NVIDIA / NVIDIA/TensorRT

Deploy DeBERTa to Triton Inference Server

Open
#4,202 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

triaged
Dominant language
C++
Stars
13.4k
Forks
2.4k
Avg merge
5d 3h
Merged PRs (30d)
2

Description

I followed the steps in the DeBERTa guide to create the modified onnx file with the plugin. When I try using this model with triton inference server, it says

Internal: onnx runtime error 9: Could not find an implementation for DisentangledAttention_TRT(1) node with name 'onnx_graphsurgeon_node_0'

Is there a way to get this to work in triton? I'm using triton 24.09.

I can confirm that I was able to get the onnx model working fine when using the onnxruntime package in a python script. It works fine in triton if I don't use the plugin

Slightly separate issue that might be better for another issue:
The model in fp16 is garbage, even with the layernorm in fp32.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the DeBERTa guide and the modified ONNX model, then reproduce the DisentangledAttention_TRT error on Triton 24.09 and compare it with the working ONNX Runtime Python script. Done means the plugin model runs in Triton; separately determine whether the reported FP16 output problem belongs in this issue.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
backend, machine-learning
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.