Skip to content

feat(window): add opt-in slice gather for relative bias #43

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

Merged
noelEOS merged 5 commits into from
Nov 3, 2026
Merged

feat(window): add opt-in slice gather for relative bias #43

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
9d55827
feat(window): add opt-in slice gather for relative bias
noelEOS Nov 3, 2026
d0f72e8
test(window): record slice-gather offline evidence
noelEOS Nov 3, 2026
0be041f
test(window): record live slice-gather validation
noelEOS Nov 3, 2026
1ac6b4c
docs(window): record validated slice lookup scope
noelEOS Nov 3, 2026
b2343d8
fix(window): preserve existing validation reports and statuses
noelEOS Nov 3, 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
4 changes: 2 additions & 2 deletions docs/building-blocks.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -155,7 +155,7 @@ their own.
| [`class_token_attention`](../tensor_ee/blocks/window/attention.py) | W → W | window | composite | no | Ready — window 3, heads 3, dim 12; window 14, heads 12, dim 768; scale_query=True: window 14, heads 12, dim 768; scale_query=True: window 3, heads 3, dim 18 | [evidence/class-token-offline.json](evidence/class-token-offline.json); [evidence/live-fixtures/20261029T174042Z-class_token.json](evidence/live-fixtures/20261029T174042Z-class_token.json); [evidence/class-token-query-scale-offline.json](evidence/class-token-query-scale-offline.json); [evidence/live-fixtures/20261102T122934Z-class_token_query_scale.json](evidence/live-fixtures/20261102T122934Z-class_token_query_scale.json) |
| [`naive_attention`](../tensor_ee/blocks/window/attention.py) | W → W | window | `round_each_op` | no | Ready | [building-blocks-2026-10-11.md#36-linear-layers-and-attention](history/building-blocks-2026-10-11.md#36-linear-layers-and-attention); [ee_attention/notes/findings.md](history/ee_attention/notes/findings.md); [evidence/segformer_b0-live.json](evidence/segformer_b0-live.json) |
| [`split_heads`](../tensor_ee/blocks/window/attention.py) | W → W | pointwise | none (no arithmetic) | no | Ready | [building-blocks-2026-10-11.md#36-linear-layers-and-attention](history/building-blocks-2026-10-11.md#36-linear-layers-and-attention); [swin_block/notes/findings.md](history/swin_block/notes/findings.md) |
| [`relative_position_bias`](../tensor_ee/blocks/window/attention.py) | V → V | pointwise | `double_round_once` | no | Ready — windows 5, 7, 8, 12 | [building-blocks-2026-10-11.md#36-linear-layers-and-attention](history/building-blocks-2026-10-11.md#36-linear-layers-and-attention); [swin_block/notes/findings.md](history/swin_block/notes/findings.md); [evidence/p01-offline.json](evidence/p01-offline.json); [evidence/live-fixtures/20261013T074301Z-p01a.json](evidence/live-fixtures/20261013T074301Z-p01a.json); [evidence/live-fixtures/20261013T074330Z-p01b.json](evidence/live-fixtures/20261013T074330Z-p01b.json); [evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json](evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json) |
| [`relative_position_bias`](../tensor_ee/blocks/window/attention.py) | V → V | pointwise | `double_round_once` | no | Ready — windows 5, 7, 8, 12; gather=slices: windows 7, 12, 14, heads 2 | [building-blocks-2026-10-11.md#36-linear-layers-and-attention](history/building-blocks-2026-10-11.md#36-linear-layers-and-attention); [swin_block/notes/findings.md](history/swin_block/notes/findings.md); [evidence/p01-offline.json](evidence/p01-offline.json); [evidence/live-fixtures/20261013T074301Z-p01a.json](evidence/live-fixtures/20261013T074301Z-p01a.json); [evidence/live-fixtures/20261013T074330Z-p01b.json](evidence/live-fixtures/20261013T074330Z-p01b.json); [evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json](evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json); [evidence/window-relative-bias-slices-offline.json](evidence/window-relative-bias-slices-offline.json); [evidence/live-fixtures/20261103T173919Z-relative_bias_slices.json](evidence/live-fixtures/20261103T173919Z-relative_bias_slices.json) |
