How to build TensorRT-LLM engine on host and deploy to Jetson Orin Nano Super?

[Sesameisgod]

Hi, I’m currently working with TensorRT-LLM and trying to deploy a model (e.g., Qwen2-VL-2B-Instruct) on a Jetson Orin Nano Super. However, due to limited memory on the Nano, I’m unable to build the TensorRT engine directly on the device.

Is there any official or recommended approach to build the TensorRT-LLM engine on a more powerful host machine (with sufficient memory and GPU), and then transfer the generated engine file to the Jetson Orin Nano Super for inference?

If so, are there any considerations or compatibility issues I should be aware of when cross-building the engine on x86 and deploying it on Jetson (aarch64)?

Thanks in advance!

[juney-nvidia]

@Sesameisgod
Hi, TensorRT-LLM has two backends now, one based on TensorRT(the first workflow supported in TensorRT-LLM) and the other based on PyTorch(the new supported workflow since 0.17 release).

For TensorRT workflow, it requires AoT(ahead-of-time) phase tuning phase to select the best combination of kernel sequence, so though technically it is possible to build the TensorRT engine on another GPU sharing the similar HW architecture, it is not the most recommended way.

June

[juney-nvidia]

• @sunnyqgg for vis in case she may have more inputs on this question.

June

[Sesameisgod]

Thank you for your response!

I’d like to follow up and ask — is there any recommended approach for building a TensorRT engine for Qwen2-VL-2B-Instruct directly on a Jetson Orin Nano Super (8GB RAM)?

I’ve tested running the model via Hugging Face Transformers on the Nano and it works successfully, which suggests that the model can run on the device.

However, the OOM issue occurs during the TensorRT engine building phase, are there any strategies (e.g., using swap) to make engine building feasible directly on the Nano?

[juney-nvidia]

@Sesameisgod I am not aware that TensorRT engine building process can use swap memory during the offline engine building process.

An alternative way is that you can try to run Qwen2-VL model in the newly introduced PyTorch workflow:

• https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/pytorch
• https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/pytorch/quickstart_multimodal.py#L107
• https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/_torch/models/modeling_qwen2vl.py

TensorRT-LLM PyTorch workflow has been introduced since 0.17 release and based on our internal performance evaluation, on popular models like LLaMA/Mistral/Mixtral, the PyTorch workflow performance is on-par(or even faster) with the TensorRT workflow, since the customized kernels are reused in both TensorRT workflow(as plugins) and PyTorch workflow(as torch custom op), also the existing C++ runtime building blocks(BatchManager/KV CacheManager/etc.) are also reused in both TensorRT and PyTorch workflow and more optimizations are added into the PyTorch workflow.

Currently we also shift more attention to the enhancement of the PyTorch workflow, such as the recently announced DeepSeek R1 performance numbers are all based on the PyTorch workflow.

What we cannot commit now is that up to now due to bandwidth limitation, we cannot commit official support for Jetson platform yet. So you need to try to run TensorRT-LLM on Jetson to observe the behavior.

Thanks
June

[Sesameisgod]

Got it, I’ll try running Qwen2-VL with the PyTorch workflow on the Jetson Orin Nano Super and see how it performs.
Really appreciate your help!

[juney-nvidia]

@Sesameisgod to ensure you are aware of this Qwen2.5-VL effort from @yechank-nvidia

https://github.com/NVIDIA/TensorRT-LLM/pull/3156/files

Thanks
June

[sunnyqgg]

Hi @Sesameisgod，

1. You can use swap memory during engine building process, but in my experience, if you allocate over 8GB, the system locks up. The TRT engine generation requires about 4 times the memory of the model size.
2. You can try to build W4A16(INT4) model engine.
3. As we discussed below, you can use the same TRT and TRT-LLM with Jetson orin(64GB) to build the engine and run it with Jetson Orin Nano Super
4. For saving memory usage in the inference phase, you can use -mmap, pls refer https://github.com/NVIDIA/TensorRT-LLM/blob/v0.12.0-jetson/README4Jetson.md#3-reference-memory-usage
5. Actually we have branch for Jeston device:https://github.com/NVIDIA/TensorRT-LLM/tree/v0.12.0-jetson, but unfortunately it doesn't support Qwen2-VL.

Thanks,
Sunny.

[Sesameisgod]

@juney-nvidia
That's great to hear! It's really nice to see support for the latest models coming so quickly. Qwen is truly an impressive VLM series.

@sunnyqgg
Understood, thank you. For now, we're not considering quantization yet. We're planning to explore building the engine on a Jetson device with 64GB of memory — so far, it seems like Jetson AGX Orin is the only one that meets this requirement, and we’re planning to purchase one for testing.

Also, as you mentioned, TensorRT-LLM v0.12 doesn't support Qwen2-VL, so I'm currently using the aarch64 Docker image from this repository. After running the container and executing the import tensorrt_llm command, the version appears correctly, so it seems to be running, though further testing is needed to confirm whether everything works properly.

If there’s any progress later on, I’ll be happy to share it here. Thanks again for your help!

[xiaohuihuige]

Hi @Sesameisgod , I am also doing the same thing as you, but I encountered a big problem. My board is Jetson NX (8G), and I ran Vila-3b (mlc), which is quite fast, but my understanding ability is a bit poor. Now I am trying to run qwen2.5vl and qwen2vl. Tensor-LLM only provided Jetson with v0.12 version, and the configuration environment is very troublesome for me. Unfortunately, your Docker that talks about Tensor LLM cannot be pulled in China

[Sesameisgod]

Hi @xiaohuihuige ,
Have you tried using a VPN? or maybe I can share the Docker image with you.
However, I'm not quite sure what's the best way to transfer it.

[garvitpathak]

Hi @Sesameisgod ,if pull it on my jetson orin nano 8gb can i convert the qwen 2 vl 2b to tensorrt and run the inference over it?