[wgpu] Memory management strategy too greedy?

[laggui]

Trying to run Llama 3.1 (and even TinyLlama, which I was previously able to use) on wgpu I get an error due to lack of memory.

On WSL:

Loading record...
thread 'main' panicked at /home/laggui/.cargo/registry/src/index.crates.io-6f17d22bba15001f/cubecl-runtime-0.2.0/src/memory_management/dynamic.rs:156:9:
No memory pool big enough to reserve 262144000 bytes.
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

On Windows:

Loading record...
thread 'main' panicked at C:\Users\guila\.cargo\registry\src\index.crates.io-6f17d22bba15001f\wgpu-22.1.0\src\backend\wgpu_core.rs:3411:5:
wgpu error: Validation Error

Caused by:
  In Device::create_buffer
    Not enough memory left.


note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
error: process didn't exit successfully: `target\release\examples\chat.exe` (exit code: 101)

My notebook only has 6GB VRAM but I can run with any other backend no problem.
I know other users on discord have reported a similar issue even with 4090s.