Query regarding BLAS implementation of GGML

[vikasbalaga]

I am trying to implement an application with GGML library on x86_64 architecture. I have built the "default" variant which will utilize GNU libraries for low level math functions.

But when I wanted to use "BLAS" as a backend library, I can see in the source code that it supports only a few operations, even for a basic operation like "MUL", I am seeing the following error

ggml-blas.cpp:250: ggml_backend_blas_graph_compute: unsupported op MUL

@ggerganov, does it mean that I can't use BLAS as a backend library or can you please confirm if there is a way to overcome these issues?

[ggerganov]

You need to use both the CPU and BLAS backends and a ggml_backend_sched to distribute the operations to the correct backends based on what they support. See the gpt-2, main-sched.cpp example and start from there.