| [`window_attention`](../tensor_ee/blocks/window/attention.py) | W → W | window | composite | no | Ready — windows 5, 7, 8, 12 | [building-blocks-2026-10-11.md#36-linear-layers-and-attention](history/building-blocks-2026-10-11.md#36-linear-layers-and-attention); *parts: linear (double_round_once), naive_attention (round_each_op)*; [swin_block/notes/findings.md](history/swin_block/notes/findings.md); [swin_pair/notes/findings.md](history/swin_pair/notes/findings.md); [evidence/p01-offline.json](evidence/p01-offline.json); [evidence/live-fixtures/20261013T074301Z-p01a.json](evidence/live-fixtures/20261013T074301Z-p01a.json); [evidence/live-fixtures/20261013T074330Z-p01b.json](evidence/live-fixtures/20261013T074330Z-p01b.json); [evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json](evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json) |
| [`shifted_window_mask`](../tensor_ee/blocks/window/attention.py) | W → W | window | `double_round_once` | yes | Ready — windows 5, 7, 8, 12; shift ⌊w/2⌋ | [building-blocks-2026-10-11.md#36-linear-layers-and-attention](history/building-blocks-2026-10-11.md#36-linear-layers-and-attention); [swin_pair/notes/findings.md](history/swin_pair/notes/findings.md); [evidence/p01-offline.json](evidence/p01-offline.json); [evidence/live-fixtures/20261013T074301Z-p01a.json](evidence/live-fixtures/20261013T074301Z-p01a.json); [evidence/live-fixtures/20261013T074330Z-p01b.json](evidence/live-fixtures/20261013T074330Z-p01b.json); [evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json](evidence/live-fixtures/20261013T075542Z-window-relative_position_bias-w12.json) |
| [`depthwise_conv3x3_window`](../tensor_ee/blocks/window/conv.py) | W → W | window + ring of 1 | `double_round_once` | no | Ready — 3 × 3, stride1, padding1, dilation1, multiplier1; window7; D4 optional bias, D1024 biased; D16 biased in C4/D16 Mix-FFN | [evidence/segformer_b0-offline.json](evidence/segformer_b0-offline.json); [evidence/segformer_b0-live.json](evidence/segformer_b0-live.json) |
Expand Down Expand Up @@ -256,7 +256,7 @@ Each block's `validated` record, the status in machine-readable form: 88 records
| [`class_token_attention`](../tensor_ee/blocks/window/attention.py) | W → W | {dim 12; heads 3; scale_query False; window 3} ∪ {dim 768; heads 12; scale_query False; window 14} ∪ {dim 768; heads 12; scale_query True; window 14} ∪ {dim 18; heads 3; scale_query True; window 3} | — |
| [`naive_attention`](../tensor_ee/blocks/window/attention.py) | W → W | any accepted call | — |
| [`split_heads`](../tensor_ee/blocks/window/attention.py) | W → W | any accepted call | — |
| [`relative_position_bias`](../tensor_ee/blocks/window/attention.py) | V → V | {window 5 or 7 or 8 or 12} | — |
| [`relative_position_bias`](../tensor_ee/blocks/window/attention.py) | V → V | {gather 'one_hot'; window 5 or 7 or 8 or 12} ∪ {gather 'slices'; heads 2; window 7 or 12 or 14} | — |
| [`window_attention`](../tensor_ee/blocks/window/attention.py) | W → W | {window 5 or 7 or 8 or 12} | — |
| [`shifted_window_mask`](../tensor_ee/blocks/window/attention.py) | W → W | {shift 2 or None; value -100.0; window 5} ∪ {shift 3 or None; value -100.0; window 7} ∪ {shift 4 or None; value -100.0; window 8} ∪ {shift 6 or None; value -100.0; window 12} | — |
| [`depthwise_conv3x3_window`](../tensor_ee/blocks/window/conv.py) | W → W | {bias True or False; channels 4; dilation 1; k 3; multiplier 1; padding 1; stride 1; window 7} ∪ {bias True; channels 16 or 1024; dilation 1; k 3; multiplier 1; padding 1; stride 1; window 7} | `DepthwiseConv3x3Constants` |
Expand Down
Loading
Loading