chenjun

the test demo

1

Thank you for your excellent work. would you please put out the test demo.

和hugging face上的效果不一样。

2

同一张图像，同样的参数，用仓库的代码和hugging face demo推理的可视化结果不一样，hugging face是做了什么优化修改嘛？

nice work! I have some questions. 1. Did you train the epe model by yourself? 2. Can you open source your pre-training model? thanks！！！

waiting for the code

9

nice work, waiting the code.

i can't find the data on https://huggingface.co/datasets, could you give the link for the data. thanks.

[ERROR] "Unable to load backend. dlerror(): libcdsprpc.so: cannot open shared object file: No such file or directory"

2

I have already run the Qwen2-7B model on my mobile phone using genie-t2t-run, and now I want to simulate running it on my computer, but I encountered an error. ```...

8gen4为啥比8gen1+npu还要慢。

1

我用qnn_net_run分别在8gen4和8gen1+上跑同一个模型yolox_l，量化方式为A8W16。发现8gen4更慢。这是我实测的数据。 | 序号 | 设备 | 模型 | 输入分辨率 | 量化类型 | converter engine | 执行 engine | 模型格式 | 推理耗时/ms | | ---- | ------ | ------- | ----------...

[bug] 骁龙8gen4 设备端运行qnn报错

18

在linux上交叉编译之后，在骁龙8gen4上测试报错 linux上编译 ``` cmake .. -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-28 -DCMAKE_C_FLAGS="-march=armv8.7a" -DCMAKE_CXX_FLAGS="-march=armv8.7a" -DGGML_OPENMP=OFF -DGGML_LLAMAFILE=OFF -DGGML_QNN=ON -DGGML_QNN_DEFAULT_LIB_SEARCH_PATH=/data/local/tmp ``` 我用的qnn_sdk_version=2.31.0.250130。将高通的动态库push到设备端。 在设备端测试 ``` export LD_LIBRARY_PATH=/data/local/tmp/mllm/install-android/lib:/data/local/tmp/mllm/qnn-lib ./llama-cli -m ../../models/Qwen2.5-0.5B-Instruct-F16.gguf ``` 报错： ``` llama_context: n_ctx_per_seq (4096) <...