Skip to content

Vectorize promo logit computation #2

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
mcognetta wants to merge 3 commits into
base: main
Choose a base branch
from
+7 −15
Open

Vectorize promo logit computation #2

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e93fe08
vectorize promo logit computation
mcognetta May 25, 2026
ed4a0ad
add shape information
mcognetta May 25, 2026
f24d553
remove trailing newline
mcognetta May 25, 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
22 changes: 7 additions & 15 deletions maia3/models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -333,6 +333,8 @@ def __init__(self, cfg):

self.promo_bias_proj = nn.Linear(cfg.head_hid_dim, 4, bias=False) # 4 promotion types: q, r, b, n

self.rank7_indices = [chess.square(file, 6) for file in range(8)] # squares 48-55
self.rank8_indices = [chess.square(file, 7) for file in range(8)] # squares 56-63

def interpolate_elo(self, elos):

Expand Down Expand Up @@ -375,26 +377,16 @@ def forward(self, tokens, self_elos, oppo_elos):
scores_base = torch.einsum("bid,bjd->bij", sq_from, sq_to) / math.sqrt(self.cfg.head_hid_dim)
scores_flat = scores_base.reshape(x.size(0), 64 * 64) # (B, 4096)

rank7_indices = [chess.square(file, 6) for file in range(8)] # squares 48-55
rank8_indices = [chess.square(file, 7) for file in range(8)] # squares 56-63

rank8_features = sq_to[:, rank8_indices, :] # (B, 8, head_hid_dim)
rank8_features = sq_to[:, self.rank8_indices, :] # (B, 8, head_hid_dim)
promo_biases = self.promo_bias_proj(rank8_features) * math.sqrt(self.cfg.head_hid_dim) # (B, 8, 4) for q,r,b,n

promotion_logits = []
for from_file in range(8): # source file (a-h)
from_sq = rank7_indices[from_file]
for to_file in range(8): # target file (a-h)
to_sq = rank8_indices[to_file]
base_score = scores_base[:, from_sq, to_sq] # (B,)
for piece_idx in range(4): # q=0, r=1, b=2, n=3
bias = promo_biases[:, to_file, piece_idx] # (B,)
promotion_logits.append((base_score + bias).unsqueeze(1))
promotion_logits = torch.cat(promotion_logits, dim=1) # (B, 256)
base = scores_base[:, self.rank7_indices][:, :, self.rank8_indices] # (B, 8, 8)
Copy link
Copy Markdown
Author

@mcognetta mcognetta May 25, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The rank_7/rank_8 could also just be replaced directly with a slice like:

base = scores_base[:, 48:56][:, :, 56:65].

This avoids a new allocation, etc. since it is just a contiguous view. Doesn't give much speedup on my machine though, so maybe not worth the loss in clarity.

promotion_logits = (base.unsqueeze(-1) + promo_biases.unsqueeze(1)).reshape(x.size(0), 256) # (B, 256)

logits_move = torch.cat([scores_flat, promotion_logits], dim=1) # (B, 4352)

x = self.last_ln(x.mean(dim=1))
logits_value = self.fc_value(F.relu(self.fc_value_hid(x))) # (B, 3)
logits_ponder = self.fc_ponder(F.relu(self.fc_ponder_hid(x))) # (B, 1)

return logits_move, logits_value, logits_ponder.squeeze(1) # (B, 4352), (B, 3), (B,)
return logits_move, logits_value, logits_ponder.squeeze(1) # (B, 4352), (B, 3), (B,)