Arthur Brussee

Ah didn't realise Bevy MSRV isn't 1.75 yet, that makes this a bit pointless. Also WASM doesn't build - trait fns need to be optionally Send, only on non-wasm platforms....

> > Ah didn't realise Bevy MSRV isn't 1.75 yet, that makes this a bit pointless. > > Shouldn't be a blocker, I think we can bump it. Oh ok...

So it seems there _is_ a way to make conditional Send bounds, and more people do it [this way](https://github.com/gfx-rs/wgpu/blob/trunk/wgpu-types/src/lib.rs#L7023). Have made an attempt with a "FutureSend" (could be MaybeSend? ConditionalSend?)...

Thanks a lot for that! Really nice to hear your design thoughts :) Great that the Wasm solution sounds workable! I've renamed FutureSend to WasmNotSend as per giooschi's suggestion on...

Oh neat! Yeah prefer that too, made it `ConditionalSend` now. There could even be a generic feature for bevy tasks to support a work-stealing scheduler or not (which then just...

Have updated to latest main in case this is still the way to go :) Some weird CI failures atm that I think are unrelated, other PRs seem to have...

Wellll seems like I fucked the merge pretty bad :') I'll redo this PR at head, it was getting a bit messy anyway.

Also just to chime in - I imagine training massive LLMs or even big convnets in the browsers is niche. However, consider NeRF and gaussian splat like models, training in...

Hm, converting this to draft for a sec as it seems tests are timing out waiting for data :/ Maybe poll isn't enough after all, will investigate.

You can already do this! Something like: ```rust // Get underlying jit tensor. let jit_tensor = my_tensor.into_primitive().tensor(); // Get the 'resource' from the client let resource = jit_tensor.client.get_resource(jit_tensor.handle.clone().binding()); // Which...