Training Acceleration Trick Deployable via Weight Fusion
torchcompiled · x · 2026-07-15
This repost/quote discusses an interesting training trick: parameterizing weights as a "learned weighted blend of symexp and regular linear weights," which reportedly yields a 1.42x wall-clock training speedup.
The author also notes:
- These parameterized weights can be fused back into standard weights for easier deployment.
- Another highlighted aspect is asymmetric init, which slightly suppresses negative values.
- Regarding why this works, current speculation suggests it may relate to early symmetry breaking and providing an initial bias.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- Stanford launches GPU-centric supercomputer Marlowe for data science and AI research — StanfordHAI · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27