Repository navigation
Add Reflection Beam 256k Miles LoRA recipe - #762
Merged
Merged
Conversation
|
Documentation preview: https://modal-labs-training-gym--training-gym-previews-preview-r-78da35.modal.run/762/docs |
Andrew Hinh (andrewhinh)
approved these changes
Oct 27, 2026
Andrew Hinh (andrewhinh)
left a comment
Collaborator
There was a problem hiding this comment.
i left non-blocking comments re hygiene. and plz clean up the docstrings/comments.
Merged
This branch was successfully deployed
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Adds Reflection Beam 1.0 Miles LoRA recipe, following Miles #4149.
Uses 32 H200s across 4 nodes: 16 for training and 16 for rollout, TP8/PP1/CP2/EP16, expert TP1, rank 16 / alpha 32, and a 256k context length (224k prompt + 32k response). Training uses upstream's 32 prompts × 8 completions (global batch size 256), microbatch size 1, FlashAttention, full activation recomputation, and asynchronous trainer/rollout execution.
Upstream LoRA defaults to async and supports synchronous execution with
--no-async-training. This Dojo recipe likewise supportsasync_mode=Falsewith the same separate 16-trainer/16-rollout GPU allocation.Autoconfig: Beam 256k DAPO-math
Batch-256 validation generated all 256 responses from 32 unpadded DAPO-math prompts with normal EOS stopping (2,286,539 output tokens; 240 natural completions, 16 at the 32k cap). Training then exposed a CP/TP padding-alignment assertion. Follow-up fix #765 preserves batch 256 and validates training against that saved batch. This tests the 256k context cap without forcing every sample to fill the window; the earlier batch-eight runs below validate full-window capacity.
Validation runs:
Retained recipe-specific patches:
cp_comm_typethrough the Miles bridge and normalize singleton CLI lists for Beam's sliding-window attention.