Skip to content

chore(renderers): advance to latest main (PR #38, closes #2492)#2518

Merged
hallerite merged 1 commit into
mainfrom
chore/bump-renderers
May 16, 2026
Merged

chore(renderers): advance to latest main (PR #38, closes #2492)#2518
hallerite merged 1 commit into
mainfrom
chore/bump-renderers

Conversation

@hallerite
Copy link
Copy Markdown
Member

@hallerite hallerite commented May 16, 2026

Summary

Bumps the deps/renderers submodule from 87084dc17d0584 (latest renderers main, now tagged renderers-v0.1.8.dev3).

Closes #2492

Confirms PrimeIntellect-ai/renderers#33 (sampled_mask AND in build_training_sample) is current. The SFT renderer path no longer trains on <|im_start|>assistant\n scaffolding — build_training_sample ANDs role attribution with sampled_mask, keeping SFT byte-aligned with the RL trajectory mask. (The upstream fix was already live at the prior pin 87084dc; this bump just keeps us current.)

Picks up

PrimeIntellect-ai/renderers#38 — per-message / per-role analytics on RenderedTokens:

  • New methods: tokens_per_message, tokens_by_role, message_token_spans, role_token_spans
  • New field: message_roles (every renderer now populates)
  • Fixes attribution loss in every bridge_to_next_turn — bridges previously discarded msg_idx / is_sampled even though local emit helpers received them. Bridges now populate message_indices (relative to new_messages), sampled_mask (uniformly False — bridge output is a prompt), and message_roles. Consumers can run the new analytics on bridge output for incremental per-message accounting without re-rendering.

Upstream tag

Cut renderers-v0.1.8.dev3 at commit 17d0584f to unblock CI: renderers uses hatch-vcs with the default guess-next-dev scheme, which refuses to derive a version for untagged commits past a .devN (N>0) tag. The first CI run failed with:

ValueError: Error getting the version from source `vcs`: choosing custom
numbers for the `.devX` distance is not supported.
The 0.1.8.dev2 can't be bumped
Please drop the tag or create a new supported one ending in .dev0

Tagging 17d0584f as v0.1.8.dev3 puts HEAD on a tag so hatch-vcs uses the tag verbatim — no bumping needed. This continues the existing v0.1.8.devN pre-release ladder leading up to the eventual v0.1.8 release.

Surface impact

  • No prime-rl callers touch the new RenderedTokens APIs (grep confirms).
  • No verifiers bump needed — verifiers' renderers>=0.1.8.dev0 constraint is satisfied; its consumers (renderer_client.py) only read bridged.token_ids and bridged.multi_modal_data, both unchanged.
  • pyproject.toml / uv.lock unchanged — renderers is a uv workspace member, so the submodule SHA change doesn't propagate to the lock.
  • Existing SFT dataset tests pass locally (tests/unit/train/sft/test_sft_dataset.py: 17/17).

🤖 Generated with Claude Code


Note

Low Risk
Low risk because this PR only advances the deps/renderers submodule pointer; behavior changes are confined to the external renderers dependency and won’t affect this repo unless callers rely on updated renderer outputs.

Overview
Advances the deps/renderers git submodule to a newer upstream commit/tag.

This pulls in upstream renderer changes around mask/role attribution and new per-message/per-role token analytics on RenderedTokens, without modifying any first-party code in this repo.

Reviewed by Cursor Bugbot for commit d31376d. Bugbot is set up for automated code reviews on this repo. Configure here.

Picks up:
- PrimeIntellect-ai/renderers#38: per-message / per-role analytics
  on RenderedTokens (tokens_per_message, tokens_by_role,
  message_token_spans, role_token_spans, message_roles field).
  Also fixes attribution loss in bridge_to_next_turn so bridge
  output is included in per-message accounting.

Confirms PrimeIntellect-ai/renderers#33 (sampled_mask AND in
build_training_sample) is current — closes #2492. The fix was
already live at the prior pin (87084dc); this bump just keeps
us current with renderers main.

No prime-rl callers touch the new RenderedTokens APIs; bump is
purely additive. pyproject.toml / uv.lock unchanged (renderers
is a workspace member; submodule SHA change doesn't affect the
lock).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@hallerite hallerite marked this pull request as ready for review May 16, 2026 14:20
@hallerite hallerite merged commit cd2f3da into main May 16, 2026
21 of 22 checks passed
@hallerite hallerite deleted the chore/bump-renderers branch May 16, 2026 15:40
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.

SFT renderer path trains on <|im_start|>assistant\n scaffolding (use renderers sampled_mask)

2 participants