Edge-aware attention cuts graph-generation GPU memory about 65%
A FlexAttention implementation kept generation quality near the baseline while improving training and sampling throughput.
Stories per day · 30 days
Relatedattention1distributed-training1flexattention1floating-point1graph-generation1huffman1knowledge-distillation1lossless-compression1model-checkpoints1neuralzip1pytorch1qwen1training1
3 stories, newest first · one card per story
Stories tagged #gpu-memory
DisaggKD splits teacher and student across nodes for distillation
On the Qwen3.5-35B-A3B to 9B pair, the split matches colocated LM loss and is faster.
NeuralZip: Reusable Setup for Fast Lossless Compression
NeuralZip reuses a precomputed exponent setup to losslessly compress model checkpoints up to 21.33 times faster than baselines.