#![recursion_limit = "256"]
mod dense_causal_attention;
mod fusion_compat;
mod profiling;
mod recurrent;
mod relu_lowrank;
mod sequence;
pub mod api {
pub use crate::kernels::{attention, projection, recurrent};
pub mod expert {
pub use crate::kernels;
}
}
pub mod kernels {
pub mod sequence {
pub use crate::sequence::{linear, mamba3};
}
pub mod attention {
pub use crate::dense_causal_attention::{
CompiledDenseCausalAttentionPlan, supports_dense_causal_attention_backend,
try_fused_dense_causal_attention_wgpu, try_fused_dense_causal_attention_wgpu_with_plan,
};
}
pub mod recurrent {
pub use crate::recurrent::{
CompiledRecurrentAttentionPlan, RecurrentAttentionOutput, RecurrentProfileSnapshot,
recurrent_profile_reset, recurrent_profile_snapshot,
supports_backend as supports_recurrent_backend, try_fused_recurrent_attention_wgpu,
try_fused_recurrent_attention_wgpu_with_plan,
};
}
pub mod projection {
pub use crate::relu_lowrank::{
LowrankForwardRouteProfileSnapshot, LowrankGradInputExecutor,
LowrankProjectionProfileSnapshot, relu_lowrank_forward_profile_reset,
relu_lowrank_forward_profile_snapshot, relu_lowrank_forward_route_profile_reset,
relu_lowrank_forward_route_profile_snapshot, relu_lowrank_grad_input_profile_reset,
relu_lowrank_grad_input_profile_snapshot, relu_lowrank_grad_weight_profile_reset,
relu_lowrank_grad_weight_profile_snapshot, supports_relu_lowrank_projection_backend,
try_fused_relu_lowrank_projection_wgpu,
try_fused_relu_lowrank_projection_wgpu_with_executor,
};
}
}