存决定。大多数基于数据库的 JSON 响应带有强 ETag；在 If-None-Match 中回传该值，数据未变化时将收到无响应体的 304。浏览器最多可复用这些响应 60 秒。新数据导入后，缓存刷新期间可能再返回一次旧响应。CollectiveX 响应会一直缓存，直到结果上传时被清除；框架版本使用一小时共享缓存。
- 标识符
- 基准结果 ID 和 GitHub 运行 ID 为正整数。批量诊断端点接受以逗号分隔并去重的 ID。
BenchmarkRow 与指标
理解主要基准测试响应数据及其中的实测字段。
- DWDP 预填充拓扑
可选字段 metrics.prefill_dwdp_size 记录每个分布式权重数据并行预填充组的 rank 数。prefill_tp 保持为 1，prefill_num_workers 统计组数。两个大小为 8 的组使用 16 颗预填充芯片。同组导航行提供可空字段 prefill_dwdp_size。缺失或为 1 表示未启用 DWDP。
结构BenchmarkRows示例{ "prefill_tp": 1, "prefill_num_workers": 2, "metrics": { "prefill_dwdp_size": 8 } }- BenchmarkRow
配置字段与 metrics 映射位于同一层级，时间指标以秒为单位。以 _per_gpu 结尾的吞吐量字段保留历史名称，实际表示每颗物理芯片每秒处理的 token 数，TPU 也使用此单位。num_prefill_gpu 和 num_decode_gpu 表示物理芯片数，与逻辑 TP 独立；聚合部署可能在两列中重复记录同一芯片数；只有 prefill_num_workers 与 decode_num_workers 均非零、或两列数值不同的分离式部署才应将两列相加；2025-09-29 至 2025-12-05 期间未记录 worker 的旧分离式数据同样在两列中重复记录整套部署的芯片数。可选字段 metrics.dp 记录产物中提供的数据并行度。判断 P99 inter-token latency 是否达标时，应使用以秒为单位的 p99_itl（乘以 1000 可换算为毫秒）。p99_tpot 表示请求内每个输出 token 的平均耗时；不能用 p99_intvty 的倒数代替 p99_itl。
结构BenchmarkRows示例{ "id": 421, "hardware": "h200_sxm", "framework": "vllm", "model": "dsr1", "precision": "fp8", "spec_method": "none", "disagg": false, "is_multinode": false, "prefill_tp": 8, "prefill_ep": 1, "prefill_dp_attention": false, "prefill_num_workers": 1, "decode_tp": 8, "decode_ep": 1, "decode_dp_attention": false, "decode_num_workers": 1, "num_prefill_gpu": 0, "num_decode_gpu": 8, "benchmark_type": "single_turn", "isl": 1024, "osl": 1024, "conc": 32, "offload_mode": "off", "image": "vllm/vllm-openai:v0.10.2", "recipe_fingerprint": "7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d", "metrics": { "median_ttft": 0.42, "median_tpot": 0.018, "tput_per_gpu": 128.4, "power_valid": 1, "power_metric_schema_version": 2, "avg_power_w": 678.5, "joules_per_output_token": 5.3, "joules_per_total_token": 2.65, "avg_temp_c": 61.2 }, "date": "2026-08-08", "run_url": "https://github.com/semianalysis/inference-benchmarks/actions/runs/123456789" }滚动查看完整示例
- 实测功率
基准测试数据行可能包含实测功率、能耗和 GPU 遥测指标（avg_power_w、avg_total_gpu_power_w、total_gpu_energy_j、p75_power_w、p75_total_gpu_power_w、p90_power_w、p90_total_gpu_power_w、joules_per_*、avg_temp_c、peak_temp_c、avg_util_pct、avg_mem_used_mb）。power_valid 有三种状态：1 表示测量窗口已通过验证；0 表示验证失败，生产端会移除实测值，摄取端也会再次清除，若仍有残留，应视为不可靠；缺失表示当前响应未提供验证结论。旧数据可能缺少该字段，但仅凭字段缺失，既无法判断缺失原因，也无法判断数据新旧或测量是否无效。power_metric_schema_version == 2 规定所有无前缀的 joules_per_* 字段均按整个部署统计能耗。未标注版本的分离式部署数据中，这些字段曾记录单个角色的能耗，因此其统计口径不明确。多节点和分离式运行中，各 worker 的功率和遥测明细位于 workers[]。power_invalid_reasons 列出生产端的原因码。power_audit 可在有效与无效行上提供测量窗口、设备和采样数量、生产端标识及保留的审计产物引用。缺少审计信息不能证明测量的新旧或有效性。NVL72（GB200/GB300）数据行可能附带 Grace 侧和计算模块指标，它们在同一测量窗口内积分：avg_cpu_socket_power_w、avg_total_cpu_power_w、total_cpu_energy_j，在每个 socket 都有模块传感器时的 avg_total_module_power_w / total_module_energy_j，以及在每个 socket 都上报相应供电轨时的 CPU 核心轨和 SysIO 轨分项 avg_total_cpu_rail_power_w / total_cpu_rail_energy_j、avg_total_cpu_sysio_power_w / total_cpu_sysio_energy_j。供电轨分项已包含在 Grace 侧总功耗内，不能再与之相加。单节点 AMD MI300X/MI325X/MI355X 数据行可能附带的 CPU 侧指标仅有 avg_cpu_socket_power_w、avg_total_cpu_power_w 和 total_cpu_energy_j，均来自 EPYC 的 RAPL 封装能耗计数器（sensor_kind 为 rapl_package）；该计数器没有 DRAM 域。cpu_power_valid 是这些指标独立于 power_valid 的三态验证结论：CPU 侧指标仅在 cpu_power_valid != 1 时被清除，与 power_valid 无关；power_audit.cpu 记录传感器类型、采集来源、socket 覆盖情况和原因码。模块读数已包含 GPU 板卡功耗。查询实测功率时，使用 powerValid=strictV2，仅保留 power_valid == 1 且 power_metric_schema_version == 2 的行。这是唯一支持的功率筛选值。常规基准测试请求应省略 powerValid，以保留缺少有效功率测量的结果。
结构BenchmarkRows示例{ "power_valid": 1, "power_metric_schema_version": 2, "avg_power_w": 678.5, "joules_per_output_token": 5.3, "cpu_power_valid": 1, "avg_cpu_socket_power_w": 250.5, "avg_total_cpu_power_w": 501 }- 以 ID 为键的映射
批量诊断响应是以十进制基准结果 ID 为键的 JSON 对象。缺少某个键表示该 ID 没有可用值。
结构Record<string, value>示例{ "421": true }- CollectiveX 版本
CollectiveX 读取需要明确指定受支持的契约版本。受支持版本：1。
结构CollectiveXDataset
端点参考
展开任一操作，即可查看参数、状态码和完整示例。
54 项操作
仪表板只读视图
通过显式筛选参数读取仪表板数据和计算结果。
GET测试版/api/v1/views/options获取 views API 的全部选项域
views API 的静态发现端点：模型（含数据库键、分类与发布日期）、序列、精度、硬件（含厂商、架构、TDP 及各供应商每小时成本）、推理框架、投机解码方法、百分位、x 轴与坐标缩放模式、图表指标（中英文标签、单位、极性）、快捷筛选域、可靠性时间范围，以及 overview/calculator/fleet 的选项集，并附仪表盘实际使用的默认值。所有取值直接来自仪表盘控件渲染所用的注册表，因此该响应是其余 /api/v1/views 端点的权威输入目录。分类为 deprecated 的模型会在推理图表的来源署名后添加 (已弃用)，英文页面显示 (Deprecated)；PNG 导出和非官方 run overlay 中也保留该后缀，不再显示“已弃用”背景水印。此变更仅影响展示，不改变视图数据。不访问数据库。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| format | query | enum | 可选 | 响应编码。该发现端点仅支持 JSON。 | json |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/options'响应
ViewOptions200
各 views 端点接受的全部选项域，以及与仪表盘一致的默认值。每个模型的 deprecatedSequences 包含全局弃用和该模型单独弃用的序列。Qwen3.5 的 8K/1K 已弃用；Qwen3.5 AgentX 保持启用，仍可显式选择历史结果。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"options"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"properties": {
"format": {
"type": "string",
"enum": [
"json"
]
}
},
"required": [
"format"
],
"additionalProperties": false
},
"models": {
"type": "array",
"items": {
"type": "object",
"properties": {
"name": {
"type": "string"
},
"dbKeys": {
"type": "array",
"items": {
"type": "string"
}
},
"category": {
"type": "string"
},
"releaseDate": {
"type": [
"string",
"null"
]
},
"compareSlug": {
"type": [
"string",
"null"
]
},
"deprecatedSequences": {
"type": "array",
"items": {
"type": "string"
},
"description": "Sequence keys categorized as deprecated for this model, including globally deprecated sequences. Historical selections remain available; this does not retire the model."
}
},
"required": [
"name",
"dbKeys",
"category",
"releaseDate",
"compareSlug",
"deprecatedSequences"
],
"additionalProperties": false
},
"description": "Models in selector order within each category, with Kimi-K3.1-Pro first. The bare inference page opens Kimi-K3.1-Pro; this navigation default does not change defaults.model or remove the inference API requirement for an explicit model."
},
"sequences": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"labelZh": {
"type": "string"
},
"urlSegment": {
"type": [
"string",
"null"
]
},
"isl": {
"type": [
"integer",
"null"
]
},
"osl": {
"type": [
"integer",
"null"
]
},
"kind": {
"type": "string"
},
"deprecated": {
"type": "boolean"
}
},
"required": [
"key",
"label",
"labelZh",
"urlSegment",
"isl",
"osl",
"kind",
"deprecated"
],
"additionalProperties": false
}
},
"precisions": {
"type": "array",
"items": {
"type": "string"
}
},
"hardware": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"vendor": {
"type": "string"
},
"arch": {
"type": "string"
},
"tdpW": {
"type": "number",
"description": "Per-chip rated TDP in watts, shared with chip pages. B300 uses 1100 W."
},
"costPerHour": {
"type": "object",
"properties": {
"h": {
"type": "number"
},
"r": {
"type": "number"
}
},
"required": [
"h",
"r"
],
"additionalProperties": false
}
},
"required": [
"key",
"label",
"vendor",
"arch",
"tdpW",
"costPerHour"
],
"additionalProperties": false
}
},
"frameworks": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"family": {
"type": [
"string",
"null"
]
}
},
"required": [
"key",
"label",
"family"
],
"additionalProperties": false
}
},
"specMethods": {
"type": "array",
"items": {
"type": "string"
}
},
"percentiles": {
"type": "array",
"items": {
"type": "string"
}
},
"xAxisModes": {
"type": "array",
"items": {
"type": "string"
}
},
"fixedSequenceStatistics": {
"type": "array",
"items": {
"type": "string"
}
},
"scaleModes": {
"type": "array",
"items": {
"type": "string"
}
},
"metrics": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"configKey": {
"type": "string"
},
"label": {
"type": "string"
},
"labelZh": {
"type": "string"
},
"unit": {
"type": [
"string",
"null"
]
},
"polarity": {
"type": [
"string",
"null"
]
},
"group": {
"type": [
"string",
"null"
]
},
"source": {
"type": "string"
}
},
"required": [
"key",
"configKey",
"label",
"labelZh",
"unit",
"polarity",
"group",
"source"
],
"additionalProperties": false
}
},
"quickFilters": {
"type": "object",
"properties": {
"vendors": {
"type": "array",
"items": {
"type": "string"
}
},
"frameworkFamilies": {
"type": "array",
"items": {
"type": "string"
}
},
"deployments": {
"type": "array",
"items": {
"type": "string"
}
},
"specModes": {
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"vendors",
"frameworkFamilies",
"deployments",
"specModes"
],
"additionalProperties": false
},
"reliabilityRanges": {
"type": "array",
"items": {
"type": "string"
}
},
"overview": {
"type": "object",
"additionalProperties": true
},
"calculator": {
"type": "object",
"additionalProperties": true
},
"fleet": {
"type": "object",
"additionalProperties": true
},
"defaults": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"view",
"apiVersion",
"params",
"models",
"sequences",
"precisions",
"hardware",
"frameworks",
"metrics",
"defaults"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "options",
"apiVersion": "v1",
"params": {
"format": "json"
},
"models": [
{
"name": "Kimi-K3",
"dbKeys": [
"kimik3"
],
"category": "deprecated",
"releaseDate": "2026-07-27",
"compareSlug": "kimi-k3",
"deprecatedSequences": [
"1k/1k",
"1k/8k"
]
},
{
"name": "DeepSeek-V4-Pro",
"dbKeys": [
"dsv4"
],
"category": "deprecated",
"releaseDate": "2026-06-30",
"compareSlug": "deepseek-v4-pro",
"deprecatedSequences": [
"1k/1k",
"1k/8k",
"8k/1k"
]
},
{
"name": "MiniMax-M3",
"dbKeys": [
"minimaxm3"
],
"category": "deprecated",
"releaseDate": "2026-06-07",
"compareSlug": "minimax-m3",
"deprecatedSequences": [
"1k/1k",
"1k/8k",
"8k/1k"
]
}
],
"sequences": [
{
"key": "8k/1k",
"label": "8k/1k",
"labelZh": "8k/1k",
"urlSegment": "8k-1k",
"isl": 8192,
"osl": 1024,
"kind": "fixed",
"deprecated": false
}
],
"precisions": [
"fp4",
"fp4fp8",
"fp8",
"bf16",
"int4"
],
"hardware": [
{
"key": "h200",
"label": "H200",
"vendor": "NVIDIA",
"arch": "Hopper",
"tdpW": 700,
"costPerHour": {
"h": 2.29,
"n": 2.99,
"r": 3.11
}
}
],
"frameworks": [
{
"key": "trt",
"label": "TRTLLM",
"family": "trt"
}
],
"specMethods": [
"mtp",
"none"
],
"percentiles": [
"p75",
"p90"
],
"xAxisModes": [
"interactivity",
"ttft",
"e2e",
"e2e-normalized-interactivity"
],
"fixedSequenceStatistics": [
"median",
"mean"
],
"scaleModes": [
"auto",
"linear",
"log"
],
"metrics": [
{
"key": "tokensPerDollarN",
"configKey": "y_tokensPerDollarN",
"label": "Tokens per Dollar (Neocloud)",
"labelZh": "每美元 Token 数（Neocloud）",
"unit": null,
"polarity": "higher",
"group": "Cost",
"source": "benchmark"
}
],
"quickFilters": {
"vendors": [
"NVIDIA",
"AMD"
],
"frameworkFamilies": [
"vllm",
"sglang",
"trt",
"atom"
],
"deployments": [
"single-node",
"multi-node",
"disagg"
],
"specModes": [
"mtp",
"stp"
]
},
"reliabilityRanges": [
"last-3-days",
"last-7-days",
"last-month",
"last-3-months",
"all-time"
],
"defaults": {
"model": "DeepSeek-V4-Pro",
"sequence": "8k/1k",
"metric": "y_tokensPerDollarH",
"percentile": "p90",
"xmode": "interactivity",
"xstat": "median",
"serviceCompare": false,
"roleShare": false,
"powerFit": false,
"powerBreakdown": false,
"headroom": false
}
}滚动查看完整示例
400
format 参数无效——该端点仅支持 JSON。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unsupported format: csv",
"param": "format",
"allowed": [
"json"
]
}500
选项数据组装失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/inference获取主推理图表视图
返回 /infx/inference 散点图所用的序列：按硬件配置分组，在所选指标、序列、精度与 x 轴模式下给出各并发档位的 x/y 数据点，并复用仪表板代码计算边界与 best-per-SKU 标记。筛选参数与仪表板快捷筛选一致（gpus、vendors、框架系列、部署模式、投机解码）。设置 optimal=true 可只保留边界点，best=true 可只保留每个 GPU SKU 的最优曲线。实测功耗使用较高功耗侧的外包络：frontier.direction 描述这一边界，metric.direction 则保留 best-per-SKU 选择所用的优化方向。整体实测功耗及能耗支持遥测已验证、拓扑受支持的 single_turn 和 AgentX 数据；AgentX 采用功耗模型的智能体工作负载状态，尚未针对该工作负载单独校准。整体实测指标另返回 tableRows，保留当前筛选范围和最优曲线选择内所有 GPU 遥测有效的观测点（没有可排名估算的 SKU 保留其全部观测点），不按 optimal 或坐标轴显示范围裁剪；估算不可用时 y 为 null，并给出原因。series 和 count 仍仅包含可绘制的数值点。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| chipBasis | query | string | 可选 | pool（默认）或 total，指定分离式配置的输入、输出吞吐量及其派生指标所除的芯片数：pool 下输入按预填充芯片、输出按解码芯片计；total 按部署中的芯片总数计，与聚合配置及 tput_per_gpu 口径一致。聚合配置在两种口径下结果相同。 | |
| topologies | query | string | 可选 | 逗号分隔的精确拓扑键，取自仪表板拓扑选择器或数据点的 topologyKey。按 GPU 分配和并行配置筛选，保留全部已测并发；省略时不限制拓扑。原样使用这些键，未知键不会匹配数据点。 | |
| allPoints | query | string | 可选 | 布尔值，默认为 false。包括被图表范围裁剪的数据点；optimal 和 best 仍独立生效。 | |
| best | query | boolean | 可选 | 默认为 true：仅在选定精度内，按当前坐标轴对每个 SKU 的候选框架进行排名。每个候选按其实测 Pareto 前沿所覆盖的面积（hypervolume，以原点为参考点）计分；越低越好的坐标轴先取倒数，因此 $/M tok 与 tok/$ 的排名一致；没有实测数据的区间不得分。未选中的精度不会挤掉当前可见的配置。best=false 返回所有选定曲线，包括同一 SKU 上的 vLLM 和 SGLang。在仪表板的模型选择器中选择另一模型时，手动选择的图例曲线会重置为目标模型的默认选择；API 请求无状态，如需保留所有选定曲线，应在每个模型的请求中显式传入 best=false。仪表板在同时显示两者前要求确认不挑起争论；只读 API 不保存交互确认状态。此排名不代表某个引擎在所有情况下都更好。 | true |
| date | query | string | 可选 | 截止日期（YYYY-MM-DD）：每个配置取该日期当天或之前的最新运行。 | 2026-08-01 |
| dates | query | string | 可选 | 最多十二个比较项，以逗号分隔，格式为 YYYY-MM-DD 或 YYYY-MM-DD~rRUN_ID。仅含日期时读取当天的逻辑快照，不按截止日期向前回溯；含运行 ID 时读取该次运行的逻辑快照。各快照独立计算。 | |
| deployment | query | string | 可选 | 逗号分隔的部署模式；agg 会展开为 single-node 与 multi-node。 | disagg |
| end | query | string | 可选 | 比较范围结束日期 YYYY-MM-DD，不得早于 start。历史视图将其用作含终点的数据范围。 | |
| format | query | enum | 可选 | 响应编码。csv 返回图表数据点；整体实测指标则导出 tableRows，不可用的 y 留空。serviceCompare、roleShare、powerFit、powerBreakdown 和 headroom 分析结果仅支持 JSON，与 CSV 同用时返回 400。 | csv |
| frameworks | query | string | 可选 | 逗号分隔的推理框架系列。 | vllm,sglang |
| gpus | query | string | 可选 | 逗号分隔的硬件键或 GPU 名称（如 h200 或完整 hwKey）。省略时，Qwen3.5 排除 MI300X、MI325X、H100 和 H200。显式列表覆盖此默认值；gpus= 包含所有硬件。 | h200,mi355x |
| metric | query | string | 可选 | Y 轴指标键或配置键，如 tpPerGpu 或 y_tokensPerDollarN（见 options → metrics）。 | y_tpPerGpu |
| model | query | string | 必填 | 前端模型显示名（见 /api/v1/views/options → models）。 | DeepSeek-V4-Pro |
| optimal | query | boolean | 可选 | 仅返回每个硬件在各精度、各快照日期上的边界点。实测功耗指标与图表一致，保留较高功耗侧的外包络；其他指标保留各自的帕累托前沿。功耗包络不构成能效推荐。 | true |
| percentile | query | enum | 可选 | agentic-traces 场景下 x 轴延迟字段使用的百分位。 | p75 |
| power | query | string | 可选 | 以逗号分隔的 certified、legacy 功率数据等级。省略时选择全部等级。 | |
| precisions | query | string | 可选 | 逗号分隔的精度列表。省略时按数据最密的精度自动选择，与仪表盘默认行为一致。 | fp8,bf16 |
| priceSource | query | string | 可选 | 利润视图支持 list、openrouter、custom；推理和历史视图支持 normalized、openrouter。响应 pricing 记录实际使用的价格。 | |
| runId | query | integer | 可选 | GitHub Actions 运行 id。返回该次运行的精确快照而非最新数据。 | 12345678 |
| sequence | query | enum | 可选 | 基准序列。也接受 8k-1k、agentic 等别名。 | agentic-traces |
| spec | query | string | 可选 | 逗号分隔的投机解码模式。 | mtp |
| start | query | string | 可选 | 比较范围起始日期 YYYY-MM-DD。与 end 一起只加入两个端点，不加入中间所有日期。历史视图将其用作含起点的数据范围。 | |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 | |
| unofficialrun | query | string | 可选 | 最多八个公开 CI 运行 ID，以逗号分隔。去除各项首尾空白和重复 ID 后，按剩余顺序分配运行索引。叠加结果与官方数据分开。 | |
| userCosts | query | string | 可选 | JSON 对象，将硬件键映射为有限非负美元/芯片小时，用于自定义成本指标。 | |
| userPowers | query | string | 可选 | JSON 对象，将硬件键映射为有限非负千瓦/芯片功率假设。 | |
| vendors | query | string | 可选 | 逗号分隔的 GPU 厂商。 | NVIDIA |
| xmetric | query | enum | 可选 | 当 x 轴为首 Token 时间（TTFT）时使用的百分位。 | p99_ttft |
| xmode | query | enum | 可选 | X 轴模式。Qwen3.5 默认为 e2e，其他模型默认为 interactivity；显式值覆盖模型默认值。e2e-normalized-interactivity 使用已持久化的 AgentX 派生指标，unofficialrun 叠加数据则使用该 run 自身报告的数值；没有合格数值的数据点不参与此视图。 | e2e |
| xstat | query | enum | 可选 | 固定长度工作负载服务轴的统计量：median（默认）或 mean。Mean streaming speed 为 1 / mean TPOT，不是各请求速度的算术平均；mean TTFT 和 E2E 使用各自记录的均值。缺失时不回退到 median。AgentX 不使用此参数，params.xstat 为 null，xAxis.statistic 返回实际分位数。 | mean |
| serviceCompare | query | boolean | 可选 | 仅通过 API 提供的分析：返回来源选项、同等服务条件下的百分比对比曲线、可选目标值对比，以及相同并发下的诊断表（matchedConcurrency）。使用筛选范围内、尚未经前沿和 best 筛选的实测点。仪表板不提供对应控件或面板。仅支持 JSON。 | true |
| serviceBaseline | query | string | 可选 | serviceSources 中的完整基准来源键。省略时按确定性顺序选择首项；显式未知键保持不可用。变化百分比为 100 ×（对比值 / 基准值 − 1）。 | serviceSources 中返回的完整键 |
| serviceComparator | query | string | 可选 | serviceSources 中的完整对比来源键。省略时按确定性顺序选择第二项。来源键保留硬件、曲线快照、测试配置指纹、镜像、拓扑和工作负载标识。同一快照内不同 telemetry producer/exporter 的 hash 不拆分来源；没有快照的行仍按原运行、日期和 hash 区分。不能用硬件名称代替。 | serviceSources 中返回的完整键 |
| serviceTarget | query | number | 可选 | 有限正数服务轴目标：streaming speed 单位为 tok/s/user，TTFT/E2E 单位为秒。省略时返回曲线，目标值对比为 null。仅在各完整来源的实测范围内做数值线性插值，不外推、不跨越缺失指标端点。并发轴不适用。 | 40 |
| roleShare | query | boolean | 可选 | 返回通过验证的分离式 prefill/decode 能耗占比，统一使用 output token 分母；rolePoints 另含各池的 W/GPU 和按本池 token 计的能耗。用同窗口总 J/output ÷ J/input 将 prefill J/input 转换为 J/output；占比分母为重建的 prefill + decode 能耗，不使用各池独立的 token 数。缺失或无效的数据会被省略或返回 null，仅支持 JSON。 | true |
| powerFit | query | boolean | 可选 | 按数据源分别返回普通最小二乘拟合：实测平均 W/GPU = P0 + m × 每个已分配 GPU 的输出 tok/s（分离式部署的输出量均摊到 prefill 和 decode GPU）。P0 是零输出截距，不是实测空载功耗；m 是每个输出 token 的边际能耗（J）。不同输出速率少于 3 个的数据源返回观测点，fit 为 null。仅支持 JSON。 | true |
| powerBreakdown | query | boolean | 可选 | 返回 powerBreakdowns：按硬件曲线（hwKey + 精度，与图例一致）将每个点的 All in Measured W/GPU 拆分为实测 GPU、实测 Grace CPU（NVL72 且插槽遥测有效时），以及建模的主机 CPU、内存、网卡、NVSwitch、存储、PCIe 及其他部件、风扇、电源损耗、Scale-out 网络分摊和 PUE 机房开销。各部分为部署分摊值 ÷ 已分配 GPU 数，总和等于 totalWattsPerGpu。缺少有效 GPU 遥测或系统模型不支持的点会被省略。仅支持 JSON。 | true |
| headroom | query | boolean | 可选 | 返回 headroom：PowerX 的实测功耗 vs TDP 余量表。按硬件曲线（hwKey + 精度）给出每芯片实测平均与 P90 功耗、P90 占注册表 TDP 的百分比以及余量 = 100 − 该百分比，在 interactivityField 上的 25 / 50 / 100 / 300 tok/s/user 处用图表自身的插值（吞吐 Pareto 前沿、单调 Steffen 样条）读取。超出曲线实测范围的单元为 null，不做外推；缺少 P90 功耗、吞吐、注册表 TDP 或不足两个点的曲线会被省略。仅 GPU 板卡功耗。仅支持 JSON。 | true |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/inference?model=Kimi-K3&metric=y_tokensPerDollarN&sequence=agentic-traces'响应
InferenceView200
按硬件配置分组的图表就绪序列，每个数据点带帕累托前沿与 best-per-SKU 标记。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"inference"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"metric": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"configKey": {
"type": "string"
},
"label": {
"type": "string"
},
"labelZh": {
"type": "string"
},
"unit": {
"type": [
"string",
"null"
]
},
"polarity": {
"type": [
"string",
"null"
]
},
"direction": {
"type": [
"string",
"null"
],
"description": "Configured optimization direction, also used by best-per-SKU selection."
}
},
"required": [
"key",
"configKey",
"label",
"labelZh"
],
"additionalProperties": false
},
"xAxis": {
"type": "object",
"properties": {
"mode": {
"type": "string"
},
"field": {
"type": "string"
},
"label": {
"type": "string"
},
"statistic": {
"type": "string"
}
},
"required": [
"mode",
"field",
"label",
"statistic"
],
"additionalProperties": false
},
"frontier": {
"type": "object",
"properties": {
"direction": {
"type": [
"string",
"null"
],
"description": "Selected boundary direction. Measured-power gauges use upper_right for interactivity or upper_left for latency, independently of metric.direction."
},
"points": {
"type": "integer"
}
},
"required": [
"direction",
"points"
],
"additionalProperties": false
},
"hardware": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"vendor": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
}
},
"series": {
"type": "array",
"items": {
"type": "object",
"properties": {
"hwKey": {
"type": "string"
},
"gpu": {
"type": "string"
},
"framework": {
"type": "string"
},
"specMethod": {
"type": "string"
},
"label": {
"type": "string",
"description": "Chart display label. If any point in a plain SGLang/vLLM series has router_name=smg, the entire series is labeled SMG SGLang/SMG vLLM. Canonical framework and hwKey are unchanged."
},
"vendor": {
"type": "string"
},
"deployment": {
"type": "string"
},
"kvOffload": {
"type": "boolean"
},
"bestPerSku": {
"type": "boolean"
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"precision": {
"type": "string"
},
"x": {
"type": "number"
},
"y": {
"type": "number"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"tp": {
"type": "number"
},
"date": {
"type": "string",
"format": "date"
},
"runId": {
"type": "integer"
},
"frontier": {
"type": "boolean"
},
"bestPerSku": {
"type": "boolean"
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
}
}
},
"required": [
"x",
"y",
"concurrency",
"topologyKey",
"tp",
"date",
"frontier",
"bestPerSku",
"metrics"
],
"additionalProperties": false
}
}
},
"required": [
"hwKey",
"gpu",
"framework",
"specMethod",
"label",
"deployment",
"kvOffload",
"bestPerSku",
"points"
],
"additionalProperties": false
}
},
"tableRows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"precision": {
"type": "string"
},
"x": {
"type": [
"number",
"null"
],
"description": "Selected service axis; null when unavailable."
},
"y": {
"type": [
"number",
"null"
],
"description": "Selected all-in metric; null when unavailable."
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"tp": {
"type": "number"
},
"date": {
"type": "string",
"format": "date"
},
"runId": {
"type": "integer"
},
"frontier": {
"type": "boolean"
},
"bestPerSku": {
"type": "boolean"
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
}
},
"hwKey": {
"type": "string"
},
"gpu": {
"type": "string"
},
"framework": {
"type": "string"
},
"specMethod": {
"type": "string"
},
"label": {
"type": "string",
"description": "Point-specific chart display label, including SMG SGLang/SMG vLLM when this point has router_name=smg. Canonical framework and hwKey are unchanged."
},
"vendor": {
"type": "string"
},
"deployment": {
"type": "string"
},
"kvOffload": {
"type": "boolean"
},
"measuredGpuWatts": {
"type": "number",
"description": "Validated measured mean W/GPU."
},
"status": {
"type": "string",
"enum": [
"available",
"unavailable"
]
},
"unavailableReason": {
"oneOf": [
{
"type": "string",
"enum": [
"workload",
"hardware",
"telemetry",
"gpu-count",
"topology",
"role-power",
"model-domain",
"energy",
"model-unavailable"
]
},
{
"type": "null"
}
]
}
},
"required": [
"id",
"precision",
"hwKey",
"gpu",
"framework",
"specMethod",
"label",
"deployment",
"kvOffload",
"x",
"y",
"concurrency",
"topologyKey",
"tp",
"date",
"frontier",
"bestPerSku",
"metrics",
"measuredGpuWatts",
"status",
"unavailableReason"
],
"additionalProperties": false
},
"description": "All in Measured only: GPU-valid observations in the selected scope, including unavailable estimates. Honors hardware/best selection; under best=true, SKUs with no rankable estimate keep their rows. Not clipped by optimal or chart axes. Also returned per comparison and overlay. count remains the numeric series point count."
},
"count": {
"type": "integer"
},
"serviceSources": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
}
},
"equalServiceComparison": {
"type": [
"object",
"null"
],
"properties": {
"target": {
"type": "number"
},
"xField": {
"type": "string"
},
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
},
"reason": {
"type": "string"
},
"metrics": {
"type": "object",
"properties": {
"meanWattsPerGpu": {
"type": "object",
"properties": {
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"changePercent": {
"type": [
"number",
"null"
]
},
"reason": {
"type": "string"
}
},
"required": [
"baseline",
"comparator",
"changePercent",
"reason"
],
"additionalProperties": false,
"description": "Mean measured GPU board W/GPU."
},
"outputTokensPerSecond": {
"type": "object",
"properties": {
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"changePercent": {
"type": [
"number",
"null"
]
},
"reason": {
"type": "string"
}
},
"required": [
"baseline",
"comparator",
"changePercent",
"reason"
],
"additionalProperties": false,
"description": "Whole-deployment output tokens/s; disaggregated GPU-count normalization preserves the workload denominator."
},
"joulesPerOutputToken": {
"type": "object",
"properties": {
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"changePercent": {
"type": [
"number",
"null"
]
},
"reason": {
"type": "string"
}
},
"required": [
"baseline",
"comparator",
"changePercent",
"reason"
],
"additionalProperties": false,
"description": "Validated measured GPU joules per output token."
}
},
"required": [
"meanWattsPerGpu",
"outputTokensPerSecond",
"joulesPerOutputToken"
],
"additionalProperties": false
}
},
"required": [
"target",
"xField",
"baseline",
"comparator",
"reason",
"metrics"
],
"additionalProperties": false
},
"equalServiceCurve": {
"type": "array",
"items": {
"type": "object",
"properties": {
"target": {
"type": "number"
},
"xField": {
"type": "string"
},
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
},
"reason": {
"type": "string"
},
"metrics": {
"type": "object",
"properties": {
"meanWattsPerGpu": {
"type": "object",
"properties": {
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"changePercent": {
"type": [
"number",
"null"
]
},
"reason": {
"type": "string"
}
},
"required": [
"baseline",
"comparator",
"changePercent",
"reason"
],
"additionalProperties": false,
"description": "Mean measured GPU board W/GPU."
},
"outputTokensPerSecond": {
"type": "object",
"properties": {
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"changePercent": {
"type": [
"number",
"null"
]
},
"reason": {
"type": "string"
}
},
"required": [
"baseline",
"comparator",
"changePercent",
"reason"
],
"additionalProperties": false,
"description": "Whole-deployment output tokens/s; disaggregated GPU-count normalization preserves the workload denominator."
},
"joulesPerOutputToken": {
"type": "object",
"properties": {
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"value": {
"type": "number"
},
"interpolated": {
"type": "boolean"
},
"endpoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"value": {
"type": "number"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"value",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"value",
"interpolated",
"endpoints"
],
"additionalProperties": false
},
"changePercent": {
"type": [
"number",
"null"
]
},
"reason": {
"type": "string"
}
},
"required": [
"baseline",
"comparator",
"changePercent",
"reason"
],
"additionalProperties": false,
"description": "Validated measured GPU joules per output token."
}
},
"required": [
"meanWattsPerGpu",
"outputTokensPerSecond",
"joulesPerOutputToken"
],
"additionalProperties": false
}
},
"required": [
"target",
"xField",
"baseline",
"comparator",
"reason",
"metrics"
],
"additionalProperties": false
}
},
"roleEnergyShares": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"sourceKey": {
"type": "string"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
},
"prefill": {
"type": "number",
"description": "Prefill energy in J/output token."
},
"decode": {
"type": "number",
"description": "Decode energy in J/output token."
},
"total": {
"type": "number",
"description": "Reconstructed prefill + decode energy in J/output token."
},
"prefillShare": {
"type": "number",
"description": "Prefill percentage of reconstructed total."
},
"decodeShare": {
"type": "number",
"description": "Decode percentage of reconstructed total."
}
},
"required": [
"x",
"sourceKey",
"point",
"prefill",
"decode",
"total",
"prefillShare",
"decodeShare"
],
"additionalProperties": false
}
},
"matchedConcurrency": {
"type": "object",
"properties": {
"baseline": {
"type": [
"object",
"null"
],
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
},
"comparator": {
"type": [
"object",
"null"
],
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
},
"interactivityField": {
"type": "string"
},
"reason": {
"type": "string",
"enum": [
"same-source",
"unknown-source"
]
},
"rows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"concurrency": {
"type": "integer"
},
"baseline": {
"type": "object",
"properties": {
"status": {
"type": "string",
"enum": [
"observed",
"missing",
"ambiguous"
],
"description": "observed: one agreeing observation; missing: none at this concurrency; ambiguous: conflicting observations, none chosen."
},
"values": {
"type": "object",
"properties": {
"joulesPerOutputToken": {
"type": [
"number",
"null"
]
},
"meanWattsPerGpu": {
"type": [
"number",
"null"
]
},
"interactivity": {
"type": [
"number",
"null"
]
}
},
"required": [
"joulesPerOutputToken",
"meanWattsPerGpu",
"interactivity"
],
"additionalProperties": false
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
}
},
"required": [
"status"
],
"additionalProperties": false
},
"comparator": {
"type": "object",
"properties": {
"status": {
"type": "string",
"enum": [
"observed",
"missing",
"ambiguous"
],
"description": "observed: one agreeing observation; missing: none at this concurrency; ambiguous: conflicting observations, none chosen."
},
"values": {
"type": "object",
"properties": {
"joulesPerOutputToken": {
"type": [
"number",
"null"
]
},
"meanWattsPerGpu": {
"type": [
"number",
"null"
]
},
"interactivity": {
"type": [
"number",
"null"
]
}
},
"required": [
"joulesPerOutputToken",
"meanWattsPerGpu",
"interactivity"
],
"additionalProperties": false
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
}
},
"required": [
"status"
],
"additionalProperties": false
},
"changePercent": {
"type": "object",
"properties": {
"joulesPerOutputToken": {
"type": [
"number",
"null"
]
},
"meanWattsPerGpu": {
"type": [
"number",
"null"
]
},
"interactivity": {
"type": [
"number",
"null"
]
}
},
"required": [
"joulesPerOutputToken",
"meanWattsPerGpu",
"interactivity"
],
"additionalProperties": false,
"description": "100 × (comparator / baseline − 1); null unless both values were observed."
}
},
"required": [
"concurrency",
"baseline",
"comparator",
"changePercent"
],
"additionalProperties": false
}
}
},
"required": [
"baseline",
"comparator",
"interactivityField",
"reason",
"rows"
],
"additionalProperties": false
},
"rolePoints": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"sourceKey": {
"type": "string"
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
},
"prefillWattsPerGpu": {
"type": [
"number",
"null"
],
"description": "Mean W/GPU inside the prefill pool."
},
"decodeWattsPerGpu": {
"type": [
"number",
"null"
],
"description": "Mean W/GPU inside the decode pool."
},
"prefillJoulesPerInputToken": {
"type": [
"number",
"null"
],
"description": "Prefill-pool joules per input token."
},
"decodeJoulesPerOutputToken": {
"type": [
"number",
"null"
],
"description": "Decode-pool joules per output token."
},
"energy": {
"type": [
"object",
"null"
],
"properties": {
"prefill": {
"type": "number"
},
"decode": {
"type": "number"
},
"total": {
"type": "number"
},
"prefillShare": {
"type": "number"
}
},
"required": [
"prefill",
"decode",
"total",
"prefillShare"
],
"additionalProperties": false,
"description": "Both pools on the output-token denominator (J/output token, share %)."
