ggjt v2 models don't load (or error gracefully)

[akx]

I freshly pulled 7e4ea5b and make clean && maked and it fails to load a model converted from pytorch using the tools from revision 63d2046 (using https://github.com/akx/ggify):

llama.cpp: loading model from models/ausboss-llama-30b-supercot-q8_0.bin
error loading model: llama.cpp: tensor '�+� ��s��93:�a-�%��Y��8Ɓ0�&�M,�9�4������"/�@�չ�"*+c�5�������9�>+n��!������O...' should not be 2563577093-dimensional
llama_init_from_file: failed to load model
llama_init_from_gpt_params: error: failed to load model 'models/ausboss-llama-30b-supercot-q8_0.bin'
main: error: unable to load model

I re-converted the model with 7e4ea5b; apparently the old file had been

llama_model_load_internal: format     = ggjt v2 (latest)

and the new one is

llama_model_load_internal: format     = ggjt v3 (latest)

(and 6% smaller!)

It would be nice if there was an error saying that ggjt v2 is not supported, instead of dumping out garbage tensor names and mind-bendingly large tensor dimensionalities 😁 but I suppose this doesn't necessarily need any action right now.

This seems to be related to

• [BUG] CLBLast generate garbage text on q8.0 models #1525 (comment)
• Feature: Short but helpful incompatibility information on ggml file version #1537

[akx]

EDIT: Right, this is obviously just #1508 🤦

Closing.