octopusmegalopod/some-glmocr-ggufs

GLM-OCR converted for use with llama.cpp. As of b8094, it may crash in llama.cpp if flash-attn is on or certain backends are used (e.g: CPU).

Regrettably, does not seem to perform well for single CJK line texts out of the box. Or perhaps it expects the image in a specific resolution, due to its dependence on PaddleLayout.

GLM-OCR in llama.cpp behaves differently depending on the maximum context for some reason. Setting -c 2000 gives wrong outputs typically, for example. Whereas -c 9000 has produced much better results so far in my tests.

So far in my tests:

• performs poorly on Korean horizontal text lines.
• performs decently on Japanese horizontal text lines.
• performs poorly on Japanese vertical text lines.
• performs excellently on Chinese horizontal text lines.
• relatively robust to quantization