sabreshao

Global function pointer array doesn't work in GPU kernel

7

Linking fails if GPU kernel call function from a global function pointer array. Error message: `error: :0:0: in function _Z8vadd_hipN8hip_impl17Empty_launch_parmEPKfS2_Pfi void (%"struct.std::less", float*, float*, float*, i32): unsupported indirect call to...

[Diffusion] add rocm support

5

## Motivation Enable sglang diffusion on AMD hardware. ## Modifications 1, Add hip path for custom_op. 2, Use Aiter as default attention backend on ROCm. ## Accuracy Tests Verified on...