}
},
"required": [
"x",
"sourceKey",
"point",
"prefillWattsPerGpu",
"decodeWattsPerGpu",
"prefillJoulesPerInputToken",
"decodeJoulesPerOutputToken",
"energy"
],
"additionalProperties": false
}
},
"powerFits": {
"type": "array",
"items": {
"type": "object",
"properties": {
"source": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"key",
"label"
],
"additionalProperties": false
},
"tdpWatts": {
"type": [
"number",
"null"
],
"description": "Rated board TDP per GPU; P0 ÷ TDP = fit.intercept / tdpWatts."
},
"fit": {
"type": [
"object",
"null"
],
"properties": {
"intercept": {
"type": "number",
"description": "P0, W/GPU at zero output (extrapolated)."
},
"slope": {
"type": "number",
"description": "m, J per output token."
},
"rSquared": {
"type": [
"number",
"null"
]
},
"n": {
"type": "integer"
},
"xMin": {
"type": "number"
},
"xMax": {
"type": "number"
}
},
"required": [
"intercept",
"slope",
"rSquared",
"n",
"xMin",
"xMax"
],
"additionalProperties": false
},
"reason": {
"type": "string",
"enum": [
"too-few-points"
]
},
"observations": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number",
"description": "Output tok/s per allocated GPU."
},
"y": {
"type": "number",
"description": "Measured mean W/GPU."
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"x",
"y",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"source",
"tdpWatts",
"fit",
"reason",
"observations"
],
"additionalProperties": false
}
},
"powerBreakdowns": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string",
"description": "hwKey|precision."
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"label": {
"type": "string",
"description": "Chart legend label, e.g. B200 (vLLM) FP4."
},
"categories": {
"type": "array",
"items": {
"type": "string",
"enum": [
"gpu",
"measuredCpu",
"cpu",
"memory",
"nic",
"nvswitch",
"storage",
"other",
"fans",
"psuLoss",
"network",
"facility"
],
"description": "Parts with watts on at least one point, in stack order. gpu and measuredCpu are measured; the rest are modeled."
}
},
"tdpWattsPerGpu": {
"type": [
"number",
"null"
],
"description": "Rated board TDP per GPU (reference, not a measurement)."
},
"allInProvisionedWattsPerGpu": {
"type": [
"number",
"null"
],
"description": "Registry all-in provisioned W/GPU, PUE included (reference, not a measurement)."
},
"modelRevision": {
"type": "string",
"description": "Pinned power_model commit."
},
"interactivityField": {
"type": "string",
"description": "Interactivity field the points are placed on: the chart x-axis when it is an interactivity axis, otherwise median_intvty."
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"concurrency": {
"type": "integer"
},
"interactivity": {
"type": "number",
"description": "tok/s/user on interactivityField; points are sorted by it."
},
"configLabel": {
"type": "string",
"description": "Date, topology and recipe of the point."
},
"wattsPerGpu": {
"type": "object",
"properties": {
"gpu": {
"type": "number"
},
"measuredCpu": {
"type": "number"
},
"cpu": {
"type": "number"
},
"memory": {
"type": "number"
},
"nic": {
"type": "number"
},
"nvswitch": {
"type": "number"
},
"storage": {
"type": "number"
},
"other": {
"type": "number"
},
"fans": {
"type": "number"
},
"psuLoss": {
"type": "number"
},
"network": {
"type": "number"
},
"facility": {
"type": "number"
}
},
"required": [
"gpu",
"measuredCpu",
"cpu",
"memory",
"nic",
"nvswitch",
"storage",
"other",
"fans",
"psuLoss",
"network",
"facility"
],
"additionalProperties": false
},
"totalWattsPerGpu": {
"type": "number",
"description": "All in Measured W/GPU."
},
"pue": {
"type": "number"
},
"cpuBasis": {
"type": [
"string",
"null"
],
"description": "`measured` or `modeled` Grace sockets on NVL72 trays; null otherwise."
},
"chassisBasis": {
"type": "string",
"enum": [
"full",
"extrapolated"
]
},
"point": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"sourceKey": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"date": {
"type": "string"
},
"runUrl": {
"type": [
"string",
"null"
]
},
"recipeFingerprint": {
"type": [
"string",
"null"
]
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"sourceKey",
"hwKey",
"precision",
"concurrency",
"topologyKey",
"date",
"runUrl",
"recipeFingerprint",
"image"
],
"additionalProperties": false
}
},
"required": [
"concurrency",
"interactivity",
"configLabel",
"wattsPerGpu",
"totalWattsPerGpu",
"pue",
"cpuBasis",
"chassisBasis",
"point"
],
"additionalProperties": false
}
}
},
"required": [
"key",
"hwKey",
"precision",
"label",
"categories",
"tdpWattsPerGpu",
"allInProvisionedWattsPerGpu",
"modelRevision",
"interactivityField",
"points"
],
"additionalProperties": false
}
},
"headroom": {
"type": "object",
"properties": {
"interactivities": {
"type": "array",
"items": {
"type": "number"
}
},
"interactivityField": {
"type": "string",
"description": "Interactivity field the cells are read on: the chart x-axis when it is an interactivity axis, otherwise median_intvty."
},
"rows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string",
"description": "hwKey|precision."
},
"hwKey": {
"type": "string"
},
"precision": {
"type": "string"
},
"label": {
"type": "string",
"description": "Chart legend label, e.g. B200 (vLLM) FP4."
},
"tdpWatts": {
"type": "number",
"description": "Rated board TDP per GPU from the registry."
},
"interactivityMin": {
"type": "number"
},
"interactivityMax": {
"type": "number"
},
"cells": {
"type": "array",
"items": {
"oneOf": [
{
"type": "object",
"properties": {
"interactivity": {
"type": "number"
},
"avgWatts": {
"type": "number",
"description": "Interpolated measured average W per chip."
},
"p90Watts": {
"type": "number",
"description": "Interpolated measured P90 fleet W per chip."
},
"percentTdpP90": {
"type": "number",
"description": "P90 ÷ TDP × 100."
},
"headroomPercent": {
"type": "number",
"description": "100 − percentTdpP90."
}
},
"required": [
"interactivity",
"avgWatts",
"p90Watts",
"percentTdpP90",
"headroomPercent"
],
"additionalProperties": false
},
{
"type": "null"
}
]
}
}
},
"required": [
"key",
"hwKey",
"precision",
"label",
"tdpWatts",
"interactivityMin",
"interactivityMax",
"cells"
],
"additionalProperties": false
}
}
},
"required": [
"interactivities",
"interactivityField",
"rows"
],
"additionalProperties": false
},
"pricing": {
"type": [
"object",
"null"
],
"additionalProperties": true
},
"comparisons": {
"type": "array",
"items": {
"type": "object",
"properties": {
"tableRows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"precision": {
"type": "string"
},
"x": {
"type": [
"number",
"null"
],
"description": "Selected service axis; null when unavailable."
},
"y": {
"type": [
"number",
"null"
],
"description": "Selected all-in metric; null when unavailable."
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"tp": {
"type": "number"
},
"date": {
"type": "string",
"format": "date"
},
"runId": {
"type": "integer"
},
"frontier": {
"type": "boolean"
},
"bestPerSku": {
"type": "boolean"
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
}
},
"hwKey": {
"type": "string"
},
"gpu": {
"type": "string"
},
"framework": {
"type": "string"
},
"specMethod": {
"type": "string"
},
"label": {
"type": "string",
"description": "Point-specific chart display label, including SMG SGLang/SMG vLLM when this point has router_name=smg. Canonical framework and hwKey are unchanged."
},
"vendor": {
"type": "string"
},
"deployment": {
"type": "string"
},
"kvOffload": {
"type": "boolean"
},
"measuredGpuWatts": {
"type": "number",
"description": "Validated measured mean W/GPU."
},
"status": {
"type": "string",
"enum": [
"available",
"unavailable"
]
},
"unavailableReason": {
"oneOf": [
{
"type": "string",
"enum": [
"workload",
"hardware",
"telemetry",
"gpu-count",
"topology",
"role-power",
"model-domain",
"energy",
"model-unavailable"
]
},
{
"type": "null"
}
]
}
},
"required": [
"id",
"precision",
"hwKey",
"gpu",
"framework",
"specMethod",
"label",
"deployment",
"kvOffload",
"x",
"y",
"concurrency",
"topologyKey",
"tp",
"date",
"frontier",
"bestPerSku",
"metrics",
"measuredGpuWatts",
"status",
"unavailableReason"
],
"additionalProperties": false
}
}
},
"additionalProperties": true
}
},
"overlays": {
"type": "array",
"items": {
"type": "object",
"properties": {
"tableRows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": [
"integer",
"null"
]
},
"precision": {
"type": "string"
},
"x": {
"type": [
"number",
"null"
],
"description": "Selected service axis; null when unavailable."
},
"y": {
"type": [
"number",
"null"
],
"description": "Selected all-in metric; null when unavailable."
},
"concurrency": {
"type": "number"
},
"topologyKey": {
"type": "string"
},
"tp": {
"type": "number"
},
"date": {
"type": "string",
"format": "date"
},
"runId": {
"type": "integer"
},
"frontier": {
"type": "boolean"
},
"bestPerSku": {
"type": "boolean"
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
}
},
"hwKey": {
"type": "string"
},
"gpu": {
"type": "string"
},
"framework": {
"type": "string"
},
"specMethod": {
"type": "string"
},
"label": {
"type": "string",
"description": "Point-specific chart display label, including SMG SGLang/SMG vLLM when this point has router_name=smg. Canonical framework and hwKey are unchanged."
},
"vendor": {
"type": "string"
},
"deployment": {
"type": "string"
},
"kvOffload": {
"type": "boolean"
},
"measuredGpuWatts": {
"type": "number",
"description": "Validated measured mean W/GPU."
},
"status": {
"type": "string",
"enum": [
"available",
"unavailable"
]
},
"unavailableReason": {
"oneOf": [
{
"type": "string",
"enum": [
"workload",
"hardware",
"telemetry",
"gpu-count",
"topology",
"role-power",
"model-domain",
"energy",
"model-unavailable"
]
},
{
"type": "null"
}
]
}
},
"required": [
"id",
"precision",
"hwKey",
"gpu",
"framework",
"specMethod",
"label",
"deployment",
"kvOffload",
"x",
"y",
"concurrency",
"topologyKey",
"tp",
"date",
"frontier",
"bestPerSku",
"metrics",
"measuredGpuWatts",
"status",
"unavailableReason"
],
"additionalProperties": false
}
}
},
"additionalProperties": true
}
}
},
"required": [
"view",
"apiVersion",
"params",
"metric",
"xAxis",
"frontier",
"series",
"count"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "inference",
"apiVersion": "v1",
"params": {
"model": "DeepSeek-V4-Pro",
"sequence": "8k/1k",
"precisions": [
"fp8"
],
"metric": "y_tpPerGpu",
"xmode": "interactivity",
"xstat": "median",
"xmetric": "p90_ttft",
"percentile": "p90",
"date": null,
"runId": null,
"gpus": [],
"vendors": [],
"frameworks": [],
"deployment": [],
"spec": [],
"topologies": [],
"optimal": true,
"best": true,
"format": "json",
"serviceCompare": false,
"serviceBaseline": null,
"serviceComparator": null,
"serviceTarget": null,
"roleShare": false,
"powerFit": false,
"powerBreakdown": false,
"headroom": false
},
"metric": {
"key": "tpPerGpu",
"configKey": "y_tpPerGpu",
"label": "Output Throughput per GPU (tok/s/gpu)",
"labelZh": "单 GPU 输出吞吐（tok/s/gpu）",
"unit": "tok/s/gpu",
"polarity": "higher",
"direction": "upper_left"
},
"xAxis": {
"mode": "interactivity",
"field": "median_intvty",
"statistic": "median",
"label": "Median Interactivity (tok/s/user)"
},
"frontier": {
"direction": "upper_left",
"points": 14
},
"hardware": [
{
"key": "h200_trt",
"label": "H200 (TRTLLM)",
"vendor": "NVIDIA"
}
],
"series": [
{
"hwKey": "h200_trt",
"gpu": "h200",
"framework": "trt",
"specMethod": "none",
"label": "H200 (TRTLLM)",
"vendor": "NVIDIA",
"deployment": "single-node",
"kvOffload": false,
"bestPerSku": true,
"points": [
{
"x": 12.5,
"y": 450.5,
"concurrency": 64,
"topologyKey": "Single|GPU=8|DP=?|TP=8|EP=1|PP=?|DCP=?|PCP=?|DPA=0|offload=off",
"tp": 8,
"date": "2026-08-20",
"runId": 12345678,
"frontier": true,
"bestPerSku": true,
"metrics": {
"tpPerGpu": 450.5,
"outputTputPerGpu": 400.2,
"inputTputPerGpu": 50.3
}
}
]
}
],
"count": 1
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
hwKey,gpu,framework,specMethod,label,vendor,deployment,kvOffload,x,y,concurrency,topologyKey,tp,date,runId,frontier,bestPerSku,metric_tpPerGpu
h200_trt,h200,trt,none,H200 (TRTLLM),NVIDIA,single-node,false,12.5,450.5,64,Single|GPU=8|DP=?|TP=8|EP=1|PP=?|DCP=?|PCP=?|DPA=0|offload=off,8,2026-08-20,12345678,true,true,450.5400
参数无效。响应体会给出参数名，枚举参数还会列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown xmode: bogus",
"param": "xmode",
"allowed": [
"interactivity",
"ttft",
"e2e",
"e2e-normalized-interactivity"
]
}500
基准数据查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/historical获取历史趋势视图
返回服务端计算的 Historical Trends 仪表盘趋势线：在每个基准快照日期上，对每个硬件配置（选择多个精度时按精度分线）在目标交互速率处对所选指标插值，再组装为按日期排序的趋势线。插值与仪表盘使用相同的单调样条数学。曲线补齐至 extendToDate（默认当前 UTC 日期），补齐点标为 synthetic: true。需要可复现的快照时请显式指定 extendToDate。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| chipBasis | query | string | 可选 | pool（默认）或 total，指定分离式配置的输入、输出吞吐量及其派生指标所除的芯片数：pool 下输入按预填充芯片、输出按解码芯片计；total 按部署中的芯片总数计，与聚合配置及 tput_per_gpu 口径一致。聚合配置在两种口径下结果相同。 | |
| chart | query | enum | 可选 | line 返回原始趋势线。waterfall 还返回基线、刷新最佳纪录的变化及最佳观测值，并附源运行记录及严格匹配的 PR 链接。不包含补齐点及性能恢复。输出 token 指标会排除缺少输出吞吐量实测记录的曲线，excludedPointCount 报告省略的日期数。越低越好的指标使用下降变化量。链接标识源曲线，不代表单个变更的独立效果。 | waterfall |
| deployment | query | string | 可选 | 逗号分隔的部署模式；agg 会展开为 single-node 与 multi-node。 | single-node |
| end | query | string | 可选 | 包含的最晚快照日期（YYYY-MM-DD）。 | 2026-08-01 |
| extendToDate | query | string | 可选 | 历史曲线补齐至此日期，格式为 YYYY-MM-DD；默认当前 UTC 日期，与仪表板一致。 | |
| format | query | enum | 可选 | 响应编码。csv 为每个趋势点或瀑布图步骤返回一行，并包含 targetInteractivity、interactivityStatistic 和 interactivityField。瀑布图的 runUrls 和 prLinks 列为 JSON 编码的数组。 | csv |
| frameworks | query | string | 可选 | 逗号分隔的推理框架系列。 | vllm,trt |
| gpus | query | string | 可选 | 逗号分隔的硬件键或 GPU 名称。 | h200,mi355x |
| metric | query | string | 可选 | 趋势指标键或配置键，如 tpPerGpu 或 y_tokensPerDollarN（见 options → metrics）。 | y_tpPerGpu |
| model | query | string | 必填 | 前端模型显示名（见 /api/v1/views/options → models）。 | DeepSeek-V4-Pro |
| precisions | query | string | 可选 | 逗号分隔的精度列表。省略时按数据最密的精度自动选择。选择多个精度时，每个硬件会按精度分线（键为 hwKey__precision）。 | fp8,bf16 |
| priceSource | query | string | 可选 | 利润视图支持 list、openrouter、custom；推理和历史视图支持 normalized、openrouter。响应 pricing 记录实际使用的价格。 | |
| sequence | query | enum | 可选 | 基准序列。也接受 8k-1k、agentic 等别名。 | agentic-traces |
| start | query | string | 可选 | 包含的最早快照日期（YYYY-MM-DD）。 | 2026-01-01 |
| target | query | number | 可选 | 每个快照日期上的指标按该目标交互性（tok/s/user）插值。AgentX 使用 P90 全响应 ITL（秒）的倒数；仅当全响应 ITL 的均值和分位数字段均缺失时才使用旧版 ITL。P90 延迟缺失或无效的数据点不参与计算，不会改用中位数，也不会直接使用 token 速率的 P90。单轮工作负载仍使用中位交互性。 | 50 |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 | |
| vendors | query | string | 可选 | 逗号分隔的 GPU 厂商。 | AMD |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/historical?model=DeepSeek-V4-Pro&metric=y_tpPerGpu&target=35'响应
HistoricalView200
按硬件配置给出的按日期排序趋势线，在目标交互性处插值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"historical"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"metric": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"configKey": {
"type": "string"
},
"label": {
"type": "string"
},
"labelZh": {
"type": "string"
}
},
"required": [
"key",
"configKey",
"label",
"labelZh"
],
"additionalProperties": false
},
"target": {
"type": "number"
},
"interactivity": {
"type": "object",
"properties": {
"statistic": {
"type": "string",
"enum": [
"median",
"p90"
]
},
"field": {
"type": "string",
"enum": [
"median_intvty",
"p90_intvty"
]
},
"unit": {
"type": "string",
"enum": [
"tok/s/user"
]
}
},
"required": [
"statistic",
"field",
"unit"
],
"additionalProperties": false
},
"hwKeysWithData": {
"type": "array",
"items": {
"type": "string"
}
},
"series": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"precision": {
"type": [
"string",
"null"
]
},
"label": {
"type": "string"
},
"vendor": {
"type": [
"string",
"null"
]
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"date": {
"type": "string",
"format": "date"
},
"value": {
"type": "number"
},
"synthetic": {
"type": "boolean"
}
},
"required": [
"date",
"value"
],
"additionalProperties": false
}
},
"steps": {
"type": "array",
"items": {
"type": "object",
"properties": {
"kind": {
"type": "string",
"enum": [
"baseline",
"gain",
"best"
]
},
"date": {
"type": "string"
},
"from": {
"type": "number"
},
"to": {
"type": "number"
},
"delta": {
"type": "number"
},
"sources": {
"type": "array",
"items": {
"type": "object",
"properties": {
"date": {
"type": "string"
},
"runId": {
"type": "integer"
},
"model": {
"type": "string"
},
"precision": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"benchmarkType": {
"type": "string"
},
"runUrl": {
"type": "string"
}
},
"required": [
"date",
"runId",
"model",
"precision",
"hwKey",
"benchmarkType",
"runUrl"
],
"additionalProperties": false
}
},
"prLinks": {
"type": "array",
"items": {
"type": "string"
}
},
"linksAvailable": {
"type": "boolean"
}
},
"required": [
"kind",
"date",
"from",
"to",
"delta",
"sources",
"prLinks",
"linksAvailable"
],
"additionalProperties": false
}
},
"excludedPointCount": {
"type": "integer"
}
},
"required": [
"key",
"hwKey",
"label",
"points"
],
"additionalProperties": false
}
},
"count": {
"type": "integer"
}
},
"required": [
"view",
"apiVersion",
"params",
"metric",
"target",
"interactivity",
"hwKeysWithData",
"series",
"count"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "historical",
"apiVersion": "v1",
"params": {
"model": "DeepSeek-V4-Pro",
"sequence": "8k/1k",
"metric": "y_tpPerGpu",
"target": 35,
"precisions": [
"fp8"
],
"gpus": [],
"vendors": [],
"frameworks": [],
"deployment": [],
"start": null,
"end": null,
"format": "json"
},
"metric": {
"key": "tpPerGpu",
"configKey": "y_tpPerGpu",
"label": "Output Throughput per GPU (tok/s/gpu)",
"labelZh": "单 GPU 输出吞吐（tok/s/gpu）"
},
"target": 35,
"interactivity": {
"statistic": "median",
"field": "median_intvty",
"unit": "tok/s/user"
},
"hwKeysWithData": [
"h200_trt"
],
"series": [
{
"key": "h200_trt",
"hwKey": "h200_trt",
"precision": null,
"label": "H200 (TRTLLM)",
"vendor": "NVIDIA",
"points": [
{
"date": "2026-07-01",
"value": 310.4
},
{
"date": "2026-08-20",
"value": 355.2
},
{
"date": "2026-08-25",
"value": 355.2,
"synthetic": true
}
]
}
],
"count": 3
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
key,hwKey,precision,label,vendor,date,value,synthetic,targetInteractivity,interactivityStatistic,interactivityField
h200_trt,h200_trt,,H200 (TRTLLM),NVIDIA,2026-07-01,310.4,false,35,median,median_intvty400
参数无效。响应体会给出参数名，枚举参数还会列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "target must be >= 1",
"param": "target"
}500
历史基准数据查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/calculator插值计算器工作点
在服务端计算吞吐计算器：基准测试扫描按硬件配置分组，取帕累托前沿，并用与仪表盘相同的单调 Steffen–Hermite 插值在所请求目标处求值。可选地按固定功率为每个配置估算集群规模（mw），并给出每个配置在 $/M token 成本上限内的最大交互性（costcap）。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| chipBasis | query | string | 可选 | pool（默认）或 total，指定分离式配置的输入、输出吞吐量及其派生指标所除的芯片数：pool 下输入按预填充芯片、输出按解码芯片计；total 按部署中的芯片总数计，与聚合配置及 tput_per_gpu 口径一致。聚合配置在两种口径下结果相同。 | |
| costProvider | query | enum | 可选 | 成本口径：costh（自有）或 costr（租赁）；tcoBasis 选择 internal 或 external 假设。 | costh |
| costType | query | enum | 可选 | 成本与吞吐字段所用的 token 口径。 | total |
| costcap | query | number | 可选 | 成本上限（$/M token）。设置后响应附带 costCap 部分，给出各配置在上限内的最大交互性。 | 2.5 |
| date | query | date | 可选 | 使用 YYYY-MM-DD 当日或之前的数据。省略则使用最新数据。 | 2026-08-08 |
| format | query | enum | 可选 | 响应编码。 | json |
| gpus | query | CSV 列表 | 可选 | 以逗号分隔的硬件键；可匹配完整 hwKey（b300_sglang）或基础芯片（b300）。 | b300,mi355x |
| hideSkuAboveConfigLimit | query | string | 可选 | 布尔值，默认为 true。排除目标超出配置实测上限的硬件。 | |
| mode | query | enum | 可选 | 插值方向。 | interactivity-to-throughput |
| model | query | string | 必填 | 模型显示名称或对比页 slug。 | DeepSeek-V4-Pro |
| mw | query | number | 可选 | 设施功率预算（MW）。设置后每个结果附带 fleet 字段（芯片数、总 tok/s、并发用户数、成本）。 | 100 |
| percentile | query | enum | 可选 | agentic traces 的交互性分位数；固定序列使用中位数。 | p90 |
| precisions | query | CSV 列表 | 可选 | 以逗号分隔的精度列表。省略时选择官方数据中曲线最多的精度，并纳入 unofficial-run 叠加数据中的精度。 | fp4,fp8 |
| runId | query | string | 可选 | 用于固定数据快照的 GitHub workflow 运行编号（数字）。 | 123456789 |
| sequence | query | enum | 可选 | 负载序列。支持 1k/1k、1k/8k、8k/1k（或连字符形式）以及 agentic-traces（别名 agentic）。 | 1k/1k |
| target | query | number | 可选 | 输入轴上的目标值：交互性（tok/s/user）；当 mode 为 throughput-to-interactivity 时为吞吐（tok/s/GPU）。 | 35 |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 | |
| unofficialrun | query | string | 可选 | 最多八个公开 CI 运行 ID，以逗号分隔。去除各项首尾空白和重复 ID 后，按剩余顺序分配运行索引。叠加结果与官方数据分开。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/calculator?model=DeepSeek-V4-Pro&sequence=1k/1k'响应
CalculatorView200
每个硬件配置在所请求目标下经 Steffen–Hermite 插值得到的工作点。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"calculator"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"generatedAt": {
"type": [
"string",
"null"
],
"description": "Latest run date among the rows the view was computed from."
},
"params": {
"type": "object",
"additionalProperties": true
},
"hardware": {
"type": "array",
"items": {
"type": "object",
"properties": {
"hwKey": {
"type": "string"
},
"resultKey": {
"type": "string"
},
"source": {
"type": "string"
},
"precision": {
"type": [
"string",
"null"
]
},
"label": {
"type": "string"
},
"value": {
"type": "number"
},
"inputThroughput": {
"type": "number"
},
"outputThroughput": {
"type": "number"
},
"cost": {
"type": "object",
"properties": {
"total": {
"type": "number"
},
"input": {
"type": "number"
},
"output": {
"type": "number"
}
},
"required": [
"total",
"input",
"output"
],
"additionalProperties": false
},
"tpPerMw": {
"type": "number"
},
"inputTpPerMw": {
"type": "number"
},
"outputTpPerMw": {
"type": "number"
},
"concurrency": {
"type": "number"
},
"cacheHitRate": {
"type": [
"number",
"null"
]
},
"inputTokenShare": {
"type": [
"number",
"null"
]
},
"clamped": {
"type": "boolean"
},
"clampedAbove": {
"type": "boolean"
},
"clampedBelow": {
"type": "boolean"
},
"nearest": {
"type": "object",
"properties": {
"below": {
"type": [
"object",
"null"
],
"properties": {
"interactivity": {
"type": "number"
},
"throughput": {
"type": "number"
},
"concurrency": {
"type": "number"
}
},
"additionalProperties": false
},
"above": {
"type": [
"object",
"null"
],
"properties": {
"interactivity": {
"type": "number"
},
"throughput": {
"type": "number"
},
"concurrency": {
"type": "number"
}
},
"additionalProperties": false
}
},
"required": [
"below",
"above"
],
"additionalProperties": false
},
"fleet": {
"type": [
"object",
"null"
],
"description": "Present only when mw is set.",
"properties": {
"chips": {
"type": "number"
},
"totalTokPerSec": {
"type": "number"
},
"concurrentUsers": {
"type": "number"
},
"costPerHour": {
"type": "number"
},
"costPerMonth": {
"type": "number"
}
},
"additionalProperties": false
}
},
"required": [
"hwKey",
"resultKey",
"precision",
"label",
"value",
"inputThroughput",
"outputThroughput",
"cost",
"tpPerMw",
"inputTpPerMw",
"outputTpPerMw",
"concurrency",
"cacheHitRate",
"inputTokenShare",
"clamped",
"clampedAbove",
"clampedBelow",
"nearest"
],
"additionalProperties": false
}
},
"costCap": {
"type": "array",
"description": "Present only when costcap is set.",
"items": {
"type": "object",
"properties": {
"hwKey": {
"type": "string"
},
"resultKey": {
"type": "string"
},
"label": {
"type": "string"
},
"maxInteractivity": {
"type": [
"number",
"null"
]
},
"throughput": {
"type": [
"number",
"null"
]
},
"concurrentUsers": {
"type": [
"number",
"null"
]
}
},
"required": [
"hwKey",
"resultKey",
"label",
"maxInteractivity",
"throughput",
"concurrentUsers"
],
"additionalProperties": false
}
},
"count": {
"type": "integer"
}
},
"required": [
"view",
"apiVersion",
"generatedAt",
"params",
"hardware",
"count"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "calculator",
"apiVersion": "v1",
"generatedAt": "2026-08-20",
"params": {
"model": "DeepSeek-V4-Pro",
"sequence": "1k/1k",
"precisions": [
"fp4"
],
"target": 35,
"mode": "interactivity-to-throughput",
"costProvider": "costh",
"costType": "total",
"percentile": "p90",
"gpus": [],
"format": "json"
},
"hardware": [
{
"hwKey": "b300_sglang",
"resultKey": "b300_sglang",
"precision": "fp4",
"label": "B300 (SGLang)",
"value": 1234.5,
"inputThroughput": 820.1,
"outputThroughput": 414.4,
"cost": {
"total": 1.21,
"input": 1.83,
"output": 3.62
},
"tpPerMw": 890123.4,
"inputTpPerMw": 591234.5,
"outputTpPerMw": 298888.9,
"concurrency": 24,
"cacheHitRate": null,
"inputTokenShare": 0.5,
"clamped": false,
"clampedAbove": false,
"clampedBelow": false,
"nearest": {
"below": {
"interactivity": 30,
"throughput": 1500,
"concurrency": 32
},
"above": {
"interactivity": 50,
"throughput": 900,
"concurrency": 16
}
}
}
],
"count": 1
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string",
"description": "One flat row per hardware result."
}响应示例
hwKey,resultKey,precision,label,value,inputThroughput,outputThroughput,costTotal,costInput,costOutput,tpPerMw,concurrency,clamped,clampedAbove,clampedBelow
b300_sglang,b300_sglang,fp4,B300 (SGLang),1234.5,820.1,414.4,1.21,1.83,3.62,890123.4,24,false,false,false400
参数无效。响应体给出参数名；枚举类参数还会给出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
},
"param": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"error": "Unknown mode: sideways",
"param": "mode",
"allowed": [
"interactivity-to-throughput",
"throughput-to-interactivity"
]
}500
计算器视图构建失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
},
"param": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/fleet集群生命周期经济性
在默认 35 tok/s/user 目标下，按设施功率预算为每种芯片确定固定集群规模，并沿模型生命周期投影收入、成本与利润。每种芯片沿其实测的最优演进阶梯前进：某一运行日期的最佳配置若超过此前所有配置，即构成一个台阶，并按 ramp 逐步生效。价格默认取最便宜集群的盈亏平衡价，输出:输入价格比固定为 4。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| cache | query | number | 可选 | 按缓存读取折扣计费的缓存输入百分比。仅适用于 agentic traces。 | 10 |
| costProvider | query | string | 可选 | costh 为自有成本，costr 为租赁成本。利润视图还支持 custom，需配合 customCosts。不支持 costn。 | |
| costType | query | string | 可选 | total（默认）、input 或 output token。分离式输入和输出速率保留各自的 prefill、decode 分母。 | |
| format | query | enum | 可选 | 响应编码。 | json |
| gpus | query | CSV 列表 | 可选 | 以逗号分隔的硬件键；可匹配完整 hwKey（b300_sglang）或基础芯片（b300）。 | b300,mi355x |
| horizon | query | number | 可选 | 投影时长（月）。默认在最后一次实测扫描后再留一小段。 | 12 |
| metric | query | enum | 可选 | 每个点的 value 字段所对应的指标。 | margin |
| model | query | string | 必填 | 模型显示名称或对比页 slug。 | DeepSeek-V4-Pro |
| mtbi | query | number | 可选 | 平均中断间隔（天）。 | 24 |
| mw | query | number | 必填 | 设施功率预算（MW），必须大于 0。 | 100 |
| oprice | query | number | 可选 | 输出 token 价格（$/M token）。默认取输入价格的 4 倍。 | 1.68 |
| percentile | query | enum | 可选 | agentic traces 的交互性分位数；固定序列使用中位数。 | p90 |
| precisions | query | CSV 列表 | 可选 | 以逗号分隔的精度列表。省略时自动选择数据中曲线最密的精度。 | fp4 |
| price | query | number | 可选 | 输入 token 价格（$/M token）。默认取最便宜集群的盈亏平衡价；只给 oprice 时按 4 倍比例推导。 | 0.42 |
| ramp | query | number | 可选 | 每次配置上线达到满负荷所需的月数。默认与仪表板生命周期面板一致（0.5）。 | 0.5 |
| recovery | query | number | 可选 | 每次中断的恢复时间（小时）。 | 12 |
| sequence | query | enum | 可选 | 负载序列。支持 1k/1k、1k/8k、8k/1k（或连字符形式）以及 agentic-traces（别名 agentic）。 | 8k/1k |
| target | query | string | 可选 | 正数工作点目标；交互性转吞吐量模式下单位为 tok/s/user。两个利润视图按模型设置默认值：Kimi K3 为 60、GLM-5.2/5.3 为 135、MiniMax M3 为 115、DeepSeek V4 Pro 为 170、DeepSeek V4.1 Flash 为 150 tok/s/user。仪表板链接通过 /<model>/target/<value> 路径片段传递目标值；本 API 仍使用 target 查询参数。这些默认值是用于扩大 SKU 覆盖范围的可修改假设，并非与 OpenRouter P90 等价的实测值。 | |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/fleet?model=DeepSeek-V4-Pro&mw=100'响应
FleetView200
每种芯片一条生命周期经济性序列：按功率预算确定固定集群规模，并沿模型生命周期投影每次实测到的配置改进。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"fleet"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"generatedAt": {
"type": [
"string",
"null"
],
"description": "Latest run date among the history rows the view was computed from."
},
"params": {
"type": "object",
"additionalProperties": true
},
"assumptions": {
"type": "object",
"additionalProperties": true
},
"series": {
"type": "array",
"items": {
"type": "object",
"properties": {
"hwKey": {
"type": "string",
"description": "Base chip key, e.g. b300."
},
"label": {
"type": "string"
},
"hwKeysUsed": {
"type": "array",
"items": {
"type": "string"
},
"description": "Every config that contributed a rung, in takeover order."
},
"disagg": {
"type": "boolean"
},
"gpus": {
"type": "integer"
},
"provisionedMw": {
"type": "number"
},
"costPerHour": {
"type": "number"
},
"concurrentUsersNow": {
"type": "number"
},
"availability": {
"type": "number"
},
"breakEvenPricePerMTok": {
"type": [
"number",
"null"
]
},
"improvementFactor": {
"type": [
"number",
"null"
]
},
"improvementCount": {
"type": "integer"
},
"paybackMonth": {
"type": [
"number",
"null"
]
},
"lifetimeMargin": {
"type": "number"
},
"revenuePerDay": {
"type": "number"
},
"costPerDay": {
"type": "number"
},
"marginPerDay": {
"type": "number"
},
"startMonth": {
"type": "number"
},
"endMonth": {
"type": "number"
},
"rampEndMonth": {
"type": [
"number",
"null"
]
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"month": {
"type": "number",
"description": "Months since the anchor date (model release)."
},
"value": {
"type": "number",
"description": "The selected metric at this point."
},
"revenue": {
"type": "number",
"description": "$/day"
},
"cost": {
"type": "number",
"description": "$/day"
},
"margin": {
"type": "number",
"description": "$/day"
},
"revenuePerMw": {
"type": "number",
"description": "$/MW/day"
},
"marginPerMw": {
"type": "number",
"description": "$/MW/day"
},
"cumulative": {
"type": "number",
"description": "Cumulative margin, $"
},
"cumulativeRevenue": {
"type": "number",
"description": "Cumulative revenue, $"
},
"isStep": {
"type": "boolean"
},
"isRamp": {
"type": "boolean"
}
},
"required": [
"month",
"value",
"revenue",
"cost",
"margin",
"revenuePerMw",
"marginPerMw",
"cumulative",
"cumulativeRevenue",
"isStep",
"isRamp"
],
"additionalProperties": false
}
}
},
"required": [
"hwKey",
"label",
"hwKeysUsed",
"disagg",
"gpus",
"provisionedMw",
"costPerHour",
"concurrentUsersNow",
"availability",
"breakEvenPricePerMTok",
"points"
],
"additionalProperties": true
}
},
"count": {
"type": "integer"
}
},
"required": [
"view",
"apiVersion",
"generatedAt",
"params",
"assumptions",
"series",
"count"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "fleet",
"apiVersion": "v1",
"generatedAt": "2026-08-20",
"params": {
"model": "DeepSeek-V4-Pro",
"sequence": "8k/1k",
"precisions": [
"fp4"
],
"mw": 100,
"price": 0.42,
"oprice": 1.68,
"ramp": 0.5,
"cache": 10,
"mtbi": 24,
"recovery": 12,
"horizon": 7,
"metric": "margin",
"percentile": "p90",
"gpus": [],
"format": "json"
},
"assumptions": {
"target": 35,
"costProvider": "costh",
"costType": "total",
"mtbiDays": 24,
"recoveryHours": 12,
"inputPricePerMTok": 0.42,
"outputPricePerMTok": 1.68,
"outputPriceMultiple": 4,
"rampMonths": 0.5,
"cachedInputPct": 10,
"cacheReadRatio": 1,
"availability": 0.98,
"breakEvenPricePerMTok": 0.42,
"anchorDate": "2026-04-24",
"horizonMonths": 7
},
"series": [
{
"hwKey": "b300",
"label": "B300",
"hwKeysUsed": [
"b300_sglang"
],
"disagg": false,
"gpus": 70921,
"provisionedMw": 99.99,
"costPerHour": 212763,
"concurrentUsersNow": 2500000,
"availability": 0.98,
"breakEvenPricePerMTok": 0.42,
"improvementFactor": 1.33,
"improvementCount": 1,
"paybackMonth": null,
"lifetimeMargin": 0,
"points": [
{
"month": 1.25,
"value": -1250000,
"revenue": 3860000,
"cost": 5110000,
"margin": -1250000,
"revenuePerMw": 38600,
"marginPerMw": -12500,
"cumulative": -38000000,
"cumulativeRevenue": 117000000,
"isStep": true,
"isRamp": true
}
]
}
],
"count": 1
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string",
"description": "One row per lifecycle point."
