Conversation
samsja
approved these changes
May 9, 2026
Merged
mikasenghaas
added a commit
that referenced
this pull request
Sep 1, 2026
## Summary - advance the `deps/verifiers` submodule from `e2103d6` to `d4a2177` - include merged Bash and RLM context compaction from verifiers #2454 and #2459 - pick up nano-rlm compaction commit `4ef3438` through the default RLM harness pin - refresh `uv.lock` for verifiers' `aiohttp>=3.14.1` requirement - migrate three RLM examples to `compaction.summarize_at_tokens`; use the former range's `98_304` midpoint - include the optional ACP semantic-edge and shared harness utility changes already on verifiers `main` Companions: [verifiers #2454](PrimeIntellect-ai/verifiers#2454), [verifiers #2459](PrimeIntellect-ai/verifiers#2459), and [nano-rlm #147](PrimeIntellect-ai/nano-rlm#147). ## Breaking - RLM harness configs must move `summarize_at_tokens` to `compaction.summarize_at_tokens`. - Leave `compaction` unset to disable proactive and reactive compaction. ## Verification - `git diff --check` - `uv lock` - `uv lock --check` - `uv run pytest -q tests/unit/test_configs.py` — 133 passed - `git submodule status deps/verifiers` — `d4a217794fc0bfd70369a8230e56653c193da8ce` - verified that `d4a2177` descends from the previous `e2103d6` pin <!-- CURSOR_SUMMARY --> --- > [!NOTE] > **Medium Risk** > Submodule bump plus a breaking harness config shape affects all RLM runs still using the old `summarize_at_tokens` field; example migrations are mechanical but custom configs must be updated. > > **Overview** > Advances the **`deps/verifiers`** submodule to pick up merged Bash/RLM **context compaction** (and related harness defaults), with **`uv.lock`** refreshed so **`aiohttp`** meets the new **`>=3.14.1`** floor. > > **Breaking for RLM harness TOML:** flat **`env.agent.harness.summarize_at_tokens`** is replaced by **`env.agent.harness.compaction.summarize_at_tokens`**. The three advanced examples (**`glm-4.5-air/search`**, **`glm-4.5-air/terminal`**, **`nemotron-3-super/swe`**) are updated accordingly—train sources that used a two-threshold list **`[65536, 131072]`** now use a single **`98304`** threshold; eval sources keep **`98304`** but under the nested **`compaction`** key. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit 7429aba. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup> <!-- /CURSOR_SUMMARY -->
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Note
Medium Risk
Modifies inference weight-loading logic for MoE expert tensor slicing and adds a new config invariant that can block runs; issues could cause runtime weight-shape mismatches or failed config validation when EPLB/quantized transfer are enabled.
Overview
Adds a new
RLConfigvalidator that requiresinference.enable_eplbto be paired with NCCL weight broadcast andquantize_in_weight_transfer=true, failing fast on unsupported combinations.Updates vLLM kernel weight transfer to handle EPLB expert remapping: expert tensor slicing now uses source expert indices derived from vLLM’s EPLB
logical_to_physical_map, with extra validation for invalid/missing expert mappings before doing in-placecopy_updates.Reviewed by Cursor Bugbot for commit 109974f. Bugbot is set up for automated code reviews on this repo. Configure here.