SkyRL Q3 2026 Roadmap

[erictang000]

Note to Community

This roadmap is a living document, and we welcome community input and feedback. We will update this roadmap to link to specific Issues and PRs for each sub-task.

Overview

SkyRL's primary focuses in Q3 2026 will continue to be pushing performance for async RL on large scale MoE models, and improving support for SkyRL's native Tinker API compatible training engine.

Tinker Engine

• [P0] Improve sqlite DB path support for higher concurrency
• [P0] Avoid inference engine initialization for the SFT code path
• [P1] Expose metrics for observability on the Tinker Engine (i.e. concurrent requests, number of tenants, etc...)
• [P1] Provide support for external postgres deployments as an alternative for sqlite for high concurrency/prod deployments
• [P1] Router replay (R3) support for the tinker engine
• [P2] Support additional Tinker LoraConfig params: [tinker] Support additional Tinker LoraConfig parameters #1632
• [P2] Support ref model in SkyRLTrainBackend

old tracker: #1380

Large Scale MoE Training (Megatron)

Megatron

• [P0] Add a DAPO recipe with MoE + megatron to CI: [CI] Add a DAPO recipe with the Megatron backend to CI #1322 @erictang000
• [P0] Enable Full FP8 Training @jinghanyao1-hub
• [P1] Low-precision expert quantization (like in link)
• [P2] Megatron dynamic context parallel support ([megatron][perf] Integrate megatron dynamic context parallelism #1019)
• [P2] Support megatron native FSDP

old tracker: #1392

Router Replay

• [P0] Add router replay test to CI: [wip] add router replay tests to CI #1770
• [P0] Validate multi-node inference for R3 with new vLLM implementation + evaluate data transfer with HTTP
• [P1] Optimize data transfer between training and inference for R3

Large Scale Inference (vLLM)

• [P0] Remove old inference: [refactor] Remove legacy inference codepath #1835
• [P0] MTP Support: [feat] Add Spec Decoding and MTP training support #1832 @zanderjiang

Weight Syncing

• [P0] Support Ray RDT as a weight syncing backend with sharded weight transfer: [WIP] RDT skyrl integration #1753 @hao-aaron
• [P0] Delta Weight updates @SumanthRH
• [P2] Inference pool fault tolerance

Model Support

• GLM 5.2 @j316chuck
• Nemotron Ultra @dyurk-lila @erictang000
• Kimi K2.6: [tx] Kimi K2.5 and Kimi K2.6 support #1566

LoRA

• [P0] Debug logprob diff mismatch growing for LoRA on MoE models (Qwen3.6, GLM-4.7-Flash): [LoRA] Investigate LoRA training inference mismatch for Qwen3.6 MoE and GLM-4.7-Flash #1776
• [P1] PiSSA Support @j316chuck
• [P1] Support in memory LoRA only weight sync for megatron + lora ([train] Support LoRA-only weight syncing for Megatron backend #1336)

Agent Integration

• [P0] Add TITO proxy (handling rollout logprobs, router replay expert indices) for easier multi-turn agent harness integrations @CharlieFRuan

SFT

• [P0] Curriculum learning sampler for SFT [RFC] Stateful dataloader and custom samplers for SFTTrainer #1839 @avigyabb

Algorithm

• Merge ECHO into SkyRL examples: [feature] Add ECHO patch to SkyRL #1715
• Support adaptive entropy training methods: [algorithm] Add adaptive entropy methods (ADAPO, REPO-R, MAI-Thinking) #1845

Miscellaneous

• AMD Support
• Add cuda 13 support
• Bump ray version to latest (2.56): [dependencies] upgrade ray to 2.55 #1545

[casper-hansen]

For weight syncing LoRA adapters, it seems a good goal would be to ensure we run merge_lora=False so we only sync the adapter and not the full model.
#1852

[benrhodes26]

In this blog post, your roadmap section says "Multimodal support is still early and we have several major features planned, including sequence packing, Megatron backend support, long context training with context parallelism, and step-wise training"

Just wondering if any of those sit on the Q3 roadmap, or are more like Q4?

This is a great project btw, and very excited for the next quarter!