NVIDIA / FasterTransformer

Transformer related optimization, including BERT, GPT

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

is_return_output_log_probs doesn't return logits for T5 model

swairshah opened this issue · comments

Branch/Tag/Commit

main

Docker Image Version

763104351884.dkr.ecr.us-east-1.amazonaws.com/pytorch-training:2.0.1-gpu-py310-cu118-ubuntu20.04-ec2

GPU name

A10

CUDA Driver

450.119.03

Reproduced Steps

For any FasterTransformer t5 model hosted on Sagemaker using the DJL images https://sagemaker.readthedocs.io/en/stable/frameworks/djl/sagemaker.djl_inference.html#fastertransformermodel (the model i am using is Flan-t5-xxl) pass the config:
 = {
    "max_length" : 64,
    "min_length" : 0,
    "temperature" : 0,
    "is_return_output_log_probs" : True,
}
It only returns generated text and not the logits.