google/tunix

View on GitHub

A Lightweight LLM Post-Training Library

Stars: 2.5k
Forks: 345
Open beginner issues: 0
Indexed issues: 43
Avg merge: 1d 7h
Merged PRs (30d): 240

Dominant language: Python
License: Apache-2.0
Last GitHub push: Sep 15, 2026
Latest indexed: Sep 16, 2026
Contributing guide: Contributing guide
Code of conduct: Code of conduct
Beginner labels: No beginner labels indexed

Open All

43 open issues indexed

Qwen example notebook Open

type:support

google/tunix#45 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Adding linting configuration file Open

type:feature/enhancement

google/tunix#93 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Qwen Example relies on HF Transformer and Accelerate, which is removed Open

type:build/install

google/tunix#94 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Wandb Open

type:feature/enhancement

google/tunix#137 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Question: how to run tunix with vLLM? Open

type:bug

google/tunix#256 · 9 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
RL Bug/Issue Tracker Open

type:bug

google/tunix#390 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Failed build: Tunix Package Tests Open

type:build/install

google/tunix#489 · 14 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Add larger Qwen 2.5 models Open

type:feature/enhancement

google/tunix#509 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
How to train on multi host tpu pod Open

type:support

google/tunix#560 · 25 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Gemma3 12B loading error Open

type:bug

google/tunix#611 · 8 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
A unified API for model loading Open

type:feature/enhancement

google/tunix#612 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Failed build: Tunix Package Tests Open

build failed

google/tunix#667 · 290 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
"RuntimeError: cannot enter context" in grpo_demo.ipynb Open

type:bug

google/tunix#686 · 6 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
grpo_demo.ipynb: Excessive output Open

type:feature/enhancement

google/tunix#687 · 4 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
GRPO training fails in multi-process mode Open

type:bug

google/tunix#698 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
vLLM sampler accesses attribute that does not exist Open

type:bug

google/tunix#780 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 16, 2026
[Bug] Fail to run train_deepscaler_nb.py Open

type:bug

google/tunix#802 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Single-item batches truncate sampler token_buffer instead of populating it Open

type:bug

google/tunix#809 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
GRPO Evaluation Fails When TRAIN_MICRO_BATCH_SIZE < num_generations Open

type:bug

google/tunix#816 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
This repo is trying to do too much Open

type:others

google/tunix#832 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Train GRPO through SGLangJax in Single or Multi Hosts with Pathways Open

type:feature/enhancement

google/tunix#874 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Gemma 3 1b model loading asyncio error on kaggle notebook Open

type:bug

google/tunix#884 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Feature request: Add Contrastive Representation Distillation (CRD) Open

type:feature/enhancement

google/tunix#887 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Clarify differences between examples/rl/grpo GSM8K recipes and scripts/grpo_demo_llama3_qwen2.py Open

type:support

google/tunix#939 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Qwen2.5-7B gets ~50% on GSM8K eval while Qwen2.5-3B gets ~80% using scripts/grpo_demo_llama3_qwen2.py Open

type:performance

google/tunix#943 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Wrong Python version causing Attribute Error is_integer() in utils.py Open

type:bug

google/tunix#944 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Add GDPO Support Open

type:feature/enhancement

google/tunix#977 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Natural Language to SQL GRPO Example Open

type:feature/enhancement

google/tunix#979 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 16, 2026
[Bug] Fix run_prod for scripts/grpo_demo_llama3_qwen2.py with sglang_jax with LoRA Open

type:bug

google/tunix#995 · 1 comment · 0 reactions · 1 assignee · Indexed Sep 16, 2026
[Tiny Feat] make log_level be configurable Open

type:feature/enhancement

google/tunix#1040 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
`.sow()` intermediate values don't enable Dead Code Elimination of unused outputs Open

type:bug type:performance

google/tunix#1049 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Qwen3-VL Open

type:feature/enhancement

google/tunix#1063 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Unsafe use of eval() in calculate_reward enables code execution if tasks are untrusted Open

type:feature/enhancement

google/tunix#1196 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Support configurable advantage estimation in AgenticRLLearner (RLOO, DrGRPO) Open

type:feature/enhancement

google/tunix#1378 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
How to enable audio and vision modalities in tunix's gemma4? Open

type:feature/enhancement

google/tunix#1380 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Clarification on (external) contribution policy Open

type:support

google/tunix#1384 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
[bug] batch iterator exception when running Qwen example Open

type:bug

google/tunix#1474 · 3 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
add BEFT as a new SFT/PEFT method Open

type:feature/enhancement

google/tunix#1490 · 4 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Failed build: Tunix Nightly Regression Tests Open

build failed

google/tunix#1585 · 116 comments · 0 reactions · 0 assignees · Indexed Sep 16, 2026
Adopt OpenTelemetry behind the MetricsLogger interface Open

type:feature/enhancement

google/tunix#1685 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
Hard dependency on `jax[tpu]` blocks or degrades installation on all non-TPU platforms Open

dependencies type:feature/enhancement

google/tunix#2047 · 2 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
ToolEnvironment crashes on explicit `finish` tool calls: arguments is a JSON string, not a dict Open

type:bug

google/tunix#2050 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026
GRPO hangs indefinitely on 2 RTX PRO 6000 Blackwell GPUs Open

type:bug

google/tunix#2196 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 16, 2026

google/tunix

Get new issues in your inbox