Vllm w load maybe ? - #9
Merged
Merged
Conversation
Signed-off-by: Sami Jaghouar <sami.jaghouar@gmail.com>
Signed-off-by: Sami Jaghouar <sami.jaghouar@gmail.com>
Signed-off-by: Sami Jaghouar <sami.jaghouar@gmail.com>
Closed
seanbell
added a commit
to clouddatalabs/scalerl-prime-rl
that referenced
this pull request
Apr 26, 2026
User's explicit stance from the snowflake_poc_critique PrimeIntellect-ai#9: "We should always train on full batches. If some groups are dropped from low variance, we should sample more groups until we have signal." Prior behavior: scheduler.generate_batch retried up to 3 times only if ALL rollouts were filtered (n_trainable == 0). A partial filter that left, say, 32 of 256 rollouts surviving proceeded as-is; n_trainable > 0 was the success threshold. New behavior (default mode): keep sampling and accumulating ADDITIONAL groups until the surviving cohort reaches batch_size, up to 20 attempts. compute_advantages is idempotent (it recomputes from `reward`), apply_filters resets is_filtered/filters at function entry, and apply_batch_advantage_normalization's scaling is keyed off compute_advantages's fresh output — re-running them on the cumulative rollout list each iteration is therefore correct. Soft-fail on max_attempts exhausted: train on whatever survived if n_trainable > 0 (don't crash). Only the truly-empty case (zero surviving) is fatal — that signals env-level corruption rather than just slow-start dynamics. Paper-faithful mode (paper_faithful_empty_batch=True) preserved: one attempt, no refill, all-filtered → raise (matches ScaleRL §3.4 "drop, don't refill"). Bumped MAX_EMPTY_BATCH_ATTEMPTS 3 -> 20 to give the dynamic-sampling path enough budget to converge on full-trainable batches during slow-start (TB cold-start: docker pulls + first-rollout images) and late-training "everything is easy" plateaus. Tests pass: 295/295 orchestrator + configs + environments.
snimu
added a commit
that referenced
this pull request
Jun 4, 2026
- enabled_losses=None now validated as the full term list, so >1 echo term per env is caught at config time instead of at rollout time. [review #6] - loss_overrides keys validated against `losses`; non-echo overrides rejected. [#7] - warn (don't fail) when prompt-role echo is configured with renderer=None (MITO), where prompt_attribution is unavailable so it would silently no-op. [#8] - token_export: add echo_mask/echo_weight columns + export sequences trained only via echo (gate on loss_mask OR echo_mask). [#9] - doc notes: echo CE uses the rollout temperature (scale alpha to compensate, kept as-is); negative alpha is intentional (suppresses tokens). [#1, #10] - tests for the new config validators. Deferred to a follow-up pass (per the review): full per-sample primary routing / rl-disable [#2b] + the <=1-primary validation it enables [#5], and the multi-run losses fingerprint [#3]. Not run locally; ruff + py_compile clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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.
No description provided.