E2E speed sprint: video training loader, token fast path, honest e2e reruns, CUDA self-hosted CI - #16

Merged

ALJainProjects merged 8 commits into

mainfrom

Aug 1, 2026

Commits on Aug 1, 2026

perf: CudaPrefetcher — overlapped H2D double-buffering for training loops
Arnav Jain
committed
31cd4ff
perf: TokenDataLoader zero-alloc pinned ring + overlapped device= path
Arnav Jain
committed
469f8ac
bench: token-to-device pipeline benchmark; honest CudaPrefetcher docstring
Arnav Jain
committed
b6aa514
feat: VideoDatasetLoader — labeled clip batches from a video directory (CUDA)
Arnav Jain
committed
151d59c
bench: video dataset download — browser UA + blender.org fallbacks (403 fix)
Arnav Jain
committed
4a7bd1f
bench: encode dataset segments from fresh ndarray frames (EINVAL fix)
Arnav Jain
committed
ec3625e
bench: class indices from dirs only (labels were 2/3); CUDA self-hosted workflow
Arnav Jain
committed
347b260
docs: e2e speed sprint results; fix quickstart_demo syntax (comma eaten by comment)
Arnav Jain
committed
f17c93f