Local Mode fails on GGML models

[bluecoconut]

Via the CTransformers library we're using ggml library

For increasing context length, which is necessary for local-mode CPU verison of StarCoder, sketch fails and can crash dropping the full kernel.

Raised issue in ggml, and hopefully this will be transparent to fix through ctransformers
Note: from the thread about quantization support: marella/ctransformers#1 if the new fix for ggml is after the quantization changes, and ctransformers doesn't update, we might be "stuck" for a bit.

Issue in ggml: ggerganov/ggml#158