FP16 model compilation failure with TensorRT 11.5.0.63 on 4090 RTX

未关闭
#4,945 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
4/5
预计耗时
3-5 天
新手友好度
45/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
cpp

调研方向

Start by reproducing with the attached model.onnx.zip using trtexec --onnx=model.onnx --saveEngine=engine.trt on TensorRT 11.5.0.63; compare the result with TensorRT 10.16.1.11 and inspect the attached failure log. The report narrows the failure to an FP16 grouped Conv in a minimal ONNX graph. Done means identifying and fixing the TensorRT compilation failure while preserving the reported FP32 and older-version behavior.

由索引模型根据 Issue 内容生成。

描述

Module:ONNX

Some models with FP16 precision fail to compile from ONNX format with error like this:

[11/18/2026-21:41:19] [W] [TRT] No low-precision conv kernel available for this strongly-typed Conv/ConvTranspose. Consider keeping Conv/ConvTranspose and Q/DQ ops in FP32 (e.g. via modelopt.onnx.autocast.convert_to_f16 with op_block_list=['Conv','ConvTranspose','QuantizeLinear','DequantizeLinear']). Failing node: conv
[11/18/2026-21:41:19] [E] Error[10]: IBuilder::buildSerializedNetworkToStream: Error Code 10: Internal Error (Could not find any implementation for node conv. In computeCosts at /_src/optimizer/common/tactic/optimizer.cpp:4235)

TensorRT 10.16.1.11 compiles these models successfully. Also, FP32 works normally in TensorRT 11.5.0.63.

I reproduced the error with the simplest ONNX sample consisted of Input->Cast(FP16)->Conv(Grouped)->Cast(FP32)->Output (attachead model.onnx.zip)
Probably problem related to group Conv optimizations.

Full failure log attached: 11.5.0.63.log

I had this issue since TensorRT 11.2.1.2 (I haven't tried versions between 10.16.1.11 and 11.2.1.2), and I thought that it is the same issue as earlier reported #4838, but I can confirm that previous issue was fixed in latest TensorRT.

Environment

TensorRT Version: 11.5.0.63

NVIDIA GPU: 4090 RTX

NVIDIA Driver Version: 620.70.06

CUDA Version: 13.3

Operating System: Ubuntu 22.04

Steps To Reproduce

Commands or scripts: trtexec --onnx=model.onnx --saveEngine=engine.trt

Have you tried the latest release?: Yes

Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt):Yes

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。