Skip to content

docs(WO79): fp16 IO-boundary diagnosis (ruby/ts cannot run fp16 artifacts) - #326

Merged
ronaldtse merged 1 commit into
mainfrom
wo79-fp16-diagnosis
Oct 22, 2026
Merged

ronaldtse merged 1 commit into
mainfrom
wo79-fp16-diagnosis

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Full diagnosis from the C1 matrix first run + the recommended exporter-side fix (Cast-to-fp32 on fp16 graph outputs; weights stay fp16) and the alternatives rejected. Contains an owner decision: re-exporting changes artifact sha256s (locator and ids unchanged). C1 matrix PRs stay open until this lands — they are correctly red against the current artifacts.

…IO-boundary fix

The C1 conformance matrix caught it on first run: ruby cannot execute
fp16 artifacts (numo has no Float16), ts corrupts the fp16 boundary,
py shows fp16 near-tie threading sensitivity. Recommended fix: the
exporter wraps fp16 graph outputs in Cast(to=FLOAT) — weights stay
fp16, IO becomes fp32, every runtime matches the reference path.
Owner decision recorded in the WO.
@ronaldtse
ronaldtse merged commit ece107e into main Oct 22, 2026
12 checks passed
@ronaldtse
ronaldtse deleted the wo79-fp16-diagnosis branch October 22, 2026 06:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant