Trait Mamba3BackendExt 

pub trait Mamba3BackendExt: Backend {
    // Provided method
    fn ssd_serial_recalculated(
        v_bnlrhp: FloatTensor<Self>,
        _da_bnlh: FloatTensor<Self>,
        b_bnlrhn: FloatTensor<Self>,
        c_bnlrhn: FloatTensor<Self>,
        initial_state_bhpr: FloatTensor<Self>,
        da_cumsum_bhnl: FloatTensor<Self>,
    ) -> (FloatTensor<Self>, FloatTensor<Self>) { ... }
}

Extends the backend for the memory-efficient serial recalculated SSD.

The default implementation runs K2-K5 using standard tensor operations. Backends that support a custom memory-efficient backward can override this.

Provided Methods

fn ssd_serial_recalculated( v_bnlrhp: FloatTensor<Self>, _da_bnlh: FloatTensor<Self>, b_bnlrhn: FloatTensor<Self>, c_bnlrhn: FloatTensor<Self>, initial_state_bhpr: FloatTensor<Self>, da_cumsum_bhnl: FloatTensor<Self>, ) -> (FloatTensor<Self>, FloatTensor<Self>)

Memory-efficient MIMO serial SSD.

Arguments

• v_bnlrhp: [batch, nchunks, chunk_len, R, nheads, per_head_dim]
• da_bnlh: [batch, nchunks, chunk_len, nheads] — pre-combined Δ·A
• b_bnlrhn: [batch, nchunks, chunk_len, R, nheads, state_rank]
• c_bnlrhn: [batch, nchunks, chunk_len, R, nheads, state_rank]
• initial_state_bhpr: [batch, nheads, per_head_dim, state_rank]
• da_cumsum_bhnl: [batch, nheads, nchunks, chunk_len] — pre-computed by K1

Returns

• y_bnlrhp: [batch, nchunks, chunk_len, R, nheads, per_head_dim]
• final_state_bhpr: [batch, nheads, per_head_dim, state_rank]

Dyn Compatibility

This trait is not dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety", so this trait is not object safe.

Implementations on Foreign Types

impl<B: Backend + Mamba3BackendExt> Mamba3BackendExt for Autodiff<B>

Available on crate feature autodiff only.

Autodiff-wrapped backends inherit the inner backend’s Mamba3BackendExt impl.

Implementors