Kristian Klemon
Kristian Klemon
Many thanks for your awesome project! While it works perfectly using the provided command line script, it's not straightforward to use it directly from code, e.g. when the FID computation...
**Bug description** The usage example in the project's README uses the length of the original protein sequence to determine the slicing indices for the model output with respect to padding....
[FlashAttention](https://arxiv.org/abs/2205.14135) is an IO-aware, highly optimized implementation of the Attention mechanism for Nvidia GPUs. Since it's implemented in CUDA, it should in principle also compile and run with ZLUDA, right?...
Inspired by this work, I implemented the Perceiver architecture with out-of-the-box FlashAttention support. It offers a great speedup over a naive implementation and up to 16x increased input sequence lengths...
[FlashAttention](https://github.com/Dao-AILab/flash-attention) is an approximation-free fast and memory efficient implementation of multi-head attention. In particular, it reduces the memory requirement from $\mathcal{O}(n^2)$ to $\mathcal{O}(n)$ and would therefore be highly beneficial for...
## Motivation Many users and workflows rely on custom nodes and extensions which correspondingly should also be available in a RunPod worker. [ComfyUI Manager](https://github.com/ltdrdata/ComfyUI-Manager), perhaps the most popular ComfyUI management...
## Motivation The ComfyUI "Save Image" node automatically creates parent folders if subfolders are provided in the save name. The ComfyUI API correspondingly provides the subfolder for a given output...
It would be nice if the package could be published to PyPI. Otherwise, it's really difficult to include it as a dependency in other packages that should be published, since...
Not really an issue, but I made a PyTorch implementation for the discrete model, available [here](https://github.com/kklemon/bgan-pytorch), if anyone is interested.
# Motivation All the routes of the worker test server (e.g. `/run`, `/sync`, etc.) start at the root-level. This does not align with the public API schema where worker routes...