}响应示例
hwKey,label,month,value,revenue,cost,margin,revenuePerMw,marginPerMw,cumulative,cumulativeRevenue,isStep,isRamp
b300,B300,1.25,-1250000,3860000,5110000,-1250000,38600,-12500,-38000000,117000000,true,true400
参数无效或缺失。响应体给出参数名；枚举类参数还会给出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
},
"param": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"error": "mw is required (facility power budget in MW, > 0)",
"param": "mw"
}500
集群视图构建失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
},
"param": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/evaluation获取评测图表视图
返回 /infx/evaluation 页面为指定模型、基准和运行日期渲染的聚合评测柱状数据：每个配置取最新数据，同配置的重试合并为一行（score 为平均值，stderr 覆盖最小/最大误差范围，n 为参与平均的运行次数）。benchmarks 列出该模型的全部可用基准。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| unofficialrun | query | string | 可选 | 最多八个公开 CI 运行 ID，以逗号分隔。去除各项首尾空白和重复 ID 后，按剩余顺序分配运行索引。叠加结果与官方数据分开。 | |
| benchmark | query | string | 可选 | 评测任务键（例如 gsm8k）。默认取该模型可用基准中按字母序的第一个。传入未知值会返回 400，并列出可用基准。 | gsm8k |
| date | query | date | 可选 | 请求的评测运行日期（YYYY-MM-DD），会像页面一样解析到最接近的可用日期。默认使用最新可用日期。 | 2026-08-20 |
| format | query | enum | 可选 | 响应编码。csv 为每个配置返回一行平面数据，标签中的换行会被展平。 | json |
| gpus | query | string | 可选 | 以逗号分隔的硬件或配置键；gpu-metrics 使用数字 GPU 索引。省略时选择全部。 | |
| model | query | string | 必填 | 模型显示名称（不区分大小写）或对比页 slug。 | DeepSeek-V4-Pro |
| precisions | query | CSV 列表 | 可选 | 以逗号分隔的精度过滤（fp4、fp4fp8、fp8、bf16、int4）。默认包含该模型评测数据中出现的全部精度。 | fp8 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/evaluation?model=DeepSeek-V4-Pro&benchmark=gsm8k&format=json'响应
EvaluationView200
按解析后的模型、基准和日期聚合的评测图表数据行。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"evaluation"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"benchmark": {
"type": [
"string",
"null"
]
},
"date": {
"type": [
"string",
"null"
],
"format": "date"
},
"precisions": {
"type": "array",
"items": {
"type": "string"
}
},
"gpus": {
"type": "array",
"items": {
"type": "string"
}
},
"unofficialrun": {
"type": [
"string",
"null"
]
},
"format": {
"type": "string",
"enum": [
"json",
"csv"
]
}
},
"required": [
"model",
"benchmark",
"date",
"precisions",
"gpus",
"unofficialrun",
"format"
],
"additionalProperties": false
},
"benchmarks": {
"type": "array",
"items": {
"type": "string"
}
},
"rows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"source": {
"type": "string"
},
"hwKey": {
"type": "string"
},
"label": {
"type": "string"
},
"score": {
"type": "number"
},
"stderr": {
"type": "number"
},
"n": {
"type": "integer",
"description": "How many repeated runs (retries/reruns) the row averages."
},
"precision": {
"type": "string"
},
"framework": {
"type": "string"
},
"date": {
"type": "string",
"format": "date"
}
},
"required": [
"source",
"hwKey",
"label",
"score",
"stderr",
"n",
"precision",
"framework",
"date"
],
"additionalProperties": false
}
}
},
"required": [
"view",
"apiVersion",
"params",
"benchmarks",
"rows"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "evaluation",
"apiVersion": "v1",
"params": {
"model": "DeepSeek-V4-Pro",
"benchmark": "gsm8k",
"date": "2026-08-20",
"precisions": [
"fp8"
],
"format": "json",
"gpus": [],
"unofficialrun": null
},
"benchmarks": [
"aime25",
"gsm8k"
],
"rows": [
{
"source": "official",
"hwKey": "h200_sglang",
"label": "H200 (SGLang)\nC128 T8 E1",
"score": 0.85,
"stderr": 0.01,
"n": 2,
"precision": "fp8",
"framework": "sglang",
"date": "2026-08-20"
}
]
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
hwKey,label,score,stderr,n,precision,framework,date
h200_sglang,H200 (SGLang) C128 T8 E1,0.85,0.01,2,fp8,sglang,2026-08-20400
model、benchmark、date、precisions 或 format 参数无效。响应体会列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown benchmark for DeepSeek-V4-Pro: mmlu",
"param": "benchmark",
"allowed": [
"aime25",
"gsm8k"
]
}500
评测数据查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/reliability获取可靠性图表视图
返回 /infx/reliability 页面图表渲染的各硬件基准运行成功率聚合结果，按滚动时间范围预设分桶。成功率为百分比，保留 2 位小数；generatedFrom 给出底层运行统计数据的最早与最晚日期。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| asOf | query | string | 可选 | 用于复现滚动可靠性或镜像日期差计算的参考日期 YYYY-MM-DD。 | |
| format | query | enum | 可选 | 响应编码。csv 为每个硬件返回一行平面数据。 | json |
| gpus | query | string | 可选 | 以逗号分隔的硬件或配置键；gpu-metrics 使用数字 GPU 索引。省略时选择全部。 | |
| range | query | enum | 可选 | 聚合使用的滚动时间范围预设，与 /infx/reliability 页面的预设一致。 | last-7-days |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/reliability?range=last-7-days&format=json'响应
ReliabilityView200
所选时间范围内每个硬件的聚合成功率。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"reliability"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"properties": {
"range": {
"type": "string",
"enum": [
"last-3-days",
"last-7-days",
"last-month",
"last-3-months",
"all-time"
]
},
"format": {
"type": "string",
"enum": [
"json",
"csv"
]
},
"gpus": {
"type": "array",
"items": {
"type": "string"
}
},
"asOf": {
"type": [
"string",
"null"
],
"format": "date"
}
},
"required": [
"range",
"format",
"gpus",
"asOf"
],
"additionalProperties": false
},
"range": {
"type": "string",
"enum": [
"last-3-days",
"last-7-days",
"last-month",
"last-3-months",
"all-time"
]
},
"hardware": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"successRate": {
"type": "number",
"description": "Success percentage over the range, rounded to 2 decimal places."
},
"successes": {
"type": "integer"
},
"total": {
"type": "integer"
}
},
"required": [
"key",
"label",
"successRate",
"successes",
"total"
],
"additionalProperties": false
}
},
"generatedFrom": {
"type": "object",
"properties": {
"firstDate": {
"type": [
"string",
"null"
],
"format": "date"
},
"lastDate": {
"type": [
"string",
"null"
],
"format": "date"
}
},
"required": [
"firstDate",
"lastDate"
],
"additionalProperties": false
}
},
"required": [
"view",
"apiVersion",
"params",
"range",
"hardware",
"generatedFrom"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "reliability",
"apiVersion": "v1",
"params": {
"range": "last-7-days",
"format": "json",
"gpus": [],
"asOf": null
},
"range": "last-7-days",
"hardware": [
{
"key": "h200",
"label": "H200",
"successRate": 93.33,
"successes": 28,
"total": 30
},
{
"key": "b200",
"label": "B200",
"successRate": 90,
"successes": 27,
"total": 30
}
],
"generatedFrom": {
"firstDate": "2025-11-03",
"lastDate": "2026-08-26"
}
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
range,key,label,successRate,successes,total
last-7-days,h200,H200,93.33,28,30400
range 或 format 参数无效。响应体会列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown range: last-year",
"param": "range",
"allowed": [
"last-3-days",
"last-7-days",
"last-month",
"last-3-months",
"all-time"
]
}500
可靠性数据查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/gpu-specs获取 GPU 规格视图
返回 /infx/gpu-specs 页面背后的静态芯片规格表——可用显存、带宽、稠密 Tensor Core TFLOP/s，以及 scale-up/scale-out 互连细节——并附带可作图指标的元数据。不读取数据库；算力为稠密（不含稀疏）值，显存容量为驱动可用值。NVL72 包含网卡及单 GPU 网络带宽（GB200：ConnectX-7 400G NDR；GB300：ConnectX-8 800G XDR），机柜间交换机和拓扑仍为 null。MI355X 使用第四代 Infinity Fabric。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| format | query | enum | 可选 | 响应编码。csv 为每个芯片返回一行平面数据。 | json |
| metric | query | enum | 可选 | 图表指标键。设置后响应会附加 ranking 数组，按该指标降序排列芯片（无该指标值的芯片不参与排名）。 | memoryBandwidth |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/gpu-specs?metric=memoryBandwidth&format=json'响应
GpuSpecsView200
静态芯片规格与图表指标元数据；仅在设置 metric 时返回 ranking。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"gpu-specs"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"properties": {
"metric": {
"type": [
"string",
"null"
]
},
"format": {
"type": "string",
"enum": [
"json",
"csv"
]
}
},
"required": [
"metric",
"format"
],
"additionalProperties": false
},
"chips": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"name": {
"type": "string"
},
"vendor": {
"type": "string",
"enum": [
"nvidia",
"amd"
]
},
"memory": {
"type": "string"
},
"memoryType": {
"type": "string"
},
"memoryBandwidth": {
"type": "string"
},
"fp4": {
"type": [
"number",
"null"
]
},
"fp8": {
"type": "number"
},
"bf16": {
"type": "number"
},
"scaleUpTech": {
"type": "string"
},
"scaleUpBandwidth": {
"type": "string"
},
"scaleUpWorldSize": {
"type": "integer"
},
"scaleOutBandwidth": {
"type": [
"string",
"null"
]
},
"scaleOutTech": {
"type": [
"string",
"null"
]
},
"nic": {
"type": [
"string",
"null"
],
"description": "Installed NIC model and per-GPU port rate; GB200 NVL72 uses ConnectX-7 400G and GB300 NVL72 uses ConnectX-8 800G."
},
"scaleOutSwitch": {
"type": [
"string",
"null"
]
},
"scaleOutTopology": {
"type": [
"string",
"null"
],
"description": "Null means the topology is unspecified; it does not imply that a known NIC or network is absent."
},
"scaleUpTopology": {
"type": "string"
},
"scaleUpSwitch": {
"type": [
"string",
"null"
]
},
"memoryGB": {
"type": [
"number",
"null"
]
},
"memoryBandwidthTBs": {
"type": [
"number",
"null"
]
},
"fp4Tflops": {
"type": [
"number",
"null"
]
},
"fp8Tflops": {
"type": "number"
},
"bf16Tflops": {
"type": "number"
},
"scaleUpBandwidthGBs": {
"type": [
"number",
"null"
]
},
"domainMemoryTB": {
"type": "number"
},
"domainMemoryBandwidthTBs": {
"type": "number"
},
"scaleOutBandwidthGbits": {
"type": [
"number",
"null"
]
}
},
"required": [
"key",
"label",
"name",
"vendor",
"memory",
"memoryType",
"memoryBandwidth",
"fp4",
"fp8",
"bf16",
"scaleUpTech",
"scaleUpBandwidth",
"scaleUpWorldSize",
"scaleOutBandwidth",
"scaleOutTech",
"nic",
"scaleOutSwitch",
"scaleOutTopology",
"scaleUpTopology",
"scaleUpSwitch",
"memoryGB",
"memoryBandwidthTBs",
"fp4Tflops",
"fp8Tflops",
"bf16Tflops",
"scaleUpBandwidthGBs",
"domainMemoryTB",
"domainMemoryBandwidthTBs",
"scaleOutBandwidthGbits"
],
"additionalProperties": false
}
},
"metrics": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"unit": {
"type": "string"
}
},
"required": [
"key",
"label",
"unit"
],
"additionalProperties": false
}
},
"ranking": {
"type": "array",
"items": {
"type": "object",
"properties": {
"chip": {
"type": "string"
},
"label": {
"type": "string"
},
"value": {
"type": "number"
},
"rank": {
"type": "integer"
}
},
"required": [
"chip",
"label",
"value",
"rank"
],
"additionalProperties": false
}
}
},
"required": [
"view",
"apiVersion",
"params",
"chips",
"metrics"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "gpu-specs",
"apiVersion": "v1",
"params": {
"metric": "memoryBandwidth",
"format": "json"
},
"chips": [
{
"key": "b200-sxm",
"label": "B200 SXM",
"name": "B200 SXM",
"vendor": "nvidia",
"memory": "180 GB",
"memoryType": "HBM3e",
"memoryBandwidth": "8 TB/s",
"fp4": 9000,
"fp8": 4500,
"bf16": 2250,
"scaleUpTech": "NVLink 5.0",
"scaleUpBandwidth": "900 GB/s",
"scaleUpWorldSize": 8,
"scaleOutBandwidth": "400 Gbit/s",
"scaleOutTech": "gIB RoCEv2 Ethernet",
"nic": "ConnectX-7 400GbE",
"scaleOutSwitch": "12.8T Whitebox Leaf Tomahawk3 & 25.6T Whitebox Tomahawk4",
"scaleOutTopology": "4-rail optimized",
"scaleUpTopology": "Switched 2-rail Optimized",
"scaleUpSwitch": "28.8Tbit/s NVSwitch Gen 4.0",
"memoryGB": 180,
"memoryBandwidthTBs": 8,
"fp4Tflops": 9000,
"fp8Tflops": 4500,
"bf16Tflops": 2250,
"scaleUpBandwidthGBs": 900,
"domainMemoryTB": 1.44,
"domainMemoryBandwidthTBs": 64,
"scaleOutBandwidthGbits": 400
}
],
"metrics": [
{
"key": "memory",
"label": "Memory",
"unit": "GB"
},
{
"key": "memoryBandwidth",
"label": "Mem BW",
"unit": "TB/s"
}
],
"ranking": [
{
"chip": "b200-sxm",
"label": "B200 SXM",
"value": 8,
"rank": 1
}
]
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
key,label,name,vendor,memory,memoryType,memoryBandwidth,fp4,fp8,bf16
b200-sxm,B200 SXM,B200 SXM,nvidia,180 GB,HBM3e,8 TB/s,9000,4500,2250400
metric 或 format 参数无效。响应体会列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown metric: tdp",
"param": "metric",
"allowed": [
"memory",
"memoryBandwidth",
"fp4",
"fp8",
"bf16",
"scaleUpBandwidth",
"scaleUpWorldSize",
"domainMemory",
"domainMemoryBandwidth",
"scaleOutBandwidth"
]
}滚动查看完整示例
500
GPU 规格数据组装失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/overview获取总览成本矩阵视图
以数据形式返回 /overview 页面的矩阵：对每个精选「模型-场景」行，给出各硬件在请求交互性档位下最优的社区（或全部引擎）推理配置，包含每百万 token 成本、每 GPU 吞吐、相对参照 GPU 或历史窗口的差值，以及胜出配置。与页面不同，无效参数会返回 400，而不是被静默归一化。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| compare | query | string | 可选 | 对比模式：hardware 表示与参照 GPU 对比；7d/30d/60d/90d 表示与相应天数之前的同一单元格对比。 | 30d |
| engine | query | string | 可选 | 推理引擎范围：仅社区框架，或包含厂商自研栈在内的全部引擎。 | community |
| format | query | string | 可选 | 响应格式。CSV 为每个「模型-场景-硬件」单元格返回一行平面数据。 | json |
| hwrows | query | string | 可选 | 硬件单元格范围：仅有定价测量值的单元格，或全部硬件列。 | all |
| models | query | string | 可选 | 模型范围：默认精选集合，或全部有数据的模型。 | default |
| ref | query | string | 可选 | compare=hardware 时用于计算差值的参照 GPU。 | b200 |
| rows | query | string | 可选 | 历史窗口下的行范围：仅窗口内发生变化的行，或全部行。 | all |
| tier | query | integer | 可选 | 交互性档位，单位为每用户每秒输出 token。所有单元格按该档位读取。 | 100 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/overview?tier=50&engine=community&compare=hardware&ref=b200'响应
OverviewView200
请求档位下的总览成本矩阵：每个「模型-场景」一行，每个硬件列一个单元格。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"overview"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"generatedAt": {
"type": [
"string",
"null"
],
"format": "date"
},
"params": {
"type": "object",
"properties": {
"tier": {
"type": "integer"
},
"engine": {
"type": "string"
},
"compare": {
"type": "string"
},
"ref": {
"type": "string"
},
"models": {
"type": "string"
},
"rows": {
"type": "string"
},
"hwrows": {
"type": "string"
},
"format": {
"type": "string",
"enum": [
"json",
"csv"
]
}
},
"required": [
"tier",
"engine",
"compare",
"ref",
"models",
"rows",
"hwrows",
"format"
],
"additionalProperties": false
},
"tiers": {
"type": "array",
"items": {
"type": "integer"
}
},
"scenarios": {
"type": "array",
"items": {
"type": "string"
}
},
"referenceHardware": {
"type": "string"
},
"historicalWindow": {
"type": [
"object",
"null"
],
"properties": {
"key": {
"type": "string"
},
"snapshotDate": {
"type": "string"
},
"targetDate": {
"type": "string"
},
"earliestDate": {
"type": "string"
}
},
"additionalProperties": true
},
"unchangedRowCount": {
"type": "integer"
},
"emptyRowCount": {
"type": "integer"
},
"rows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"modelLabel": {
"type": "string"
},
"category": {
"type": "string"
},
"scenario": {
"type": "string",
"enum": [
"single_turn_8k1k",
"agentx"
]
},
"cells": {
"type": "array",
"items": {
"type": "object",
"properties": {
"hardware": {
"type": "string"
},
"hardwareLabel": {
"type": "string"
},
"costPerMtok": {
"type": [
"number",
"null"
]
},
"throughputPerGpu": {
"type": [
"number",
"null"
]
},
"estimated": {
"type": "boolean"
},
"deltaVsRefPct": {
"type": [
"number",
"null"
]
},
"missingReason": {
"type": [
"string",
"null"
]
},
"config": {
"type": [
"object",
"null"
],
"properties": {
"framework": {
"type": "string"
},
"frameworkLabel": {
"type": "string"
},
"precision": {
"type": "string"
},
"specMethod": {
"type": "string"
},
"specLabel": {
"type": [
"string",
"null"
]
},
"disagg": {
"type": "boolean"
},
"multinode": {
"type": "boolean"
},
"latestDate": {
"type": [
"string",
"null"
]
}
},
"additionalProperties": false
},
"history": {
"type": "object",
"properties": {
"status": {
"type": "string"
},
"baselineCostPerMtok": {
"type": [
"number",
"null"
]
},
"costDeltaPct": {
"type": [
"number",
"null"
]
},
"baselineDate": {
"type": [
"string",
"null"
]
}
},
"additionalProperties": false
}
},
"required": [
"hardware",
"hardwareLabel",
"costPerMtok",
"throughputPerGpu",
"estimated",
"deltaVsRefPct",
"missingReason",
"config"
],
"additionalProperties": false
}
}
},
"required": [
"model",
"modelLabel",
"category",
"scenario",
"cells"
],
"additionalProperties": false
}
}
},
"required": [
"view",
"apiVersion",
"generatedAt",
"params",
"tiers",
"scenarios",
"referenceHardware",
"historicalWindow",
"unchangedRowCount",
"emptyRowCount",
"rows"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "overview",
"apiVersion": "v1",
"generatedAt": "2026-08-20",
"params": {
"tier": 50,
"engine": "community",
"compare": "hardware",
"ref": "b200",
"models": "default",
"rows": "all",
"hwrows": "all",
"format": "json"
},
"tiers": [
30,
50,
75,
100,
150,
200
],
"scenarios": [
"single_turn_8k1k",
"agentx"
],
"referenceHardware": "b200",
"historicalWindow": null,
"unchangedRowCount": 0,
"emptyRowCount": 0,
"rows": [
{
"model": "DeepSeek-V4-Pro",
"modelLabel": "DeepSeekv4 Pro 0813 1.6T",
"category": "frontier",
"scenario": "agentx",
"cells": [
{
"hardware": "b200",
"hardwareLabel": "B200",
"costPerMtok": 0.42,
"throughputPerGpu": 1315.2,
"estimated": false,
"deltaVsRefPct": null,
"missingReason": null,
"config": {
"framework": "sglang",
"frameworkLabel": "SGLang",
"precision": "fp8",
"specMethod": "mtp",
"specLabel": "MTP",
"disagg": true,
"multinode": false,
"latestDate": "2026-08-20"
}
}
]
}
]
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
model,scenario,tier,hardware,cost_per_mtok,throughput_per_gpu,estimated,delta_vs_ref_pct,missing_reason,framework,precision,spec_method,disagg,multinode,history_status,baseline_cost_per_mtok,history_delta_pct,baseline_date
DeepSeek-V4-Pro,agentx,50,b200,0.42,1315.2,false,,,SGLang,fp8,mtp,true,false,,,,400
参数无效。响应体会指出参数名并列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown tier: 42",
"param": "tier",
"allowed": [
"30",
"50",
"75",
"100",
"150",
"200"
]
}500
总览数据组装失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/rankings获取 GPU 排名视图
以数据形式返回 /rankings 页面。按模型与场景，在该条目的交互性档位下依最优社区推理配置对硬件排序。fastest-gpu 按每 GPU tok/s 排序，cheapest-gpu 按每百万 token 成本排序。每行包含胜出的框架、精度和 disagg 标志，芯片字段将各硬件对应到 /chips 注册表中的 slug。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| format | query | string | 可选 | 响应格式。CSV 为每个上榜硬件返回一行平面数据。 | json |
| kind | query | string | 可选 | 排名类型：fastest-gpu 按每 GPU tok/s 降序，cheapest-gpu 按每百万 token 成本升序。 | cheapest-gpu |
| model | query | string | 可选 | 模型显示名称（不区分大小写）或对比页 slug。默认为全部有排名的模型；无可测数据的模型在全量结果中会被省略，显式指定时则保留。 | DeepSeek-V4-Pro |
| scenario | query | string | 可选 | 工作负载场景。别名：8k-1k 对应 single_turn_8k1k，agentic 对应 agentx。默认为每个模型在 /overview 矩阵中精选的全部场景。 | agentx |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/rankings?kind=cheapest-gpu&model=DeepSeek-V4-Pro&scenario=agentx'响应
RankingsView200
社区引擎范围下，每个「模型-场景」的硬件排名。每个条目都会给出排名所用的交互性档位（tok/s/user），默认为 125。只有当 50、75、100、150、175 或 200 中某一档的实测芯片比 125 档更多时才改用，此时取实测芯片最多的档位，数量相同时取较低档位。没有档位达到两款实测芯片时，该条目不含排名行。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"rankings"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"generatedAt": {
"type": [
"string",
"null"
],
"format": "date"
},
"params": {
"type": "object",
"properties": {
"kind": {
"type": "string"
},
"model": {
"type": "string"
},
"scenario": {
"type": "string"
},
"engine": {
"type": "string"
},
"format": {
"type": "string",
"enum": [
"json",
"csv"
]
}
},
"required": [
"kind",
"model",
"scenario",
"engine",
"format"
],
"additionalProperties": false
},
"kind": {
"type": "string",
"enum": [
"fastest-gpu",
"cheapest-gpu"
]
},
"entries": {
"type": "array",
"items": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"modelSlug": {
"type": "string"
},
"modelLabel": {
"type": "string"
},
"scenario": {
"type": "string",
"enum": [
"single_turn_8k1k",
"agentx"
]
},
"tier": {
"type": "integer"
},
"rows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"rank": {
"type": "integer"
},
"hardware": {
"type": "string"
},
"hardwareLabel": {
"type": "string"
},
"chip": {
"type": [
"string",
"null"
]
},
"value": {
"type": [
"number",
"null"
]
},
"unit": {
"type": "string",
"enum": [
"tokens_per_second_per_gpu",
"usd_per_million_tokens"
]
},
"framework": {
"type": [
"string",
"null"
]
},
"precision": {
"type": [
"string",
"null"
]
},
"disagg": {
"type": [
"boolean",
"null"
]
}
},
"required": [
"rank",
"hardware",
"hardwareLabel",
"chip",
"value",
"unit",
"framework",
"precision",
"disagg"
],
"additionalProperties": false
}
}
},
"required": [
"model",
"modelSlug",
"modelLabel",
"scenario",
"tier",
"rows"
],
"additionalProperties": false
}
}
},
"required": [
"view",
"apiVersion",
"generatedAt",
"params",
"kind",
"entries"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "rankings",
"apiVersion": "v1",
"generatedAt": "2026-08-20",
"params": {
"kind": "cheapest-gpu",
"model": "DeepSeek-V4-Pro",
"scenario": "agentx",
"engine": "community",
"format": "json"
},
"kind": "cheapest-gpu",
"entries": [
{
"model": "DeepSeek-V4-Pro",
"modelSlug": "deepseek-v4",
"modelLabel": "DeepSeekv4 Pro 0813 1.6T",
"scenario": "agentx",
"tier": 50,
"rows": [
{
"rank": 1,
"hardware": "b200",
"hardwareLabel": "B200",
"chip": "b200",
"value": 0.42,
"unit": "usd_per_million_tokens",
"framework": "SGLang",
"precision": "fp8",
"disagg": true
}
]
}
]
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
kind,model,model_slug,scenario,tier,rank,hardware,hardware_label,chip,value,unit,framework,precision,disagg
cheapest-gpu,DeepSeek-V4-Pro,deepseek-v4,agentx,50,1,b200,B200,b200,0.42,usd_per_million_tokens,SGLang,fp8,true400
kind、model、scenario 或 format 参数无效。响应体会指出参数名并列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown kind: slowest-gpu",
"param": "kind",
"allowed": [
"fastest-gpu",
"cheapest-gpu"
]
}500
排名数据查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/compare获取 GPU 组合对比视图
以数据形式返回 /compare 页面：对一个模型和两块 GPU，使用与页面相同的插值管线给出各交互性档位下的吞吐、成本和能效，并标注每档位胜者与整体对比摘要。支持自定义档位以及 per-dollar、precision、spec-decode 变体。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| format | query | string | 可选 | 响应格式。CSV 为每个档位返回一行平面数据。 | json |
| gpus | query | CSV GPU 列表 | 可选 | 恰好 2 个不同的 GPU 基础键，以逗号分隔（例如 b200,mi355x）。未知键返回 400 并列出允许的取值。 | b200,mi355x |
| model | query | string | 可选 | 模型显示名称（不区分大小写）或对比页模型 slug。与 gpus 搭配使用，可替代 slug。 | DeepSeek-V4-Pro |
| scenario | query | string | 可选 | 工作负载序列。别名：8k-1k、1k-1k、1k-8k、agentic。默认为页面选取的组合默认值（AgentX 精选模型回退到 agentic-traces，其余为 8k/1k）。 | 8k/1k |
| slug | query | string | 可选 | 对比页 slug，格式为 <model>-<gpuA>-vs-<gpuB>。slug 与 model+gpus 二选一；GPU 顺序会像页面跳转一样归一化。 | deepseek-v4-b200-vs-mi355x |
| tiers | query | CSV 数字列表 | 可选 | 自定义交互性目标（每用户 tok/s），逗号分隔，最多 12 个。默认为页面在共同测量范围内选取的 3 个目标。超出测量范围的档位会被忽略。 | 25,50,75 |
| variant | query | string | 可选 | 视图变体。per-dollar 将胜负与差值的比较基准从吞吐切换为每百万 token 成本；precision 增加按精度的对比摘要；spec-decode 增加按投机解码方法的拆分（仅固定序列场景）。 | per-dollar |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/compare?slug=deepseek-v4-b200-vs-mi355x'响应
CompareView200
在生效场景与精度下该 GPU 组合的逐档位插值对比表，以及双方摘要。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"compare"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"generatedAt": {
"type": [
"string",
"null"
],
"format": "date"
},
"params": {
"type": "object",
"properties": {
"slug": {
"type": "string"
},
"model": {
"type": "string"
},
"gpus": {
"type": "array",
"items": {
"type": "string"
}
},
"scenario": {
"type": "string"
},
"variant": {
"type": "string"
},
"tiers": {
"oneOf": [
{
"type": "string",
"enum": [
"default"
]
},
{
"type": "array",
"items": {
"type": "number"
}
}
]
},
"format": {
"type": "string",
"enum": [
"json",
"csv"
]
}
},
"additionalProperties": false
},
"model": {
"type": "object",
"properties": {
"slug": {
"type": "string"
},
"displayName": {
"type": "string"
},
"label": {
"type": "string"
}
},
"required": [
"slug",
"displayName",
"label"
],
"additionalProperties": false
},
"gpus": {
"type": "array",
"items": {
"type": "string"
}
},
"scenario": {
"type": [
"string",
"null"
]
},
"precision": {
"type": [
"string",
"null"
]
},
"variant": {
"type": "string"
},
"tiers": {
"type": "array",
"items": {
"type": "number"
}
},
"interactivityRange": {
"type": "object",
"properties": {
"min": {
"type": "number"
},
"max": {
"type": "number"
}
},
"required": [
"min",
"max"
],
"additionalProperties": false
},
"dataRange": {
"type": "object",
"properties": {
"oldest": {
"type": [
"string",
"null"
]
},
"newest": {
"type": [
"string",
"null"
]
}
},
"required": [
"oldest",
"newest"
],
"additionalProperties": false
},
"table": {
"type": "array",
"items": {
"type": "object",
"properties": {
"tier": {
"type": "number"
},
"a": {
"type": [
"object",
"null"
],
"properties": {
"hardware": {
"type": "string"
},
"configKey": {
"type": "string"
},
"throughputPerGpu": {
"type": "number"
},
"inputThroughputPerGpu": {
"type": "number"
},
"outputThroughputPerGpu": {
"type": "number"
},
"costPerMtok": {
"type": "number"
},
"costPerMtokInput": {
"type": "number"
},
"costPerMtokOutput": {
"type": "number"
},
"throughputPerMw": {
"type": "number"
},
"concurrency": {
"type": "number"
},
"precision": {
"type": [
"string",
"null"
]
},
"clamped": {
"type": "boolean"
}
},
"additionalProperties": false
},
"b": {
"type": [
"object",
"null"
],
"properties": {
"hardware": {
"type": "string"
},
"configKey": {
"type": "string"
},
"throughputPerGpu": {
"type": "number"
},
"inputThroughputPerGpu": {
"type": "number"
},
"outputThroughputPerGpu": {
"type": "number"
},
"costPerMtok": {
"type": "number"
},
"costPerMtokInput": {
"type": "number"
},
"costPerMtokOutput": {
"type": "number"
},
"throughputPerMw": {
"type": "number"
},
"concurrency": {
"type": "number"
},
"precision": {
"type": [
"string",
"null"
]
},
"clamped": {
"type": "boolean"
}
},
"additionalProperties": false
},
"basis": {
"type": "string",
"enum": [
"throughputPerGpu",
"costPerMtok"
]
},
"deltaPct": {
"type": [
"number",
"null"
]
},
"winner": {
"type": [
"string",
"null"
]
}
},
"required": [
"tier",
"a",
"b",
"basis",
"deltaPct",
"winner"
],
"additionalProperties": false
}
},
"summary": {
"type": "object",
"properties": {
"a": {
"type": "object",
"properties": {
"hardware": {
"type": "string"
},
"configCount": {
"type": "integer"
},
"bestThroughputPerGpu": {
"type": [
"number",
"null"
]
},
"bestMedianTtft": {
"type": [
"number",
"null"
]
},
"bestMedianTpot": {
"type": [
"number",
"null"
]
}
},
"additionalProperties": true
},
"b": {
"type": "object",
"properties": {
"hardware": {
"type": "string"
},
"configCount": {
"type": "integer"
},
"bestThroughputPerGpu": {
"type": [
"number",
"null"
]
},
"bestMedianTtft": {
"type": [
"number",
"null"
]
},
"bestMedianTpot": {
"type": [
"number",
"null"
]
}
},
"additionalProperties": true
},
"headToHead": {
"type": [
"object",
"null"
],
"properties": {
"faster": {
"type": "string"
},
"slower": {
"type": "string"
},
"tputPct": {
"type": "number"
},
"cheaper": {
"type": "string"
},
"pricier": {
"type": "string"
},
"costPct": {
"type": "number"
}
},
"additionalProperties": false
},
"byPrecision": {
"type": "array",
"items": {
"type": "object",
"properties": {
"precision": {
"type": "string"
},
"tiers": {
"type": "array",
"items": {
"type": "number"
}
},
"headToHead": {
"type": [
"object",
"null"
],
"properties": {
"faster": {
"type": "string"
},
"slower": {
"type": "string"
},
"tputPct": {
"type": "number"
},
"cheaper": {
"type": "string"
},
"pricier": {
"type": "string"
},
"costPct": {
"type": "number"
}
},
"additionalProperties": false
}
},
"required": [
"precision",
"tiers",
"headToHead"
],
"additionalProperties": false
}
},
"bySpecDecode": {
"type": "array",
"items": {
"type": "object",
"properties": {
"specMethod": {
"type": "string"
},
"tier": {
"type": "number"
},
"a": {
"type": [
"object",
"null"
],
"properties": {
"hardware": {
"type": "string"
},
"configKey": {
"type": "string"
},
"throughputPerGpu": {
"type": "number"
},
"inputThroughputPerGpu": {
"type": "number"
},
"outputThroughputPerGpu": {
"type": "number"
},
"costPerMtok": {
"type": "number"
},
"costPerMtokInput": {
"type": "number"
},
"costPerMtokOutput": {
"type": "number"
},
"throughputPerMw": {
"type": "number"
},
"concurrency": {
"type": "number"
},
"precision": {
"type": [
"string",
"null"
]
},
"clamped": {
"type": "boolean"
}
},
"additionalProperties": false
},
"b": {
"type": [
"object",
"null"
],
"properties": {
"hardware": {
"type": "string"
},
"configKey": {
"type": "string"
},
"throughputPerGpu": {
"type": "number"
},
"inputThroughputPerGpu": {
"type": "number"
},
"outputThroughputPerGpu": {
"type": "number"
},
"costPerMtok": {
"type": "number"
},
"costPerMtokInput": {
"type": "number"
},
"costPerMtokOutput": {
"type": "number"
},
"throughputPerMw": {
"type": "number"
},
"concurrency": {
"type": "number"
},
"precision": {
"type": [
"string",
"null"
]
},
"clamped": {
"type": "boolean"
}
},
"additionalProperties": false
}
},
"required": [
"specMethod",
"tier",
"a",
"b"
],
"additionalProperties": false
}
}
},
"required": [
"a",
"b",
"headToHead"
],
"additionalProperties": false
}
},
"required": [
"view",
"apiVersion",
"generatedAt",
"params",
"model",
"gpus",
"scenario",
"precision",
"variant",
"tiers",
"interactivityRange",
"dataRange",
"table",
"summary"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "compare",
"apiVersion": "v1",
"generatedAt": "2026-08-20",
"params": {
"slug": "deepseek-v4-b200-vs-mi355x",
"model": "DeepSeek-V4-Pro",
"gpus": [
"b200",
"mi355x"
],
"scenario": "auto",
"variant": "default",
"tiers": "default",
"format": "json"
},
"model": {
"slug": "deepseek-v4",
"displayName": "DeepSeek-V4-Pro",
"label": "DeepSeekv4 Pro 0813 1.6T"
},
"gpus": [
"b200",
"mi355x"
],
"scenario": "agentic-traces",
"precision": "fp8",
"variant": "default",
"tiers": [
21,
34,
47
],
"interactivityRange": {
"min": 12,
"max": 55
},
"dataRange": {
"oldest": "2026-05-02",
"newest": "2026-08-20"
},
"table": [
{
"tier": 34,
"a": {
"hardware": "b200",
"configKey": "b200_sglang",
"throughputPerGpu": 1315.2,
"inputThroughputPerGpu": 9821.4,
"outputThroughputPerGpu": 1315.2,
"costPerMtok": 0.42,
"costPerMtokInput": 0.06,
"costPerMtokOutput": 0.48,
"throughputPerMw": 985000,
"concurrency": 96,
"precision": "fp8",
"clamped": false
},
"b": {
"hardware": "mi355x",
"configKey": "mi355x_sglang",
"throughputPerGpu": 1104.7,
"inputThroughputPerGpu": 8455.1,
"outputThroughputPerGpu": 1104.7,
"costPerMtok": 0.47,
"costPerMtokInput": 0.07,
"costPerMtokOutput": 0.53,
