mirekphd

icon indicating a website link https://hub.docker.com/u/latestml

icon location EU icon location AI/ML Engineer / AI/ML Ops / Dev Ops / sys admin

Results 83 comments of mirekphd

Docker build fails because of maybe-uninitialized warning

> I can report that changing the `build-type` to Release results in an successful build. I confirm this result, thanks @mapa17, but a successful compilation seems to be version-dependent and...

Support for QWEN 2.5 VL

More specifically, the currently latest version of `autoawq` pins `transformers` to an older version from December 2024, that did not support [Qwen2.5-VL-72B-Instruct)](https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct) (at least `transformers==4.49.0` is required for this model)....

Build flash-attn takes a lot of time

Confirmed. In my `htop` I see unexpected NVIDIA compiler flags for Hopper architecture (`ptxas -arch sm_90`) when I specify Ampere arch. alone using `export TORCH_CUDA_ARCH_LIST="8.0"` (I'd expect only `ptxas -arch...