Can you add details of hardware?

[dongxiaok]

Can the description information include test indicators related to hardware configuration and output speed?
This project is characterized by client-side deployment and fast CPU operation, but it runs very slowly on my own machine and takes several minutes to output the results.(8 core 16GB laptop)
I would like to know what configuration of the machine this project should be on to get the so-called "fast output"

[MonolithFoundation]

It's relatively fast on my side (M1 mac), be note that Namo-R1 used native input solution, which means if your images it's big, then try set max size constraint otherwise it will use native resolution. (To compare this is unfair as many models resize it to 224 in the first place). And the overall computation should be increase as the tokens feed into LLM could be huge if size too much.

Could you try the CLI demo_bare.py to debug deeper for input size?

Namo-R1 uses native input resolution and for the LLM part it is actually the same, but the number of tokens depends on inputs.

This will be resolved in v2 I suppose, as the author states he is fine-tuning on siglip-v2.

Also, int8/6 bit/4 bit on the CPU should be on the way.