Skip to content

Add feature to initialize hidden state in Selective Scan #806

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add feature to initialize hidden state in Selective Scan #806

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 19 additions & 0 deletions mamba_ssm/ops/selective_scan_interface.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,6 +111,25 @@ def selective_scan_fn(u, delta, A, B, C, D=None, z=None, delta_bias=None, delta_
"""
return SelectiveScanFn.apply(u, delta, A, B, C, D, z, delta_bias, delta_softplus, return_last_state)

def select_scan_fn_init_hidden(u, delta, A, B, C, D=None, z=None, delta_bias=None, delta_softplus=False,
return_last_state=False, init_state=None):
"""
Selective Scan with hidden state initialization
It uses the base kernel method but applies a transformation to B to mimic the initialization of the hidden state.
Memory and time overhead is minimal compared to ref method.
"""
delta2 = delta + delta_bias[..., None].float()
v = F.softplus(delta2[:,:,0]) * u[:,:,0] + 1e-7
B = B.clone()
sig = (init_state[:,:,:] * torch.exp(F.softplus(delta2[:,:,0]).unsqueeze(-1) * A.unsqueeze(0)) ) / v.unsqueeze(-1)
dim = A.shape[0]
B = repeat(B, "B G N L -> B (G H) N L", H=dim // B.shape[1])
B = B.clone()
B[:,:,:,0] = B[:,:,:,0] + sig
C = C.clone()
C = repeat(C, "B G N L -> B (G H) N L", H=dim // C.shape[1])

return SelectiveScanFn.apply(u, delta, A, B, C, D, z, delta_bias, delta_softplus, return_last_state)

def selective_scan_ref(u, delta, A, B, C, D=None, z=None, delta_bias=None, delta_softplus=False,
return_last_state=False):
Expand Down