Engine build failure of TensorRT 11.4 (Windows ARM64 wheel) when building ONNX models on GPU NVIDIA RTX Spark N1X (sm121): missing Myelin resource DLL

未关闭
#4,904 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 6 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
8/100
Issue 类型
缺陷
描述清晰度
需要澄清
活跃度
活跃
技术栈
cpp, python

调研方向

The failure is a missing Myelin resource DLL (myelin_resource64_sm110_2.dll) raised at resource_wrc.h:45 during codeGenerator.cpp compilation, and the installed wheel ships only nvinfer_builder_resource_sm120_11.dll, with no sm121 builder resource. The repro is a Python tensorrt-cu13 install on Windows ARM64 building any ONNX model. Fixing it means changing NVIDIA's packaging or build artifacts, not this repository's code, so a newcomer cannot close it from here.

由索引模型根据 Issue 内容生成。

描述

Module:Engine Build

Description

On a Windows 11 ARM64 laptop with an NVIDIA RTX Spark N1X GPU (compute capability 12.1), the TensorRT Windows ARM64 wheel installs and imports, but every engine build fails in Myelin:

Internal Error: MyelinCheckException: resource_wrc.h:45: CHECK(hMod_) failed. Failed to load resource DLL: myelin_resource64_sm110_2.dll, error: 2
[TRT] [E] IBuilder::buildSerializedNetwork: Error Code 1: Myelin ([myelin_graph.h:1295: attachExceptionMsgToGraph] MyelinCheckException: resource_wrc.h:45: CHECK(hMod_) failed. Failed to load resource DLL: myelin_resource64_sm110_2.dll, error: 2 In nvinfer1::builder::MyelinGraphTranslatorBase::compileGraph at C:\_src\optimizer\myelin\codeGenerator.cpp:2011)

The installed packages contain only these TensorRT DLLs:

nvinfer_11.dll                        230 MB
nvinfer_builder_resource_sm120_11.dll 225 MB
nvinfer_plugin_11.dll                  12 MB

So there is a builder resource for sm120 only (none for sm121), and no myelin_resource* DLL at all. The error also asks for an sm110 resource although torch.cuda.get_device_capability() reports (12, 1).

Workaround on the same machine: under WSL2 (Ubuntu 24.04 ARM64), nvidia-smi sees the GPU and the Linux aarch64 wheel tensorrt-cu13 11.4.0.106 builds and runs strongly typed engines for the same model (FP32 8.3 ms, FP16 4.0 ms, INT8 QDQ 4.1 ms per 640x640 image, batch 16). So GPU, driver and model are fine; only the native Windows ARM64 package seems to be missing the sm121 / Myelin resources.

Is the RTX Spark (12.1) on Windows ARM64 meant to be supported by the standard TensorRT wheels? If so, could the sm121 builder resource and the Myelin resources be added?

Environment

TensorRT Version: 11.4.0.106 (tensorrt-cu13, tensorrt_cu13_libs, tensorrt_cu13_bindings, pip from pypi.nvidia.com, win_arm64)

NVIDIA GPU: NVIDIA RTX Spark N1X (6144-core Blackwell RTX GPU), compute capability 12.1

NVIDIA Driver Version: 615.83 (KMD 617.74, as reported by nvidia-smi)

CUDA Version: 13.4

CUDNN Version: none installed separately

Operating System: Windows 11 ARM64 (build 28000)

Python Version (if applicable): 3.12.16 (ARM64)

Tensorflow Version (if applicable): n/a

PyTorch Version (if applicable): 2.14.0+cu134 (CUDA works)

Baremetal or Container (if so, version): Baremetal (native Windows)

Relevant Files

Model link: YOLO11l (Ultralytics 8.4.192) fine-tuned on BDD100K, exported to ONNX (opset 18, static batch 16, 640x640). The failure happens at the Myelin compile step, before anything model-specific, so I expect any similar ONNX model to fail; a public yolo11n.pt export was not tested separately.

Steps To Reproduce

Commands or scripts:

python -m pip install tensorrt-cu13 --extra-index-url https://pypi.nvidia.com
python -c "import tensorrt as trt; print(trt.__version__)"     # 11.4.0.106
# build any ONNX model, e.g. via Ultralytics:
yolo export model=yolo11l.pt format=engine quantize=16 device=0
# or the TensorRT Python API: OnnxParser + builder.build_serialized_network(...) -> fails as above

Have you tried the latest release?: 11.4.0.106 is the newest wheel pip resolved on 2026-10-30.

Can this model run on other frameworks?: Yes. ONNX Runtime CPU runs it; TensorRT-RTX 1.6 (onnxruntime-ep-nv-tensorrt-rtx-cu13 0.4.0) builds and runs FP32/FP16 for it on the same Windows machine; Linux TensorRT 11.4 in WSL2 on the same machine builds FP32/FP16/INT8 engines.

主要语言
C++
星标
13.5k
派生
2.4k
平均合并
6 天 21 小时
30 天内合并 PR
3

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

NVIDIA/TensorRT 的其他 Issue

查看 NVIDIA/TensorRT 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。