Clean up -jit suffix in feature flags and modules

[laggui]

Checklist

• Confirmed that run-checks all script has been executed.

Changes

Feature flags:

• "cuda-jit" -> "cuda"
• "hip-jit" -> "hip"
• "wgpu-spirv" -> "vulkan" (for burn and burn-core, which enables burn-wgpu/spirv)

Modules:

• burn::backend::cuda_jit -> burn::backend::cuda
• burn::backend::hip_jit -> burn::backend::hip

Backends:

• CudaJit -> Cuda
• HipJit -> Hip

[syl20bnr]

Nice!
Tested on the AMD runner.

You can update the commented tests in burn-hip/src/lib.rs to match the CUDA ones. I did not execute them all, I just verified that there was no compilation issues. Keep them commented.

mod tests {
    use burn_jit::JitBackend;

    pub type TestRuntime = cubecl::hip::HipRuntime;
    pub use half::f16;

    burn_jit::testgen_all!([f16, f32], [i8, i16, i32, i64], [u8, u32]);
}

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 83.20%. Comparing base (f630b3b) to head (5e63c00).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2705   +/-   ##
=======================================
  Coverage   83.20%   83.20%           
=======================================
  Files         819      819           
  Lines      106814   106814           
=======================================
  Hits        88870    88870           
  Misses      17944    17944           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[nathanielsimard]

I think we should create a type alias for vulkan.

pub type Vulkan = Wgpu<Compiler = Spirv>;

Also, I would keep a feature flag spirv in burn_wgpu, so that users can customize the compiler used for Wgpu without using the type alias.

Also, the Wgpu backend should not have a default compiler, I think. We could create a WebGPU type alias like Vulkan with the default compiler being Wgsl.

So, we can have a hierarchy of types, where we don't compromise on flexibility but make it clearer for Burn users.

New feature flags:

vulkan = ["burn_wgpu", "spirv"]
webgpu = ["burn_wgpu", "burn_wgpu/wgsl"]
wgpu = ["burn_wgpu"]

We have to make it clear that the Wgpu backend is decoupled from the compiler; users can create their own feature flag to choose the compiler if they want it to change depending on the platform. We could (and probably should) create a composed compiler that will become the default Wgpu compiler, where each compiler feature flag activates a new compiler option to be determined at runtime. However, for now, we can't do that just yet, and users won't have the best performance by default because of that, and it's not clear how to do that right now.

Also, having the possibility to just have one compiler activated will always be a good option to reduce binary size and ease deployment.