acsweet

I've added a few tests, how do they look to you? The changes caused one test around buffers to very occasionally fail (`tests/array_tests.cpp "test array shared buffer"`), I think related...

I ran a few benchmarks, apologies for the delay! @awni Results are below, prompt tps and generated tps. ### Prompt TPS | Model | This PR | Current MLX Release...

It does seem to be slightly slower (with some high variance on the prompt tps). Not sure where to go from here. If this makes the PR a no go,...

Thank you! This might be safe to close! results on linux with build for latest mlx release (Intel i9-9980HK, Ubuntu 24.04.2 LTS): ``` (conv3d) max_diff: 0.0 (conv3d) total_absolute_diff: 0.0 (conv2d)...

@fbadine Were these tests passing now for you? I'm still getting similar errors for MLX `0.26.3` with `Ubuntu 24.04.2 LTS`