OpenPipe/ART

View on GitHub

Agent Reinforcement Trainer: train multi-step agents for real-world tasks using GRPO. Give your agents on-the-job training. Reinforcement learning for Qwen3.6, GPT-OSS, Llama, and more!

Stars: 10.8k
Forks: 989
Open beginner issues: 0
Indexed issues: 72
Avg merge: 6h 29m
Merged PRs (30d): 85

Dominant language: Python
License: Apache-2.0
Last GitHub push: Sep 19, 2026
Latest indexed: Sep 20, 2026
Contributing guide: Contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Beginner friendly Everything

Open All

72 issues indexed so far

Tokenization throughput: process pool capped at 4 workers, GIL-bound thread path (benchmark: caladan 059 batches) Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 48/100

OpenPipe/ART#926 · Opened Sep 19, 2026
Support combined expert and data parallelism in TrainerRank Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

OpenPipe/ART#917 · Opened Sep 16, 2026
Support pipeline parallelism in TrainerRank without exposing stage scheduling to callers Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

OpenPipe/ART#916 · Opened Sep 16, 2026
Investigate output and gradient variability between matched same-source Qwen TrainerRank actors Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#902 · Opened Sep 15, 2026
Investigate large hidden-state differences when independent requests are co-packed (Qwen3.6, no-grad) Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

OpenPipe/ART#901 · Opened Sep 15, 2026
Retained sampled suffix can carry logprobs under a different causal prefix Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#883 · 2 comments · Opened Sep 10, 2026
TrainerRank reusable-cache admission can leave insufficient physical CUDA memory for backward library allocations Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

OpenPipe/ART#870 · 8 comments · Opened Sep 9, 2026
Determine whether a large dynamics RL batch refusal is necessary or avoidable by execution layout Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#869 · 3 comments · Opened Sep 9, 2026
Qwen3.5-35B-A3B at EP1 with CP>1 segfaults in Transformer Engine's grouped GEMM on real-data routing (likely empty local experts) Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

OpenPipe/ART#851 · Opened Sep 4, 2026
TrainerRank MoE memory admission: backward-recompute OOM and profile calibration failures Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

OpenPipe/ART#848 · 28 comments · Opened Sep 4, 2026
Dedicated LocalBackend vLLM server can become unreachable during LoRA adapter reload Open

Difficulty 4/5 3-5 days Newbie friendliness 50/100

OpenPipe/ART#678 · 1 comment · Opened May 12, 2026
Regression: forked GRPO run can produce malformed checkpoint outputs on newer ART ref Open

bug

Difficulty 5/5 Over a week Newbie friendliness 35/100

OpenPipe/ART#672 · Opened May 8, 2026
LoRA adapters are not properly syncing when using LocalBackend - stale adapters are being used in rollouts Open

Difficulty 3/5 1-2 days Newbie friendliness 62/100

OpenPipe/ART#661 · Opened Apr 24, 2026
LocalBackend fork_checkpoint doesn't update vLLM's initial LoRA Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

OpenPipe/ART#651 · Opened Apr 13, 2026
Add from_entity parameter to _experimental_fork_checkpoint Open

Difficulty 2/5 1-3 hours Newbie friendliness 72/100

OpenPipe/ART#649 · Opened Apr 10, 2026
auto_rl example uses Open

Difficulty 3/5 1-2 days Newbie friendliness 56/100

OpenPipe/ART#641 · Opened Apr 3, 2026
ServerlessBackend creates a new W&B run per training step, causing "Run finished" alerts to fire on every step Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

OpenPipe/ART#638 · Opened Apr 2, 2026
[Security] Hardcoded API Key in vLLM Server Configuration Allows Authentication Bypass Open

Difficulty 3/5 1-2 days Newbie friendliness 52/100

OpenPipe/ART#628 · Opened Mar 23, 2026
Running out of memory with version v0.5.17 Open

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#617 · Opened Mar 16, 2026
Training only processes one trajectory at a time? Open

question

Difficulty 3/5 1-2 days Newbie friendliness 42/100

OpenPipe/ART#533 · 1 reaction · Opened Jan 25, 2026
RFC: Deprecate allow_training_without_logprobs option Open

discussion

Difficulty 5/5 Over a week Newbie friendliness 30/100

OpenPipe/ART#528 · Opened Jan 21, 2026
W&B offline Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#493 · 1 comment · Opened Dec 19, 2025
Unsloth + vLLM multi-GPU setup in ART crashes when restricting CUDA_VISIBLE_DEVICES or using tensor parallelism >1 on H100 GPUs Open

bug

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#478 · Opened Dec 11, 2025
RULER produces scores >1, arbitrary outputs (7, 10, 15…), nondeterministic behavior, and malformed JSON when judge model is Qwen3-32B running on Ollama Open

bug

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#477 · Opened Dec 11, 2025
RULER + Litellm with Ollama: `ollama_chat` recommended, but `/api/chat` cannot generate required JSON schema Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 35/100

