Is GGML_OP_OUT_PROD still used in Llama.cpp/GGML?

[taronaeo]

I noticed that while developing my IBM zDNN backend, the GGML_OP_OUT_PROD operation does not seem to get called at all. Verified this with ggml-cpu as well and same result, no hits to that operation.

Thinking out loud, is the operation still used? Or has it been deprecated? Or is it model-dependent where only a specific few models will utilise the operation?

[slaren]

It's not important for inference, but during training it's used in the backwards pass of GGML_OP_MUL_MAT.