TensorRT-LLM icon indicating copy to clipboard operation
TensorRT-LLM copied to clipboard

Fix nmt weight conversion

Open Pzzzzz5142 opened this issue 1 year ago • 0 comments

For WMT14 model, it shares the vocab across the encoder and decoder. So it wouldn't trigger this error. However, for language pair which has large differences like zh-en, usually we don't share the vocab. So here we need set the vocab size for decoder correctly.

Pzzzzz5142 avatar May 24 '24 05:05 Pzzzzz5142