mamba-mini

An efficient implementation of selective scan in one file, works with both cpu and gpu, with corresponding mathematical derivation. It is probably the code which is the most close to selective_scan_cuda in mamba.

update!

20240304: New implementation with new derivations! we now support a new approach to implement selective_scan chunk-parallely: selective_scan_easyv3. It is faster than selective_scan_easy when d_state=1, but still slower than mamba_ssm with cuda. We would implement it in triton and test the speed in the future.

mathematical derivation to `chunk-naive version`

code is in selective_scan_easy and SelectiveScanEasy.

mathematical derivation to `chunk-parallel version`

This is the chunk parallel version of selective scan, with support to some different branches. code is in selective_scan_easyv3.

naive code

import torch
def selective_scan_easy(us, dts, As, Bs, Cs, Ds, delta_bias=None, delta_softplus=False, return_last_state=False, chunksize=64):
    """
    # B: batch_size, G: groups, D: dim, N: state dim, L: seqlen
    us: B, G * D, L 
    dts: B, G * D, L
    As: G * D, N
    Bs: B, G, N, L
    Cs: B, G, N, L
    Ds: G * D
    delta_bias: G * D
    # chunksize can be any as you like. But as the chunksize raises, hs may get None, as exp(sum(delta) A) is really small
    """
    def selective_scan_chunk(us, dts, As, Bs, Cs, hprefix):
        """
        partial(h) / partial(t) = Ah + Bu; y = Ch + Du;
        => partial(h*exp(-At)) / partial(t) = Bu*exp(-At);
        => h_t = h_0 + sum_{0}_{t}_{Bu*exp(A(t-v)) dv};
        => h_b = exp(A(dt_a + ... + dt_{b-1})) * (h_a + sum_{a}_{b-1}_{Bu*exp(-A(dt_a + ... + dt_i)) dt_i});
           y_i = C_i*h_i + D*u_i
        """
        """
        us, dts: (L, B, G, D) # L is chunk_size
        As: (G, D, N)
        Bs, Cs: (L, B, G, N)
        Ds: (G, D)
        hprefix: (B, G, D, N)
        """
        ts = dts.cumsum(dim=0)
        Ats = torch.einsum("gdn,lbgd->lbgdn", As, ts).exp()
        scale = Ats[-1].detach()
        rAts = Ats / scale
        duts = dts * us
        dtBus = torch.einsum("lbgd,lbgn->lbgdn", duts, Bs)
        hs_tmp = rAts * (dtBus / rAts).cumsum(dim=0) 
        hs = hs_tmp + Ats * hprefix.unsqueeze(0)
        ys = torch.einsum("lbgn,lbgdn->lbgd", Cs, hs) 
        return ys, hs
    
    inp_dtype = us.dtype
    has_D = Ds is not None

    dts = dts.float()
    if delta_bias is not None:
        dts = dts + delta_bias.view(1, -1, 1).float()
    if delta_softplus:
        dts = torch.nn.functional.softplus(dts)
    
    if len(Bs.shape) == 3:
        Bs = Bs.unsqueeze(1)
    if len(Cs.shape) == 3:
        Cs = Cs.unsqueeze(1)
    B, G, N, L = Bs.shape
    us = us.view(B, G, -1, L).permute(3, 0, 1, 2).float()
    dts = dts.view(B, G, -1, L).permute(3, 0, 1, 2).float()
    As = As.view(G, -1, N).float()
    Bs = Bs.permute(3, 0, 1, 2).float()
    Cs = Cs.permute(3, 0, 1, 2).float()
    Ds = Ds.view(G, -1).float() if has_D else None
    D = As.shape[1]
    
    oys = []
    # ohs = []
    hprefix = us.new_zeros((B, G, D, N), dtype=torch.float)
    for i in range(0, L - 1, chunksize):
        ys, hs = selective_scan_chunk(
            us[i:i + chunksize], dts[i:i + chunksize], 
            As, Bs[i:i + chunksize], Cs[i:i + chunksize], hprefix, 
        )
        oys.append(ys)
        # ohs.append(hs)
        hprefix = hs[-1]

    oys = torch.cat(oys, dim=0)
    # ohs = torch.cat(ohs, dim=0)
    if has_D:
        oys = oys + Ds * us
    oys = oys.permute(1, 2, 3, 0).view(B, -1, L)
    oys = oys.to(inp_dtype)
    # hprefix = hprefix.to(inp_dtype)

    return oys if not return_last_state else (oys, hprefix.view(B, G * D, N))

to test

pytest test_selective_scan.py

MzeroMiko/mamba-mini