German Abramov
German Abramov
same, have you solved this issue?
solved this. you should install `setup.py` from `SegFormer` folder with `pip install -e .` . If you have installed not `mmcv` but `mmcv-full` you should change this in `requirements/readthedocs.txt`
I guess we can take ViT weights from `ViT-Pytorch` library, but NLP part of the model contains many Transformer Blocks, so we don't have weights for them. Looks like we...
had some problem on ubuntu `pip install tornado==5.1.1` (after this you need restart jupyter notebook with `jupyter notebook` helped me / 15th July 2020
yes this command helped me with Ubuntu 20.04 and Chromium / 15th July 2020 `pip install notebook==5.6.0`
same issue here, triton.compiler.CompilationError: at 78:24: during training
change triton_flash_attention to flash_attention in yaml is a solution of this problem~
torch 1.13 right, not 2.0
it is works only with triton dev version? official triton 2.0 doesn’t works?
great stuff, I just want to say thanks, the triton flash attention works with: `torch==1.13.1+cu117` `triton-pre-mlir @ git+https://github.com/vchiley/triton.git@2dd3b957698a39bbca615c02a447a98482c144a3#subdirectory=python` and `flash-attn==v1.0.3.post0` installed everything from `setup.py`(gpu setup) with this docker `mosaicml/pytorch:latest`