Concept detail
Deferred normalization
0 tagged videos · 1 connection videos · degree 0
Publication timeline
- 2026-091 video
Source videos
Observed neighbors
CUDA stream concurrency (tensor cores vs. CUDA cores)
Splitting the scalar divide enables the matmul and normalization to overlap on separate streams.
1 source videos
Weight Folding, CUDA Streams, and the Bug That Made My Model Speak Backwards — Filip MakraduliRelated tagged insights
No core insights are tagged with this concept.