Upload ./QA.md with huggingface_hub
Browse files
QA.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ❓ Q&A for Building TensorRT Engines
|
| 2 |
+
|
| 3 |
+
### 1. Error: `trtexec: not found`
|
| 4 |
+
|
| 5 |
+
**Solution:** Try the following command
|
| 6 |
+
|
| 7 |
+
```shell
|
| 8 |
+
source trt/activate.sh
|
| 9 |
+
```
|
| 10 |
+
|
| 11 |
+
This command is used to configure the environment variables required to build the TensorRT Engine.
|
| 12 |
+
|
| 13 |
+
### 2. Error: `Unable to open library: libnvinfer_plugin.so.9 due to libcublas.so.11: cannot open shared object file`
|
| 14 |
+
|
| 15 |
+
This error occurs because the `libnvinfer_plugin.so.9` library depends on the `libcublas.so.11` library.
|
| 16 |
+
However, you may use a different version of cuda library (see `/usr/local/cuda`), such as cuda-12. It is recommended
|
| 17 |
+
to use cuda-11 to build the TensorRT Engine.
|
| 18 |
+
|
| 19 |
+
### 3. The size of the engine file generated by each build is different
|
| 20 |
+
|
| 21 |
+
**Explanation:** During the process of building the engine, the building program will automatically select the
|
| 22 |
+
optimal kernel and other parameters. Therefore, the size of the engine file generated by each build may be different.
|
| 23 |
+
Therefore, please ensure that the GPU is idle during the build to ensure that the building program correctly selects
|
| 24 |
+
the optimal kernel.
|
| 25 |
+
|
| 26 |
+
|