Benchmark Atlas

亚洲模型实验室 · arXiv 论文

Mixed-precision Training Lifecycle

复现 DeepSeek-V3 Figure 6:把线性层前向、激活反传、权重反传和优化器更新放进一个显式精度转换闭环。

deepseekfp8mixed-precisiontraining-systemdataflow
FP8 GEMMs with BF16 boundaries and FP32 state复现 DeepSeek-V3 Figure 6:把线性层前向、激活反传、权重反传和优化器更新放进一个显式精度转换闭环。 示意数据,仅复现信息结构,不代表最新榜单结果。 FP8 GEMMs with BF16 boundaries and FP32 stateDeepSeek-V3 Technical Report · Figure 6 / PDF p.15DEMO DATALINEAR OPERATOR · PRECISION DOMAINSInputBF16Fprop GEMMFP8 → FP32OutputBF16Input gradBF16Dgrad GEMMFP8 → FP32Output gradBF16Wgrad GEMMFP8 → FP32Weight gradFP32OptimizerFP32 statesMaster WFP32to FP8to BF16to FP8to BF16cached activation · FP8to FP8accumulate FP32updateretain FP32to FP8 weightFP8 for compute-dense GEMMs · BF16 at activation boundaries · FP32 for accumulation and optimizer stateprecision is encoded on every conversion edge, not inferred from node color alone
Illustrative demo data · renderer: mixedPrecisionLifecycle

Information grammar

它如何组织信息

Fprop/Dgrad/Wgrad 三 GEMM 闭环 + 每条转换边的精度协议 + BF16 激活边界 + FP32 master/optimizer 状态域

Visual system
VS-82 DeepSeek precision-domain lifecycle
Component ID
deepseek-fp8-precision-lifecycle

Use when

适用判断

适合审计混合精度训练的数据类型边界;节点和转换边都必须标明精度,不能仅靠颜色暗示。

Figure-level evidence

Figure 6 · PDF p.15

Fprop、Dgrad、Wgrad 使用 FP8 GEMM;激活边界保留 BF16,累加、master weight、gradient 与 optimizer state 保留 FP32。

Verified
2026-07-13
Paper source
DeepSeek-V3 Technical Report

Structured data

可替换的数据模型

示意数据,仅复现信息结构,不代表最新榜单结果。

{
  "domains": [
    "FP8 GEMM",
    "BF16 activation",
    "FP32 accumulation",
    "FP32 optimizer state"
  ]
}

Related components

同一图表家族的其他语法