Hermes Trismegistus

For example running the following code yields an error that the two tensors are on different devices when trying to add them together. It would be great if this was...

I like option 2 as it's explicit, perhaps allows the compiler to automatically choose the correct device in some cases, and enforces an extra layer of reliability as long as...

Does anyone know if the limited buffer size in wgpu will be alleviated eventually? Even if f16 gets supported the buffer size limits will still be a barrier to running...

Would it be reasonable to use the pub fn [limits](https://docs.rs/wgpu/latest/wgpu/struct.Adapter.html#method.limits)(&self) -> [Limits](https://docs.rs/wgpu/latest/wgpu/struct.Limits.html) function on the adapter to get the best limits that the adapter offers instead of relying on defaults?...

My bad, I accidentally tested with tch.

I get the following panic when trying to run stable diffusion: `thread panicked at 'Error in Queue::submit: Validation Error Caused by: Parent device is lost ', /home/hermes/.cargo/registry/src/index.crates.io-6f17d22bba15001f/wgpu-0.17.0/src/backend/direct.rs:2289:30 note: run with...

Setting `MAX_TASKS` to 1 enabled inference to work, but it was very slow compared to the tch run, ~5 minutes for 1 image with wgpu vs ~15 seconds for two...

Try `result = (power * base.log()).exp()`. This is equivalent to base^power.

I've been able to produce a cache bug on wgpu as well. See branch https://github.com/Gadersd/whisper-burn/tree/wgpu_cache_bug. Running with wgpu with a modification of one line of code results in transcription failure....

For the latter example the query cache is modified when it shouldn't be. Replacing `let wv = qkv_attention(q, k2, v2, None, self.n_head);` with `let wv = qkv_attention(q.clone(), k2, v2, None,...