"throughputPerMw": 912000,
"concurrency": 88,
"precision": "fp8",
"clamped": false
},
"basis": "throughputPerGpu",
"deltaPct": 19.1,
"winner": "b200"
}
],
"summary": {
"a": {
"hardware": "b200",
"configCount": 12,
"bestThroughputPerGpu": 1840.3,
"bestMedianTtft": 0.42,
"bestMedianTpot": 8.1
},
"b": {
"hardware": "mi355x",
"configCount": 9,
"bestThroughputPerGpu": 1512.8,
"bestMedianTtft": 0.51,
"bestMedianTpot": 9.4
},
"headToHead": {
"faster": "B200",
"slower": "MI355X",
"tputPct": 19,
"cheaper": "B200",
"pricier": "MI355X",
"costPct": 11
}
}
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
model,scenario,tier,basis,delta_pct,winner,a_hardware,a_throughput_per_gpu,a_cost_per_mtok,a_concurrency,a_clamped,b_hardware,b_throughput_per_gpu,b_cost_per_mtok,b_concurrency,b_clamped
DeepSeek-V4-Pro,agentic-traces,34,throughputPerGpu,19.1,b200,b200,1315.2,0.42,96,false,mi355x,1104.7,0.47,88,false400
slug、model、gpus、scenario、variant、tiers 或 format 参数无效。响应体会指出参数名，并在适用时列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown compare slug: not-a-pair. Expected <model>-<gpuA>-vs-<gpuB>, e.g. deepseek-v4-b200-vs-mi355x.",
"param": "slug"
}500
对比数据组装失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/operatorx获取 OperatorX 视图
/opx 页面背后的 OperatorX 算子级实测数据：在各 GPU 上计时的 GEMM、MoE 与注意力算子，按用例跨硬件对齐。每个用例与 GPU 一行，取该 GPU 最新存储的运行结果。用例来自一个工作负载，或某个模型贡献的全部用例，并限定在一种设备切分（parallel）下。flops 与 bytes 为每 GPU 的有效工作量（每个数据并行组的工作量除以 tp），因此 tflops 与 tb_s 可跨切分比较；算子未定义计数时为 null。roofline_us 为该用例在该 GPU 上的最短可能时间：每个阶段（投影、注意力核心、专家、路由等）按其自身精度的峰值算力或内存带宽取较慢者，再加上经 scale-up 链路的集合通信；roofline_share 为 roofline_us 与 latency_us 之比。run_id 与 result_index 标识存储的结果。在新结果入库前保持缓存；缺少无法读取的运行时不缓存。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| format | query | enum | 可选 | 响应编码。csv 为每个用例与 GPU 返回一行平面数据；sources 与 dims 合并为文本。 | json |
| hardware | query | string | 可选 | 以逗号分隔的 GPU 键，取自 options.hardware（h200、b200、mi355x 等）。省略时选择全部。 | b200,h200 |
| model | query | string | 可选 | options.models 中的模型名称，不区分大小写：返回该模型贡献的所有用例，跨工作负载。不能与 workload 同时使用。 | DeepSeek-V4-Pro |
| op | query | enum | 可选 | 算子类别：gemm（默认）、moe 或 attention（包含所有注意力模块类型）。 | attention |
| page | query | string | 可选 | JSON 页码，从 0 开始，每页 100 行。CSV 返回全部匹配行。 | 0 |
| parallel | query | string | 可选 | options.parallel 中的设备切分键：1 表示单卡（tp=dp=ep=dcp=1），其余为大于 1 的维度，例如 tp8 或 dp8·ep8。默认单卡；没有单卡数据时选择最小的切分。 | 1 |
| status | query | enum | 可选 | 结果状态：ok（默认）、unsupported、error、missing 或 all。未运行某用例的 GPU 不返回该用例的行。 | ok |
| workload | query | string | 可选 | options.workloads 中的工作负载 ID，例如 attention:Kimi-K3。默认选择覆盖最广的工作负载。不能与 model 同时使用。 | attention:DeepSeek-V4-Pro |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/operatorx?op=attention&model=DeepSeek-V4-Pro'响应
OperatorXView200
解析后的选择、可选项，以及每个用例与 GPU 一行的数据（该 GPU 最新存储的结果）。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"operatorx"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"properties": {
"op": {
"type": "string"
},
"workload": {
"type": [
"string",
"null"
]
},
"model": {
"type": [
"string",
"null"
]
},
"parallel": {
"type": "string"
},
"hardware": {
"type": "array",
"items": {
"type": "string"
}
},
"status": {
"type": "string"
},
"page": {
"type": "integer"
},
"format": {
"type": "string",
"enum": [
"json",
"csv"
]
}
},
"required": [
"op",
"workload",
"model",
"parallel",
"hardware",
"status",
"page",
"format"
],
"additionalProperties": false
},
"options": {
"type": "object",
"properties": {
"ops": {
"type": "array",
"items": {
"type": "string"
}
},
"workloads": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "string"
},
"label": {
"type": "string"
},
"cases": {
"type": "integer"
},
"hardware": {
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"id",
"label",
"cases",
"hardware"
],
"additionalProperties": false
}
},
"models": {
"type": "array",
"items": {
"type": "string"
}
},
"parallel": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"cases": {
"type": "integer"
}
},
"required": [
"key",
"label",
"cases"
],
"additionalProperties": false
}
},
"hardware": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "string"
},
"runner": {
"type": "string"
},
"runs": {
"type": "array",
"items": {
"type": "object",
"properties": {
"runId": {
"type": "string"
},
"generatedAt": {
"type": "string"
}
},
"required": [
"runId",
"generatedAt"
],
"additionalProperties": false
}
}
},
"required": [
"id",
"runner",
"runs"
],
"additionalProperties": false
}
}
},
"required": [
"ops",
"workloads",
"models",
"parallel",
"hardware"
],
"additionalProperties": false
},
"url": {
"type": "string"
},
"total": {
"type": "integer"
},
"pageSize": {
"type": "integer"
},
"rows": {
"type": "array",
"items": {
"type": "object",
"properties": {
"case_key": {
"type": "string"
},
"op_type": {
"type": "string",
"enum": [
"gemm",
"moe",
"mla",
"mla_dsa",
"dsv4_attn",
"gqa",
"qsa",
"gdn",
"kda"
]
},
"parallel": {
"type": "string"
},
"testlist": {
"type": "string"
},
"sources": {
"type": "array",
"items": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"roles": {
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"model",
"roles"
],
"additionalProperties": false
}
},
"shape": {
"type": "string"
},
"precision": {
"type": "string"
},
"compute_precision": {
"type": "string",
"enum": [
"fp4",
"fp8",
"bf16",
"other"
]
},
"x": {
"type": [
"number",
"null"
]
},
"dims": {
"type": "object",
"additionalProperties": {
"type": "number"
}
},
"flops": {
"type": [
"number",
"null"
]
},
"bytes": {
"type": [
"number",
"null"
]
},
"hardware": {
"type": "string"
},
"status": {
"type": "string",
"enum": [
"ok",
"unsupported",
"error",
"missing"
]
},
"latency_us": {
"type": [
"number",
"null"
]
},
"tflops": {
"type": [
"number",
"null"
]
},
"tb_s": {
"type": [
"number",
"null"
]
},
"roofline_us": {
"type": [
"number",
"null"
]
},
"roofline_share": {
"type": [
"number",
"null"
]
},
"kernel": {
"type": [
"string",
"null"
]
},
"cuda_graph": {
"type": [
"boolean",
"null"
]
},
"run_id": {
"type": [
"string",
"null"
]
},
"result_index": {
"type": [
"integer",
"null"
]
},
"revision": {
"type": [
"string",
"null"
]
}
},
"required": [
"case_key",
"op_type",
"parallel",
"testlist",
"sources",
"shape",
"precision",
"compute_precision",
"x",
"dims",
"flops",
"bytes",
"hardware",
"status",
"latency_us",
"tflops",
"tb_s",
"roofline_us",
"roofline_share",
"kernel",
"cuda_graph",
"run_id",
"result_index",
"revision"
],
"additionalProperties": false
}
}
},
"required": [
"view",
"apiVersion",
"params",
"options",
"url",
"total",
"pageSize",
"rows"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "operatorx",
"apiVersion": "v1",
"params": {
"op": "attention",
"workload": null,
"model": "DeepSeek-V4-Pro",
"parallel": "1",
"hardware": [
"b200",
"h200"
],
"status": "ok",
"page": 0,
"format": "json"
},
"options": {
"ops": [
"gemm",
"moe",
"attention"
],
"workloads": [
{
"id": "attention:DeepSeek-V4-Pro",
"label": "DeepSeek-V4-Pro attention",
"cases": 99,
"hardware": [
"b200",
"h200"
]
}
],
"models": [
"DeepSeek-V4-Pro"
],
"parallel": [
{
"key": "1",
"label": "1 GPU",
"cases": 99
}
],
"hardware": [
{
"id": "b200",
"runner": "b200-dgxc",
"runs": [
{
"runId": "36372535030",
"generatedAt": "2026-09-27T12:00:00Z"
}
]
}
]
},
"url": "/opx/model?model=DeepSeek-V4-Pro",
"total": 1,
"pageSize": 100,
"rows": [
{
"case_key": "{\"batch\":{\"groups\":[{\"count\":1,\"ctx\":1024,\"q\":1}]},...}|vllm",
"op_type": "dsv4_attn",
"parallel": "1",
"testlist": "attn_deepseek_v4",
"sources": [
{
"model": "DeepSeek-V4-Pro",
"roles": [
"attn"
]
}
],
"shape": "V4 attention ÷4 · 1×q1 ctx1024",
"precision": "proj e4m3 (128×128 ue8m0) · kv fp8",
"compute_precision": "fp8",
"x": 1,
"dims": {
"tokens": 1,
"hidden": 7168,
"heads": 128,
"compressRatio": 4
},
"flops": 1200000000,
"bytes": 110000000,
"hardware": "b200",
"status": "ok",
"latency_us": 41.2,
"tflops": 29.1,
"tb_s": 2.67,
"roofline_us": 18.3,
"roofline_share": 0.44,
"kernel": "flash_mla_sparse_fwd",
"cuda_graph": true,
"run_id": "36372535030",
"result_index": 66,
"revision": "63"
}
]
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
case_key,op_type,parallel,testlist,sources,shape,precision,compute_precision,x,dims,flops,bytes,hardware,status,latency_us,tflops,tb_s
...,dsv4_attn,1,attn_deepseek_v4,DeepSeek-V4-Pro: attn,V4 attention ÷4 · 1×q1 ctx1024,proj e4m3 (128×128 ue8m0) · kv fp8,fp8,1,tokens=1 hidden=7168 heads=128 compressRatio=4,1200000000,110000000,b200,ok,41.2,29.1,2.67400
参数未知、重复或超出范围。响应体会在可能时列出允许的取值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown parallel: tp4",
"param": "parallel",
"allowed": [
"1",
"tp8"
]
}503
无法读取 OperatorX 数据库。不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}500
视图组装失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/first-token首 token 达标配置
只读首 token 达标配置，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| chipBasis | query | string | 可选 | pool（默认）或 total，指定分离式配置的输入、输出吞吐量及其派生指标所除的芯片数：pool 下输入按预填充芯片、输出按解码芯片计；total 按部署中的芯片总数计，与聚合配置及 tput_per_gpu 口径一致。聚合配置在两种口径下结果相同。 | |
| caps | query | string | 可选 | 一至八个有限正数，表示首 token 时间上限（秒）；排序去重后使用，默认 2,5,10,15,20。任一值无效时返回 400。 | |
| costProvider | query | string | 可选 | costh 为自有成本，costr 为租赁成本。利润视图还支持 custom，需配合 customCosts。不支持 costn。 | |
| costType | query | string | 可选 | total（默认）、input 或 output token。分离式输入和输出速率保留各自的 prefill、decode 分母。 | |
| date | query | string | 可选 | 快照截止日期，格式为 YYYY-MM-DD。指定 runId 时改为读取该次运行的逻辑快照。 | |
| gpus | query | string | 可选 | 以逗号分隔的硬件或配置键；gpu-metrics 使用数字 GPU 索引。省略时选择全部。 | |
| minInteractivity | query | string | 可选 | 最低 tok/s/user，AgentX 默认 150，固定长度工作负载默认 35。 | |
| model | query | string | 必填 | 模型显示名称或比较页 slug，不区分大小写。基于基准测试的视图需要此参数。镜像视图还接受当前镜像列表中的模型名，去除首尾空白后匹配，默认 all；未知模型返回 400。 | DeepSeek-V4-Pro |
| percentile | query | string | 可选 | AgentX 延迟分位数为 p75 或 p90；CollectiveX 支持 p50、p95、p99。 | |
| precisions | query | string | 可选 | 以逗号分隔的精度键；省略时按可用测试曲线数量选择。计算器扩展视图选择官方数据中曲线最多的精度，并纳入 unofficial-run 叠加数据中的精度。 | |
| runId | query | string | 可选 | 正安全整数格式的工作流运行 ID。不填时选择默认运行；GPU 指标必须填写。 | 123 |
| sequence | query | string | 可选 | 工作负载：1k/1k、1k/8k、8k/1k 或 agentic-traces。AgentX 扩展视图默认使用 agentic-traces。 | |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 | |
| unofficialrun | query | string | 可选 | 最多八个公开 CI 运行 ID，以逗号分隔。去除各项首尾空白和重复 ID 后，按剩余顺序分配运行索引。叠加结果与官方数据分开。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/first-token?model=DeepSeek-V4-Pro'响应
FirstTokenView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"first-token"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"data": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "first-token",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/cache-reuse缓存复用曲线
只读缓存复用曲线，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| config | query | string | 可选 | configurations 中的缓存复用配置键。省略时采用仪表板默认配置。 | |
| date | query | string | 可选 | 快照截止日期，格式为 YYYY-MM-DD。指定 runId 时改为读取该次运行的逻辑快照。 | |
| gpus | query | string | 可选 | 以逗号分隔的硬件或配置键；gpu-metrics 使用数字 GPU 索引。省略时选择全部。 | |
| model | query | string | 必填 | 模型显示名称或比较页 slug，不区分大小写。基于基准测试的视图需要此参数。镜像视图还接受当前镜像列表中的模型名，去除首尾空白后匹配，默认 all；未知模型返回 400。 | DeepSeek-V4-Pro |
| percentile | query | string | 可选 | AgentX 延迟分位数为 p75 或 p90；CollectiveX 支持 p50、p95、p99。 | |
| precisions | query | string | 可选 | 以逗号分隔的精度键；省略时按可用测试曲线数量选择。计算器扩展视图选择官方数据中曲线最多的精度，并纳入 unofficial-run 叠加数据中的精度。 | |
| recipe | query | string | 可选 | data.recipes 中的方案键，对应仪表板 c_recipe。省略或未知键按缓存层级数据行数、总行数和标签依次选择默认方案；仅在没有官方数据时使用叠加运行的方案。params.recipe 返回实际选择的键。 | |
| runId | query | string | 可选 | 正安全整数格式的工作流运行 ID。不填时选择默认运行；GPU 指标必须填写。 | 123 |
| sequence | query | string | 可选 | 工作负载：1k/1k、1k/8k、8k/1k 或 agentic-traces。AgentX 扩展视图默认使用 agentic-traces。 | |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 | |
| unofficialrun | query | string | 可选 | 最多八个公开 CI 运行 ID，以逗号分隔。去除各项首尾空白和重复 ID 后，按剩余顺序分配运行索引。叠加结果与官方数据分开。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/cache-reuse?model=DeepSeek-V4-Pro'响应
CacheReuseView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"cache-reuse"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"configurations": {
"type": "object",
"additionalProperties": true
},
"data": {
"type": [
"object",
"null"
],
"additionalProperties": true,
"properties": {
"recipe": {
"type": [
"string",
"null"
]
},
"recipes": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"label": {
"type": "string"
},
"rows": {
"type": "integer"
},
"tiered": {
"type": "integer"
}
},
"required": [
"key",
"label",
"rows",
"tiered"
]
}
}
}
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "cache-reuse",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/profit-estimator每芯片小时利润
只读每芯片小时利润，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| cacheHitMode | query | string | 可选 | actual 或 theoretical（默认，界面显示为“理论上限”）。Actual 使用基准测试中服务端实测的 cache 命中率；theoretical 使用 trace 计算的无限缓存理论命中率，缺失时回退到实测值。 | |
| cachedInputPrice | query | string | 可选 | 自定义缓存输入价格，单位 USD/百万 token，默认 0.1。 | |
| costProvider | query | string | 可选 | costh 为自有成本，costr 为租赁成本。利润视图还支持 custom，需配合 customCosts。不支持 costn。 | |
| customCosts | query | string | 可选 | JSON 对象，将基础硬件键映射为有限非负美元/芯片小时，最多 100 项。 | |
| date | query | string | 可选 | 快照截止日期，格式为 YYYY-MM-DD。指定 runId 时改为读取该次运行的逻辑快照。 | |
| dates | query | string | 可选 | 最多十二个比较项，以逗号分隔，格式为 YYYY-MM-DD 或 YYYY-MM-DD~rRUN_ID。仅含日期时读取当天的逻辑快照，不按截止日期向前回溯；含运行 ID 时读取该次运行的逻辑快照。各快照独立计算。 | |
| end | query | string | 可选 | 比较范围结束日期 YYYY-MM-DD，不得早于 start。历史视图将其用作含终点的数据范围。 | |
| gpus | query | string | 可选 | 以逗号分隔的硬件或配置键；gpu-metrics 使用数字 GPU 索引。省略时选择全部。 | |
| inputPrice | query | string | 可选 | 自定义未缓存输入价格，单位 USD/百万 token，默认 1。 | |
| labCut | query | string | 可选 | 模型许可或收入分成百分比，范围 0 至 100，默认值随模型变化。 | |
| model | query | string | 必填 | 模型显示名称或比较页 slug。仪表板提供具有可用 AgentX 测试曲线和 token 定价的模型；本 API 也接受其他已知模型，但可能返回空的估算结果。可通过 /api/v1/availability 查询 agentic_traces 数据。 | DeepSeek-V4-Pro |
| outputPrice | query | string | 可选 | 自定义输出价格，单位 USD/百万 token，默认 1。 | |
| percentile | query | string | 可选 | AgentX 延迟分位数为 p75 或 p90；CollectiveX 支持 p50、p95、p99。 | |
| powerBasis | query | string | 可选 | provisioned（默认，整体预配功耗）、modeled（整体实测功耗）或 compare。按 GW 年估算时，modeled 先筛选满足系统功耗要求的数据点，再在同一目标值上构建曲线，不外推，也不借用历史数据。compare 的两种功耗方案使用同一条有效曲线的吞吐量；没有有效曲线覆盖目标时，保留原曲线的预配估算。provisioned 单独使用时仍沿用原性能曲线。整体实测功耗在实测输入上叠加组件估算和 PUE，并非墙上电表读数。NVL72 需要通过验证的 GPU 遥测，每个计算 tray 平均分摊建模液冷机架的功耗；带有完整且通过验证的 Grace socket 遥测时使用实测 Grace 功耗（不接受仅有 CPU rail 或模块传感器的读数），否则使用 Grace 功耗估算值。插值的两个数据点须采用相同功耗口径。powerSource 记录拓扑（chassis 或 nvl72-tray）、cpu（measured 或 modeled）、PUE 和固定的功耗模型版本；powerLabel 标明配对估算和整机外推，skipped 保留覆盖缺失原因。 | |
| precisions | query | string | 可选 | 以逗号分隔的精度键；省略时按可用测试曲线数量选择。计算器扩展视图选择官方数据中曲线最多的精度，并纳入 unofficial-run 叠加数据中的精度。 | |
| priceSource | query | string | 可选 | 利润视图支持 list、openrouter、custom；推理和历史视图支持 normalized、openrouter。响应 pricing 记录实际使用的价格。 | |
| runId | query | string | 可选 | 正安全整数格式的工作流运行 ID。不填时选择默认运行；GPU 指标必须填写。 | 123 |
| sequence | query | string | 可选 | 工作负载：1k/1k、1k/8k、8k/1k 或 agentic-traces。AgentX 扩展视图默认使用 agentic-traces。 | |
| start | query | string | 可选 | 比较范围起始日期 YYYY-MM-DD。与 end 一起只加入两个端点，不加入中间所有日期。历史视图将其用作含起点的数据范围。 | |
| target | query | string | 可选 | 正数工作点目标；交互性转吞吐量模式下单位为 tok/s/user。两个利润视图按模型设置默认值：Kimi K3 为 60、GLM-5.2/5.3 为 135、MiniMax M3 为 115、DeepSeek V4 Pro 为 170、DeepSeek V4.1 Flash 为 150 tok/s/user。仪表板链接通过 /<model>/target/<value> 路径片段传递目标值；本 API 仍使用 target 查询参数。这些默认值是用于扩大 SKU 覆盖范围的可修改假设，并非与 OpenRouter P90 等价的实测值。 | |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 | |
| unofficialrun | query | string | 可选 | 最多八个公开 CI 运行 ID，以逗号分隔。去除各项首尾空白和重复 ID 后，按剩余顺序分配运行索引。叠加结果与官方数据分开。 | |
| utilization | query | string | 可选 | 利用率百分比，范围 0 至 100，默认 60。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/profit-estimator?model=DeepSeek-V4-Pro'响应
ProfitEstimatorView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"profit-estimator"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"pricing": {
"type": "object",
"additionalProperties": true
},
"data": {
"type": "object",
"additionalProperties": true
},
"overlays": {
"type": "object",
"additionalProperties": true
},
"comparisons": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "profit-estimator",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/profit-estimator-per-gigawatt每吉瓦年利润
只读每吉瓦年利润，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| cacheHitMode | query | string | 可选 | actual 或 theoretical（默认，界面显示为“理论上限”）。Actual 使用基准测试中服务端实测的 cache 命中率；theoretical 使用 trace 计算的无限缓存理论命中率，缺失时回退到实测值。 | |
| cachedInputPrice | query | string | 可选 | 自定义缓存输入价格，单位 USD/百万 token，默认 0.1。 | |
| costProvider | query | string | 可选 | costh 为自有成本，costr 为租赁成本。利润视图还支持 custom，需配合 customCosts。不支持 costn。 | |
| customCosts | query | string | 可选 | JSON 对象，将基础硬件键映射为有限非负美元/芯片小时，最多 100 项。 | |
| date | query | string | 可选 | 快照截止日期，格式为 YYYY-MM-DD。指定 runId 时改为读取该次运行的逻辑快照。 | |
| dates | query | string | 可选 | 最多十二个比较项，以逗号分隔，格式为 YYYY-MM-DD 或 YYYY-MM-DD~rRUN_ID。仅含日期时读取当天的逻辑快照，不按截止日期向前回溯；含运行 ID 时读取该次运行的逻辑快照。各快照独立计算。 | |
| end | query | string | 可选 | 比较范围结束日期 YYYY-MM-DD，不得早于 start。历史视图将其用作含终点的数据范围。 | |
| gpus | query | string | 可选 | 以逗号分隔的硬件或配置键；gpu-metrics 使用数字 GPU 索引。省略时选择全部。 | |
| inputPrice | query | string | 可选 | 自定义未缓存输入价格，单位 USD/百万 token，默认 1。 | |
| labCut | query | string | 可选 | 模型许可或收入分成百分比，范围 0 至 100，默认值随模型变化。 | |
| model | query | string | 必填 | 模型显示名称或比较页 slug。仪表板提供具有可用 AgentX 测试曲线和 token 定价的模型；本 API 也接受其他已知模型，但可能返回空的估算结果。可通过 /api/v1/availability 查询 agentic_traces 数据。 | DeepSeek-V4-Pro |
| outputPrice | query | string | 可选 | 自定义输出价格，单位 USD/百万 token，默认 1。 | |
| percentile | query | string | 可选 | AgentX 延迟分位数为 p75 或 p90；CollectiveX 支持 p50、p95、p99。 | |
| powerBasis | query | string | 可选 | provisioned（默认，整体预配功耗）、modeled（整体实测功耗）或 compare。按 GW 年估算时，modeled 先筛选满足系统功耗要求的数据点，再在同一目标值上构建曲线，不外推，也不借用历史数据。compare 的两种功耗方案使用同一条有效曲线的吞吐量；没有有效曲线覆盖目标时，保留原曲线的预配估算。provisioned 单独使用时仍沿用原性能曲线。整体实测功耗在实测输入上叠加组件估算和 PUE，并非墙上电表读数。NVL72 需要通过验证的 GPU 遥测，每个计算 tray 平均分摊建模液冷机架的功耗；带有完整且通过验证的 Grace socket 遥测时使用实测 Grace 功耗（不接受仅有 CPU rail 或模块传感器的读数），否则使用 Grace 功耗估算值。插值的两个数据点须采用相同功耗口径。powerSource 记录拓扑（chassis 或 nvl72-tray）、cpu（measured 或 modeled）、PUE 和固定的功耗模型版本；powerLabel 标明配对估算和整机外推，skipped 保留覆盖缺失原因。 | |
| precisions | query | string | 可选 | 以逗号分隔的精度键；省略时按可用测试曲线数量选择。计算器扩展视图选择官方数据中曲线最多的精度，并纳入 unofficial-run 叠加数据中的精度。 | |
| priceSource | query | string | 可选 | 利润视图支持 list、openrouter、custom；推理和历史视图支持 normalized、openrouter。响应 pricing 记录实际使用的价格。 | |
| runId | query | string | 可选 | 正安全整数格式的工作流运行 ID。不填时选择默认运行；GPU 指标必须填写。 | 123 |
| sequence | query | string | 可选 | 工作负载：1k/1k、1k/8k、8k/1k 或 agentic-traces。AgentX 扩展视图默认使用 agentic-traces。 | |
| start | query | string | 可选 | 比较范围起始日期 YYYY-MM-DD。与 end 一起只加入两个端点，不加入中间所有日期。历史视图将其用作含起点的数据范围。 | |
| target | query | string | 可选 | 正数工作点目标；交互性转吞吐量模式下单位为 tok/s/user。两个利润视图按模型设置默认值：Kimi K3 为 60、GLM-5.2/5.3 为 135、MiniMax M3 为 115、DeepSeek V4 Pro 为 170、DeepSeek V4.1 Flash 为 150 tok/s/user。仪表板链接通过 /<model>/target/<value> 路径片段传递目标值；本 API 仍使用 target 查询参数。这些默认值是用于扩大 SKU 覆盖范围的可修改假设，并非与 OpenRouter P90 等价的实测值。 | |
| tcoBasis | query | string | 可选 | internal（默认）或 external，使用与界面相同的自有或租赁成本口径。 | |
| unofficialrun | query | string | 可选 | 最多八个公开 CI 运行 ID，以逗号分隔。去除各项首尾空白和重复 ID 后，按剩余顺序分配运行索引。叠加结果与官方数据分开。 | |
| utilization | query | string | 可选 | 利用率百分比，范围 0 至 100，默认 60。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/profit-estimator-per-gigawatt?model=DeepSeek-V4-Pro'响应
ProfitEstimatorPerGigawattView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"profit-estimator-per-gigawatt"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"pricing": {
"type": "object",
"additionalProperties": true
},
"data": {
"type": "object",
"additionalProperties": true
},
"overlays": {
"type": "object",
"additionalProperties": true
},
"comparisons": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "profit-estimator-per-gigawatt",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/collectivexCollectiveX EP、KV 与 swap 视图
只读CollectiveX EP、KV 与 swap 视图，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| activeSeries | query | string | 可选 | 以逗号分隔、带运行命名空间的 EP 曲线 ID。省略时选择全部。 | |
| kvSeries | query | string | 可选 | 以逗号分隔的 KV run:case ID。省略时选择全部。 | |
| swapSeries | query | string | 可选 | 以逗号分隔的 swap 曲线 ID。省略时选择全部。 | |
| backend | query | string | 可选 | 后端筛选；省略时选择全部。 | |
| epSize | query | string | 可选 | EP 大小，有 8 时默认选 8，否则选首个可用值。 | |
| kvOp | query | string | 可选 | KV 操作：push（默认）或 pull。 | |
| kvX | query | string | 可选 | KV 展示方式：isl（默认）、batch、frontier 或 overlap。 | |
| kvY | query | string | 可选 | KV 指标：latency（默认）或 bandwidth。 | |
| modes | query | string | 可选 | 以逗号分隔的 EP 模式键，默认选择全部可用模式。 | |
| operation | query | string | 可选 | EP 操作：pair_period（运行包含该指标时为默认，否则为 roundtrip；连续成对的稳态周期）、roundtrip（单次）、dispatch 或 combine。 | |
| overlapIsl | query | string | 可选 | 重叠测试输入长度：auto（默认，取批大小扫描最宽的 ISL）或正整数。 | |
| pageTokens | query | string | 可选 | 已弃用，将被忽略：KV 数据按整行缓存搬运，不再区分页大小。 | |
| percentile | query | string | 可选 | AgentX 延迟分位数为 p75 或 p90；CollectiveX 支持 p50、p95、p99。 | |
| phase | query | string | 可选 | CollectiveX 阶段，默认 decode；视频阶段为 measurement、startup 或 warmup。 | |
| precision | query | string | 可选 | 精度键；CollectiveX 有 fp8 时默认使用 fp8。镜像视图去除首尾空白、不区分大小写，默认 all。 | |
| runs | query | string | 可选 | 按顺序排列的运行 ID，以逗号分隔，最多八个。省略时选择最新有实测数据的运行；空值表示不选。 | |
| sku | query | string | 可选 | CollectiveX 芯片 SKU 筛选，默认 all。 | |
| suite | query | string | 可选 | 运行列表筛选：all、ep、kv 或 swap。 | |
| swapDirection | query | string | 可选 | h2d（默认）、d2h 或 d2d。 | |
| swapLayout | query | string | 可选 | contiguous（默认）或 random。 | |
| swapMetric | query | string | 可选 | bandwidth（默认）或 latency。 | |
| swapPercentile | query | string | 可选 | p50（默认）、p95 或 p99。 | |
| version | query | string | 可选 | CollectiveX 数据集 schema 版本，默认 1。 | |
| yAxis | query | string | 可选 | CollectiveX 支持 latency、tokens-per-second、activation-rate、payload-rate；视频支持 dollar、clipsGpu、secondsGpu、clipsAllocatedGpu、secondsAllocatedGpu、energy。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/collectivex'响应
CollectivexView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"collectivex"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"runs": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"ep": {
"type": "object",
"additionalProperties": true
},
"kv": {
"type": "object",
"additionalProperties": true
},
"swap": {
"type": "object",
"additionalProperties": true
},
"options": {
"type": "object",
"additionalProperties": true
},
"knownSupport": {
"type": "object",
"additionalProperties": true
},
"knownSupportFootnotes": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "collectivex",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/submissions提交表格与每周数量
只读提交表格与每周数量，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| direction | query | string | 可选 | asc 或 desc。提交记录默认 desc，GPU 统计默认 asc。 | |
| limit | query | string | 可选 | 提交表格行数上限，默认 100，最大 10000。 | |
| lines | query | string | 可选 | 提交图曲线，以逗号分隔，可选 nvidia、amd、openai、google、huawei、unknown 和 total。默认与界面一致，包含所有有提交记录的厂商及 total。 | |
| mode | query | string | 可选 | 提交图支持 weekly（默认）、cumulative；计算器支持 interactivity_to_throughput、throughput_to_interactivity。 | |
| offset | query | string | 可选 | 提交表格偏移量，默认 0。 | |
| onChangeOnly | query | string | 可选 | 布尔值，默认 true。仅周图按仪表板规则应用按变更运行的统计起点。 | |
| search | query | string | 可选 | 提交记录搜索，不区分大小写，覆盖芯片、模型、框架、精度、投机解码及厂商。 | |
| sort | query | string | 可选 | 表格排序列；提交记录默认 date，GPU 统计默认 gpuIndex。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/submissions'响应
SubmissionsView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"submissions"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"rows": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"stats": {
"type": "object",
"additionalProperties": true
},
"series": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "submissions",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/current-inferencex-image当前推理镜像
只读当前推理镜像，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。 ageSummary 按最近秩法计算当前筛选行在 asOf 的提交年龄 P90（排除未来日期行）。includeHistory=true 基于同一批配置的带镜像成功提交，重建 90 个每日 ageHistory 点，因此最后一个点与 ageSummary 一致。当前表格中没有的配置（例如已被更新 run 取代的投机解码或节点类型变体）不计入。每个展示配置计一次；即使镜像标签未变，重新提交仍会重置年龄。两项指标均不代表镜像构建年龄。镜像标签中的版本号低于对应基础引擎（vLLM 或 SGLang）的最新 release 时，该行标记为 outdated。nightly 镜像的 commit 若能在 GitHub 上解析，则仅当该 commit 是此 release 的祖先提交时才标记；其余 nightly 镜像比较构建日期与 release 发布日期，构建日期取标签中的日期，标签无日期时取该构建在镜像目录中最早的提交日期。其他标签以及没有 release 来源的引擎不作标记。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| asOf | query | string | 可选 | 用于复现滚动可靠性或镜像日期差计算的参考日期 YYYY-MM-DD。 | |
| includeHistory | query | string | 可选 | true 或 false（默认）。返回截至 asOf 的 90 个 UTC 日期点，包含最近秩法计算的 p90AgeDays 和 configurationCount。统计的配置与 ageSummary 相同，即当前筛选结果中日期不晚于 asOf 的行，因此最后一个点与 ageSummary 一致。每个配置的历史由其带镜像的成功提交重建，截至当前行的日期，并保留窗口开始前的最近一次记录。无数据时 p90 为 null，不用未来记录向前回填。所有镜像筛选条件均适用。此指标是提交年龄，不是镜像构建年龄。 | |
| frameworks | query | string | 可选 | 以逗号分隔的框架系列键；省略时选择全部。 | |
| hardware | query | string | 可选 | 镜像视图硬件键，去除首尾空白、不区分大小写，默认 all。 | |
| model | query | string | 可选 | 模型显示名称或比较页 slug，不区分大小写。基于基准测试的视图需要此参数。镜像视图还接受当前镜像列表中的模型名，去除首尾空白后匹配，默认 all；未知模型返回 400。 | DeepSeek-V4-Pro |
| nodeType | query | string | 可选 | 镜像部署类型：all（默认）、single 或 disagg。 | |
| precision | query | string | 可选 | 精度键；CollectiveX 有 fp8 时默认使用 fp8。镜像视图去除首尾空白、不区分大小写，默认 all。 | |
| sequence | query | string | 可选 | 镜像场景。默认为 agentic-traces（Agentic）；所选模型无此场景时，回退到其首个有效场景。 | |
| spec | query | string | 可选 | 投机解码筛选；镜像视图使用单个值，去除首尾空白、不区分大小写，默认 all。推理视图使用逗号分隔列表。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/current-inferencex-image'响应
CurrentInferencexImageView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"current-inferencex-image"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"rows": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"options": {
"type": "object",
"additionalProperties": true
},
"releases": {
"type": "object",
"additionalProperties": true
},
"ageSummary": {
"type": "object",
"properties": {
"date": {
"type": "string",
"format": "date"
},
"p90AgeDays": {
"type": [
"integer",
"null"
],
"minimum": 0
},
"configurationCount": {
"type": "integer",
"minimum": 0
}
},
"required": [
"date",
"p90AgeDays",
"configurationCount"
],
"additionalProperties": false
},
"ageHistory": {
"type": "array",
"items": {
"type": "object",
"properties": {
"date": {
"type": "string",
"format": "date"
},
"p90AgeDays": {
"type": [
"integer",
"null"
],
"minimum": 0
},
"configurationCount": {
"type": "integer",
"minimum": 0
}
},
"required": [
"date",
"p90AgeDays",
"configurationCount"
],
"additionalProperties": false
}
},
"ageMethodology": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "current-inferencex-image",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/gpu-metricsGPU 遥测与全记录统计
只读GPU 遥测与全记录统计，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。 每个 run 只从一个来源读取遥测：已入库的 run 读取数据库中存储的序列，其余 run 读取 GitHub 产物。数据库 coverage 描述整个 run：complete 表示记录的产物与文件清单均已入库，partial 表示部分已记录的遥测缺失，unknown 表示未保留完整清单；missingArtifacts 仅列出整体缺失的产物；若只是文件不完整，此列表可能为空。GitHub 数据的 coverage 为 null，表示尚未评估。响应使用 private, no-store；上游数据库故障保留 HTTP 503，不作为空数据返回。各文件、主机的序列身份独立保留。原始数据行和图表按所选 GPU 索引筛选。统计则覆盖所选文件或主机序列的全部 GPU，包含启动阶段和 warmup，每次读取时由该序列的样本计算；若已存序列读取到的样本数与记录不符，则不返回统计。均值按样本加权，P50/P95/P99 在 p*(N-1) 处线性插值，标准差以 N 为分母。这些统计不是 serving-window 功率或 J/token。折线时间以所有 GPU 的首个采样为起点；缺失指标读数会被跳过，不补零。相关性图要求两个指标均有读数。本接口投影原始 GPU-metrics 浏览器，不是推理页的 Power Timeline。时间线通过现有 series=power 读取和基准测试 power_audit 获取角色与窗口证据；i_ptaxis、i_ptlines、i_ptwindow、i_ptfocus、i_ptutility 和 i_ptconc 分享字段仅控制客户端显示，本接口不接受这些参数。它们不改变原始遥测或全记录统计。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| artifact | query | string | 可选 | GPU 指标使用 artifacts[] 返回的产物名称；一个产物含多个文件时，格式为 <artifact>/<file>。视频使用数字产物 ID，需配合 run。 | |
| chartView | query | string | 可选 | chart（默认）或 correlation，两者均返回未降采样来源行。 | |
| corrXMetric | query | string | 可选 | 相关性图 x 轴 GPU 指标，默认 power。 | |
| corrYMetric | query | string | 可选 | 相关性图 y 轴 GPU 指标。默认 temperature；若该产物未采集温度，则改用 corrXMetric 以外第一个已采集的指标，没有此类指标时相关性数据为空。显式指定的指标即使未采集也照常使用，此时相关性数据为空。 | |
| direction | query | string | 可选 | asc 或 desc。提交记录默认 desc，GPU 统计默认 asc。 | |
| downsample | query | string | 可选 | 布尔值，默认 true；将渲染提示 maxInteractivePoints 设为 2000，false 时为 null。返回的原始行和统计量不降采样。 | |
| gpus | query | string | 可选 | 以逗号分隔的硬件或配置键；gpu-metrics 使用数字 GPU 索引。省略时选择全部。 | |
| metric | query | string | 可选 | 各视图对应的指标键。GPU 指标支持功耗、温度、时钟、利用率及可用 AMD 指标。 | |
| runId | query | string | 必填 | 正安全整数格式的工作流运行 ID。不填时选择默认运行；GPU 指标必须填写。 | 123 |
| sort | query | string | 可选 | 表格排序列；提交记录默认 date，GPU 统计默认 gpuIndex。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/gpu-metrics?runId=123'响应
GpuMetricsView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"gpu-metrics"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"runInfo": {
"type": "object",
"additionalProperties": true
},
"artifacts": {
"type": "array",
"items": {
"type": "string"
}
},
"coverage": {
"type": [
"object",
"null"
],
"properties": {
"status": {
"type": "string",
"enum": [
"complete",
"partial",
"unknown"
]
},
"missingArtifacts": {
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"status",
"missingArtifacts"
],
"additionalProperties": false
},
"rows": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"stats": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
},
"description": "Per-GPU statistics for the selected file/host series, including startup and warmup. For stored runs they are computed on each read from the stored samples: a missing reading stays missing, and a series whose stored samples do not match its recorded sample count returns no statistics. Count is the finite-reading count after first-wins timestamp/GPU deduplication. Mean is sample-weighted, P50/P95/P99 use linear interpolation at p*(N-1), and standard deviation divides by N. Values use the selected metric unit; rows carry no metric field, since params.metric identifies it. GPU visibility and chart downsampling do not change this population. These are not serving-window power or J/token."