OpenPipe/ART#476 · Opened Dec 11, 2025
RULER scoring + training tightly coupled to Litellm/OpenAI, cannot cleanly use ChatOllama/ChatNVIDIA as judge/inference models Open

bug

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#475 · Opened Dec 11, 2025
Context window incorrectly capped at 8192 tokens when using Ollama (token count exceeded 8192) Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 35/100

OpenPipe/ART#473 · Opened Dec 11, 2025
Training hangs at 0/12 with Qwen2.5-14B + Unsloth/vLLM (KV cache reported as 0.0 GB despite free VRAM) Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 35/100

OpenPipe/ART#472 · Opened Dec 11, 2025
Is it not possible to set the rollout engine to multiple gpus? Open

question

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#470 · Opened Dec 10, 2025
Training hangs after a few steps Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#469 · 3 comments · 3 reactions · Opened Dec 8, 2025
Training gets stuck after completing just one trajectory Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#459 · 1 comment · 1 reaction · Opened Nov 18, 2025
Cuda out of Memory Open

question

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#452 · 1 comment · Opened Nov 9, 2025
Torchtune development has been discontinued. Have you considered switching to a different multi-GPU train backend? Open

question

Difficulty 5/5 Over a week Newbie friendliness 20/100

OpenPipe/ART#450 · 2 comments · Opened Nov 5, 2025
[Bug] Unable to create Qwen3 MoE model Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 35/100

OpenPipe/ART#444 · 4 comments · 1 reaction · Opened Oct 23, 2025
Cuda error in RULER notebook Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#436 · 1 comment · 1 reaction · Opened Oct 14, 2025
Minio Support Open

enhancement

Difficulty 3/5 1-2 days Newbie friendliness 45/100

OpenPipe/ART#435 · Opened Oct 9, 2025
Why Do Results from MCP-Trained Models Differ Greatly Between generate_benchmarks.py and train.py Open

question

Difficulty 4/5 3-5 days Newbie friendliness 32/100

OpenPipe/ART#417 · Opened Sep 22, 2025
uv venv error initializing art backend on skypilot/aws Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#416 · 4 comments · Opened Sep 18, 2025
AttributeError: 'LLM' object has no attribute 'engine_step' with Unsloth and vLLM Open

bug

Difficulty 3/5 1-2 days Newbie friendliness 35/100

OpenPipe/ART#411 · 1 comment · Opened Sep 17, 2025
[Feature Request] Azure OpenAI Support Open

enhancement

Difficulty 3/5 1-2 days Newbie friendliness 45/100

OpenPipe/ART#409 · 1 comment · Opened Sep 17, 2025
RuntimeError: torch.cuda.MemPool doesn't currently support expandable_segments during vLLM model initialization Open

bug

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#405 · 3 comments · 1 reaction · Opened Sep 16, 2025
Always Block when train Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#401 · 3 comments · Opened Sep 12, 2025
[Feature Request] Add integration to smolagents Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 35/100

OpenPipe/ART#398 · 2 comments · Opened Sep 9, 2025
About appropriate groups_per_step parameter setting given training and validation dataset volume Open

question

Difficulty 5/5 Over a week Newbie friendliness 20/100

OpenPipe/ART#397 · Opened Sep 9, 2025
Allow running multiple `art.LocalBackend` sessions on different processes Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 42/100

OpenPipe/ART#396 · Opened Sep 8, 2025
[Question] About 2048 scoring Open

question

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#394 · 3 comments · Opened Sep 7, 2025
Training crashes after sometime Open

bug

Difficulty 4/5 3-5 days Newbie friendliness 25/100

OpenPipe/ART#386 · 1 comment · Opened Aug 30, 2025
Feature: Dr. GRPO Support Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 35/100

OpenPipe/ART#364 · 1 comment · 3 reactions · Opened Aug 25, 2025
Support `exclude` parameter for SkyPilotBackend `_experimental_pull_from_s3` Open

enhancement

Difficulty 2/5 1-3 hours Newbie friendliness 52/100

OpenPipe/ART#350 · Opened Aug 20, 2025
Add support for SFT through ART directly Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 25/100

OpenPipe/ART#348 · 1 comment · 3 reactions · Opened Aug 19, 2025

OpenPipe/ART

Get new issues in your inbox