cdutr

I tried this, but the cookie logic stops working. Apparently the lib depends on the `site.com/cookies` URL being available and sending to the cookies page.

Yes, I would like to! I'll probably need some help, but I'll start working on it

I implemented and benchmarked selective attention from the [ICLR 2025 paper](https://arxiv.org/abs/2410.02703). Unfortunately, it doesn't work well for diffusion models. ### The Problem The paper's speedup comes from **KV-cache eviction** in...

Hey @kashif! I've prepared a documentation update with a new Performance section covering: - Attention backend benchmarks (from your tests) - torch.compile speedup (~2.4x) - Variable-length prompt handling with CFG...

Thanks @sayakpaul! I removed the benchmark script, moved all tests to this [gist](https://gist.github.com/cdutr/bea337e4680268168550292d7819dc2f). ## torch.compile test Also tested the performance with `torch.compile`, and results were similar, the details are below....

Hey. I would like to work on the last inputs and comments. I was off last week attending a conference but should go back to it tomorrow. Any additional feedback...

Hi, I am back, will reply the points below: ## @dxqb ### Comment 1: encoder_hidden_states_mask Type Hints Updated to use `torch.BoolTensor` type hint and added documentation clarifying that only boolean...

Hey @kashif, the comparisons on your PR are really nice\! Yes, I can point my PR to yours. Let me know once you've removed the overlapping parts and I'll rebase...

Thanks a lot for the MVP rewards, I really appreciate it and I'm really enjoying working on this project. It makes total sense to move forward with merging the other...

@sayakpaul, I will close this PR, since we are working in the other one, ok?