},
"rendering": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "gpu-metrics",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/videoVideoGenX 仪表板与已发布对比证据
view=overview 与 /vgx 结果页共用 resultProjection，只准入同一份计划 20 条视频的 C1 工作负载，各硬件统一至少 10 个有效样本。响应包含实测指标及分别标注的参与/已分配 GPU 成本和估算利润；分配证据缺失时保留 null。point.videoSecondsPerHour 表示有效测量请求每小时生成的实际解码视频秒数，排除 warmup；视频秒吞吐量、两种 GPU 口径的每视频秒成本及估算收入共用此输出速率。有效输出时长不完整或旧 History 缺少此可选字段时，这些指标及利润均为 null；durationSeconds 保留计划请求时长，仅用于工作负载身份与筛选，不作替代。默认数据点须有发布批准。unofficialrun 只把指定运行放入 overlayCells（所选 GPU 数）和 overlayDeploymentCells（所有匹配的 GPU 数），并将其排除在正式结果、部署和工作负载构成之外。已存储的运行保留实际批准状态；不在 History 中的运行从其公开的 h3-results 产物实时读取文档，视为未批准，且不附带媒体；其视频仍会附上同一运行尝试的可信 h3-quality 产物中的 PSNR 与未评分原因。默认优先选择有已批准结果的工作负载；仅由指定运行提供的工作负载须明确选择，或在没有匹配的已批准结果时作为回退。清除叠加运行后也不会把未批准数据转为正式结果。移除时，工作负载选择仅在已批准选项中仍然有效时保留。overview 默认返回 JSON；format=csv 导出当前筛选的正式及指定运行数据点，包含来源标识、单位和两种 GPU 成本口径。duration_s 保留计划请求时长；紧邻 clips_per_s 的 video_seconds_per_s 为每测量秒生成的有效解码视频秒数。CSV 响应设置 Cache-Control: private, no-store；缺失值以空单元格表示。质量、阶段计时、内存和启动证据缺失时保留 null。播放、拖动、差异显示、主题和展开控件只影响展示；媒体使用只读已发布路径。只读VideoGenX 仪表板与已发布对比证据，使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。 默认 dashboard 模式复用已发布历史 → videoPoints → dashboardCells → metrics/plot/compare/evidence。与界面一样最多读取前五页历史，返回覆盖范围和截断状态，选择一个 canonical workload 并应用相同样本门槛。v_hidden 只影响图表和表格。view=compare 额外读取最多两份已发布媒体产物，复用界面配对函数，仅返回相对媒体路径，不返回 asset URL；媒体缺失或读取失败会明确标记。run/artifact 或显式 discovery/results/tradeoff 保留原有契约。发现列表包括标准 run-sweep 运行，其中可能没有视频产物。单次运行的产物列表读取 GitHub 全部分页，兼容旧名称和原生矩阵的唯一名称。已分配 GPU 数须有保留的调度器分配证据，原生功率设备数不能替代；缺失的分配数量及按其归一化的指标保持 null。不代表日期区间、完整历史、功率/质量验收或瓶颈因果判断。不读取本地数据包或任意 URL。HTTP 响应仍为 private, no-store。已发布 History 和媒体元数据以 video 标签跨实例缓存，读取时不调用 GitHub；错误、索引或批准回执读取失败的 History 页和缺失媒体不保留，发布、History 补建或 ?scope=video 会立即使缓存失效。页面首屏数据和真实首帧预读仅影响展示，不改变选择项、指标或来源归属。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| r_model | query | string | 可选 | 完整模型标识；未指定时选择硬件覆盖最多的合格工作负载。 | |
| r_workload | query | string | 可选 | options.workloads 中的短 key，内部解析为完整 canonical workloadKey。同一工作负载任一数据点的 ID 也能解析，因此之后有新结果发布时，已分享的链接仍然有效。所有硬件共用计划 20 条视频的 C1 spec，每种硬件至少 10 个有效样本。 | |
| r_resolution | query | string | 可选 | 封印计划中的完整宽×高，例如 1344×768。 | |
| r_duration | query | string | 可选 | 计划中的请求视频时长，单位秒，必须为正数；不是解码输出时长。 | |
| r_mode | query | string | 可选 | 明确记录的生成模式；不从模型名称推断缺失模式。 | |
| r_steps | query | string | 可选 | 记录的去噪步数。 | |
| r_gpus | query | string | 可选 | 已弃用的概览控件参数，仍可传入但会被忽略。始终包含全部 GPU 数量，params.gpuCount 返回 null。 | |
| r_engine | query | string | 可选 | 记录的引擎名称，与 runtime revision 分开。 | |
| r_date | query | string | 可选 | 截至该执行日期查看结果，格式 YYYY-MM-DD；不以发布日期替代。每个正式部署显示其在该日期或之前测得的最新运行；指定运行叠加不受此筛选。runs 列出该工作负载的全部正式运行，run 为当前选中的运行。 | |
| r_run | query | string | 可选 | r_date 当天要查看截至哪一次运行的 GitHub Actions 运行 ID；默认为当天最后一次运行。 | |
| r_configs | query | string | 可选 | 以逗号分隔的芯片配置（chipConfigs 中的 hardware:GPU 数:tp:ulysses 键，最多 4 个）；history 返回这些配置的正式运行，用于对比历史趋势图。 | |
| r_from | query | string | 可选 | history 返回的运行的最早执行日期，格式 YYYY-MM-DD；默认不限。 | |
| r_to | query | string | 可选 | history 返回的运行的最晚执行日期，格式 YYYY-MM-DD；默认不限。 | |
| r_hidden | query | string | 可选 | 以逗号分隔的隐藏硬件键，作用于概览图表、表格、部署详情和视频。 | |
| r_deadline | query | string | 可选 | 交付 deadline，单位秒，默认 120。达标率为按时交付的有效视频占比；goodput 为按时交付的有效视频数除以测量窗口时长。 | |
| r_tier | query | string | 可选 | 硬件注册表 TCO 档位 h（自有，默认）或 r（租赁），同时返回两种 GPU 成本口径。 | |
| r_api | query | string | 可选 | 生成模式未记录时可指定的非负视频秒场景价格，单位美元；默认 null。 | |
| r_price_t2v | query | string | 可选 | 可选非负 T2V 视频秒场景价格，单位美元；默认 null。 | |
| r_price_i2v | query | string | 可选 | 可选非负 I2V 视频秒场景价格，单位美元；默认 null。 | |
| r_price_r2v | query | string | 可选 | 可选非负 R2V 视频秒场景价格，单位美元；默认 null。 | |
| r_utilization | query | string | 可选 | 估算利用率，范围 0–1，默认 1。收入随利用率变化，预留 GPU 每小时成本不变。 | |
| r_x | query | string | 可选 | 已弃用的概览控件参数，仍可传入但会被忽略。图表始终使用单条视频中位延迟，params.xMetric 返回 p50。 | |
| r_y | query | string | 可选 | 图表 Y 轴指标：videoSecondsPerDollar（默认）、videosPerGpuHour 或 kjPerClip。仅影响展示；无效值使用默认值。 | |
| r_basis | query | string | 可选 | 已弃用的概览控件参数，仍可传入但会被忽略。图表始终采用参与计算的 GPU 口径，params.costBasis 返回 participating。证据数据和 CSV 仍保留两种成本口径。 | |
| r_optimal | query | string | 可选 | Optimal Only：为 0 时同时绘制被支配的部署，其他值只绘制各硬件的前沿部署。仅影响展示。 | |
| r_point | query | string | 可选 | 打开的部署详情，格式为 runId:pointId。未知或已被筛除的键回退到概览页。仅影响展示。 | |
| unofficialrun | query | string | 可选 | 单个运行 ID，可用 unofficialrun 或分享链接的 unofficialruns 别名；多个 ID 或同时使用两个别名返回 400。所选运行只放入 overview 中独立的 overlayCells 和 overlayDeploymentCells，并从正式结果中排除。已存储的运行保留实际批准状态；不在前五页 History 中的运行，会从其公开的原生 h3-results 产物实时读取 JSON 文档，始终视为未批准，publishedAt 为 null，且不附带媒体。未知、无法读取或未达准入要求的运行不返回叠加结果。该参数不会触发存储或发布。 | |
| unofficialruns | query | string | 可选 | 单个运行 ID，可用 unofficialrun 或分享链接的 unofficialruns 别名；多个 ID 或同时使用两个别名返回 400。所选运行只放入 overview 中独立的 overlayCells 和 overlayDeploymentCells，并从正式结果中排除。已存储的运行保留实际批准状态；不在前五页 History 中的运行，会从其公开的原生 h3-results 产物实时读取 JSON 文档，始终视为未批准，publishedAt 为 null，且不附带媒体。未知、无法读取或未达准入要求的运行不返回叠加结果。该参数不会触发存储或发布。 | |
| artifact | query | string | 可选 | GPU 指标使用 artifacts[] 返回的产物名称；一个产物含多个文件时，格式为 <artifact>/<file>。视频使用数字产物 ID，需配合 run。 | |
| cell | query | string | 可选 | serving cell ID。未知 ID 与界面一致，回退到首个 cell；evidence.cell 记录解析后的选择。 | |
| compare | query | string | 可选 | 视频：最多八组已发布证据的 run:artifact，以逗号分隔。 | |
| costs | query | string | 可选 | JSON 对象，将视频数据点 ID 映射为 {hourly,source,date} 字符串；hourly 为非负 USD/deployment-hour。 | |
| format | query | string | 可选 | 默认返回 JSON。CSV 仅适用于 view=overview，按当前选择导出正式及指定运行的数据点，包含来源标识、两种 GPU 成本口径和单位。duration_s 为计划请求时长，video_seconds_per_s 为每测量秒生成的有效解码视频秒数。CSV 空单元格表示数据缺失。 | |
| gpuBasis | query | string | 可选 | 视频分母：participating（默认）或 allocated GPU。 | |
| page | query | string | 可选 | 视频 CI 发现页码从 1 开始（view=discovery）。视频 dashboard/compare 最多读取前五页已发布历史。 | |
| phase | query | string | 可选 | CollectiveX 阶段，默认 decode；视频阶段为 measurement、startup 或 warmup。 | |
| run | query | string | 可选 | 公开视频 CI 运行 ID；保留原有运行/产物选择方式。省略时读取仪表板；使用 view=discovery 列出运行。 | |
| selected | query | string | 可选 | 选中的视频权衡图数据点 ID。 | |
| slot | query | string | 可选 | 媒体或保真度 slot ID。未知 ID 回退到界面默认值；evidence.slot 记录解析后的选择。 | |
| source | query | string | 可选 | 已发布视频来源 ID，默认首个来源。 | |
| v_x | query | string | 可选 | 仪表板 X 轴指标：p90Latency（默认）或 p50Latency，单位秒。无效值按界面默认值处理。 | |
| v_y | query | string | 可选 | 仪表板 Y 轴指标：videosPerDollar（默认）、dollarsPerVideo、videosPerGpuHour 或 kjPerVideo。无效值按界面默认值处理。 | |
| v_tier | query | string | 可选 | TCO 成本档位：h（自有，默认）或 r（租赁），使用与仪表板相同的硬件成本表；无效值使用 h。 | |
| v_view | query | string | 可选 | chart（默认）或 table；仅影响展示，在 params.displayView 中返回，不改变数据集。 | |
| v_optimal | query | string | 可选 | 除值恰为 0 外，仅保留每种硬件的 Pareto 部署。作用于图表和表格；排队 cell 始终不进入这两者。 | |
| v_api | query | string | 可选 | 正数 API 视频秒参考价，单位 USD/video-second，四舍五入至四位小数；无效值回退至界面带日期的参考价。仅用于乘以视频时长，不影响 TCO 或图表效率。 | |
| v_hidden | query | string | 可选 | 以逗号分隔的隐藏硬件键，仅接受视频硬件列表中的键；默认空，排序去重，忽略未知键。只筛选图表和表格，不改变 KPI、Compare 或 Evidence。 | |
| v_base | query | string | 可选 | 基线硬件键；无效或无可用测量时回退到界面默认值（符合条件的代表部署中最慢的一种）。两侧始终为不同硬件：若一侧与另一侧相同，则改用另一侧的默认值。 | |
| v_cand | query | string | 可选 | 候选硬件键；无效或无可用测量时回退到界面默认值（符合条件的代表部署中最快的一种）。两侧始终为不同硬件：若一侧与另一侧相同，则改用另一侧的默认值。 | |
| v_case | query | string | 可选 | 匹配用例索引从零开始，范围 0–9999（默认 0）。view=compare 按可用配对范围收敛，并在 cases.resolvedCaseIndex 返回实际索引；dashboard 不加载媒体。 | |
| view | query | string | 可选 | 视频 overview（图表、结果表格和部署详情，使用 r_* 参数）、dashboard（未指定 run/artifact 时默认）、compare（已发布匹配用例），或原有 discovery/results/tradeoff。指定 run/artifact 时默认 results。不同视图的参数不能混用。 | |
| workload | query | string | 可选 | 视频 workloads 中的工作负载分组键。不同工作负载不合并比较。 | |
| xAxis | query | string | 可选 | 视频延迟：p90 或 median。serving 仅在 p90 不可用时默认 median。 | |
| yAxis | query | string | 可选 | CollectiveX 支持 latency、tokens-per-second、activation-rate、payload-rate；视频支持 dollar、clipsGpu、secondsGpu、clipsAllocatedGpu、secondsAllocatedGpu、energy。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/video'响应
VideoView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"video"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"cells": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"deploymentCells": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"workloadMix": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"overlayCells": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"overlayDeploymentCells": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"clipQuality": {
"type": "object",
"description": "Per-clip evidence from the same selected artifacts and matching helper as the UI. No cohort PSNR aggregate; missing pairs remain null.",
"x-description-zh": "与界面共用所选产物和匹配函数的逐条视频证据。不计算样本组 PSNR 汇总值；缺失配对保留 null。",
"properties": {
"status": {
"type": "string",
"enum": [
"ready",
"partial",
"unavailable"
]
},
"groups": {
"type": "array",
"items": {
"type": "object",
"properties": {
"key": {
"type": "string"
},
"scope": {
"type": "string",
"enum": [
"ci",
"matched-noise-diagnostic"
]
},
"prompt": {
"type": "string"
},
"seed": {
"type": "number"
},
"take": {
"type": [
"integer",
"null"
]
},
"clips": {
"type": "array",
"items": {
"type": "object",
"properties": {
"hardware": {
"type": "string"
},
"pointId": {
"type": [
"string",
"null"
],
"description": "Performance point identity for CI clips; null for diagnostic reruns."
},
"contextPointId": {
"type": [
"string",
"null"
],
"description": "Original CI context for a diagnostic, never its media or performance identity."
},
"slotId": {
"type": [
"string",
"null"
]
},
"caseId": {
"type": [
"string",
"null"
]
},
"mediaPath": {
"type": [
"string",
"null"
]
},
"quality": {
"type": [
"object",
"null"
],
"properties": {
"attachmentSha256": {
"type": "string"
},
"scope": {
"type": "string",
"enum": [
"matched-noise-diagnostic"
]
},
"origin": {
"type": "object",
"additionalProperties": true,
"description": "Exact original CI candidate identity retained only as diagnostic context."
},
"candidate": {
"type": "object",
"additionalProperties": true,
"description": "Exact CI candidate identity, or the separately sealed native diagnostic run/receipt/media identity with captured-noise and topology metadata. Diagnostic clips do not inherit CI latency, energy or cost."
},
"reference": {
"type": "object",
"additionalProperties": true,
"description": "Reference case from the public descriptor: receipt/plan/media hashes, official Diffusers/model revisions, denoising_updates, sigma_points, and the stored MP4 url (empty for an unstored named run) with optional previewUrl. Reference execution is not a hardware performance result."
},
"scorer": {
"type": "object",
"additionalProperties": true,
"description": "evaluator.media.compare_media source revision and SHA256; method decoded-rgb24-all-frames."
},
"status": {
"type": "string",
"enum": [
"measured",
"identical",
"incompatible"
]
},
"psnrDb": {
"type": [
"number",
"null"
],
"description": "Decoded RGB24, every pixel/frame, peak=1 after division by 255, full-stream MSE; dB. Null for exact decoded equality or incompatible pairs; status distinguishes them. No quality pass/fail threshold."
},
"comparedFrames": {
"type": "integer",
"minimum": 0
},
"totalFrames": {
"type": "integer",
"minimum": 0
},
"checks": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"notes": {
"type": "array",
"items": {
"type": "string"
}
},
"qualityStatus": {
"type": "string",
"enum": [
"unjudged"
]
},
"releaseQualified": {
"type": "boolean",
"enum": [
false
]
}
},
"required": [
"candidate",
"reference",
"scorer",
"status",
"psnrDb",
"comparedFrames",
"totalFrames",
"qualityStatus",
"releaseQualified"
]
},
"unscored": {
"type": [
"object",
"null"
],
"description": "Why the scorer did not compare this clip; unreadable when it could not read the artifact. Null when the clip has a pair or no quality attachment exists.",
"properties": {
"reason": {
"type": "string",
"enum": [
"invalid_candidate",
"checksum_mismatch",
"no_reference",
"settings_mismatch",
"unreadable"
]
},
"fields": {
"type": "array",
"items": {
"type": "string"
},
"description": "Settings that differ from the reference."
}
},
"required": [
"reason",
"fields"
]
}
},
"required": [
"hardware",
"pointId",
"slotId",
"caseId",
"mediaPath",
"quality",
"unscored"
]
}
}
},
"required": [
"key",
"scope",
"prompt",
"seed",
"take",
"clips"
]
}
}
},
"required": [
"status",
"groups"
]
},
"options": {
"type": "object",
"additionalProperties": true
},
"state": {
"type": "object",
"additionalProperties": true
},
"workloadKey": {
"type": [
"string",
"null"
]
},
"runs": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
},
"description": "Official runs of the selected workload in measurement order, each { id, date } with the GitHub Actions run ID and execution date.",
"x-description-zh": "所选工作负载的正式运行，按测量时间排序；每项为 { id, date }，即 GitHub Actions 运行 ID 和执行日期。"
},
"run": {
"type": [
"object",
"null"
],
"description": "The run the official cells are shown as of (r_date and r_run); null when no official run matches.",
"x-description-zh": "正式结果所截至的运行（由 r_date 和 r_run 决定）；没有匹配的正式运行时为 null。"
},
"chipConfigs": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
},
"description": "Every official deployment of the selected workload as { key, point }, the r_configs choices.",
"x-description-zh": "所选工作负载的全部正式部署，每项为 { key, point }，即 r_configs 的可选项。"
},
"history": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
},
"description": "Cells for every official run of the r_configs chip configs within r_from and r_to, each run contributing its newest publication of each deployment, plus the named overlay run's cells for those configs; empty when r_configs is empty.",
"x-description-zh": "r_configs 所选芯片配置在 r_from 至 r_to 期间每次正式运行的结果（每次运行中的每个部署取最新发布的结果），以及指定叠加运行中这些配置的结果；r_configs 为空时为空数组。"
},
"excludedCount": {
"type": "integer",
"minimum": 0
},
"sampleFloor": {
"type": "integer",
"const": 10
},
"scheduledClips": {
"type": "integer",
"const": 20
},
"roster": {
"type": "array",
"items": {
"type": "string"
},
"description": "Default GB200, B200, MI355X and H200 slots, plus H100 when qualified approved or exact named diagnostic evidence is available.",
"x-description-zh": "默认包含 GB200、B200、MI355X 和 H200；存在符合门槛的已批准 H100 结果或明确指定的诊断运行时增加 H100。"
},
"metricDefinitions": {
"type": "object",
"additionalProperties": true
},
"plot": {
"type": "object",
"additionalProperties": true
},
"rows": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"kpis": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"comparison": {
"type": "object",
"additionalProperties": true
},
"provenance": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"workload": {
"type": [
"string",
"null"
]
},
"otherWorkloads": {
"type": "integer",
"minimum": 0
},
"coverage": {
"type": "object",
"properties": {
"pagesRead": {
"type": "integer",
"minimum": 1,
"maximum": 5
},
"maxPages": {
"type": "integer",
"enum": [
5
]
},
"nextPage": {
"type": [
"integer",
"null"
]
},
"truncated": {
"type": "boolean"
}
},
"required": [
"pagesRead",
"maxPages",
"nextPage",
"truncated"
]
},
"sources": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"points": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"curves": {
"type": "object",
"additionalProperties": true
},
"evidence": {
"type": [
"object",
"null"
],
"additionalProperties": true
},
"discovery": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "video",
"apiVersion": "v1",
"params": {}
}其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
channel,approved,published_at,run_id,artifact_id,source_manifest_sha256,point_id,cell_id,workload_key,workload_sha256,model,hardware,hardware_name,engine,mode,resolution,duration_s,fps,steps,observed_at,participating_gpus,allocated_gpus,scheduled_clips,completed_clips,valid_clips,failed_clips,attempted_clips,warmup_s,client_p50_s,p95_s,clips_per_s,video_seconds_per_s,deadline_s,deadline_attainment,deadline_goodput_clips_per_s,gpu_board_energy_j_per_clip,gpu_board_mean_w,power_valid,power_samples,cost_tier,participating_cost_usd_per_video,participating_cost_usd_per_video_s,participating_cost_usd_per_hour,allocated_cost_usd_per_video,allocated_cost_usd_per_video_s,allocated_cost_usd_per_hour,participating_profit_usd_per_hour,allocated_profit_usd_per_hour
204
暂无已发布的视频产物，响应不含正文。
响应结构
{
"type": "null"
}响应示例
null404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/ubenchxuBenchX 微基准测试及各设备的测试协议和带宽参考值（mem-bw、sm-l2-distance、tpc-skyline 和 core-memory-profiles）
只读uBenchX 微基准测试及各设备的测试协议和带宽参考值（mem-bw、sm-l2-distance、tpc-skyline 和 core-memory-profiles），使用仪表板的数据读取和计算函数。未知或重复查询键返回 400；响应包含解析后的参数，保留缺失数据。仅影响样式的控件不作为 API 参数。 tpc-skyline 另外返回 skuConfigurations，与 NVIDIA 实测的 gpus 分开列出。gpu=all（默认）或 Huawei 950PR 时返回所提供的全部九种 950PR 配置；指定完整的 Ascend 950PR_<suffix> 时只返回该 SKU。每行保留 Cube/Vector/AICPU 数量、频率（MHz）、内存（GiB）、L2（MiB）、virtualisationOptions，以及原始 ddrRate 数值（unit=null）。来源元数据链接到所提供的截图，并标记 vendorSourceVerified=false；这些配置数据不是拓扑实测结果。test=all 时，tpcSkyline 中返回同一数组，并遵循显式指定的 gpu 筛选。 core-memory-profiles 默认 gpu=all，先返回 Ascend 950PR_9579，再返回 B200 SXM；gpu 可单独选择设备。B200 保留 orderedSms、smInfo、gpcBounds、matrix、stats 和 metadata，单位为 clock64 cycles，使用独立 colorLimits。旧版 sm-l2-distance API 仍可使用，响应结构不变。Ascend 返回各面板的实际 coreIds、dieGroups、完整矩阵、原始 SYS_CNT tick 单位及去重核心对的组内和组间统计；没有对应核心对时为 null。colorLimits 和 limitations 保留统一色标及解释限制。这些数据是访存特征差异，不是核心间通信延迟或物理布局图。 mem-bw 包含 Huawei FP16 轮转缓冲池图执行结果。metadata.protocol=rotating-pool-npu-graph 时，timeMs 为每次拷贝的摊销时间。peakBandwidthSource 说明该次测量采用的归一化参考值；行数据可保留 observations、bandwidthP10Gbps、bandwidthP90Gbps、processBandwidthGbps、slots 和 combinedPoolBytes。缺失大小列在 metadata.missingSizesBytes。Huawei 现覆盖 8 B 至 16 GiB。metadata.smallMessageExtension 记录 8 B–2 KiB 的独立 CANN 9.2.1 测量、4096 字节槽间距、归档摘要及缓存影响说明；原测量元数据适用于 4 KiB 及以上。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| gpu | query | string | 可选 | 加速器键，如 "H100 SXM"、"Huawei 950PR"、"B200 SXM" 或 "Ascend 950PR_9579"。core-memory-profiles 和 tpc-skyline 默认 "all"，其他测试默认首个可用设备。"all" 返回全部设备；显式选择设备时，没有数据的测试保留空 gpus 数组。 | |
| test | query | string | 可选 | uBenchX 测试选择：mem-bw、sm-l2-distance、tpc-skyline、core-memory-profiles 或 all（默认），决定返回哪种基准测试数据。 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/ubenchx'响应
UbenchxView200
解析后的选择与公开视图数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"ubenchx"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"additionalProperties": true
},
"gpus": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"tests": {
"type": "array",
"items": {
"type": "string"
}
},
"memBw": {
"type": "object",
"additionalProperties": true
},
"smL2Distance": {
"type": "object",
"additionalProperties": true
},
"tpcSkyline": {
"type": "object",
"additionalProperties": true
},
"skuConfigurations": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"coreMemoryProfiles": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"view",
"apiVersion",
"params"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"view": "ubenchx",
"apiVersion": "v1",
"params": {}
}404
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}429
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}503
保留上游状态码，移除内部详情，失败响应不缓存。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Source data unavailable"
}400
查询无效或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Invalid parameter",
"param": "model"
}500
数据读取或视图计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/views/model-architectures读取模型架构规格
返回英文、中文模型页面和仪表盘架构图共用的静态架构注册表。不传 model 时返回全部条目，包括没有规范 compare slug 的模型（slug 为 null）。传入完整模型显示名称或规范 compare slug 可筛选条目；已知模型尚无架构数据时，models 数组为空；别名、空值、未知或重复参数均返回 400。未提供的可选规格保持省略。参数量以十亿为单位，维度和上下文长度使用精确整数；sourceUrl 指向厂商来源。架构图展开仅影响展示。该目录不读取数据库、基准测试结果、私有数据或非官方 run；模型页的基准测试控件请使用 inference 视图。公开访问，无需认证，仅返回 JSON。成功响应使用 Cache-Control: public, max-age=0, s-maxage=300。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| model | query | string | 可选 | 完整模型显示名称或规范 compare slug。省略时返回全部架构；不接受空值或别名。 | kimi-k3 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/views/model-architectures?model=kimi-k3'响应
ModelArchitecturesView200
架构规格、解析后的模型选择及规范 slug。
响应结构
{
"type": "object",
"properties": {
"view": {
"type": "string",
"enum": [
"model-architectures"
]
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"params": {
"type": "object",
"properties": {
"model": {
"type": [
"string",
"null"
]
}
},
"required": [
"model"
],
"additionalProperties": false
},
"models": {
"type": "array",
"items": {
"type": "object",
"properties": {
"slug": {
"type": [
"string",
"null"
]
},
"architecture": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"totalParams": {
"type": "number",
"description": "Billions of parameters; see model-specific features for excluded auxiliary modules."
},
"activeParams": {
"type": "number",
"description": "Billions of active parameters per token; decode count when activeParamsPrefill is present."
},
"activeParamsPrefill": {
"type": "number"
},
"conditionalMemoryParams": {
"type": "number"
},
"conditionalMemoryName": {
"type": "string"
},
"architectureType": {
"type": "string",
"enum": [
"dense",
"moe"
]
},
"attentionType": {
"type": "string"
},
"attentionLabel": {
"type": "string"
},
"numLayers": {
"type": "integer"
},
"hiddenSize": {
"type": "integer"
},
"numHeads": {
"type": "integer"
},
"numKVHeads": {
"type": "integer"
},
"headDim": {
"type": "integer"
},
"vocabSize": {
"type": "integer"
},
"ffnDim": {
"type": "integer"
},
"numExperts": {
"type": "integer"
},
"activeExperts": {
"type": "integer"
},
"hasSharedExpert": {
"type": "boolean"
},
"sharedExperts": {
"type": "integer"
},
"denseFFNLayers": {
"type": "integer"
},
"denseFFNDim": {
"type": "integer"
},
"hashRoutedLayers": {
"type": "integer"
},
"slidingWindow": {
"type": "integer"
},
"hyperConnections": {
"type": "integer"
},
"contextWindow": {
"type": "integer"
},
"advertisedContextWindow": {
"type": "integer"
},
"features": {
"type": "array",
"items": {
"type": "string"
}
},
"featuresZh": {
"type": "array",
"items": {
"type": "string"
}
},
"developer": {
"type": "string"
},
"sourceUrl": {
"type": "string"
},
"attentionExpandable": {
"type": "boolean"
},
"alternatingAttentionExpandable": {
"type": "boolean"
},
"alternatingNote": {
"type": "string"
},
"alternatingNoteZh": {
"type": "string"
},
"denseLayerAttentionLabel": {
"type": "string"
},
"ffnVariant": {
"type": "string"
},
"ffnGateActivation": {
"type": "string"
},
"ffnSwiGLUOAI": {
"type": "boolean"
},
"ffnClamp": {
"type": "number"
},
"routedExpertHiddenSize": {
"type": "integer"
},
"residualType": {
"type": "string",
"enum": [
"AttnRes",
"gated"
]
},
"alternatingLayers": {
"type": "array",
"items": {
"type": "object",
"properties": {
"label": {
"type": "string"
},
"description": {
"type": "string"
},
"count": {
"type": "integer"
},
"colorKey": {
"type": "string"
},
"slidingWindow": {
"type": "integer"
},
"attentionShape": {
"type": "string"
}
},
"required": [
"label",
"description",
"count",
"colorKey"
],
"additionalProperties": false
}
},
"encoderDecoder": {
"type": "object",
"properties": {
"encoderLayers": {
"type": "integer"
},
"decoderLayers": {
"type": "integer"
},
"globalKVSource": {
"type": "string",
"enum": [
"final-encoder-hidden-states"
]
},
"encoderSlidingOnlyLayers": {
"type": "integer"
},
"decoderAttention": {
"type": "string",
"enum": [
"joint-local-swa-and-shared-global-kv"
]
}
},
"required": [
"encoderLayers",
"decoderLayers",
"globalKVSource",
"encoderSlidingOnlyLayers",
"decoderAttention"
],
"additionalProperties": false
}
},
"required": [
"model",
"totalParams",
"activeParams",
"architectureType",
"attentionType"
],
"additionalProperties": true
}
},
"required": [
"slug",
"architecture"
],
"additionalProperties": false
}
}
},
"required": [
"view",
"apiVersion",
"params",
"models"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"view": "model-architectures",
"apiVersion": "v1",
"params": {
"model": "Kimi-K3"
},
"models": [
{
"slug": "kimi-k3",
"architecture": {
"model": "Kimi-K3",
"totalParams": 2800,
"activeParams": 104,
"architectureType": "moe",
"attentionType": "Hybrid",
"numLayers": 93,
"hiddenSize": 7168,
"numExperts": 898,
"activeExperts": 16,
"sourceUrl": "https://huggingface.co/moonshotai/Kimi-K3"
}
}
]
}滚动查看完整示例
400
模型无效、model 为空、参数未知或重复。
响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
},
"param": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}滚动查看完整示例
响应示例
{
"error": "Unknown model: missing",
"param": "model"
}500
架构目录组装失败。
响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}核心基准数据
基准结果、可用配置、工作流来源、评估与可靠性数据。
GET稳定/api/v1/availability列出可用的基准配置
返回已有基准数据的模型、序列、精度、硬件、框架、投机解码方式、基准类型和日期组合。
参数
无需参数
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/availability'响应
AvailabilityRows200
可用配置行。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"isl": {
"type": [
"number",
"null"
]
},
"osl": {
"type": [
"number",
"null"
]
},
"precision": {
"type": "string"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"disagg": {
"type": "boolean"
},
"benchmark_type": {
"type": "string"
},
"date": {
"type": "string",
"format": "date"
}
},
"required": [
"model",
"isl",
"osl",
"precision",
"hardware",
"framework",
"spec_method",
"disagg",
"benchmark_type",
"date"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
[
{
"model": "dsr1",
"isl": 1024,
"osl": 1024,
"precision": "fp8",
"hardware": "h200_sxm",
"framework": "vllm",
"spec_method": "none",
"disagg": false,
"benchmark_type": "single_turn",
"date": "2026-08-08"
}
]滚动查看完整示例
500
可用配置查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/benchmarks读取基准结果
返回指定展示模型的原始基准测试数据行。使用 date 可获取截至指定日期的快照；exact=true 仅返回该日期的数据；runId 用于限定最新结果的查询范围；将 exactRun=true 与数值型 runId 搭配使用，可获取该次运行对应的逻辑快照。AgentX 按模型、硬件、框架、精度和工作负载划分曲线；常规运行会整体替换同一曲线中此前的所有拓扑、推测解码和 offload 变体。显式标记为 append-only 的运行可继承此前使用相同镜像的快照，并保留各数据点来源运行的 ID 和 URL。TPU 提示框和数据点详情页头部不显示 GitHub 运行链接，API 的 run_url 仍保留来源运行的原始 URL。非 CI 发布的官方预览结果以其来源页面作为 run_url，或为 null。历史快照仍可查询。view=calculator 返回页面专用的裁剪投影（会移除实测功率指标和 workers，其允许列表可能变化）。powerValid=strictV2 仅返回采用 schema v2 且功率测量通过验证的数据行，不能与 view=calculator 组合使用。省略 powerValid 则保留常规基准测试结果，不按功率有效性筛选。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| model | query | string | 必填 | 展示模型名称。GLM-5.3-Flash 仅对应 glm5.3flash，与 GLM-5.2/5.3 分开，支持 8k/1k 和 agentic-traces。注册模型不代表已有公开发布的实测数据，也不代表硬件已通过验证。非官方运行叠加数据与已发布的基准结果分开。网页界面仅在启用该功能开关后显示 GLM-5.3-Flash。该开关只控制界面显示，不用于访问控制；只读 API 的模型参数和返回数据不受其影响。 | DeepSeek-R1-0528 |
| date | query | date | 可选 | 返回该日期（YYYY-MM-DD）当日或之前的最新数据，exact 为 true 时仅限当日。空值表示最新数据；不是有效日期的取值（如 latest）返回 400。 | 2026-08-08 |
| exact | query | boolean | 可选 | 设为 true 时严格匹配所给日期。仅接受 true 或 false。 | false |
| runId | query | integer | 可选 | GitHub Actions 运行 ID，须为正整数。其他取值返回 400。 | 123456789 |
| exactRun | query | boolean | 可选 | 与数字 runId 一起使用时，返回该次运行对应的逻辑快照；append-only 运行的快照还会包含此前运行的数据，前提是两者共有的拓扑使用同一镜像。仅接受 true 或 false，且为 true 时必须提供 runId。 | false |
| view | query | enum | 可选 | calculator 会将每行裁剪为吞吐量计算器所需的页面专用指标允许列表并移除 workers；此视图不包含实测功率指标。需要同时提供 sequence。省略则返回全部已存储指标，包括实测功率。其他取值返回 400 Unknown view。 | calculator |
| sequence | query | enum | 可选 | 当 view=calculator 时必填，其余情况会被忽略。未知值返回 400 Unknown calculator sequence。 | 1k/1k |
| powerValid | query | enum | 可选 | 仅接受 strictV2：保留 metrics.power_valid 为数字 1、且 metrics.power_metric_schema_version 为数字 2 的数据行，其能耗指标采用整个部署的统计口径。省略此参数则不按功率筛选，即使功率缺失或无效，也会保留吞吐量和延迟结果。其他取值（包括空值）返回 400 Unknown powerValid filter。不能与 view=calculator 组合使用。 | strictV2 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/benchmarks?model=DeepSeek-R1-0528'响应
BenchmarkRows200
基准行，标量指标位于 metrics 对象中。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"model": {
"type": "string"
},
"precision": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"disagg": {
"type": "boolean"
},
"is_multinode": {
"type": "boolean"
},
"prefill_tp": {
"type": "integer"
},
"prefill_ep": {
"type": "integer"
},
"prefill_dp_attention": {
"type": "boolean"
},
"prefill_num_workers": {
"type": "integer"
},
"decode_tp": {
"type": "integer"
},
"decode_ep": {
"type": "integer"
},
"decode_dp_attention": {
"type": "boolean"
},
"decode_num_workers": {
"type": "integer"
},
"num_prefill_gpu": {
"type": "integer",
"description": "Physical prefill chips; aggregate engines may mirror their single chip count in both role columns."
},
"num_decode_gpu": {
"type": "integer",
"description": "Physical decode chips, independent of logical TP and DP. Sum role counts only for disaggregated engines."
},
"benchmark_type": {
"type": "string"
},
"isl": {
"type": [
"number",
"null"
]
},
"osl": {
"type": [
"number",
"null"
]
},
"conc": {
"type": "integer"
},
"offload_mode": {
"type": "string"
},
"image": {
"type": [
"string",
"null"
]
},
"recipe_fingerprint": {
"type": [
"string",
"null"
]
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
},
"description": "Scalar metric map. Time metrics, including p99_itl and p99_tpot, are in seconds. p99_itl measures inter-token latency; p99_tpot measures per-request time per output token. Use the actual p99_itl field for an inter-token latency requirement, not the reciprocal of p99_intvty. Throughput metrics use tokens per second per GPU unless their name states otherwise; output_tput_per_gpu counts output tokens. On disaggregated rows tput_per_gpu divides by every chip, input_tput_per_gpu by prefill chips and output_tput_per_gpu by decode chips, for fixed-sequence and agentic runs alike. Agentic disaggregated rows also keep the original per-total-chip rates from the producer as input_tput_per_total_gpu and output_tput_per_total_gpu. Keys evolve independently; measured power / energy / GPU-telemetry keys are typed below.",
"properties": {
"power_valid": {
"type": "number",
"description": "Publication verdict: 1 = validated measurement window; 0 = failed validation — measured power/energy values are withheld from this row end-to-end, so treat any that remain as unreliable; absent = no validation verdict is available in this response. Absence alone establishes neither the reason, the measurement age, nor invalidity."
},
"power_metric_schema_version": {
"type": "number",
"description": "Power schema version. Version 2 defines every unprefixed joules_per_* field as whole-deployment energy, including on disaggregated runs."
},
"cpu_power_valid": {
"type": "number",
"description": "CPU-side leg verdict (NVL72 Grace sockets or single-node AMD EPYC RAPL packages), independent of power_valid: 1 = the CPU-side (and module, when present) keys were integrated over the validated window with full socket coverage; 0 = the leg failed and no CPU-side keys are emitted; absent = the run had no CPU-side leg. Ingest withholds the CPU-side keys unless this verdict is 1, and power_valid = 0 does not remove them. GPU power and energy never change because of this verdict."
