ERNIEFrom the builder
ERNIE 4.5 Gets a Major Inference Speed Boost
Baidu has shipped a long-context speedup for ERNIE-4.5-21B-A3B and ERNIE-4.5-300B-A47B, currently for the Paddle versions deployed with FastDeploy. It relies on PLAS, a pluggable sparse attention method that splits text into blocks and uses a small learnable module to pick only the most relevant blocks, without touching the original weights. Baidu says accuracy barely changed in long-document tests averaging about 113K input tokens.
The best part is its “pluggable” nature. We can add PLAS to a fully trained model without changing the original weights, ensuring that the model’s core knowledge remains intact.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.