Operator aten::count_nonzero.dim_IntList is not implemented

[stuaxo]

Looks like this isn't implemented yet - found when testing comfyui

comfy/samplers.py:966: UserWarning: The operator 'aten::count_nonzero.dim_IntList' is not currently supported on the ocl backend. Please open an issue at for requesting support https://github.com/artyom-beilis/pytorch_dlprim/issues (Triggered internally at /home/artik/Projects/dlprim_build/pytorch_dlprim/src/tensor_ops.cpp:415.)

[JingliangGao]

Yes. Some ATen operators have not been implemented in this project. BTW, when I want to infer a LLM, like QWen2, the operator 'aten::index_selected' is not currently supported. Try to implement the missing op by yourself if possible, or ask the author @artyom-beilis for help :）

[stuaxo]

LlamaCPP (also MIT licensed) has some OpenCL operators, it could be worth porting some of them.

[octaveoclx]

https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-opencl/kernels/
This is the place where it happens. This comes with only forward/inference kernels. I believe @artyom-beilis has implemented many of them in a better form while flash attention is indeed missing. Right now it is mainly updated for Adreno devices from the employees of a company.