},
"avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts during the measured load window."
},
"avg_total_gpu_power_w": {
"type": "number",
"description": "Mean total GPU power draw in watts across the measured deployment."
},
"total_gpu_energy_j": {
"type": "number",
"description": "Total GPU energy integrated over the measured deployment load window, in joules."
},
"p75_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p75_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"p90_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p90_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"joules_per_successful_query": {
"type": "number",
"description": "Whole-deployment energy in joules divided by successful requests."
},
"joules_per_output_token": {
"type": "number",
"description": "Energy per generated output token in joules; cluster-wide on schema-version-2 rows, including disaggregated runs."
},
"joules_per_total_token": {
"type": "number",
"description": "Total system energy divided by input plus output tokens; a workload-shape-fair view that does not treat prompt tokens as free."
},
"prefill_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across prefill workers; emitted only for deployments with distinct prefill and decode roles."
},
"decode_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across decode workers; emitted only for deployments with distinct prefill and decode roles."
},
"joules_per_input_token": {
"type": "number",
"description": "Energy per input token in joules; cluster-wide on schema-version-2 rows."
},
"prefill_joules_per_input_token": {
"type": "number",
"description": "Role-local prefill energy per input token in joules."
},
"decode_joules_per_output_token": {
"type": "number",
"description": "Role-local decode energy per generated output token in joules."
},
"avg_temp_c": {
"type": "number",
"description": "Mean per-GPU temperature in degrees Celsius during the load window."
},
"peak_temp_c": {
"type": "number",
"description": "Maximum instantaneous per-GPU temperature in degrees Celsius during the load window."
},
"avg_util_pct": {
"type": "number",
"description": "Mean per-GPU utilization percentage (0-100) during the load window."
},
"avg_mem_used_mb": {
"type": "number",
"description": "Mean per-GPU memory used in MB during the load window."
},
"avg_cpu_socket_power_w": {
"type": "number",
"description": "Mean over Grace sockets of each socket’s window-mean Grace-side power in watts (ACPI hwmon Grace Power Socket sensor, which NVIDIA defines only as the Grace socket power; modeled as the CPU and SysIO rails, LPDDR5X and socket regulation) over the same validated window as GPU energy. On single-node AMD rows (power_audit.cpu.sensor_kind = rapl_package) it is instead each EPYC package’s RAPL energy-counter power, which has no DRAM domain."
},
"avg_total_cpu_power_w": {
"type": "number",
"description": "Sum over every CPU socket in the deployment of window-mean socket power in watts (Grace side on NVL72, RAPL package on single-node AMD); dividing by avg_cpu_socket_power_w recovers the socket count (two per compute tray or AMD host)."
},
"total_cpu_energy_j": {
"type": "number",
"description": "CPU-side energy in joules (Grace side on NVL72, RAPL package on single-node AMD) integrated over the validated window across every socket. Additive to total_gpu_energy_j, never included in it."
},
"avg_total_module_power_w": {
"type": "number",
"description": "Sum over sockets of window-mean compute-module power in watts (Grace + its Blackwell GPUs + HBM + LPDDR5X + regulator loss, ACPI hwmon Module Power Socket sensor); emitted only when the module sensor exists on every socket. It already contains the GPU-board watts, so do not add avg_total_gpu_power_w to it."
},
"total_module_energy_j": {
"type": "number",
"description": "Compute-module energy in joules over the validated window across every socket; emitted with avg_total_module_power_w and includes the GPU energy."
},
"avg_total_cpu_rail_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean CPU core rail power in watts (ACPI CPU Power Socket sensor, no SysIO); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_rail_energy_j": {
"type": "number",
"description": "CPU core rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_rail_power_w and already included in total_cpu_energy_j."
},
"avg_total_cpu_sysio_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean SysIO rail power in watts (ACPI SysIO Power Socket sensor); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_sysio_energy_j": {
"type": "number",
"description": "SysIO rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_sysio_power_w and already included in total_cpu_energy_j."
}
}
},
"workers": {
"type": "array",
"items": {
"type": "object",
"properties": {
"role": {
"type": "string"
},
"worker_idx": {
"type": "integer"
},
"hosts": {
"type": "array",
"items": {
"type": "string"
}
},
"num_gpus": {
"type": "integer"
},
"avg_power_w": {
"type": "number"
},
"avg_temp_c": {
"type": "number"
},
"peak_temp_c": {
"type": "number"
},
"avg_util_pct": {
"type": "number"
},
"avg_mem_used_mb": {
"type": "number"
}
},
"required": [
"role",
"worker_idx",
"num_gpus",
"avg_power_w"
],
"additionalProperties": false
}
},
"power_invalid_reasons": {
"type": "array",
"items": {
"type": "string"
},
"description": "Producer snake_case reason codes explaining a withheld measurement, present when metrics.power_valid == 0. Absent on legacy rows and validated rows."
},
"power_audit": {
"type": "object",
"properties": {
"window_start_unix": {
"type": "number"
},
"window_end_unix": {
"type": "number"
},
"expected_gpu_count": {
"type": "integer"
},
"observed_gpu_count": {
"type": "integer"
},
"sample_count": {
"type": "integer"
},
"max_sample_gap_s": {
"type": "number"
},
"producer_sha": {
"type": [
"string",
"null"
]
},
"exporter_image_sha256": {
"type": [
"string",
"null"
]
},
"source": {
"type": "string"
},
"observed_gpu_ids": {
"type": "array",
"items": {
"type": "string"
}
},
"cpu": {
"type": "object",
"properties": {
"sensor_kind": {
"type": "string",
"enum": [
"module",
"grace_socket",
"dcgm_cpu_rail",
"rapl_package"
]
},
"source": {
"type": "string",
"description": "Collector that produced the CPU-side samples: acpi (hwmon), dcgm, or rapl (AMD EPYC package energy counter)."
},
"expected_sockets": {
"type": "integer"
},
"observed_sockets": {
"type": "integer"
},
"sample_row_count": {
"type": "integer"
},
"reason_codes": {
"type": "array",
"items": {
"type": "string"
}
}
},
"additionalProperties": true,
"description": "CPU-side leg provenance emitted alongside cpu_power_valid: the sensor kind that fed the CPU-side keys (on NVL72, module preferred over grace_socket over dcgm_cpu_rail; rapl_package on single-node AMD EPYC hosts), its collector, expected vs. observed sockets (two per compute tray or AMD host), stored sample rows, and the leg’s reason codes (empty when valid). Absent on runs without the CPU-side leg."
}
},
"additionalProperties": true,
"description": "Compact power measurement-window audit emitted alongside the power_valid verdict: window bounds, expected vs. observed GPU counts, sample statistics, and producer identity (producer_sha / exporter_image_sha256 are null for single-node telemetry without an srt-slurm producer). source names the validation document: a relative path within the source run artifact bundle or, on AgentX rows whose retained window ingest or backfill recovered, the alias power_validation_<stem>_conc<N>.json for LOGS/agentic/conc_<N>/power_validation.json in the power_audit_<stem> artifact (such rows carry only source and the window bounds); observed_gpu_ids contains producer device identifiers, which may be indices rather than physical UUIDs. cpu carries the CPU-side leg’s provenance (NVL72 or single-node AMD) when that leg ran. Present on valid and invalid rows when emitted; absence alone does not establish age or validity."
},
"date": {
"type": "string",
"format": "date"
},
"workflow_run_id": {
"type": "integer"
},
"run_started_at": {
"type": [
"string",
"null"
],
"format": "date-time"
},
"run_url": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"hardware",
"framework",
"model",
"precision",
"spec_method",
"disagg",
"is_multinode",
"prefill_tp",
"prefill_ep",
"prefill_dp_attention",
"prefill_num_workers",
"decode_tp",
"decode_ep",
"decode_dp_attention",
"decode_num_workers",
"num_prefill_gpu",
"num_decode_gpu",
"benchmark_type",
"isl",
"osl",
"conc",
"offload_mode",
"image",
"recipe_fingerprint",
"metrics",
"date",
"run_url"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
[
{
"id": 421,
"hardware": "h200_sxm",
"framework": "vllm",
"model": "dsr1",
"precision": "fp8",
"spec_method": "none",
"disagg": false,
"is_multinode": false,
"prefill_tp": 8,
"prefill_ep": 1,
"prefill_dp_attention": false,
"prefill_num_workers": 1,
"decode_tp": 8,
"decode_ep": 1,
"decode_dp_attention": false,
"decode_num_workers": 1,
"num_prefill_gpu": 0,
"num_decode_gpu": 8,
"benchmark_type": "single_turn",
"isl": 1024,
"osl": 1024,
"conc": 32,
"offload_mode": "off",
"image": "vllm/vllm-openai:v0.10.2",
"recipe_fingerprint": "7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d",
"metrics": {
"median_ttft": 0.42,
"median_tpot": 0.018,
"tput_per_gpu": 128.4,
"power_valid": 1,
"power_metric_schema_version": 2,
"avg_power_w": 678.5,
"joules_per_output_token": 5.3,
"joules_per_total_token": 2.65,
"avg_temp_c": 61.2
},
"date": "2026-08-08",
"run_url": "https://github.com/semianalysis/inference-benchmarks/actions/runs/123456789"
}
]滚动查看完整示例
400
查询参数格式错误，请求不会读取数据库：模型缺失或不受支持、date 不是 YYYY-MM-DD 格式的有效日期、exact 或 exactRun 不是 true 或 false、runId 不是正整数、exactRun=true 但未提供 runId、view 不是 calculator、计算器序列未知、提供的 powerValid 取值不是 strictV2，或 powerValid=strictV2 与 view=calculator 组合使用。未知参数会被忽略。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown model"
}500
基准查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/benchmarks/history读取基准历史
返回指定模型在固定输入/输出 token 组合或 Agentic Traces 条件下的历史逻辑快照。被替换的 AgentX 变体仍保留在此前的快照中。append-only 快照包含继承的数据点，并分别提供各点的原始来源运行元数据和曲线快照元数据。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| model | query | string | 必填 | 展示模型名称。 | DeepSeek-R1-0528 |
| isl | query | integer | 可选 | 正整数输入序列长度，单位为 token；除非 benchmarkType=agentic_traces，否则必填。 | 1024 |
| osl | query | integer | 可选 | 正整数输出序列长度，单位为 token；除非 benchmarkType=agentic_traces，否则必填。 | 1024 |
| benchmarkType | query | string | 可选 | 设为 agentic_traces 可在不提供 ISL/OSL 的情况下读取 Agentic Traces 历史。 | agentic_traces |
| view | query | enum | 可选 | calculator 会将每行裁剪为吞吐量计算器所需的指标，以减小响应体积。省略后会保留历史接口提供的其他指标，包括已记录的实测功率。两种视图都不返回 ttft、tpot、itl、e2el 和 intvty 的 mean_* 与 std_* 统计值。 | calculator |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/benchmarks/history?model=DeepSeek-R1-0528&isl=1024&osl=1024'响应
BenchmarkRows200
历史基准行。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"model": {
"type": "string"
},
"precision": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"disagg": {
"type": "boolean"
},
"is_multinode": {
"type": "boolean"
},
"prefill_tp": {
"type": "integer"
},
"prefill_ep": {
"type": "integer"
},
"prefill_dp_attention": {
"type": "boolean"
},
"prefill_num_workers": {
"type": "integer"
},
"decode_tp": {
"type": "integer"
},
"decode_ep": {
"type": "integer"
},
"decode_dp_attention": {
"type": "boolean"
},
"decode_num_workers": {
"type": "integer"
},
"num_prefill_gpu": {
"type": "integer",
"description": "Physical prefill chips; aggregate engines may mirror their single chip count in both role columns."
},
"num_decode_gpu": {
"type": "integer",
"description": "Physical decode chips, independent of logical TP and DP. Sum role counts only for disaggregated engines."
},
"benchmark_type": {
"type": "string"
},
"isl": {
"type": [
"number",
"null"
]
},
"osl": {
"type": [
"number",
"null"
]
},
"conc": {
"type": "integer"
},
"offload_mode": {
"type": "string"
},
"image": {
"type": [
"string",
"null"
]
},
"recipe_fingerprint": {
"type": [
"string",
"null"
]
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
},
"description": "Scalar metric map. Time metrics, including p99_itl and p99_tpot, are in seconds. p99_itl measures inter-token latency; p99_tpot measures per-request time per output token. Use the actual p99_itl field for an inter-token latency requirement, not the reciprocal of p99_intvty. Throughput metrics use tokens per second per GPU unless their name states otherwise; output_tput_per_gpu counts output tokens. On disaggregated rows tput_per_gpu divides by every chip, input_tput_per_gpu by prefill chips and output_tput_per_gpu by decode chips, for fixed-sequence and agentic runs alike. Agentic disaggregated rows also keep the original per-total-chip rates from the producer as input_tput_per_total_gpu and output_tput_per_total_gpu. Keys evolve independently; measured power / energy / GPU-telemetry keys are typed below.",
"properties": {
"power_valid": {
"type": "number",
"description": "Publication verdict: 1 = validated measurement window; 0 = failed validation — measured power/energy values are withheld from this row end-to-end, so treat any that remain as unreliable; absent = no validation verdict is available in this response. Absence alone establishes neither the reason, the measurement age, nor invalidity."
},
"power_metric_schema_version": {
"type": "number",
"description": "Power schema version. Version 2 defines every unprefixed joules_per_* field as whole-deployment energy, including on disaggregated runs."
},
"cpu_power_valid": {
"type": "number",
"description": "CPU-side leg verdict (NVL72 Grace sockets or single-node AMD EPYC RAPL packages), independent of power_valid: 1 = the CPU-side (and module, when present) keys were integrated over the validated window with full socket coverage; 0 = the leg failed and no CPU-side keys are emitted; absent = the run had no CPU-side leg. Ingest withholds the CPU-side keys unless this verdict is 1, and power_valid = 0 does not remove them. GPU power and energy never change because of this verdict."
},
"avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts during the measured load window."
},
"avg_total_gpu_power_w": {
"type": "number",
"description": "Mean total GPU power draw in watts across the measured deployment."
},
"total_gpu_energy_j": {
"type": "number",
"description": "Total GPU energy integrated over the measured deployment load window, in joules."
},
"p75_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p75_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"p90_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p90_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"joules_per_successful_query": {
"type": "number",
"description": "Whole-deployment energy in joules divided by successful requests."
},
"joules_per_output_token": {
"type": "number",
"description": "Energy per generated output token in joules; cluster-wide on schema-version-2 rows, including disaggregated runs."
},
"joules_per_total_token": {
"type": "number",
"description": "Total system energy divided by input plus output tokens; a workload-shape-fair view that does not treat prompt tokens as free."
},
"prefill_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across prefill workers; emitted only for deployments with distinct prefill and decode roles."
},
"decode_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across decode workers; emitted only for deployments with distinct prefill and decode roles."
},
"joules_per_input_token": {
"type": "number",
"description": "Energy per input token in joules; cluster-wide on schema-version-2 rows."
},
"prefill_joules_per_input_token": {
"type": "number",
"description": "Role-local prefill energy per input token in joules."
},
"decode_joules_per_output_token": {
"type": "number",
"description": "Role-local decode energy per generated output token in joules."
},
"avg_temp_c": {
"type": "number",
"description": "Mean per-GPU temperature in degrees Celsius during the load window."
},
"peak_temp_c": {
"type": "number",
"description": "Maximum instantaneous per-GPU temperature in degrees Celsius during the load window."
},
"avg_util_pct": {
"type": "number",
"description": "Mean per-GPU utilization percentage (0-100) during the load window."
},
"avg_mem_used_mb": {
"type": "number",
"description": "Mean per-GPU memory used in MB during the load window."
},
"avg_cpu_socket_power_w": {
"type": "number",
"description": "Mean over Grace sockets of each socket’s window-mean Grace-side power in watts (ACPI hwmon Grace Power Socket sensor, which NVIDIA defines only as the Grace socket power; modeled as the CPU and SysIO rails, LPDDR5X and socket regulation) over the same validated window as GPU energy. On single-node AMD rows (power_audit.cpu.sensor_kind = rapl_package) it is instead each EPYC package’s RAPL energy-counter power, which has no DRAM domain."
},
"avg_total_cpu_power_w": {
"type": "number",
"description": "Sum over every CPU socket in the deployment of window-mean socket power in watts (Grace side on NVL72, RAPL package on single-node AMD); dividing by avg_cpu_socket_power_w recovers the socket count (two per compute tray or AMD host)."
},
"total_cpu_energy_j": {
"type": "number",
"description": "CPU-side energy in joules (Grace side on NVL72, RAPL package on single-node AMD) integrated over the validated window across every socket. Additive to total_gpu_energy_j, never included in it."
},
"avg_total_module_power_w": {
"type": "number",
"description": "Sum over sockets of window-mean compute-module power in watts (Grace + its Blackwell GPUs + HBM + LPDDR5X + regulator loss, ACPI hwmon Module Power Socket sensor); emitted only when the module sensor exists on every socket. It already contains the GPU-board watts, so do not add avg_total_gpu_power_w to it."
},
"total_module_energy_j": {
"type": "number",
"description": "Compute-module energy in joules over the validated window across every socket; emitted with avg_total_module_power_w and includes the GPU energy."
},
"avg_total_cpu_rail_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean CPU core rail power in watts (ACPI CPU Power Socket sensor, no SysIO); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_rail_energy_j": {
"type": "number",
"description": "CPU core rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_rail_power_w and already included in total_cpu_energy_j."
},
"avg_total_cpu_sysio_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean SysIO rail power in watts (ACPI SysIO Power Socket sensor); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_sysio_energy_j": {
"type": "number",
"description": "SysIO rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_sysio_power_w and already included in total_cpu_energy_j."
}
}
},
"workers": {
"type": "array",
"items": {
"type": "object",
"properties": {
"role": {
"type": "string"
},
"worker_idx": {
"type": "integer"
},
"hosts": {
"type": "array",
"items": {
"type": "string"
}
},
"num_gpus": {
"type": "integer"
},
"avg_power_w": {
"type": "number"
},
"avg_temp_c": {
"type": "number"
},
"peak_temp_c": {
"type": "number"
},
"avg_util_pct": {
"type": "number"
},
"avg_mem_used_mb": {
"type": "number"
}
},
"required": [
"role",
"worker_idx",
"num_gpus",
"avg_power_w"
],
"additionalProperties": false
}
},
"power_invalid_reasons": {
"type": "array",
"items": {
"type": "string"
},
"description": "Producer snake_case reason codes explaining a withheld measurement, present when metrics.power_valid == 0. Absent on legacy rows and validated rows."
},
"power_audit": {
"type": "object",
"properties": {
"window_start_unix": {
"type": "number"
},
"window_end_unix": {
"type": "number"
},
"expected_gpu_count": {
"type": "integer"
},
"observed_gpu_count": {
"type": "integer"
},
"sample_count": {
"type": "integer"
},
"max_sample_gap_s": {
"type": "number"
},
"producer_sha": {
"type": [
"string",
"null"
]
},
"exporter_image_sha256": {
"type": [
"string",
"null"
]
},
"source": {
"type": "string"
},
"observed_gpu_ids": {
"type": "array",
"items": {
"type": "string"
}
},
"cpu": {
"type": "object",
"properties": {
"sensor_kind": {
"type": "string",
"enum": [
"module",
"grace_socket",
"dcgm_cpu_rail",
"rapl_package"
]
},
"source": {
"type": "string",
"description": "Collector that produced the CPU-side samples: acpi (hwmon), dcgm, or rapl (AMD EPYC package energy counter)."
},
"expected_sockets": {
"type": "integer"
},
"observed_sockets": {
"type": "integer"
},
"sample_row_count": {
"type": "integer"
},
"reason_codes": {
"type": "array",
"items": {
"type": "string"
}
}
},
"additionalProperties": true,
"description": "CPU-side leg provenance emitted alongside cpu_power_valid: the sensor kind that fed the CPU-side keys (on NVL72, module preferred over grace_socket over dcgm_cpu_rail; rapl_package on single-node AMD EPYC hosts), its collector, expected vs. observed sockets (two per compute tray or AMD host), stored sample rows, and the leg’s reason codes (empty when valid). Absent on runs without the CPU-side leg."
}
},
"additionalProperties": true,
"description": "Compact power measurement-window audit emitted alongside the power_valid verdict: window bounds, expected vs. observed GPU counts, sample statistics, and producer identity (producer_sha / exporter_image_sha256 are null for single-node telemetry without an srt-slurm producer). source names the validation document: a relative path within the source run artifact bundle or, on AgentX rows whose retained window ingest or backfill recovered, the alias power_validation_<stem>_conc<N>.json for LOGS/agentic/conc_<N>/power_validation.json in the power_audit_<stem> artifact (such rows carry only source and the window bounds); observed_gpu_ids contains producer device identifiers, which may be indices rather than physical UUIDs. cpu carries the CPU-side leg’s provenance (NVL72 or single-node AMD) when that leg ran. Present on valid and invalid rows when emitted; absence alone does not establish age or validity."
},
"date": {
"type": "string",
"format": "date"
},
"workflow_run_id": {
"type": "integer"
},
"run_started_at": {
"type": [
"string",
"null"
],
"format": "date-time"
},
"run_url": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"hardware",
"framework",
"model",
"precision",
"spec_method",
"disagg",
"is_multinode",
"prefill_tp",
"prefill_ep",
"prefill_dp_attention",
"prefill_num_workers",
"decode_tp",
"decode_ep",
"decode_dp_attention",
"decode_num_workers",
"num_prefill_gpu",
"num_decode_gpu",
"benchmark_type",
"isl",
"osl",
"conc",
"offload_mode",
"image",
"recipe_fingerprint",
"metrics",
"date",
"run_url"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
[
{
"id": 421,
"hardware": "h200_sxm",
"framework": "vllm",
"model": "dsr1",
"precision": "fp8",
"spec_method": "none",
"disagg": false,
"is_multinode": false,
"prefill_tp": 8,
"prefill_ep": 1,
"prefill_dp_attention": false,
"prefill_num_workers": 1,
"decode_tp": 8,
"decode_ep": 1,
"decode_dp_attention": false,
"decode_num_workers": 1,
"num_prefill_gpu": 0,
"num_decode_gpu": 8,
"benchmark_type": "single_turn",
"isl": 1024,
"osl": 1024,
"conc": 32,
"offload_mode": "off",
"image": "vllm/vllm-openai:v0.10.2",
"recipe_fingerprint": "7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d",
"metrics": {
"median_ttft": 0.42,
"median_tpot": 0.018,
"tput_per_gpu": 128.4,
"power_valid": 1,
"power_metric_schema_version": 2,
"avg_power_w": 678.5,
"joules_per_output_token": 5.3,
"joules_per_total_token": 2.65,
"avg_temp_c": 61.2
},
"date": "2026-08-08",
"run_url": "https://github.com/semianalysis/inference-benchmarks/actions/runs/123456789"
}
]滚动查看完整示例
400
必填参数缺失或模型不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "model, isl, and osl are required"
}500
历史查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/workflow-info读取工作流溯源信息
返回指定日期的工作流运行、变更记录、可用配置，以及各次运行涵盖的配置。date 为必填参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| date | query | date | 必填 | 必填的 YYYY-MM-DD 日期。 | 2026-08-08 |
| benchmarkType | query | string | 可选 | 设为 agentic_traces 可将每次运行的配置覆盖限定为 Agentic Traces。 | agentic_traces |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/workflow-info?date=2026-08-08'响应
WorkflowInfo200
工作流溯源信息，分为四个数组。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"runs": {
"type": "array",
"items": {
"type": "object",
"properties": {
"github_run_id": {
"type": "integer"
},
"name": {
"type": "string"
},
"conclusion": {
"type": [
"string",
"null"
]
},
"run_attempt": {
"type": "integer"
},
"html_url": {
"type": [
"string",
"null"
]
},
"created_at": {
"type": "string",
"format": "date-time"
},
"date": {
"type": "string",
"format": "date"
}
},
"required": [
"github_run_id",
"name",
"conclusion",
"run_attempt",
"html_url",
"created_at",
"date"
],
"additionalProperties": false
}
},
"changelogs": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"configs": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"runConfigs": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"runs",
"changelogs",
"configs",
"runConfigs"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"runs": [
{
"github_run_id": 123456789,
"name": "nightly-h200",
"conclusion": "success",
"run_attempt": 1,
"html_url": "https://github.com/semianalysis/inference-benchmarks/actions/runs/123456789",
"created_at": "2026-08-08T03:00:00Z",
"date": "2026-08-08"
}
],
"changelogs": [],
"configs": [],
"runConfigs": []
}滚动查看完整示例
400
缺少 date，或 date 不是 YYYY-MM-DD。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Invalid date format (YYYY-MM-DD required)"
}500
工作流查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/evaluations列出评估汇总
返回最新尝试的评估结果，包含配置、任务、来源和指标值。
参数
无需参数
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/evaluations'响应
EvaluationRows200
评估结果行。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"config_id": {
"type": "integer"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"model": {
"type": "string"
},
"precision": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"spec_algorithm": {
"type": [
"string",
"null"
],
"description": "Speculative algorithm recorded by a benchmark result for the same config and workflow run, or null when none was recorded."
},
"spec_draft_model": {
"type": [
"string",
"null"
],
"description": "Draft model recorded alongside spec_algorithm (\"native\" for built-in heads), or null when none was recorded."
},
"disagg": {
"type": "boolean"
},
"is_multinode": {
"type": "boolean"
},
"prefill_tp": {
"type": "integer"
},
"prefill_ep": {
"type": "integer"
},
"prefill_dp_attention": {
"type": "boolean"
},
"prefill_num_workers": {
"type": "integer"
},
"decode_tp": {
"type": "integer"
},
"decode_ep": {
"type": "integer"
},
"decode_dp_attention": {
"type": "boolean"
},
"decode_num_workers": {
"type": "integer"
},
"num_prefill_gpu": {
"type": "integer",
"description": "Physical prefill chips; aggregate engines may mirror their single chip count in both role columns."
},
"num_decode_gpu": {
"type": "integer",
"description": "Physical decode chips, independent of logical TP and DP. Sum role counts only for disaggregated engines."
},
"task": {
"type": "string"
},
"date": {
"type": "string",
"format": "date"
},
"conc": {
"type": [
"number",
"null"
]
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
}
},
"timestamp": {
"type": "string",
"format": "date-time"
},
"run_url": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"config_id",
"hardware",
"framework",
"model",
"precision",
"spec_method",
"spec_algorithm",
"spec_draft_model",
"disagg",
"is_multinode",
"prefill_tp",
"prefill_ep",
"prefill_dp_attention",
"prefill_num_workers",
"decode_tp",
"decode_ep",
"decode_dp_attention",
"decode_num_workers",
"num_prefill_gpu",
"num_decode_gpu",
"task",
"date",
"conc",
"metrics",
"timestamp",
"run_url"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
[
{
"id": 72,
"config_id": 11,
"hardware": "h200_sxm",
"framework": "vllm",
"model": "dsr1",
"precision": "fp8",
"spec_method": "none",
"spec_algorithm": null,
"spec_draft_model": null,
"disagg": false,
"is_multinode": false,
"prefill_tp": 8,
"prefill_ep": 1,
"prefill_dp_attention": false,
"prefill_num_workers": 1,
"decode_tp": 8,
"decode_ep": 1,
"decode_dp_attention": false,
"decode_num_workers": 1,
"num_prefill_gpu": 0,
"num_decode_gpu": 8,
"task": "gpqa",
"date": "2026-08-08",
"conc": null,
"metrics": {
"accuracy": 0.78
},
"timestamp": "2026-08-08T03:00:00Z",
"run_url": null
}
]滚动查看完整示例
500
评估查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/reliability列出基准可靠性
按硬件和日期返回成功运行数与总运行数。
参数
无需参数
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/reliability'响应
ReliabilityRows200
可靠性计数行。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "object",
"properties": {
"hardware": {
"type": "string"
},
"date": {
"type": "string",
"format": "date"
},
"n_success": {
"type": "integer"
},
"total": {
"type": "integer"
}
},
"required": [
"hardware",
"date",
"n_success",
"total"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
[
{
"hardware": "h200_sxm",
"date": "2026-08-08",
"n_success": 18,
"total": 20
}
]500
可靠性查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}外部数据源
面向电子表格、版本跟踪、提交记录和运行镜像的稳定数据源。
GET测试版/api/v1/pareto读取 Pareto 前沿和最差边界
使用图表共用的支配关系算法，计算已存储基准测试观测值的两条全局边界。前沿遵循请求指定的坐标轴优劣方向；最差边界将两个方向同时反转。坐标直接使用原始指标，不转换显示单位，并按 x 升序排列；observations 保留同坐标的全部观测值及其来源、快照字段。不返回插值、填充区域或背景图。缺失或非有限的坐标会被排除并计数；有限的零值和负值仍有效。必须指定工作负载和原始模型。此接口不复现浏览器中的 Optimal Only、快捷筛选、隐藏系列、对数轴排除规则或非官方产物叠加。功率或能耗坐标轴必须指定 powerValid=strictV2。空选择返回 200 和空边界。图表 Advanced 另有独立的 Pareto Frontier 控件：i_frontier=1 表示绿色填充，2 表示场景背景，省略则关闭。已停用的 i_hinterland 分享参数会被仪表板忽略，并被此接口拒绝。不要向此端点传入图表界面参数。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| model | query | string | 必填 | 展示模型名称。 | DeepSeek-R1-0528 |
| rawModel | query | string | 必填 | model 对应的原始模型键，精确匹配，避免混合不同变体。 | dsr1 |
| sequence | query | enum | 必填 | 精确工作负载。AgentX 对应 benchmark_type=agentic_traces；固定长度工作负载对应 single_turn 和匹配的 ISL/OSL。 | 1k/1k |
| xMetric | query | string | 必填 | 原始 metrics 对象中的键，不是图表标签。未知或缺失的键不会产生有效观测值。单位遵循 BenchmarkRow.metrics。 | median_intvty |
| yMetric | query | string | 必填 | 原始 metrics 对象中的键，不是图表标签。未知或缺失的键不会产生有效观测值。单位遵循 BenchmarkRow.metrics。 | output_tput_per_gpu |
| xDirection | query | enum | 必填 | 显式指定优劣方向：max 表示越高越好，min 表示越低越好。 | max |
| yDirection | query | enum | 必填 | 显式指定优劣方向：max 表示越高越好，min 表示越低越好。 | max |
| hardware | query | string | 可选 | 单个原始键，精确匹配。省略则汇总所有匹配值；未知值产生空选择。 | h200_sxm |
| framework | query | string | 可选 | 单个原始键，精确匹配。省略则汇总所有匹配值；未知值产生空选择。 | vllm |
| precision | query | string | 可选 | 单个原始键，精确匹配。省略则汇总所有匹配值；未知值产生空选择。 | fp8 |
| date | query | date | 可选 | 截至指定日期的快照；exact=true 选择该日期的快照。观测日期可能更早。不能与 exactRun=true 同时使用。 | 2026-08-08 |
| exact | query | boolean | 可选 | 为 true 时必须提供 date。仅接受字面值 true/false。 | false |
| runId | query | string | 可选 | 正安全整数形式的 GitHub Actions 运行 ID。未指定 exactRun=true 时限定最新结果查询范围；格式错误会被拒绝。 | 123456789 |
| exactRun | query | boolean | 可选 | 与 runId 一起读取该运行已存储的逻辑快照；append-only 运行还会包含此前运行的数据，前提是两者共有的拓扑使用同一镜像。不能同时提供 date 或 exact。仅接受字面值 true/false。 | false |
| powerValid | query | enum | 可选 | 已注册的功率或能耗指标作为坐标轴时必须设为 strictV2，其他情况可选。在选择和计数前筛选原始数据行。 | strictV2 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/pareto?model=DeepSeek-R1-0528&rawModel=dsr1&sequence=1k%2F1k&xMetric=median_intvty&yMetric=output_tput_per_gpu&xDirection=max&yDirection=max'响应
ParetoBoundaries200
边界坐标及来源观测值；空结果有效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"source_url": {
"type": "string",
"description": "Relative raw benchmark request URL, including snapshot and power selectors. Resolve against the API origin."
},
"selection": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"rawModel": {
"type": "string"
},
"sequence": {
"type": "string",
"enum": [
"1k/1k",
"1k/8k",
"8k/1k",
"agentic-traces"
]
},
"xMetric": {
"type": "string"
},
"yMetric": {
"type": "string"
},
"xDirection": {
"type": "string",
"enum": [
"min",
"max"
]
},
"yDirection": {
"type": "string",
"enum": [
"min",
"max"
]
},
"hardware": {
"type": [
"string",
"null"
]
},
"framework": {
"type": [
"string",
"null"
]
},
"precision": {
"type": [
"string",
"null"
]
}
},
"required": [
"model",
"rawModel",
"sequence",
"xMetric",
"yMetric",
"xDirection",
"yDirection",
"hardware",
"framework",
"precision"
],
"additionalProperties": false
},
"counts": {
"type": "object",
"properties": {
"returned": {
"type": "integer",
"description": "Rows from the raw request after any powerValid filter."
},
"selected": {
"type": "integer",
"description": "Rows after raw-model, workload, hardware, framework and precision selection."
},
"eligible": {
"type": "integer",
"description": "Selected observations with both finite numeric axes, including coordinate ties."
},
"missing_or_nonfinite": {
"type": "integer",
"description": "Selected observations excluded for absent or non-finite axes."
}
},
"required": [
"returned",
"selected",
"eligible",
"missing_or_nonfinite"
],
"additionalProperties": false
},
"frontier": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"y": {
"type": "number"
},
"observations": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"model": {
"type": "string"
},
"precision": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"disagg": {
"type": "boolean"
},
"is_multinode": {
"type": "boolean"
},
"prefill_tp": {
"type": "integer"
},
"prefill_ep": {
"type": "integer"
},
"prefill_dp_attention": {
"type": "boolean"
},
"prefill_num_workers": {
"type": "integer"
},
"decode_tp": {
"type": "integer"
},
"decode_ep": {
"type": "integer"
},
"decode_dp_attention": {
"type": "boolean"
},
"decode_num_workers": {
"type": "integer"
},
"num_prefill_gpu": {
"type": "integer",
"description": "Physical prefill chips; aggregate engines may mirror their single chip count in both role columns."
},
"num_decode_gpu": {
"type": "integer",
"description": "Physical decode chips, independent of logical TP and DP. Sum role counts only for disaggregated engines."
},
"benchmark_type": {
"type": "string"
},
"isl": {
"type": [
"number",
"null"
]
},
"osl": {
"type": [
"number",
"null"
]
},
"conc": {
"type": "integer"
},
"offload_mode": {
"type": "string"
},
"image": {
"type": [
"string",
"null"
]
},
"recipe_fingerprint": {
"type": [
"string",
"null"
]
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
},
"description": "Scalar metric map. Time metrics, including p99_itl and p99_tpot, are in seconds. p99_itl measures inter-token latency; p99_tpot measures per-request time per output token. Use the actual p99_itl field for an inter-token latency requirement, not the reciprocal of p99_intvty. Throughput metrics use tokens per second per GPU unless their name states otherwise; output_tput_per_gpu counts output tokens. On disaggregated rows tput_per_gpu divides by every chip, input_tput_per_gpu by prefill chips and output_tput_per_gpu by decode chips, for fixed-sequence and agentic runs alike. Agentic disaggregated rows also keep the original per-total-chip rates from the producer as input_tput_per_total_gpu and output_tput_per_total_gpu. Keys evolve independently; measured power / energy / GPU-telemetry keys are typed below.",
"properties": {
"power_valid": {
"type": "number",
"description": "Publication verdict: 1 = validated measurement window; 0 = failed validation — measured power/energy values are withheld from this row end-to-end, so treat any that remain as unreliable; absent = no validation verdict is available in this response. Absence alone establishes neither the reason, the measurement age, nor invalidity."
},
"power_metric_schema_version": {
"type": "number",
"description": "Power schema version. Version 2 defines every unprefixed joules_per_* field as whole-deployment energy, including on disaggregated runs."
},
"cpu_power_valid": {
"type": "number",
"description": "CPU-side leg verdict (NVL72 Grace sockets or single-node AMD EPYC RAPL packages), independent of power_valid: 1 = the CPU-side (and module, when present) keys were integrated over the validated window with full socket coverage; 0 = the leg failed and no CPU-side keys are emitted; absent = the run had no CPU-side leg. Ingest withholds the CPU-side keys unless this verdict is 1, and power_valid = 0 does not remove them. GPU power and energy never change because of this verdict."
},
"avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts during the measured load window."
},
"avg_total_gpu_power_w": {
"type": "number",
"description": "Mean total GPU power draw in watts across the measured deployment."
},
"total_gpu_energy_j": {
"type": "number",
"description": "Total GPU energy integrated over the measured deployment load window, in joules."
},
"p75_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p75_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"p90_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p90_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"joules_per_successful_query": {
"type": "number",
"description": "Whole-deployment energy in joules divided by successful requests."
},
"joules_per_output_token": {
"type": "number",
"description": "Energy per generated output token in joules; cluster-wide on schema-version-2 rows, including disaggregated runs."
},
"joules_per_total_token": {
"type": "number",
"description": "Total system energy divided by input plus output tokens; a workload-shape-fair view that does not treat prompt tokens as free."
},
"prefill_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across prefill workers; emitted only for deployments with distinct prefill and decode roles."
},
"decode_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across decode workers; emitted only for deployments with distinct prefill and decode roles."
},
"joules_per_input_token": {
"type": "number",
"description": "Energy per input token in joules; cluster-wide on schema-version-2 rows."
},
"prefill_joules_per_input_token": {
"type": "number",
"description": "Role-local prefill energy per input token in joules."
},
"decode_joules_per_output_token": {
"type": "number",
"description": "Role-local decode energy per generated output token in joules."
},
"avg_temp_c": {
"type": "number",
"description": "Mean per-GPU temperature in degrees Celsius during the load window."
},
"peak_temp_c": {
"type": "number",
"description": "Maximum instantaneous per-GPU temperature in degrees Celsius during the load window."
