add Laguna XS-2.1 SWE example - #3186
Open
rasdani wants to merge 4 commits into
Open
Conversation
rasdani
changed the title
rasdani
marked this pull request as ready for review
August 4, 2026 01:40
rasdani
force-pushed
the
exp/laguna-xs21-swe-example
branch
from
August 4, 2026 01:56
fba1792 to
8d0f43b
Compare
mikasenghaas
reviewed
Aug 4, 2026
|
|
||
| [inference.model] | ||
| max_model_len = 131072 | ||
| tool_call_parser = "poolside_v1" |
Member
There was a problem hiding this comment.
should also add this into our auto-map if not there yet
|
|
||
| [orchestrator.train.sampling] | ||
| temperature = 1.0 | ||
| max_completion_tokens = 32768 |
| [orchestrator.train.sampling] | ||
| temperature = 1.0 | ||
| max_completion_tokens = 32768 | ||
| extra_body = { top_k = 20, min_p = 0.0 } |
Member
There was a problem hiding this comment.
top k is not supported yet, see #2979 for pr to enable replay on trainer. should acc prob block if this is extra body to prevent corrupt training
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 2 potential issues.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 1886ecb. Configure here.
| fused_lm_head_token_chunk_size = 1024 | ||
| optim_cpu_offload = true | ||
|
|
||
| [trainer.model.compile] |
There was a problem hiding this comment.
Trainer compile left enabled
High Severity
The empty [trainer.model.compile] table keeps torch.compile on (CompileConfig is the default). Disabling it requires compile = "None". This run already hit a compiled-trainer MoE collective mismatch and NCCL timeout, so leaving compile enabled is likely to reproduce that failure.
Reviewed by Cursor Bugbot for commit 1886ecb. Configure here.
| [orchestrator.train.sampling] | ||
| temperature = 1.0 | ||
| max_completion_tokens = 32768 | ||
| extra_body = { top_k = 20, min_p = 0.0 } |
There was a problem hiding this comment.
Train top_k can corrupt ratios
High Severity
orchestrator.train.sampling.extra_body sets top_k = 20, which overrides the policy-rollout default of top_k = -1. Truncated train sampling without trainer kept-set / acceptance-prob replay support makes inference and trainer logprobs disagree and can silently corrupt GRPO importance ratios.
Reviewed by Cursor Bugbot for commit 1886ecb. Configure here.
Contributor
Author
|
The full-scale compiled run reproduced the NCCL failure on Slurm job Full NCCL all-to-all timeout trace (job 1643)Rank 0 ( Rank 13 ( |
rasdani
pushed a commit
that referenced
this pull request
Aug 5, 2026
…L deadlock When torch.compile is used with Ulysses context parallelism and FSDP, the all-to-all collectives are traced into the compiled backward graph while FSDP's reduce-scatter hooks fire eagerly via the autograd engine. The compiler may reorder the all-to-all backward differently across ranks, causing rank 0 to wait on reduce_scatter while rank 1 waits on all_to_all — a classic collective ordering mismatch that deadlocks NCCL after the 1-hour watchdog timeout (PR #3186, job 1578). Wrapping _all_to_all_seq_to_head and _all_to_all_head_to_seq with torch._dynamo.disable forces a graph break around the collectives so they execute eagerly in the autograd engine, restoring a single consistent backward ordering between the two collective systems.
samsja
added a commit
that referenced
this pull request
Aug 6, 2026
…L deadlock (#3195) When torch.compile is used with Ulysses context parallelism and FSDP, the all-to-all collectives are traced into the compiled backward graph while FSDP's reduce-scatter hooks fire eagerly via the autograd engine. The compiler may reorder the all-to-all backward differently across ranks, causing rank 0 to wait on reduce_scatter while rank 1 waits on all_to_all — a classic collective ordering mismatch that deadlocks NCCL after the 1-hour watchdog timeout (PR #3186, job 1578). Wrapping _all_to_all_seq_to_head and _all_to_all_head_to_seq with torch._dynamo.disable forces a graph break around the collectives so they execute eagerly in the autograd engine, restoring a single consistent backward ordering between the two collective systems.
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.
Summary
Add a standalone
poolside/Laguna-XS-2.1RL example for ScaleSWE with online pass@1 evaluation on SWE-Bench Verified.The example captures the tested setup:
poolside_v1parsersThis is based directly on current
main, so it does not carry the earlier local parser workaround, old dependency pins, GLM configs, or cluster-specific node exclusions.Validation
uv run rl @ examples/advanced/laguna-xs-2.1/swe.toml --dry-run --output-dir <temp-dir>uv lock --checkprime-rl-configswheel outside the workspaceverifiers==0.2.2.dev76andvf.ServeConfiggit diff --checkThe dry run generated trainer, orchestrator, inference, and Slurm configs successfully with throughput-oriented DEP8 inference and the 10,000-step CLI override. The external slim-wheel import reproducing the previous CI failure also passes. A compiled two-node, 16-GPU trainer run completed five steps on current main; step 1 took 2m17s and subsequent steps took 19–22s without NCCL errors.
Historical trainer compilation failure
An earlier compiled trainer failed in Slurm job
1578on 2026-08-01. The current config restores compilation after the two-node five-step validation described above. At sequence 3750, rank 0 entered FSDP_REDUCE_SCATTER_BASEwhile rank 1 (and ranks 2–7) entered UlyssesALLTOALL_BASE. After the 3,600-second watchdog timeout, NCCL terminated the process group. The relevant rank-0 and rank-1 stderr is reproduced verbatim below.Full NCCL collective-mismatch trace (job 1578)
Note
Low Risk
Documentation-only addition of an example TOML; no runtime, library, or training code paths change.
Overview
Adds
examples/advanced/laguna-xs-2.1/swe.toml, a runnable recipe for GRPO on ScaleSWE withpoolside/Laguna-XS-2.1, using the bash harness and pass@1 eval on SWE-Bench Verified every 20 steps.The config wires a 2-node trainer (Ulysses CP=4, Muon, router replay, compile + activation checkpointing/offload) to four inference replicas at DP8 with expert parallel and routed-expert return for MoE training. Orchestration uses the laguna-xs-2.1 renderer with thinking enabled, Prime sandboxes labeled
laguna-xs21-swe(with Slurm pre-run cleanup), and vLLMpoolside_v1tool/reasoning parsers.Reviewed by Cursor Bugbot for commit a5e6abf. Bugbot is set up for automated code reviews on this repo. Configure here.