Transaction-Science/open-standards
Watch
1
0
Fork
You've already forked open-standards
 0
Code Issues Pull requests Projects Releases 15 Packages Wiki Activity Actions

wai-rs: hyper_scale_decoder floor(>>) not truncation — byte-exact vs the reference decoder #4

Merged
dcharlot merged 1 commit from hyperscale-floorfix into main 2026-07-12 12:22:07 -04:00
Conversation 0 Commits 1 Files changed 2 +37 -48
dcharlot commented 2026-07-12 12:22:06 -04:00
Owner
Copy link

The per-channel power-of-two division is an arithmetic right shift (floor) in the reference decoder; the exported ONNX Div truncates toward zero, which differs on negatives (~1 LSB). Switching decode_hyper_scale from / to >> takes it from 30765/40960 to 40960/40960 BYTE-EXACT vs the reference i_s on real test256.bin data.

Retargets the calibration test from the (truncating) onnx.reference golden to the real reference-decoder i_s — the correct target.

General rule this establishes: the exported ONNX integer Div is a lossy stand-in for >>; trust the reference source, not the ONNX.

The per-channel power-of-two division is an arithmetic right shift (floor) in the reference decoder; the exported ONNX Div truncates toward zero, which differs on negatives (~1 LSB). Switching decode_hyper_scale from `/` to `>>` takes it from 30765/40960 to **40960/40960 BYTE-EXACT** vs the reference i_s on real test256.bin data. Retargets the calibration test from the (truncating) onnx.reference golden to the real reference-decoder i_s — the correct target. General rule this establishes: the exported ONNX integer `Div` is a lossy stand-in for `>>`; trust the reference source, not the ONNX.
The fully-integer hyper_scale_decoder does its per-channel ÷2^n as an ARITHMETIC
RIGHT SHIFT (floor), matching the real reference decoder. The exported ONNX
represents it as integer `Div` (truncation toward zero), which differs from `>>` on
negatives by 1 LSB — a ~1-LSB-wrong export. decode_hyper_scale used `/` (matching the
ONNX); switch it to `>>` (matching the reference).

Impact, on the real test256.bin hyper-latent (z[1,160,4,4] → i_s[1,160,16,16]): the
old truncation matched only 30765/40960 of the reference i_s (the rest off by exactly
+1, at the negative-remainder positions surfaced by the final abs); with `>>` it is
40960/40960 BYTE-EXACT vs the reference decoder.

Retargets tests/jpegai_hyperscale_calibration.rs from the (truncating) onnx.reference
golden to the real reference-decoder i_s — the correct target — validated byte-exact
on real data. Structural lib test unaffected (shape/range/determinism).

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
Transaction-Science/open-standards!4
No description provided.