Skip to content

[WS2][Logp] Expose the lse_vocab_tile_reduction ablation switch on linear-logp ops - #317

Open
zhangj1an wants to merge 1 commit into
RL-Align:testfrom
zhangj1an:jian/logp-lse-vocab-tile-reduction-switch
Open

[WS2][Logp] Expose the lse_vocab_tile_reduction ablation switch on linear-logp ops#317
zhangj1an wants to merge 1 commit into
RL-Align:testfrom
zhangj1an:jian/logp-lse-vocab-tile-reduction-switch

Conversation

@zhangj1an

@zhangj1an zhangj1an commented Aug 17, 2026

Copy link
Copy Markdown
Collaborator

Add a simple API to turn on or off deterministic LSE vocab reduction order for logprob ops. Follows @KJLdefeated's design in PR #314.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TGgEEbTcLRZjwePXJztT9S

@coderabbitai

coderabbitai Bot commented Aug 17, 2026

Copy link
Copy Markdown

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 2f04a83d-5e02-4153-8c68-64ab0cff7b59

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@zhangj1an
zhangj1an force-pushed the jian/logp-lse-vocab-tile-reduction-switch branch from 4d714d0 to 84b1738 Compare August 17, 2026 15:41
@zhangj1an
zhangj1an requested a review from bitborne as a code owner August 17, 2026 15:41
@zhangj1an
zhangj1an changed the base branch from main to test August 17, 2026 15:41
@zhangj1an
zhangj1an force-pushed the jian/logp-lse-vocab-tile-reduction-switch branch from 84b1738 to 5a2f048 Compare August 17, 2026 16:04
The WS2 mismatch matrix addresses switches by dotted path. Take
`logp.lse_vocab_tile_reduction` as a keyword-only argument on all three
linear-logp backends so the driver can reach it whichever one the
registry resolves to on a given device.

Per call rather than per instance: `KernelRegistry` caches op instances
for the life of the process, so an attribute would leak one ablation
arm's setting into every later call.

Interface only: the argument defaults to `original`, the per-shard merge
these ops already do, so behaviour is unchanged. `fixed` raises rather
than falling back, which would make the matrix report the factor as
irrelevant.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TGgEEbTcLRZjwePXJztT9S
@zhangj1an
zhangj1an force-pushed the jian/logp-lse-vocab-tile-reduction-switch branch from 5a2f048 to 775a364 Compare August 18, 2026 13:06
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