v0.6.0

Overview

⭐️ Highlights

Trinity-RFT v0.6.0 adds SGLang support, optimizes fully asynchronous parameter synchronization and scheduling to reduce bubbles, improves MoE training stability, and upgrades the veRL backend to v0.8.0.

Explorer

1. Add SGLang inference engine support, including SGLang OpenAI API responses with prompt_token_ids, token_ids, and routed_experts.
2. Upgrade vLLM to v0.23.0 and SGLang to v0.5.13.
3. Refactor rollout model creation and management to support cross-node deployment of rollout models.
4. Optimize fully asynchronous parameter synchronization with partial rollouts.
5. Refactor weight synchronization with vLLM's WeightTransferEngine.
6. Improve MoE rollout support by returning routed experts from vLLM/SGLang OpenAI API paths.
7. Add data, pipeline, and expert parallel settings for rollout models to improve rollout performance.
8. Enhance multi-turn and multi-modal rollout support.
9. Add rollout performance benchmarking tools.

Trainer

1. Upgrade veRL to v0.8.0 while keeping v0.7.1 supported.
2. Improve Qwen3.5 compatibility with FSDP and Megatron backends.
3. Improve MoE model training stability and reduce OOM risk in FSDP/Megatron training paths.

Buffer

1. Refactor Buffer with an asynchronous implementation.
2. Update the SQL schema and add filtering support to the Experience Viewer.
3. Optimize the Experience Viewer.

Others

1. Add Connect the Dots (CoD) example.
2. Add InterCode-SQL example/benchmark.
3. Add the GiGPO and DAPO algorithms.
4. Improve Docker and remote GPU unit test workflows.

🚨 Breaking Changes

Due to database schema changes, the database files in the existing checkpoint folder cannot be recognized by the new version. To continue running previous experiments, please remove the corresponding .db files (typically located in the checkpoint_root_dir/project/name/buffer directory).

What's Changed

• Fix opd key error typo by @garyzhang99 in #527
• Update vllm to v0.19.1 by @pan-x-c in #529
• Refactor WorkflowRuner and Scheduler to handling partially completed tasks by @pan-x-c in #530
• Optimize Explorer by @pan-x-c in #531
• Add Performance tools by @pan-x-c in #532
• Fix: skip trainer config validation in bench mode by @shiweijiezero in #535
• Support SGLang Inference Engine by @pan-x-c in #533
• Fix(trainer,checkpoint,explorer): resolve Qwen3.5/vLLM compatibility bugs and harden runtime paths by @chenyushuo in #534
• SGLang OpenAI API Server returns prompt_token_ids and token_ids by @pan-x-c in #536
• Refactor Inference Model Creation and Management by @pan-x-c in #537
• Update vLLM to 0.21.0, SGLang to 0.5.12 by @pan-x-c in #538
• Train Qwen3.5 with megatron by @pan-x-c in #540
• fix(trainer): add torch.cuda.empty_cache() after FSDP update_actor by @Leon-Algo in #541
• Bug fix on update_weight for MoE model. by @chenyushuo in #543
• Support return routed experts for MoE Models by @pan-x-c in #542
• Refactor Monitor Metric Aggregation by @pan-x-c in #539
• Enhance multi-turn multi-modal sft by @chenyushuo in #545
• Update Megatron by @pan-x-c in #546
• fix(trainer): add empty_cache() after compute_ref_log_prob to prevent OOM by @Leon-Algo in #548
• Fix Megatron OOM when calculating entropy by @pan-x-c in #550
• Refactor Docker environment and unit test workflow by @pan-x-c in #554
• Fix Tinker API by @pan-x-c in #553
• Enhance multi-modal for vllm model by @chenyushuo in #547
• Update SQL Schema and Add Filter to Experience Viewer by @pan-x-c in #555
• DAPO Integration in Trinity: Algorithm Registration, Dynamic Sampling Pipeline, Overlong Handling, and Reviewer-Focused Docs by @artaasd95 in #552
• fix mm_utils by @chenyushuo in #558
• fix(sglang): fix shutdown Gloo connection error by @MengsD in #559
• docs: use TINKER_API_KEY in tinker docs by @pan-x-c with @Copilot in #564
• fix: propagate the optional 'group' config field across paths, namespaces & docs by @shiweijiezero in #562
• fix(sglang): fix sglang oom (Qwen3.5-35B-A3B) by @MengsD in #561
• Return routed experts for vLLM OpenAI API by @pan-x-c in #565
• Refactor Buffer into async implementation by @pan-x-c in #557
• explorer: fix glibc getenv SIGSEGV when creating runners at startup by @shiweijiezero in #567
• fix(example): fix dapo-math reward issue by @MengsD in #566
• Add the GiGPO algorithm [FEATURE] by @artaasd95 in #560
• fix(model): add routed_experts to dummy experiences by @MengsD in #568
• Optimize Docker unittest tool by @pan-x-c in #569
• Run unittest on remote GPU server by @pan-x-c in #571
• Update VERL to 0.8.0 by @pan-x-c in #570
• Update news by @yanxi-chen in #574
• Add InterCode-SQL example and benchmark by @PatrickG1014 in #572
• Fix verl FSDP can't save checkpoints when offloading is enable by @pan-x-c in #577
• Refactor weight synchronization with vLLM's WeightTransferEngine by @pan-x-c in #578
• Add data/pipeline/expert parallel settings for rollout model by @chenyushuo in #576
• Add fully async mode without any bubble by @pan-x-c in #579
• Optimize viewer and add tests by @pan-x-c in #581
• Add Connect the Dots (CoD) overview under examples/research_cod by @shiweijiezero in #583
• Update Connect the Dots (CoD) overview README and figures by @shiweijiezero in #585
• Update vLLM to v0.23.0, SGLang to v0.5.13 by @pan-x-c in #586
• Fix qwen3.5 for verl==0.8.0 by @chenyushuo in #589
• trinity convert supports converting multiple steps by @pan-x-c in #587
• Add CPT Algorithm && Formatter by @chenyushuo in #592
• fix(buffer): reshuffle with correct epoch in ShuffleSelector by @MengsD in #593
• patch _build_module by @chenyushuo in #595
• Release v0.6.0 by @pan-x-c in #594

New Contributors

• @Leon-Algo made their first contribution in #541
• @artaasd95 made their first contribution in #552
• @MengsD made their first contribution in #559
• @PatrickG1014 made their first contribution in #572

Full Changelog: v0.5.2...v0.6.0