},
"avg_util_pct": {
"type": "number",
"description": "Mean per-GPU utilization percentage (0-100) during the load window."
},
"avg_mem_used_mb": {
"type": "number",
"description": "Mean per-GPU memory used in MB during the load window."
},
"avg_cpu_socket_power_w": {
"type": "number",
"description": "Mean over Grace sockets of each socket’s window-mean Grace-side power in watts (ACPI hwmon Grace Power Socket sensor, which NVIDIA defines only as the Grace socket power; modeled as the CPU and SysIO rails, LPDDR5X and socket regulation) over the same validated window as GPU energy. On single-node AMD rows (power_audit.cpu.sensor_kind = rapl_package) it is instead each EPYC package’s RAPL energy-counter power, which has no DRAM domain."
},
"avg_total_cpu_power_w": {
"type": "number",
"description": "Sum over every CPU socket in the deployment of window-mean socket power in watts (Grace side on NVL72, RAPL package on single-node AMD); dividing by avg_cpu_socket_power_w recovers the socket count (two per compute tray or AMD host)."
},
"total_cpu_energy_j": {
"type": "number",
"description": "CPU-side energy in joules (Grace side on NVL72, RAPL package on single-node AMD) integrated over the validated window across every socket. Additive to total_gpu_energy_j, never included in it."
},
"avg_total_module_power_w": {
"type": "number",
"description": "Sum over sockets of window-mean compute-module power in watts (Grace + its Blackwell GPUs + HBM + LPDDR5X + regulator loss, ACPI hwmon Module Power Socket sensor); emitted only when the module sensor exists on every socket. It already contains the GPU-board watts, so do not add avg_total_gpu_power_w to it."
},
"total_module_energy_j": {
"type": "number",
"description": "Compute-module energy in joules over the validated window across every socket; emitted with avg_total_module_power_w and includes the GPU energy."
},
"avg_total_cpu_rail_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean CPU core rail power in watts (ACPI CPU Power Socket sensor, no SysIO); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_rail_energy_j": {
"type": "number",
"description": "CPU core rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_rail_power_w and already included in total_cpu_energy_j."
},
"avg_total_cpu_sysio_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean SysIO rail power in watts (ACPI SysIO Power Socket sensor); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_sysio_energy_j": {
"type": "number",
"description": "SysIO rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_sysio_power_w and already included in total_cpu_energy_j."
}
}
},
"workers": {
"type": "array",
"items": {
"type": "object",
"properties": {
"role": {
"type": "string"
},
"worker_idx": {
"type": "integer"
},
"hosts": {
"type": "array",
"items": {
"type": "string"
}
},
"num_gpus": {
"type": "integer"
},
"avg_power_w": {
"type": "number"
},
"avg_temp_c": {
"type": "number"
},
"peak_temp_c": {
"type": "number"
},
"avg_util_pct": {
"type": "number"
},
"avg_mem_used_mb": {
"type": "number"
}
},
"required": [
"role",
"worker_idx",
"num_gpus",
"avg_power_w"
],
"additionalProperties": false
}
},
"power_invalid_reasons": {
"type": "array",
"items": {
"type": "string"
},
"description": "Producer snake_case reason codes explaining a withheld measurement, present when metrics.power_valid == 0. Absent on legacy rows and validated rows."
},
"power_audit": {
"type": "object",
"properties": {
"window_start_unix": {
"type": "number"
},
"window_end_unix": {
"type": "number"
},
"expected_gpu_count": {
"type": "integer"
},
"observed_gpu_count": {
"type": "integer"
},
"sample_count": {
"type": "integer"
},
"max_sample_gap_s": {
"type": "number"
},
"producer_sha": {
"type": [
"string",
"null"
]
},
"exporter_image_sha256": {
"type": [
"string",
"null"
]
},
"source": {
"type": "string"
},
"observed_gpu_ids": {
"type": "array",
"items": {
"type": "string"
}
},
"cpu": {
"type": "object",
"properties": {
"sensor_kind": {
"type": "string",
"enum": [
"module",
"grace_socket",
"dcgm_cpu_rail",
"rapl_package"
]
},
"source": {
"type": "string",
"description": "Collector that produced the CPU-side samples: acpi (hwmon), dcgm, or rapl (AMD EPYC package energy counter)."
},
"expected_sockets": {
"type": "integer"
},
"observed_sockets": {
"type": "integer"
},
"sample_row_count": {
"type": "integer"
},
"reason_codes": {
"type": "array",
"items": {
"type": "string"
}
}
},
"additionalProperties": true,
"description": "CPU-side leg provenance emitted alongside cpu_power_valid: the sensor kind that fed the CPU-side keys (on NVL72, module preferred over grace_socket over dcgm_cpu_rail; rapl_package on single-node AMD EPYC hosts), its collector, expected vs. observed sockets (two per compute tray or AMD host), stored sample rows, and the leg’s reason codes (empty when valid). Absent on runs without the CPU-side leg."
}
},
"additionalProperties": true,
"description": "Compact power measurement-window audit emitted alongside the power_valid verdict: window bounds, expected vs. observed GPU counts, sample statistics, and producer identity (producer_sha / exporter_image_sha256 are null for single-node telemetry without an srt-slurm producer). source names the validation document: a relative path within the source run artifact bundle or, on AgentX rows whose retained window ingest or backfill recovered, the alias power_validation_<stem>_conc<N>.json for LOGS/agentic/conc_<N>/power_validation.json in the power_audit_<stem> artifact (such rows carry only source and the window bounds); observed_gpu_ids contains producer device identifiers, which may be indices rather than physical UUIDs. cpu carries the CPU-side leg’s provenance (NVL72 or single-node AMD) when that leg ran. Present on valid and invalid rows when emitted; absence alone does not establish age or validity."
},
"date": {
"type": "string",
"format": "date"
},
"workflow_run_id": {
"type": "integer"
},
"run_started_at": {
"type": [
"string",
"null"
],
"format": "date-time"
},
"run_url": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"hardware",
"framework",
"model",
"precision",
"spec_method",
"disagg",
"is_multinode",
"prefill_tp",
"prefill_ep",
"prefill_dp_attention",
"prefill_num_workers",
"decode_tp",
"decode_ep",
"decode_dp_attention",
"decode_num_workers",
"num_prefill_gpu",
"num_decode_gpu",
"benchmark_type",
"isl",
"osl",
"conc",
"offload_mode",
"image",
"recipe_fingerprint",
"metrics",
"date",
"run_url"
],
"additionalProperties": false
}
}
},
"required": [
"x",
"y",
"observations"
],
"additionalProperties": false
}
},
"hinterland": {
"type": "array",
"items": {
"type": "object",
"properties": {
"x": {
"type": "number"
},
"y": {
"type": "number"
},
"observations": {
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"model": {
"type": "string"
},
"precision": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"disagg": {
"type": "boolean"
},
"is_multinode": {
"type": "boolean"
},
"prefill_tp": {
"type": "integer"
},
"prefill_ep": {
"type": "integer"
},
"prefill_dp_attention": {
"type": "boolean"
},
"prefill_num_workers": {
"type": "integer"
},
"decode_tp": {
"type": "integer"
},
"decode_ep": {
"type": "integer"
},
"decode_dp_attention": {
"type": "boolean"
},
"decode_num_workers": {
"type": "integer"
},
"num_prefill_gpu": {
"type": "integer",
"description": "Physical prefill chips; aggregate engines may mirror their single chip count in both role columns."
},
"num_decode_gpu": {
"type": "integer",
"description": "Physical decode chips, independent of logical TP and DP. Sum role counts only for disaggregated engines."
},
"benchmark_type": {
"type": "string"
},
"isl": {
"type": [
"number",
"null"
]
},
"osl": {
"type": [
"number",
"null"
]
},
"conc": {
"type": "integer"
},
"offload_mode": {
"type": "string"
},
"image": {
"type": [
"string",
"null"
]
},
"recipe_fingerprint": {
"type": [
"string",
"null"
]
},
"metrics": {
"type": "object",
"additionalProperties": {
"type": "number"
},
"description": "Scalar metric map. Time metrics, including p99_itl and p99_tpot, are in seconds. p99_itl measures inter-token latency; p99_tpot measures per-request time per output token. Use the actual p99_itl field for an inter-token latency requirement, not the reciprocal of p99_intvty. Throughput metrics use tokens per second per GPU unless their name states otherwise; output_tput_per_gpu counts output tokens. On disaggregated rows tput_per_gpu divides by every chip, input_tput_per_gpu by prefill chips and output_tput_per_gpu by decode chips, for fixed-sequence and agentic runs alike. Agentic disaggregated rows also keep the original per-total-chip rates from the producer as input_tput_per_total_gpu and output_tput_per_total_gpu. Keys evolve independently; measured power / energy / GPU-telemetry keys are typed below.",
"properties": {
"power_valid": {
"type": "number",
"description": "Publication verdict: 1 = validated measurement window; 0 = failed validation — measured power/energy values are withheld from this row end-to-end, so treat any that remain as unreliable; absent = no validation verdict is available in this response. Absence alone establishes neither the reason, the measurement age, nor invalidity."
},
"power_metric_schema_version": {
"type": "number",
"description": "Power schema version. Version 2 defines every unprefixed joules_per_* field as whole-deployment energy, including on disaggregated runs."
},
"cpu_power_valid": {
"type": "number",
"description": "CPU-side leg verdict (NVL72 Grace sockets or single-node AMD EPYC RAPL packages), independent of power_valid: 1 = the CPU-side (and module, when present) keys were integrated over the validated window with full socket coverage; 0 = the leg failed and no CPU-side keys are emitted; absent = the run had no CPU-side leg. Ingest withholds the CPU-side keys unless this verdict is 1, and power_valid = 0 does not remove them. GPU power and energy never change because of this verdict."
},
"avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts during the measured load window."
},
"avg_total_gpu_power_w": {
"type": "number",
"description": "Mean total GPU power draw in watts across the measured deployment."
},
"total_gpu_energy_j": {
"type": "number",
"description": "Total GPU energy integrated over the measured deployment load window, in joules."
},
"p75_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p75_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P75 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"p90_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts, divided by participating GPU count, over the validated benchmark window with piecewise-linear interpolation. Not a percentile of individual-device percentiles; absent when not measured."
},
"p90_total_gpu_power_w": {
"type": "number",
"description": "Time-weighted P90 of synchronized total GPU-board watts over the validated benchmark window with piecewise-linear interpolation. Accelerator boards only, not chassis or facility power."
},
"joules_per_successful_query": {
"type": "number",
"description": "Whole-deployment energy in joules divided by successful requests."
},
"joules_per_output_token": {
"type": "number",
"description": "Energy per generated output token in joules; cluster-wide on schema-version-2 rows, including disaggregated runs."
},
"joules_per_total_token": {
"type": "number",
"description": "Total system energy divided by input plus output tokens; a workload-shape-fair view that does not treat prompt tokens as free."
},
"prefill_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across prefill workers; emitted only for deployments with distinct prefill and decode roles."
},
"decode_avg_power_w": {
"type": "number",
"description": "Mean per-GPU power draw in watts across decode workers; emitted only for deployments with distinct prefill and decode roles."
},
"joules_per_input_token": {
"type": "number",
"description": "Energy per input token in joules; cluster-wide on schema-version-2 rows."
},
"prefill_joules_per_input_token": {
"type": "number",
"description": "Role-local prefill energy per input token in joules."
},
"decode_joules_per_output_token": {
"type": "number",
"description": "Role-local decode energy per generated output token in joules."
},
"avg_temp_c": {
"type": "number",
"description": "Mean per-GPU temperature in degrees Celsius during the load window."
},
"peak_temp_c": {
"type": "number",
"description": "Maximum instantaneous per-GPU temperature in degrees Celsius during the load window."
},
"avg_util_pct": {
"type": "number",
"description": "Mean per-GPU utilization percentage (0-100) during the load window."
},
"avg_mem_used_mb": {
"type": "number",
"description": "Mean per-GPU memory used in MB during the load window."
},
"avg_cpu_socket_power_w": {
"type": "number",
"description": "Mean over Grace sockets of each socket’s window-mean Grace-side power in watts (ACPI hwmon Grace Power Socket sensor, which NVIDIA defines only as the Grace socket power; modeled as the CPU and SysIO rails, LPDDR5X and socket regulation) over the same validated window as GPU energy. On single-node AMD rows (power_audit.cpu.sensor_kind = rapl_package) it is instead each EPYC package’s RAPL energy-counter power, which has no DRAM domain."
},
"avg_total_cpu_power_w": {
"type": "number",
"description": "Sum over every CPU socket in the deployment of window-mean socket power in watts (Grace side on NVL72, RAPL package on single-node AMD); dividing by avg_cpu_socket_power_w recovers the socket count (two per compute tray or AMD host)."
},
"total_cpu_energy_j": {
"type": "number",
"description": "CPU-side energy in joules (Grace side on NVL72, RAPL package on single-node AMD) integrated over the validated window across every socket. Additive to total_gpu_energy_j, never included in it."
},
"avg_total_module_power_w": {
"type": "number",
"description": "Sum over sockets of window-mean compute-module power in watts (Grace + its Blackwell GPUs + HBM + LPDDR5X + regulator loss, ACPI hwmon Module Power Socket sensor); emitted only when the module sensor exists on every socket. It already contains the GPU-board watts, so do not add avg_total_gpu_power_w to it."
},
"total_module_energy_j": {
"type": "number",
"description": "Compute-module energy in joules over the validated window across every socket; emitted with avg_total_module_power_w and includes the GPU energy."
},
"avg_total_cpu_rail_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean CPU core rail power in watts (ACPI CPU Power Socket sensor, no SysIO); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_rail_energy_j": {
"type": "number",
"description": "CPU core rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_rail_power_w and already included in total_cpu_energy_j."
},
"avg_total_cpu_sysio_power_w": {
"type": "number",
"description": "Sum over Grace sockets of window-mean SysIO rail power in watts (ACPI SysIO Power Socket sensor); emitted only when every socket reports a gap-free rail. A breakdown inside avg_total_cpu_power_w, not additive to it. NVL72 only."
},
"total_cpu_sysio_energy_j": {
"type": "number",
"description": "SysIO rail energy in joules over the validated window across every socket; emitted with avg_total_cpu_sysio_power_w and already included in total_cpu_energy_j."
}
}
},
"workers": {
"type": "array",
"items": {
"type": "object",
"properties": {
"role": {
"type": "string"
},
"worker_idx": {
"type": "integer"
},
"hosts": {
"type": "array",
"items": {
"type": "string"
}
},
"num_gpus": {
"type": "integer"
},
"avg_power_w": {
"type": "number"
},
"avg_temp_c": {
"type": "number"
},
"peak_temp_c": {
"type": "number"
},
"avg_util_pct": {
"type": "number"
},
"avg_mem_used_mb": {
"type": "number"
}
},
"required": [
"role",
"worker_idx",
"num_gpus",
"avg_power_w"
],
"additionalProperties": false
}
},
"power_invalid_reasons": {
"type": "array",
"items": {
"type": "string"
},
"description": "Producer snake_case reason codes explaining a withheld measurement, present when metrics.power_valid == 0. Absent on legacy rows and validated rows."
},
"power_audit": {
"type": "object",
"properties": {
"window_start_unix": {
"type": "number"
},
"window_end_unix": {
"type": "number"
},
"expected_gpu_count": {
"type": "integer"
},
"observed_gpu_count": {
"type": "integer"
},
"sample_count": {
"type": "integer"
},
"max_sample_gap_s": {
"type": "number"
},
"producer_sha": {
"type": [
"string",
"null"
]
},
"exporter_image_sha256": {
"type": [
"string",
"null"
]
},
"source": {
"type": "string"
},
"observed_gpu_ids": {
"type": "array",
"items": {
"type": "string"
}
},
"cpu": {
"type": "object",
"properties": {
"sensor_kind": {
"type": "string",
"enum": [
"module",
"grace_socket",
"dcgm_cpu_rail",
"rapl_package"
]
},
"source": {
"type": "string",
"description": "Collector that produced the CPU-side samples: acpi (hwmon), dcgm, or rapl (AMD EPYC package energy counter)."
},
"expected_sockets": {
"type": "integer"
},
"observed_sockets": {
"type": "integer"
},
"sample_row_count": {
"type": "integer"
},
"reason_codes": {
"type": "array",
"items": {
"type": "string"
}
}
},
"additionalProperties": true,
"description": "CPU-side leg provenance emitted alongside cpu_power_valid: the sensor kind that fed the CPU-side keys (on NVL72, module preferred over grace_socket over dcgm_cpu_rail; rapl_package on single-node AMD EPYC hosts), its collector, expected vs. observed sockets (two per compute tray or AMD host), stored sample rows, and the leg’s reason codes (empty when valid). Absent on runs without the CPU-side leg."
}
},
"additionalProperties": true,
"description": "Compact power measurement-window audit emitted alongside the power_valid verdict: window bounds, expected vs. observed GPU counts, sample statistics, and producer identity (producer_sha / exporter_image_sha256 are null for single-node telemetry without an srt-slurm producer). source names the validation document: a relative path within the source run artifact bundle or, on AgentX rows whose retained window ingest or backfill recovered, the alias power_validation_<stem>_conc<N>.json for LOGS/agentic/conc_<N>/power_validation.json in the power_audit_<stem> artifact (such rows carry only source and the window bounds); observed_gpu_ids contains producer device identifiers, which may be indices rather than physical UUIDs. cpu carries the CPU-side leg’s provenance (NVL72 or single-node AMD) when that leg ran. Present on valid and invalid rows when emitted; absence alone does not establish age or validity."
},
"date": {
"type": "string",
"format": "date"
},
"workflow_run_id": {
"type": "integer"
},
"run_started_at": {
"type": [
"string",
"null"
],
"format": "date-time"
},
"run_url": {
"type": [
"string",
"null"
]
}
},
"required": [
"id",
"hardware",
"framework",
"model",
"precision",
"spec_method",
"disagg",
"is_multinode",
"prefill_tp",
"prefill_ep",
"prefill_dp_attention",
"prefill_num_workers",
"decode_tp",
"decode_ep",
"decode_dp_attention",
"decode_num_workers",
"num_prefill_gpu",
"num_decode_gpu",
"benchmark_type",
"isl",
"osl",
"conc",
"offload_mode",
"image",
"recipe_fingerprint",
"metrics",
"date",
"run_url"
],
"additionalProperties": false
}
}
},
"required": [
"x",
"y",
"observations"
],
"additionalProperties": false
}
}
},
"required": [
"source_url",
"selection",
"counts",
"frontier",
"hinterland"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"source_url": "/api/v1/benchmarks?model=DeepSeek-R1-0528",
"selection": {
"model": "DeepSeek-R1-0528",
"rawModel": "dsr1",
"sequence": "1k/1k",
"xMetric": "median_intvty",
"yMetric": "output_tput_per_gpu",
"xDirection": "max",
"yDirection": "max",
"hardware": null,
"framework": null,
"precision": null
},
"counts": {
"returned": 0,
"selected": 0,
"eligible": 0,
"missing_or_nonfinite": 0
},
"frontier": [],
"hinterland": []
}滚动查看完整示例
400
参数缺失、无效、为空、重复或未知，快照选择条件冲突，或功率坐标轴未指定 strictV2。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Invalid xDirection"
}500
来源基准测试查询或边界计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/tco-feed计算 TCO 数据源
为电子表格中的 TCO 模型计算帕累托前沿上的吞吐量数据点或加权分数。所有评分假设都编码在 URL 中。CSV 会以扁平表格形式返回当前所选视图中的相同数据。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| model | query | string | 可选 | 数据库模型键或展示模型名称。 | dsv4 |
| workloads | query | CSV 工作负载列表 | 可选 | 以逗号分隔的 <isl>x<osl> token 组合。 | 1024x1024,8192x1024 |
| tiers | query | CSV 数字列表 | 可选 | 正数交互性目标，单位为每用户每秒输出 token。 | 30,50,75,100 |
| date | query | date | 可选 | 使用 YYYY-MM-DD 当日或之前的数据。省略则使用最新数据。 | 2026-08-08 |
| format | query | enum | 可选 | 响应编码。 | json |
| view | query | enum | 可选 | points 针对每个硬件、工作负载与档位的组合返回一行；scores 针对每个硬件返回一行。 | points |
| weights | query | CSV 数字列表 | 可选 | 仅用于 scores。每个档位一个非负权重，并归一化为总和 1。 | 0.35,0.4,0.2,0.05 |
| workload_weights | query | CSV 数字列表 | 可选 | 仅用于 scores。每个工作负载对应一个非负权重；所有权重归一化后总和为 1，默认使用等权重。 | 0.5,0.5 |
| alpha | query | number | 可选 | 仅用于 scores。输入 token 价值比，范围为 [0, 10]。 | 0.25 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/tco-feed?model=dsv4&workloads=1024x1024,8192x1024&tiers=30,50,75,100&view=points&format=json'响应
TcoFeed200
所选 points 或 scores 视图的响应对象。
媒体类型: application/json响应结构
{
"oneOf": [
{
"type": "object",
"properties": {
"model": {
"type": "string"
},
"db_model_keys": {
"type": "array",
"items": {
"type": "string"
}
},
"date": {
"type": [
"string",
"null"
],
"format": "date"
},
"workloads": {
"type": "array",
"items": {
"type": "string"
}
},
"tiers": {
"type": "array",
"items": {
"type": "number"
}
},
"rows": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"model",
"db_model_keys",
"date",
"workloads",
"tiers",
"rows"
],
"additionalProperties": false
},
{
"type": "object",
"properties": {
"model": {
"type": "string"
},
"db_model_keys": {
"type": "array",
"items": {
"type": "string"
}
},
"date": {
"type": [
"string",
"null"
],
"format": "date"
},
"workloads": {
"type": "array",
"items": {
"type": "string"
}
},
"tiers": {
"type": "array",
"items": {
"type": "number"
}
},
"weights": {
"type": "array",
"items": {
"type": "number"
}
},
"workload_weights": {
"type": "array",
"items": {
"type": "number"
}
},
"alpha": {
"type": "number"
},
"rows": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"model",
"db_model_keys",
"date",
"workloads",
"tiers",
"weights",
"workload_weights",
"alpha",
"rows"
],
"additionalProperties": false
}
]
}滚动查看完整示例
响应示例
{
"model": "dsv4",
"db_model_keys": [
"dsv4"
],
"date": null,
"workloads": [
"1024x1024"
],
"tiers": [
50
],
"rows": [
{
"hardware": "h200_sxm",
"workload": "1024x1024",
"tier": 50,
"output_tput_per_gpu": 118.2,
"boundary": "interpolated",
"is_interpolated": true,
"frontier_points": 3,
"frontier_min_interactivity": 30,
"frontier_max_interactivity": 75,
"latest_date": "2026-08-08",
"oldest_frontier_date": "2026-08-06",
"evidence_date": {
"from": "2026-08-06",
"to": "2026-08-08"
}
}
]
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
hardware,workload,tier,output_tput_per_gpu,boundary,frontier_points,frontier_min_interactivity,frontier_max_interactivity,latest_date,oldest_frontier_date
h200_sxm,1024x1024,50,118.2,interpolated,3,30,75,2026-08-08,2026-08-06
400
模型、日期、视图、格式、工作负载、档位、权重或 alpha 值无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Invalid tiers: expected comma-separated positive numbers"
}500
TCO 计算失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/submissions读取提交覆盖
返回配置级提交汇总和每日硬件提交量。
参数
无需参数
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/submissions'响应
Submissions200
提交汇总和数量数组。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"summary": {
"type": "array",
"items": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"precision": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"disagg": {
"type": "boolean"
},
"is_multinode": {
"type": "boolean"
},
"num_prefill_gpu": {
"type": "integer",
"description": "Physical prefill chips; aggregate engines may mirror their single chip count in both role columns."
},
"num_decode_gpu": {
"type": "integer",
"description": "Physical decode chips, independent of logical TP and DP. Sum role counts only for disaggregated engines."
},
"prefill_tp": {
"type": "integer"
},
"prefill_ep": {
"type": "integer"
},
"decode_tp": {
"type": "integer"
},
"decode_ep": {
"type": "integer"
},
"date": {
"type": "string",
"format": "date"
},
"total_datapoints": {
"type": "integer"
},
"distinct_sequences": {
"type": "integer"
},
"distinct_concurrencies": {
"type": "integer"
},
"max_concurrency": {
"type": "integer"
},
"image": {
"type": [
"string",
"null"
]
}
},
"required": [
"model",
"hardware",
"framework",
"precision",
"spec_method",
"disagg",
"is_multinode",
"num_prefill_gpu",
"num_decode_gpu",
"prefill_tp",
"prefill_ep",
"decode_tp",
"decode_ep",
"date",
"total_datapoints",
"distinct_sequences",
"distinct_concurrencies",
"max_concurrency",
"image"
],
"additionalProperties": false
}
},
"volume": {
"type": "array",
"items": {
"type": "object",
"properties": {
"date": {
"type": "string",
"format": "date"
},
"hardware": {
"type": "string"
},
"datapoints": {
"type": "integer"
}
},
"required": [
"date",
"hardware",
"datapoints"
],
"additionalProperties": false
}
}
},
"required": [
"summary",
"volume"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"summary": [
{
"model": "dsr1",
"hardware": "h200_sxm",
"framework": "vllm",
"precision": "fp8",
"spec_method": "none",
"disagg": false,
"is_multinode": false,
"num_prefill_gpu": 0,
"num_decode_gpu": 8,
"prefill_tp": 1,
"prefill_ep": 1,
"decode_tp": 8,
"decode_ep": 1,
"date": "2026-08-08",
"total_datapoints": 24,
"distinct_sequences": 3,
"distinct_concurrencies": 8,
"max_concurrency": 256,
"image": "vllm/vllm-openai:v0.10.2"
}
],
"volume": [
{
"date": "2026-08-08",
"hardware": "h200_sxm",
"datapoints": 24
}
]
}滚动查看完整示例
500
提交查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/framework-releases读取最新框架版本
返回 vLLM 和 SGLang 最近 20 个 GitHub release 中版本号最高的非草稿、非预发布引擎版本（标签以 v 开头）及其发布时间。commits 列出当前基准测试镜像中不带版本号的 nightly commit，以及 GitHub 比较该 commit 与此 release 的结果；behind 表示该 commit 是此 release 的祖先提交。GitHub 无法解析的 commit 不会列出。null 表示上游查询没有可用版本。
参数
无需参数
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/framework-releases'响应
FrameworkReleases200
框架键映射到版本标签、ISO 8601 发布时间和 nightly commit 比较结果；无可用版本时为 null。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": {
"oneOf": [
{
"type": "object",
"properties": {
"tag": {
"type": "string"
},
"publishedAt": {
"type": "string",
"format": "date-time"
},
"commits": {
"type": "object",
"additionalProperties": {
"type": "string",
"enum": [
"ahead",
"behind",
"diverged",
"identical"
]
}
}
},
"required": [
"tag",
"publishedAt",
"commits"
],
"additionalProperties": false
},
{
"type": "null"
}
]
}
}滚动查看完整示例
响应示例
{
"vllm": {
"tag": "v0.10.2",
"publishedAt": "2026-08-01T12:00:00Z",
"commits": {
"af7f9488c2210d67": "behind",
"e68efc685ab1": "diverged"
}
},
"sglang": {
"tag": "v0.4.10",
"publishedAt": "2026-07-28T09:30:00Z",
"commits": {}
}
}滚动查看完整示例
500
版本查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/latest-images读取最新运行镜像
返回每个基准配置和序列最近观测到的容器镜像。
参数
无需参数
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/latest-images'响应
LatestImageRows200
最新镜像行。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "object",
"properties": {
"model": {
"type": "string"
},
"hardware": {
"type": "string"
},
"framework": {
"type": "string"
},
"precision": {
"type": "string"
},
"spec_method": {
"type": "string"
},
"disagg": {
"type": "boolean"
},
"isl": {
"type": [
"number",
"null"
]
},
"osl": {
"type": [
"number",
"null"
]
},
"benchmark_type": {
"type": "string"
},
"image": {
"type": "string"
},
"date": {
"type": "string",
"format": "date"
}
},
"required": [
"model",
"hardware",
"framework",
"precision",
"spec_method",
"disagg",
"isl",
"osl",
"benchmark_type",
"image",
"date"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
[
{
"model": "dsr1",
"hardware": "h200_sxm",
"framework": "vllm",
"precision": "fp8",
"spec_method": "none",
"disagg": false,
"isl": 1024,
"osl": 1024,
"benchmark_type": "single_turn",
"image": "vllm/vllm-openai:v0.10.2",
"date": "2026-08-08"
},
{
"model": "dsr1",
"hardware": "h200_sxm",
"framework": "vllm",
"precision": "fp8",
"spec_method": "none",
"disagg": false,
"isl": null,
"osl": null,
"benchmark_type": "agentic_traces",
"image": "vllm/vllm-openai:v0.10.2",
"date": "2026-08-08"
}
]滚动查看完整示例
500
镜像查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}数据集
数据集目录、元数据、会话索引与会话结构。
GET稳定/api/v1/datasets列出已导入的数据集
返回数据集目录卡片，不包含较大的 chart_data 字段。
参数
无需参数
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/datasets'响应
DatasetRecords200
数据集目录记录。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "object",
"properties": {
"id": {
"type": "string"
},
"slug": {
"type": "string"
},
"label": {
"type": "string"
},
"variant": {
"type": "string"
},
"description": {
"type": [
"string",
"null"
]
},
"hf_url": {
"type": [
"string",
"null"
]
},
"license": {
"type": [
"string",
"null"
]
},
"conversation_count": {
"type": "integer"
},
"summary": {
"type": "object",
"additionalProperties": true
},
"ingested_at": {
"type": "string",
"format": "date-time"
}
},
"required": [
"id",
"slug",
"label",
"variant",
"description",
"hf_url",
"license",
"conversation_count",
"summary",
"ingested_at"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
[
{
"id": "ds_01",
"slug": "cc-traces-weka",
"label": "CC Traces Weka",
"variant": "default",
"description": "Agentic coding traces",
"hf_url": "https://huggingface.co/datasets/example/cc-traces-weka",
"license": "Apache-2.0",
"conversation_count": 1200,
"summary": {
"totalIn": 8200000,
"totalOut": 1700000
},
"ingested_at": "2026-08-08T03:00:00Z"
}
]滚动查看完整示例
500
数据集目录查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/datasets/{slug}读取数据集详情
返回一条数据集目录记录及其预计算的 chart_data 分布。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| slug | path | string | 必填 | 目录中的数据集 slug。 | cc-traces-weka |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/datasets/cc-traces-weka'响应
DatasetDetail200
包含 chart_data 的数据集元数据。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"id": {
"type": "string"
},
"slug": {
"type": "string"
},
"label": {
"type": "string"
},
"variant": {
"type": "string"
},
"description": {
"type": [
"string",
"null"
]
},
"hf_url": {
"type": [
"string",
"null"
]
},
"license": {
"type": [
"string",
"null"
]
},
"conversation_count": {
"type": "integer"
},
"summary": {
"type": "object",
"additionalProperties": true
},
"ingested_at": {
"type": "string",
"format": "date-time"
},
"chart_data": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"id",
"slug",
"label",
"variant",
"description",
"hf_url",
"license",
"conversation_count",
"summary",
"ingested_at",
"chart_data"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"id": "ds_01",
"slug": "cc-traces-weka",
"label": "CC Traces Weka",
"variant": "default",
"description": "Agentic coding traces",
"hf_url": null,
"license": "Apache-2.0",
"conversation_count": 1200,
"summary": {},
"ingested_at": "2026-08-08T03:00:00Z",
"chart_data": {
"tokens": {
"bins": [
0,
1000,
2000
],
"counts": [
140,
320
]
}
}
}滚动查看完整示例
404
没有使用此 slug 的数据集。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
数据集查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/datasets/{slug}/conversations列出数据集会话
返回可搜索、排序和分页的会话索引。只包含计数，不包含完整会话结构。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| slug | path | string | 必填 | 目录中的数据集 slug。 | cc-traces-weka |
| search | query | string | 可选 | 按会话 ID 搜索，会先去除首尾空格，最多 100 个字符。 | trace-018 |
| limit | query | integer | 可选 | 每页数量；小于 1 时按 1 处理，大于 200 时按 200 处理。 | 50 |
| offset | query | integer | 可选 | 从 0 开始的行偏移。负数会变为 0。 | 0 |
| sort | query | enum | 可选 | 按 tokens、turns、subagents 或 id 排序。未知值回退到 tokens。 | tokens |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/datasets/cc-traces-weka/conversations?limit=50&offset=0&sort=tokens'响应
ConversationList200
总数和会话索引项。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"total": {
"type": "integer"
},
"items": {
"type": "array",
"items": {
"type": "object",
"properties": {
"conv_id": {
"type": "string"
},
"models": {
"type": "array",
"items": {
"type": "string"
}
},
"num_turns": {
"type": "integer"
},
"num_subagent_groups": {
"type": "integer"
},
"total_in": {
"type": "integer"
},
"total_out": {
"type": "integer"
},
"total_cached": {
"type": "integer"
}
},
"required": [
"conv_id",
"models",
"num_turns",
"num_subagent_groups",
"total_in",
"total_out",
"total_cached"
],
"additionalProperties": false
}
}
},
"required": [
"total",
"items"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"total": 1200,
"items": [
{
"conv_id": "trace-018",
"models": [
"claude-sonnet-4"
],
"num_turns": 42,
"num_subagent_groups": 3,
"total_in": 18200,
"total_out": 4200,
"total_cached": 9600
}
]
}滚动查看完整示例
400
search 超过 100 个字符。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "search too long"
}404
没有使用此 slug 的数据集。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
会话查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET稳定/api/v1/datasets/{slug}/conversations/{convId}读取会话结构
返回单个会话及可用于火焰图的嵌套结构。App Router 对每个路径值解码一次。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| slug | path | string | 必填 | 目录中的数据集 slug。 | cc-traces-weka |
| convId | path | string | 必填 | 会话 ID，必须与会话索引中列出的值完全一致。 | trace-018 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/datasets/cc-traces-weka/conversations/trace-018'响应
ConversationDetail200
会话计数和嵌套结构。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"conv_id": {
"type": "string"
},
"models": {
"type": "array",
"items": {
"type": "string"
}
},
"num_turns": {
"type": "integer"
},
"num_subagent_groups": {
"type": "integer"
},
"total_in": {
"type": "integer"
},
"total_out": {
"type": "integer"
},
"total_cached": {
"type": "integer"
},
"structure": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"conv_id",
"models",
"num_turns",
"num_subagent_groups",
"total_in",
"total_out",
"total_cached",
"structure"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"conv_id": "trace-018",
"models": [
"claude-sonnet-4"
],
"num_turns": 42,
"num_subagent_groups": 3,
"total_in": 18200,
"total_out": 4200,
"total_cached": 9600,
"structure": {
"name": "trace-018",
"children": []
}
}滚动查看完整示例
404
数据集或会话不存在。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
会话查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}CollectiveX
带版本的集合通信基准测试结果，每次结果上传对应一次运行。
GET测试版/api/v1/collectivex/latest读取最新 CollectiveX 数据集
返回最新已存储运行的带版本中立数据集。每次运行对应一次上传的 CollectiveX 结果目录，叠加在上一次最新运行之上（上传的 case 会替换该 case 已存储的结果），且至少包含一个已测量结果：run_id 为上传时间的 Unix 秒级时间戳（必要时递增至最新已存储运行之后），run_attempt 为 1，run.source_sha 为测量这些上传结果所用的 collectivex/ 提交。run.covered_skus 只列出有已测量结果的资源池。运行在运维人员上传时导入，读取时不会导入。 在 series[].points[] 中，components.*.latency_us、components.*.activation_data_rate_gbps_at_latency_percentile、components.*.payload_data_rate_gbps_at_latency_percentile 与 roundtrip_token_rate_at_latency_percentile 包含 p50、p90、p95 和 p99；若测试工具因过多样本未对齐或被主机阻塞而未发布尾部延迟，p90、p95 和 p99 为 null，p50 始终为数值。series[].library_version 为该 case 运行的后端构建版本（结果未记录时为 null）。可选的 swap_blocks 结果包含校验通过的复制延迟（微秒）和有效载荷 GB/s（字节数仅计算一次），并保留 block_bytes、num_blocks、运行环境来源及未测量组合数。多池扫描分别保留各 GPU 池的独立结果。 每个序列带有 maturity（candidate 表示目前没有推理引擎可以选用该配置），components.pair_period 为连续成对的稳态周期，即主延迟指标。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| version | query | enum | 必填 | CollectiveX 契约版本。 | 1 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/collectivex/latest?version=1'响应
