Skip to content

fix: two silent math bugs in IntegratedDiscreteDDIMStep and QK-norm attention scale #29

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
Dhakshin2007 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix: two silent math bugs in IntegratedDiscreteDDIMStep and QK-norm attention scale #29

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ac74c70
fix: use full invariant prob vector pi in IntegratedDiscreteDDIMStep
Dhakshin2007 Mar 15, 2026
683fc91
fix: use exp2 for norm_qk_scale to correctly exponentiate log-space p…
Dhakshin2007 Mar 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
7 changes: 3 additions & 4 deletions hackable_diffusion/lib/architecture/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -300,11 +300,10 @@ def __call__(
if self.normalize_qk:
scale = self.param(
"norm_qk_scale",
nn.initializers.constant(
jnp.log2(seq_len_kv**2 - seq_len_kv + SAFETY_EPSILON)
),
(1, 1, 1, 1),
nn.initializers.zeros_init(),
(1, 1, 1, 1),
)
scale = jnp.exp2(scale)

norm_q = jnp.linalg.norm(q, ord=2, axis=-1, keepdims=True)
norm_k = jnp.linalg.norm(k, ord=2, axis=-1, keepdims=True)
Expand Down
6 changes: 2 additions & 4 deletions hackable_diffusion/lib/sampling/discrete_step_sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -696,15 +696,13 @@ def update(

# Extract invariant probabilities.
pi = self.invariant_probs_vec
pi_xt = pi[xt[..., 0]][..., None] # The prior prob of the current token
# (bsz, *seq_len, 1)

# Calculate q(x_t | x_s).
q_xt_given_xs = ratio * xt_oh + (1.0 - ratio) * pi_xt
q_xt_given_xs = ratio * xt_oh + (1.0 - ratio) * pi
# (bsz, *seq_len, M)

# Calculate q(x_t | x_0)'
q_xt_given_x0 = alpha_t * xt_oh + (1.0 - alpha_t) * pi_xt
q_xt_given_x0 = alpha_t * xt_oh + (1.0 - alpha_t) * pi
# (bsz, *seq_len, M)

# Calculate integration weights: W(x_0) = p(x_0 | x_t) / q(x_t | x_0).
Expand Down