NVIDIA-NeMo/NeMo

Speed up RNNT model inference using TRT

Open

#14,531 opened on Aug 20, 2025

View on GitHub
 (1 comment) (0 reactions) (1 assignee)Python (3,421 forks)github user discovery
ASRcommunity-requesthelp wantedwaiting-on-customer

Repository metrics

Stars
 (17,298 stars)
PR merge metrics
 (Avg merge 12d) (49 merged PRs in 30d)

Description

Hi,

I previously trained an RNNT model and now want to accelerate it by converting it to TensorRT. I’ve exported the model to ONNX and have encoder.onnx and decoder.onnx.

I’m using the TensorRT 25.03 Docker image and trtexec to convert the models. The decoder works fine with --fp16, but when I use --fp16 for the encoder, some outputs return NaN and the results are incorrect.

Has anyone encountered this issue or knows how to fix it?

Are there any methods to accelerate RNNT model inference?

Contributor guide

Research direction
Investigate why FP16 inference fails for the encoder model. Check TensorRT logs for warnings, verify operator support for FP16, consider using INT8 quantization or fp16 with best flag. Also examine the ONNX export for problematic ops.
Tech stack
python
Domain
aiperformance
Issue type
Bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PythonTensorRTONNX
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
40