CollectiveXDataset200
最新的 CollectiveX 运行记录、覆盖情况、序列，以及可能存在的 KV 测试用例。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"version": {
"type": "integer"
},
"run": {
"type": "object",
"properties": {
"run_id": {
"type": "string"
},
"run_attempt": {
"type": "integer"
},
"generated_at": {
"type": "string",
"format": "date-time"
},
"conclusion": {
"type": [
"string",
"null"
]
},
"source_sha": {
"type": "string"
},
"source_repository": {
"type": "string"
},
"requested_cases": {
"type": "integer"
},
"terminal_cases": {
"type": "integer"
},
"measured_cases": {
"type": "integer"
},
"unsupported_cases": {
"type": "integer"
},
"failed_cases": {
"type": "integer"
},
"requested_points": {
"type": "integer"
},
"terminal_points": {
"type": "integer"
},
"measured_points": {
"type": "integer"
},
"covered_skus": {
"type": "array",
"items": {
"type": "string"
}
},
"swap_requested_cases": {
"type": "integer"
},
"swap_measured_cases": {
"type": "integer"
},
"kv_requested_cases": {
"type": "integer"
},
"kv_measured_cases": {
"type": "integer"
}
},
"required": [
"run_id",
"run_attempt",
"generated_at",
"conclusion",
"source_sha",
"requested_cases",
"terminal_cases",
"measured_cases",
"unsupported_cases",
"failed_cases",
"requested_points",
"terminal_points",
"measured_points",
"covered_skus"
],
"additionalProperties": false
},
"coverage": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"series": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"kv": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"swap_blocks": {
"type": "array",
"items": {
"type": "object",
"properties": {
"result_id": {
"type": "string"
},
"sku": {
"type": "string"
},
"runtime": {
"type": "object",
"additionalProperties": true
},
"timing": {
"type": "string"
},
"warmup": {
"type": "integer"
},
"iterations": {
"type": "integer"
},
"max_payload_bytes": {
"oneOf": [
{
"type": "integer"
},
{
"type": "null"
}
]
},
"skipped_points": {
"type": "integer"
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"direction": {
"type": "string",
"enum": [
"h2d",
"d2h",
"d2d"
]
},
"layout": {
"type": "string",
"enum": [
"contiguous",
"random"
]
},
"block_bytes": {
"type": "integer"
},
"num_blocks": {
"type": "integer"
},
"payload_bytes": {
"type": "integer"
},
"seed": {
"type": "integer"
},
"host_memory": {
"type": "string"
},
"api": {
"type": "string"
},
"sample_count": {
"type": "integer"
},
"latency_us": {
"type": "object",
"additionalProperties": true
},
"payload_gbps_at_latency_percentile": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"direction",
"layout",
"block_bytes",
"num_blocks",
"payload_bytes",
"seed",
"host_memory",
"api",
"sample_count",
"latency_us",
"payload_gbps_at_latency_percentile"
],
"additionalProperties": false
}
}
},
"required": [
"result_id",
"sku",
"runtime",
"timing",
"warmup",
"iterations",
"max_payload_bytes",
"skipped_points",
"points"
],
"additionalProperties": false
}
}
},
"required": [
"version",
"run",
"coverage",
"series"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"version": 1,
"run": {
"run_id": "1792089351",
"run_attempt": 1,
"generated_at": "2026-10-15T18:35:51Z",
"conclusion": "success",
"source_sha": "e65c9045320c78f1fb86200802aceafa0c6310d8",
"requested_cases": 12,
"terminal_cases": 12,
"measured_cases": 10,
"unsupported_cases": 2,
"failed_cases": 0,
"requested_points": 48,
"terminal_points": 48,
"measured_points": 40,
"covered_skus": [
"h200_sxm"
],
"kv_requested_cases": 4,
"kv_measured_cases": 4
},
"coverage": [],
"series": [],
"kv": [],
"swap_blocks": []
}滚动查看完整示例
400
version 缺失或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown version"
}404
该版本下没有已存储的运行。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
已存储运行查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/collectivex/runs列出 CollectiveX 运行
按上传时间从新到旧返回全部已存储的运行汇总。只有至少包含一个已测量结果的上传才构成运行，covered_skus 只列出有已测量结果的资源池。运行在读取前即已导入，因此 discovery_complete 始终为 true；保留该字段是为了兼容现有客户端。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| version | query | enum | 必填 | CollectiveX 契约版本。 | 1 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/collectivex/runs?version=1'响应
CollectiveXRunList200
版本、运行汇总以及 discovery_complete（始终为 true）。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"version": {
"type": "integer"
},
"runs": {
"type": "array",
"items": {
"type": "object",
"properties": {
"run_id": {
"type": "string"
},
"run_attempt": {
"type": "integer"
},
"generated_at": {
"type": "string",
"format": "date-time"
},
"conclusion": {
"type": [
"string",
"null"
]
},
"source_repository": {
"type": "string"
},
"covered_skus": {
"type": "array",
"items": {
"type": "string"
}
},
"requested_cases": {
"type": "integer"
},
"measured_cases": {
"type": "integer"
},
"requested_points": {
"type": "integer"
},
"terminal_points": {
"type": "integer"
},
"terminal_counts": {
"type": "object",
"properties": {
"measured": {
"type": "integer"
},
"unsupported": {
"type": "integer"
},
"failed": {
"type": "integer"
}
},
"required": [
"measured",
"unsupported",
"failed"
],
"additionalProperties": false
},
"swap_cases": {
"type": "object",
"properties": {
"requested": {
"type": "integer"
},
"measured": {
"type": "integer"
}
},
"required": [
"requested",
"measured"
],
"additionalProperties": false
},
"kv_cases": {
"type": "object",
"properties": {
"requested": {
"type": "integer"
},
"measured": {
"type": "integer"
}
},
"required": [
"requested",
"measured"
],
"additionalProperties": false
}
},
"required": [
"run_id",
"run_attempt",
"generated_at",
"conclusion",
"covered_skus",
"requested_cases",
"measured_cases",
"requested_points",
"terminal_points",
"terminal_counts"
],
"additionalProperties": false
}
},
"discovery_complete": {
"type": "boolean"
}
},
"required": [
"version",
"runs",
"discovery_complete"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"version": 1,
"runs": [
{
"run_id": "1792089351",
"run_attempt": 1,
"generated_at": "2026-10-15T18:35:51Z",
"conclusion": "success",
"covered_skus": [
"h200_sxm"
],
"requested_cases": 12,
"measured_cases": 10,
"requested_points": 48,
"terminal_points": 48,
"terminal_counts": {
"measured": 40,
"unsupported": 8,
"failed": 0
},
"kv_cases": {
"requested": 4,
"measured": 4
}
}
],
"discovery_complete": true
}滚动查看完整示例
400
version 缺失或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown version"
}500
已存储运行列表查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/collectivex/changelog列出 CollectiveX 变更记录
按从新到旧返回已关联到已存储运行的 InferenceX collectivex/perf-changelog.yaml 记录：每个运行与记录一行，并附上该运行改变的每个被选中用例在之前和之后的 pair-period p50。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| version | query | enum | 必填 | CollectiveX 契约版本。 | 1 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/collectivex/changelog?version=1'响应
CollectiveXChangelog200
版本与已关联的变更记录。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"version": {
"type": "integer"
},
"entries": {
"type": "array",
"items": {
"type": "object",
"properties": {
"run_id": {
"type": "string"
},
"entry_index": {
"type": "integer"
},
"pr_number": {
"type": "integer"
},
"pr_link": {
"type": "string"
},
"date": {
"type": "string"
},
"pools": {
"type": "array",
"items": {
"type": "string"
}
},
"selectors": {
"type": "object",
"additionalProperties": true
},
"description": {
"type": "string"
},
"changes": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"run_id",
"entry_index",
"pr_number",
"pr_link",
"date",
"pools",
"selectors",
"description",
"changes"
],
"additionalProperties": false
}
}
},
"required": [
"version",
"entries"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"version": 1,
"entries": [
{
"run_id": "1793344844",
"entry_index": 14,
"pr_number": 4674,
"pr_link": "https://github.com/SemiAnalysisAI/InferenceX/pull/4674",
"date": "2026-10-30",
"pools": [
"mi300x"
],
"selectors": {
"backends": [
"uccl-ep"
],
"modes": [
"low-latency"
],
"precisions": [
"fp8"
]
},
"description": "Model uccl-ep's in-kernel FP8 cast exactly in low-latency mode.",
"changes": []
}
]
}滚动查看完整示例
400
version 缺失或不受支持。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown version"
}500
变更记录查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/collectivex/runs/{runId}读取 CollectiveX 运行
以带版本的 CollectiveX 数据集返回一次已存储的运行。运行 ID 为上传时间的 Unix 秒级时间戳（必要时递增至最新已存储运行之后），与 /api/v1/collectivex/runs 列出的一致。 在 series[].points[] 中，components.*.latency_us、components.*.activation_data_rate_gbps_at_latency_percentile、components.*.payload_data_rate_gbps_at_latency_percentile 与 roundtrip_token_rate_at_latency_percentile 包含 p50、p90、p95 和 p99；若测试工具因过多样本未对齐或被主机阻塞而未发布尾部延迟，p90、p95 和 p99 为 null，p50 始终为数值。 每个序列带有 maturity（candidate 表示目前没有推理引擎可以选用该配置），components.pair_period 为连续成对的稳态周期，即主延迟指标。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| runId | path | integer | 必填 | 正整数运行 ID：上传时间的 Unix 秒级时间戳，小于 10^10。 | 1792089351 |
| version | query | enum | 必填 | CollectiveX 契约版本。 | 1 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/collectivex/runs/1792089351?version=1'响应
CollectiveXDataset200
请求的 CollectiveX 数据集。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"version": {
"type": "integer"
},
"run": {
"type": "object",
"properties": {
"run_id": {
"type": "string"
},
"run_attempt": {
"type": "integer"
},
"generated_at": {
"type": "string",
"format": "date-time"
},
"conclusion": {
"type": [
"string",
"null"
]
},
"source_sha": {
"type": "string"
},
"source_repository": {
"type": "string"
},
"requested_cases": {
"type": "integer"
},
"terminal_cases": {
"type": "integer"
},
"measured_cases": {
"type": "integer"
},
"unsupported_cases": {
"type": "integer"
},
"failed_cases": {
"type": "integer"
},
"requested_points": {
"type": "integer"
},
"terminal_points": {
"type": "integer"
},
"measured_points": {
"type": "integer"
},
"covered_skus": {
"type": "array",
"items": {
"type": "string"
}
},
"swap_requested_cases": {
"type": "integer"
},
"swap_measured_cases": {
"type": "integer"
},
"kv_requested_cases": {
"type": "integer"
},
"kv_measured_cases": {
"type": "integer"
}
},
"required": [
"run_id",
"run_attempt",
"generated_at",
"conclusion",
"source_sha",
"requested_cases",
"terminal_cases",
"measured_cases",
"unsupported_cases",
"failed_cases",
"requested_points",
"terminal_points",
"measured_points",
"covered_skus"
],
"additionalProperties": false
},
"coverage": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"series": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"kv": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"swap_blocks": {
"type": "array",
"items": {
"type": "object",
"properties": {
"result_id": {
"type": "string"
},
"sku": {
"type": "string"
},
"runtime": {
"type": "object",
"additionalProperties": true
},
"timing": {
"type": "string"
},
"warmup": {
"type": "integer"
},
"iterations": {
"type": "integer"
},
"max_payload_bytes": {
"oneOf": [
{
"type": "integer"
},
{
"type": "null"
}
]
},
"skipped_points": {
"type": "integer"
},
"points": {
"type": "array",
"items": {
"type": "object",
"properties": {
"direction": {
"type": "string",
"enum": [
"h2d",
"d2h",
"d2d"
]
},
"layout": {
"type": "string",
"enum": [
"contiguous",
"random"
]
},
"block_bytes": {
"type": "integer"
},
"num_blocks": {
"type": "integer"
},
"payload_bytes": {
"type": "integer"
},
"seed": {
"type": "integer"
},
"host_memory": {
"type": "string"
},
"api": {
"type": "string"
},
"sample_count": {
"type": "integer"
},
"latency_us": {
"type": "object",
"additionalProperties": true
},
"payload_gbps_at_latency_percentile": {
"type": "object",
"additionalProperties": true
}
},
"required": [
"direction",
"layout",
"block_bytes",
"num_blocks",
"payload_bytes",
"seed",
"host_memory",
"api",
"sample_count",
"latency_us",
"payload_gbps_at_latency_percentile"
],
"additionalProperties": false
}
}
},
"required": [
"result_id",
"sku",
"runtime",
"timing",
"warmup",
"iterations",
"max_payload_bytes",
"skipped_points",
"points"
],
"additionalProperties": false
}
}
},
"required": [
"version",
"run",
"coverage",
"series"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"version": 1,
"run": {
"run_id": "1792089351",
"run_attempt": 1,
"generated_at": "2026-10-15T18:35:51Z",
"conclusion": "success",
"source_sha": "e65c9045320c78f1fb86200802aceafa0c6310d8",
"requested_cases": 12,
"terminal_cases": 12,
"measured_cases": 10,
"unsupported_cases": 2,
"failed_cases": 0,
"requested_points": 48,
"terminal_points": 48,
"measured_points": 40,
"covered_skus": [
"h200_sxm"
],
"kv_requested_cases": 4,
"kv_measured_cases": 4
},
"coverage": [],
"series": [],
"kv": [],
"swap_blocks": []
}滚动查看完整示例
400
version 或 runId 无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Unknown version or run id"
}404
该运行未存储。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
运行查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}诊断读取
按结果提供跟踪、缓存、请求、同组结果和服务器指标诊断。
GET测试版/api/v1/agentic-aggregates读取智能体汇总百分位
按基准结果 ID 返回 ISL、OSL、KV 缓存利用率和前缀缓存命中率百分位。ID 会去重，最多接受 200 个。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| ids | query | 逗号分隔的整数 | 必填 | 1 到 200 个正整数基准结果 ID。 | 421,422 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/agentic-aggregates?ids=421,422'响应
AgenticAggregateMap200
结果 ID 映射到汇总百分位或 null 指标组。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"isl": {
"oneOf": [
{
"type": "object",
"properties": {
"mean": {
"type": "number"
},
"p50": {
"type": "number"
},
"p75": {
"type": "number"
},
"p90": {
"type": "number"
},
"p95": {
"type": "number"
},
"p99": {
"type": "number"
},
"n": {
"type": "integer"
}
},
"required": [
"mean",
"p50",
"p75",
"p90",
"p95",
"p99",
"n"
],
"additionalProperties": false
},
{
"type": "null"
}
]
},
"osl": {
"oneOf": [
{
"type": "object",
"properties": {
"mean": {
"type": "number"
},
"p50": {
"type": "number"
},
"p75": {
"type": "number"
},
"p90": {
"type": "number"
},
"p95": {
"type": "number"
},
"p99": {
"type": "number"
},
"n": {
"type": "integer"
}
},
"required": [
"mean",
"p50",
"p75",
"p90",
"p95",
"p99",
"n"
],
"additionalProperties": false
},
{
"type": "null"
}
]
},
"kvCacheUtil": {
"oneOf": [
{
"type": "object",
"properties": {
"mean": {
"type": "number"
},
"p50": {
"type": "number"
},
"p75": {
"type": "number"
},
"p90": {
"type": "number"
},
"p95": {
"type": "number"
},
"p99": {
"type": "number"
},
"n": {
"type": "integer"
}
},
"required": [
"mean",
"p50",
"p75",
"p90",
"p95",
"p99",
"n"
],
"additionalProperties": false
},
{
"type": "null"
}
]
},
"prefixCacheHitRate": {
"oneOf": [
{
"type": "object",
"properties": {
"mean": {
"type": "number"
},
"p50": {
"type": "number"
},
"p75": {
"type": "number"
},
"p90": {
"type": "number"
},
"p95": {
"type": "number"
},
"p99": {
"type": "number"
},
"n": {
"type": "integer"
}
},
"required": [
"mean",
"p50",
"p75",
"p90",
"p95",
"p99",
"n"
],
"additionalProperties": false
},
{
"type": "null"
}
]
}
},
"required": [
"id",
"isl",
"osl",
"kvCacheUtil",
"prefixCacheHitRate"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
{
"421": {
"id": 421,
"isl": {
"mean": 18320,
"p50": 16440,
"p75": 20110,
"p90": 24880,
"p95": 27940,
"p99": 31900,
"n": 512
},
"osl": null,
"kvCacheUtil": null,
"prefixCacheHitRate": null
}
}滚动查看完整示例
400
ids 缺失、格式错误或超过 200 个唯一 ID。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "ids query param is required"
}500
汇总查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/benchmark-siblings读取同组基准点
返回基准 SKU，以及同一硬件、框架、模型、精度、方法、基准类型和工作流运行中的全部点。同组结果的工作负载可能不同。is_current 仅在结果 ID 与请求的 ID 相同时为 true；false 不表示结果已过时、无效或已被替代。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| id | query | integer | 必填 | 基准结果 ID，必须为正安全整数。 | 421 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/benchmark-siblings?id=421'响应
BenchmarkSiblings200
SKU 元数据和同组导航行。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"sku": {
"type": "object",
"additionalProperties": true
},
"siblings": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"sku",
"siblings"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"sku": {
"hardware": "h200_sxm",
"framework": "vllm",
"model": "dsr1",
"precision": "fp8",
"spec_method": "none",
"benchmark_type": "agentic_traces",
"github_run_id": 123456789,
"date": "2026-08-08",
"dataset_slug": "cc-traces-weka"
},
"siblings": [
{
"id": 421,
"conc": 32,
"offload_mode": "off",
"decode_tp": 8,
"decode_ep": 1,
"decode_pp": null,
"decode_dcp_size": 8,
"decode_pcp_size": 1,
"decode_dp_attention": false,
"decode_num_workers": 1,
"prefill_tp": 8,
"prefill_ep": 1,
"prefill_pp": null,
"prefill_dcp_size": 8,
"prefill_pcp_size": 1,
"prefill_dwdp_size": null,
"prefill_dp_attention": false,
"prefill_num_workers": 1,
"num_prefill_gpu": 0,
"num_decode_gpu": 8,
"disagg": false,
"is_multinode": false,
"tput_per_gpu": 128.4,
"total_requests": 320,
"is_current": true,
"has_trace": true
}
]
}滚动查看完整示例
400
id 缺失或无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "id is required (benchmark_result_id)"
}404
没有使用此 ID 的基准结果。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
同组查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/derived-agentic-metrics读取派生智能体指标
按基准结果 ID 返回归一化交互性百分位。ID 会去重，最多接受 200 个。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| ids | query | 逗号分隔的整数 | 必填 | 1 到 200 个正整数基准结果 ID。 | 421,422 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/derived-agentic-metrics?ids=421,422'响应
DerivedAgenticMetricMap200
结果 ID 映射到 p75 和 p90 归一化交互性。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"p75_e2e_norm_intvty": {
"type": [
"number",
"null"
]
},
"p90_e2e_norm_intvty": {
"type": [
"number",
"null"
]
}
},
"required": [
"id",
"p75_e2e_norm_intvty",
"p90_e2e_norm_intvty"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
{
"421": {
"id": 421,
"p75_e2e_norm_intvty": 31.2,
"p90_e2e_norm_intvty": 24.8
}
}400
ids 缺失、格式错误或超过 200 个唯一 ID。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "ids query param is required"
}500
派生指标查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/request-timeline读取请求时间线
返回带版本号的基准测试时间窗口，以及每个请求的重放标识、来源追溯信息、调度/确认/完成/取消时间、token 指标、阶段和 worker 信息。当前版本的已存储时间线以 gzip JSON 直接流式传输，保留全部请求记录（若客户端的 Accept-Encoding 不含 gzip，则解压后返回，且不带 ETag）。时间线复制到按内容寻址的 CDN 存储后，接受 gzip 的客户端会收到 307 重定向，指向完全相同的 gzip 字节，其 Content-Type 为 application/gzip 且没有 Content-Encoding，需要自行解压（例如 curl -L ... | gunzip）。旧版时间线的响应结构不变。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| id | query | integer | 必填 | 基准结果 ID，必须为正安全整数。 | 421 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/request-timeline?id=421'响应
RequestTimeline200
时间线元数据和请求记录。纳秒事件字段是相对 startNs 的偏移。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"version": {
"type": "integer"
},
"startNs": {
"type": "integer"
},
"endNs": {
"type": "integer"
},
"durationS": {
"type": "number"
},
"requests": {
"type": "array",
"items": {
"type": "object",
"properties": {
"cid": {
"type": "string"
},
"ri": {
"type": "integer"
},
"ti": {
"type": "integer"
},
"srcTrace": {
"type": "string"
},
"srcOuter": {
"type": "integer"
},
"srcInner": {
"type": "integer"
},
"srcKind": {
"type": "string"
},
"wid": {
"type": "string"
},
"ad": {
"type": "integer"
},
"phase": {
"type": "string"
},
"credit": {
"type": "integer"
},
"start": {
"type": "integer"
},
"ack": {
"type": [
"number",
"null"
]
},
"end": {
"type": "integer"
},
"ttftMs": {
"type": [
"number",
"null"
]
},
"tpotMs": {
"type": [
"number",
"null"
]
},
"isl": {
"type": [
"number",
"null"
]
},
"osl": {
"type": [
"number",
"null"
]
},
"cancelled": {
"type": "boolean"
}
},
"required": [
"cid",
"ti",
"wid",
"ad",
"phase",
"credit",
"start",
"ack",
"end",
"ttftMs",
"tpotMs",
"isl",
"osl",
"cancelled"
],
"additionalProperties": false
}
}
},
"required": [
"version",
"startNs",
"endNs",
"durationS",
"requests"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"version": 6,
"startNs": 1000000000,
"endNs": 2400000000,
"durationS": 1.4,
"requests": [
{
"cid": "trace-018",
"ri": 0,
"ti": 0,
"wid": "7",
"ad": 0,
"phase": "profiling",
"credit": 0,
"start": 1200000,
"ack": 1800000,
"end": 420000000,
"ttftMs": 42.3,
"tpotMs": 18.1,
"isl": 18320,
"osl": 410,
"cancelled": false
}
]
}滚动查看完整示例
400
id 缺失或无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "id is required (benchmark_result_id)"
}404
该结果没有时间线。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
时间线查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/server-log读取基准服务器日志
返回某个基准测试结果 ID 对应的一份已存储的 .log/.out 文件。file 应使用 server-log-files 返回的文件名。可添加 offset 或 limit，按指定范围分块读取；添加 download=1 可将完整的所选文件作为文本附件流式下载。不带 offset、limit 或 download 时，超过 16 MiB 的文件返回 413；使用 download=1 时，超过 512 MiB 的文件返回 413，需改用 offset/limit 分段读取。已存储的日志不会变化，因此成功响应会在 CDN 缓存一年。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| id | query | integer | 必填 | 正整数基准结果 ID。 | 421 |
| file | query | string | 可选 | 产物内的相对文件名。省略时读取主文件或旧版日志文件。 | results/router.log |
| offset | query | integer | 可选 | 从零开始的字符偏移量。提供 offset 或 limit 后启用分块模式。 | 0 |
| limit | query | integer | 可选 | 分块字符数，范围为 1 到 262144，默认值为 65536。 | 65536 |
| download | query | integer | 可选 | 设为 1 时，将完整的所选文件作为文本附件流式下载；不能与 offset 或 limit 同时使用。 | 1 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/server-log?id=421&file=results%2Frouter.log&offset=0&limit=65536'响应
ServerLog | ServerLogChunk | text/plain attachment200
基准结果 ID 和服务器日志文本；分块模式还会返回范围元数据。
媒体类型: application/json响应结构
{
"oneOf": [
{
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"serverLog": {
"type": "string"
}
},
"required": [
"id",
"serverLog"
],
"additionalProperties": false
},
{
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"fileName": {
"type": "string"
},
"serverLog": {
"type": "string"
},
"offset": {
"type": "integer"
},
"nextOffset": {
"type": [
"number",
"null"
]
}
},
"required": [
"id",
"fileName",
"serverLog",
"offset",
"nextOffset"
],
"additionalProperties": false
}
]
}滚动查看完整示例
响应示例
{
"id": 421,
"fileName": "results/router.log",
"serverLog": "INFO router initialized\n",
"offset": 0,
"nextOffset": 24
}400
id 缺失或无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "id is required (benchmark_result_id)"
}404
该结果没有服务器日志。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}413
文件超过完整读取（16 MiB，JSON 响应体包含 bytes 和 maxBytes）或下载（512 MiB，纯文本响应体）的大小上限。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Log file is 951652286 bytes, above the 16777216-byte limit for this mode. Read it with offset and limit, or stream it with download=1."
}500
服务器日志查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/server-log-search搜索完整基准测试日志
在某个基准测试结果对应的所有已存储 .log/.out 文件中，按字面值执行不区分大小写的搜索。响应仅返回数量受限的上下文片段，无需传输完整日志文件。单次请求在所有文件中最多扫描 6,400 万个字符；若扫描因该上限而未到达末尾，incomplete 为 true。搜索超时则返回 504。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| id | query | integer | 必填 | 正整数基准结果 ID。 | 421 |
| q | query | string | 必填 | 按字面值匹配的搜索文本，长度为 1 到 256 个字符。 | router ready |
| limit | query | integer | 可选 | 最多返回的匹配数，范围为 1 到 100，默认值为 50。 | 50 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/server-log-search?id=421&q=router%20ready&limit=50'响应
ServerLogSearchResult200
数量有上限的匹配结果，包含文件名、字符偏移量和上下文文本。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"query": {
"type": "string"
},
"matches": {
"type": "array",
"items": {
"type": "object",
"properties": {
"fileName": {
"type": "string"
},
"offset": {
"type": "integer"
},
"before": {
"type": "string"
},
"match": {
"type": "string"
},
"after": {
"type": "string"
}
},
"required": [
"fileName",
"offset",
"before",
"match",
"after"
],
"additionalProperties": false
}
},
"truncated": {
"type": "boolean"
},
"incomplete": {
"type": "boolean"
}
},
"required": [
"id",
"query",
"matches",
"truncated",
"incomplete"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"id": 421,
"query": "router ready",
"matches": [
{
"fileName": "results/router.log",
"offset": 128,
"before": "INFO ",
"match": "router ready",
"after": " on port 8000\n"
}
],
"truncated": false,
"incomplete": false
}滚动查看完整示例
400
搜索参数无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Invalid query"
}504
搜索超时。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Search timed out. Try a more specific term."
}500
完整日志搜索失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/server-log-files列出已存储的基准测试日志文件
列出匹配服务器日志产物中保留的所有 .log 和 .out 文件名。路径相对于产物根目录返回，主文件排在首位。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| id | query | integer | 必填 | 基准结果 ID，必须为正安全整数。 | 421 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/server-log-files?id=421'响应
ServerLogFileNames200
相对于产物根目录的 .log/.out 文件名。
媒体类型: application/json响应结构
{
"type": "array",
"items": {
"type": "string"
}
}响应示例
[
"results/server.log",
"results/benchmark.log",
"results/router.log"
]400
id 缺失或无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "id is required (benchmark_result_id)"
}404
该结果没有日志文件集合。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
日志文件名查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/log-availability检查服务器日志可用性
仅返回具有已存储服务器日志的基准结果 ID。ID 会去重，最多接受 500 个。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| ids | query | 逗号分隔的整数 | 必填 | 1 到 500 个正整数基准结果 ID。 | 421,422 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/log-availability?ids=421,422'响应
LogAvailabilityMap200
可用结果 ID 映射为 true。缺失的键表示没有服务器日志。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": {
"type": "boolean"
}
}响应示例
{
"421": true
}400
ids 缺失、格式错误或超过 500 个唯一 ID。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Expected ids as comma-separated positive integers"
}500
服务器日志可用性查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/trace-availability检查跟踪可用性
仅返回具有已存储跟踪的基准结果 ID。ID 会去重，最多接受 500 个。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| ids | query | 逗号分隔的整数 | 必填 | 1 到 500 个正整数基准结果 ID。 | 421,422 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/trace-availability?ids=421,422'响应
TraceAvailabilityMap200
可用结果 ID 映射为 true。缺失的键表示没有跟踪。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": {
"type": "boolean"
}
}响应示例
{
"421": true
}400
ids 缺失、格式错误或超过 500 个唯一 ID。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "ids query param is required"
}500
跟踪可用性查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/gpu-metrics-point/stats读取功耗负载波动统计
一个基准测试点运行期间记录的 PowerX 功耗曲线的负载波动统计，基于已存的每 GPU 均值一秒桶：峰值、最小值、空闲底线（服务窗口前的 P5）、窗口均值、P5/P95、最大 1 秒与 5 秒变化、每秒上升/下降速率以及波动率（标准差 ÷ 均值），并附记录的服务窗口和被跳过（而非补插）的遥测间断数。每个关联功耗序列（单节点产物，或 sweep bundle 的一个验证窗口）各一条。PowerX 的“随时间变化”视图使用同一函数。scale=nvl72 乘以 72 换算为每机架 kW，是假设全部 GPU 同步波动的上限。format=csv 改为返回逐秒曲线。每次读取都核对实时遥测版本；响应为 no-store。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| id | query | integer | 必填 | 正安全整数基准结果 ID。 | 445805 |
| format | query | enum | 可选 | json 返回统计；csv 返回逐秒曲线（artifact、source、t_s、since_serving_s、in_validated_window、w_per_gpu 或 kw_per_nvl72_rack）。 | json |
| scale | query | enum | 可选 | gpu：每 GPU 瓦数（每桶的每 GPU 均值）。nvl72：× 72 换算为每 NVL72 机架 kW，假设全部 GPU 同步波动的上限；波动率不变。 | gpu |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/gpu-metrics-point/stats?id=445805'响应
GpuMetricsPointLoadSwingStats200
各关联功耗序列的统计，附定义与注意事项。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"benchmarkResultId": {
"type": "integer",
"minimum": 1,
"maximum": 9007199254740991
},
"apiVersion": {
"type": "string",
"enum": [
"v1"
]
},
"scale": {
"type": "string",
"enum": [
"gpu",
"nvl72"
]
},
"unit": {
"type": "string",
"enum": [
"W per GPU",
"kW per NVL72 rack"
]
},
"gpusPerRack": {
"type": "integer"
},
"definitions": {
"type": "object",
"additionalProperties": {
"type": "string"
}
},
"caveats": {
"type": "array",
"items": {
"type": "string"
}
},
"series": {
"type": "array",
"items": {
"type": "object",
"properties": {
"artifact": {
"type": "string"
},
"source": {
"type": [
"string",
"null"
]
},
"gpuCount": {
"type": "integer"
},
"window": {
"oneOf": [
{
"type": "object",
"properties": {
"startSeconds": {
"type": "number"
},
"endSeconds": {
"type": "number"
}
},
"required": [
"startSeconds",
"endSeconds"
],
"additionalProperties": false
},
{
"type": "null"
}
]
},
"stats": {
"type": "object",
"properties": {
"unit": {
"type": "string"
},
"sampleIntervalSeconds": {
"type": "number"
},
"sampleCount": {
"type": "integer"
},
"windowSampleCount": {
"type": "integer"
},
"preServingSampleCount": {
"type": "integer"
},
"gapCount": {
"type": "integer"
},
"peak": {
"type": "number"
},
"min": {
"type": "number"
},
"idleFloor": {
"type": [
"number",
"null"
]
},
"mean": {
"type": [
"number",
"null"
]
},
"p5": {
"type": [
"number",
"null"
]
},
"p95": {
"type": [
"number",
"null"
]
},
"maxDelta1s": {
"type": "number"
},
"maxDelta5s": {
"type": "number"
},
"rampUpPerSecond": {
"type": "number"
},
"rampDownPerSecond": {
"type": "number"
},
"volatility": {
"type": [
"number",
"null"
]
}
},
"required": [
"unit",
"sampleIntervalSeconds",
"sampleCount",
"windowSampleCount",
"preServingSampleCount",
"gapCount",
"peak",
"min",
"idleFloor",
"mean",
"p5",
"p95",
"maxDelta1s",
"maxDelta5s",
"rampUpPerSecond",
"rampDownPerSecond",
"volatility"
],
"additionalProperties": false
}
},
"required": [
"artifact",
"source",
"gpuCount",
"window",
"stats"
],
"additionalProperties": false
}
}
},
"required": [
"benchmarkResultId",
"apiVersion",
"scale",
"unit",
"gpusPerRack",
"definitions",
"caveats",
"series"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"benchmarkResultId": 445805,
"apiVersion": "v1",
"scale": "gpu",
"unit": "W per GPU",
"gpusPerRack": 72,
"definitions": {
"peak_min_delta_ramp": "whole run, per-GPU mean of each one-second bucket",
"mean_p5_p95_volatility": "validated serving window only",
"idle_floor": "P5 of the samples before the serving window starts",
"ramp": "largest change over a 5-second span divided by that span",
"gaps": "buckets more than 1.5× the sample interval apart are counted and never bridged"
},
"caveats": [
"One-second DCGM samples; sub-second swings are not captured.",
"Single node or NVL72 rack, inference only.",
"The NVL72 scale multiplies the per-GPU trace by 72 and assumes every GPU swings together, so it is an upper bound on rack-level swings."
],
"series": [
{
"artifact": "gpu_metrics_dsv4pro_8k1k_fp4_vllm_tp8_c64",
"source": null,
"gpuCount": 8,
"window": {
"startSeconds": 212,
"endSeconds": 1407
},
"stats": {
"unit": "W per GPU",
"sampleIntervalSeconds": 1,
"sampleCount": 1480,
"windowSampleCount": 1196,
"preServingSampleCount": 212,
"gapCount": 0,
"peak": 1012.4,
"min": 141.2,
"idleFloor": 143.8,
"mean": 871.5,
"p5": 744.1,
"p95": 968.3,
"maxDelta1s": 318.6,
"maxDelta5s": 612.9,
"rampUpPerSecond": 122.6,
"rampDownPerSecond": 97.4,
"volatility": 0.071
}
}
]
}滚动查看完整示例
其他表示格式: text/csv
响应结构
{
"type": "string"
}响应示例
artifact,source,t_s,since_serving_s,in_validated_window,w_per_gpu
gpu_metrics_dsv4pro_8k1k_fp4_vllm_tp8_c64,,0,-212,0,143.8400
id 缺失或不是正安全整数，或 format / scale 不是允许的值。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "id is required (benchmark_result_id)"
}404
该数据点没有关联遥测或没有可用的功耗序列。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
遥测查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/trace-histograms读取跟踪直方图
按基准结果 ID 返回输入和输出 token 计数数组。ID 会去重，最多接受 200 个。响应复制到按内容寻址的 CDN 存储后，接受 gzip 的客户端会收到 307 重定向，指向相同 JSON 的 gzip，其 Content-Type 为 application/gzip 且没有 Content-Encoding，需要自行解压。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| ids | query | 逗号分隔的整数 | 必填 | 1 到 200 个正整数基准结果 ID。 | 421,422 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/trace-histograms?ids=421,422'响应
TraceHistogramMap200
结果 ID 映射到原始 ISL 和 OSL 样本。
媒体类型: application/json响应结构
{
"type": "object",
"additionalProperties": {
"type": "object",
"properties": {
"id": {
"type": "integer"
},
"isl": {
"type": "array",
"items": {
"type": "number"
}
},
"osl": {
"type": "array",
"items": {
"type": "number"
}
}
},
"required": [
"id",
"isl",
"osl"
],
"additionalProperties": false
}
}滚动查看完整示例
响应示例
{
"421": {
"id": 421,
"isl": [
18220,
19340,
15110
],
"osl": [
410,
380,
512
]
}
}滚动查看完整示例
400
ids 缺失、格式错误或超过 200 个唯一 ID。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "ids query param is required"
}500
直方图查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}GET测试版/api/v1/trace-server-metrics读取跟踪服务器指标
返回数据点元数据（含 prefill/decode TP、EP、DP attention、worker 数、物理 GPU 数及 recipe_fingerprint，用于缓存复用链接），以及可直接用于图表的聚合时间序列，涵盖缓存使用率、队列深度、prefill/decode 吞吐量和 prompt token 来源。metricSources 包含各来源的描述信息；仅当用户在数据点详情界面选择某个来源时，才会加载该来源对应的数组。TokenSpeed 提供设备 KV 使用率以及运行中和排队请求数；未提供的缓存命中率和 token 吞吐量指标保留为空。
参数
| 名称 | 位置 | 类型 | 要求 | 说明 | 示例 |
|---|---|---|---|---|---|
| id | query | integer | 必填 | 基准结果 ID，必须为正安全整数。 | 421 |
请求
curl
curl -sS 'https://inferencex.semianalysis.com/api/v1/trace-server-metrics?id=421'响应
TraceServerMetrics200
点元数据、窗口边界和服务器指标序列。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"meta": {
"type": "object",
"additionalProperties": true,
"properties": {
"prefill_tp": {
"type": "integer"
},
"prefill_ep": {
"type": "integer"
},
"prefill_dp_attention": {
"type": "boolean"
},
"prefill_num_workers": {
"type": "integer"
},
"decode_tp": {
"type": "integer"
},
"decode_ep": {
"type": "integer"
},
"decode_dp_attention": {
"type": "boolean"
},
"decode_num_workers": {
"type": "integer"
},
"num_prefill_gpu": {
"type": "integer"
},
"num_decode_gpu": {
"type": "integer"
},
"recipe_fingerprint": {
"type": [
"string",
"null"
]
}
}
},
"startNs": {
"type": "integer"
},
"endNs": {
"type": "integer"
},
"durationS": {
"type": "number"
},
"timeslicesCount": {
"type": "integer"
},
"kvCacheUsage": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"prefixCacheHitRate": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"queueDepth": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"promptTokensBySource": {
"type": "object",
"additionalProperties": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"prefillTps": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"decodeTps": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"prefixCacheHitsTps": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"hostKvCacheUsage": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"kvCacheUsageByEngine": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
},
"kvCachePoolTokens": {
"type": [
"number",
"null"
]
},
"metricSources": {
"type": "array",
"items": {
"type": "object",
"additionalProperties": true
}
}
},
"required": [
"meta",
"startNs",
"endNs",
"durationS",
"timeslicesCount",
"kvCacheUsage",
"prefixCacheHitRate",
"queueDepth",
"promptTokensBySource",
"prefillTps",
"decodeTps",
"prefixCacheHitsTps",
"hostKvCacheUsage",
"kvCacheUsageByEngine",
"kvCachePoolTokens",
"metricSources"
],
"additionalProperties": false
}滚动查看完整示例
响应示例
{
"meta": {
"id": 421,
"hardware": "h200_sxm",
"framework": "vllm",
"model": "dsr1",
"conc": 32,
"date": "2026-08-08"
},
"startNs": 1000000000,
"endNs": 2400000000,
"durationS": 1.4,
"timeslicesCount": 2,
"kvCacheUsage": [
{
"t": 0,
"v": 0.44
}
],
"prefixCacheHitRate": [],
"queueDepth": [],
"promptTokensBySource": {},
"prefillTps": [],
"decodeTps": [],
"prefixCacheHitsTps": [],
"hostKvCacheUsage": [],
"kvCacheUsageByEngine": [],
"kvCachePoolTokens": 983040,
"metricSources": []
}滚动查看完整示例
400
id 缺失或无效。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "id is required (benchmark_result_id)"
}404
该结果没有服务器指标。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Not found"
}500
服务器指标查询失败。
媒体类型: application/json响应结构
{
"type": "object",
"properties": {
"error": {
"type": "string"
}
},
"required": [
"error"
],
"additionalProperties": true
}响应示例
{
"error": "Internal server error"
}