Some issues if try to use it in a real life :)
Denys88 opened this issue · comments
Hi, I tried to use your product but got a lot of small issues and found some lack of functionality.
-
Please don't expect that git address is something line this:
higgsfield/higgsfield/internal/util.py
Line 47 in 61bc432
I tried to use it with the internal github which is different from public github and got an error. -
There are a lot of cases where your error messages are useless like in first example.
I tried to use higgsfield manually and got a lot of messages like 'something is not a string'.
Quick debug helped me to find that I forgot or put wrong command line parameter. It could be improved. -
LLama and hugging face:
When I import llama loader it automatically tries to get access to the HF without any my permitions. Overall trying to access something from internet without explicit calls is a big red flag from the security of view. In my case I've already downloaded everything and don't need to connect to the HG at all. -
Would be nice to see more examples:
- very simple manually implemented architecture which supports deepspeed/zero distribution training.
- example which show how to manually run everything without github and hf access.
- ability to run your code on a single machine - single gpu and single machine multiple gpu too.
Because how do you expect people to debug their code?
I wanted to run a simple example without setting up my machines and using github and found it impossible which is a big problem in my opinion/
Overall great job and nice implementation but it could be much user friendlier.
Thanks!
Hey! Thanks a lot for your thorough inspection.
- A nice catch. We haven't thought about it.
- Yes, the errors are ill-defined for now. The case with "something is not a string" happens since we parse AST directly without any proper analysis (line numbers, etc.).
- We'll change the logic soon, so you can use the locally downloaded models, datasets with higgsfield without any implicit calls.
- Overall, we are going to provide more tutorials. We're quite sorry for not having them right now.
very simple manually implemented architecture which supports deepspeed/zero distribution training.
- Right now our API provides a support for major LLMs. You can implement your own if you're eager to.
example which shows how to manually run everything without github and hf access.
- We chose github just only for usability. It's not a big deal to make a converter into gitlab or other github-like services which has a concept of CI/CD. Yet you don't have to depend of Hugging Face when it comes to datasets or models. The current impl provides a way to do so.
ability to run your code on a single machine + single gpu and single machine + multiple gpu.
- We'll make it happen in upcoming updates.
Appreciate it very much.