Skip to content

Fix excel routed dtype latency - #13

Closed
ZZZxwx wants to merge 23 commits into
jiashaokun-1:mainfrom
ZZZxwx:fix-excel-routed-dtype-latency
Closed

Fix excel routed dtype latency#13
ZZZxwx wants to merge 23 commits into
jiashaokun-1:mainfrom
ZZZxwx:fix-excel-routed-dtype-latency

Conversation

@ZZZxwx

@ZZZxwx ZZZxwx commented May 25, 2026

Copy link
Copy Markdown
Contributor

修复 spec-based 训练 Excel 报表中 Ops sheet 的 per-op latency 展示问题。

此前最终 Step Time 计算路径已经会根据 op.component 解析 compute dtype,例如 routed expert 使用 routed_expert_compute_dtype,但 Excel exporter 在单算子 Latency (μs) 展示时没有传入该 dtype,导致 routed expert 在 FP8/FP4 配置下仍按默认 BF16 peak 显示 latency。

本 PR 将 Excel exporter 的 per-op latency 计算改为复用 _resolve_compute_dtype(op, model),并把解析出的 dtype 传给 _cost_phase_time(),使 Excel 中 routed expert 的 latency 与最终训练时间估算路径保持一致。

同时新增回归测试,覆盖 routed expert FP8 compute + FP4 weight 场景,验证 Excel Ops sheet 的 latency 使用 routed compute dtype,而不是 BF16 默认路径。

laksjdf and others added 23 commits May 21, 2026 17:45
pp_p2p统计口径问题(掩盖、bwd)
寻优脚本优化,在format_result的时候增加缓存
支持pp开低精度、开overlap
支持pp开低精度、开overlap
Fix DualPipeV VPP scheduling accounting
1.fix dsv4 PARAM_SCOPE miss 2.add xlsx test
寻优允许pp=1情况,增加hidden限制
@ZZZxwx ZZZxwx closed this May 25, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants