Drop a fixed batch proportion instead of per-sample tokens: capi author shares training trick
giffmana · x · 2026-09-09
- Tim Darcet suggests improving token dropping: replace "each sample is dropped with probability p" with "a proportion p of the batch is dropped", yielding constant shapes, good efficiency, and perfect compilation when everything runs on GPU.
- giffmana notes this is the same permute+slice trick as in tokendrop, but cautions it only works for fixed example sizes — increasingly rare in both vision and language models where you may want to fully keep or drop layers per example. Implementation referenced in the capi repo.
Related event: Meta researcher shares batch-ratio token dropping training trick(3 posts)→
More from Infra
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Running a local 9B as a coding agent: brief the big model to plan, let the small model execute — Electrical_Ninja3805 · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- Gary Marcus: With Q3 data added, Anthropic's compute-efficiency growth shows deceleration — GaryMarcus · 2026-10-27
- SpaceX locks in $40B GPU financing deal with Apollo, the largest ever committed — TechCrunch AI · 2026-10-27
- Inference is just Kubernetes with a multi-trillion-dollar market, and the communities overlap — matt_slotnick · 2026-10-27