Skip to content

Fix inter-chunk recurrence in the Zamba2 / Nemotron-H Mamba2 slow path #47250

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

Closed
Jeronymous wants to merge 4 commits into from
Closed

Fix inter-chunk recurrence in the Zamba2 / Nemotron-H Mamba2 slow path #47250

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
4415fa5
Fixes #47246 : Fix inter-chunk recurrence in Zamba2/Nemotron-H Mamba2…
Jeronymous Jul 10, 2026
c9288cd
Add regression tests for the Mamba2 slow-path inter-chunk recurrence
Jeronymous Jul 10, 2026
10f37af
Address review on the slow-path tests
Jeronymous Jul 15, 2026
65abbe6
Refine slow-path tests per review
Jeronymous Jul 15, 2026
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
5 changes: 2 additions & 3 deletions src/transformers/models/nemotron_h/modeling_nemotron_h.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -551,9 +551,8 @@ def torch_forward(self, input_states, cache_params: Cache | None=None, attention
states = torch.cat([previous_states, states], dim=1)
decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))

states_permuted = states.permute(0, 2, 1, 3, 4)
result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2)
new_states = result.permute(0, 2, 1, 3, 4)
decay_chunk = decay_chunk.transpose(1, 3)
new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
states, ssm_state = new_states[:, :-1], new_states[:, -1]

# Compute state -> output conversion per chunk
Expand Down
5 changes: 2 additions & 3 deletions src/transformers/models/zamba2/modeling_zamba2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -839,9 +839,8 @@ def torch_forward(self, input_states, cache_params: Cache | None=None, attention
states = torch.cat([previous_states, states], dim=1)
decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))

states_permuted = states.permute(0, 2, 1, 3, 4)
result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2)
new_states = result.permute(0, 2, 1, 3, 4)
decay_chunk = decay_chunk.transpose(1, 3)
new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
states, ssm_state = new_states[:, :-1], new_states[:, -1]

# Compute state -> output conversion per chunk
Expand Down
5 changes: 2 additions & 3 deletions src/transformers/models/zamba2/modular_zamba2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -627,9 +627,8 @@ def torch_forward(self, input_states, cache_params: Cache | None=None, attention
states = torch.cat([previous_states, states], dim=1)
decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))

states_permuted = states.permute(0, 2, 1, 3, 4)
result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2)
new_states = result.permute(0, 2, 1, 3, 4)
decay_chunk = decay_chunk.transpose(1, 3)
new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
states, ssm_state = new_states[:, :-1], new_states[:, -1]

# Compute state -> output conversion per chunk
Expand Down