Closed aleablu closed 6 months ago
@aleablu that 'was' a necessary install when it was added, I'm sure newer triton installs work just fine now and wouldn't cause issues with newer torch. Try using the default tritons in latest torch versions...
I'm using newer torch==2.4.0 , newer triton==3.0.0 and open_clip_torch==2.26.1 in python 3.8. And I have the same problem " AttributeError: module 'triton.language' has no attribute 'libdevice' " when I try to run convert_int8_model_to_inference_mode
.
Could you please tell me more details about the environment in int8-support notebook?
I am following the int8 tutorial at https://github.com/mlfoundations/open_clip?tab=readme-ov-file#int8-support but I cannot make it work with the latest version of open clip.
Installing the required triton version (2.0.0.post1) forces pip to also install torch 1.13 which breaks the code that loads open_clip, because I'm using the latest version of open_clip_torch (2.24.0).
I tried installing the latest triton version (2.1.0) but the code fails with:
Is there a way to solve this without reverting to a version of open_clip that supported torch 1.13?
Here is the code I'm using for reference, even though it's a mere copy-paste of the notebook: