Repository navigation

Issues

New issue

SkyRL Q3 2026 Roadmap
#1846 · erictang000 opened on Jun 29, 2026
8
SkyRL H1 2026 Roadmap
#1391 · SumanthRH opened on Mar 25, 2026
2

[model] GLM-5.3-Flash Tracker
NovaSky-AI/SkyRL#2487 · opened on Oct 14, 2026
[tinker] SkyRL-Train create_model leaks the placement group when the first model fails to build
NovaSky-AI/SkyRL#2482 · opened on Oct 14, 2026
[tinker] Per-architecture definition of LoRA groups for train_attn / train_mlp on FSDP
NovaSky-AI/SkyRL#2481 · opened on Oct 14, 2026
[skyrl-gym][text2sql] 44% of SkyRL-SQL-653 tasks give reward 1.0 to any query that returns no rows
NovaSky-AI/SkyRL#2451 · opened on Oct 9, 2026
[tracker] SkyRL VLM training
NovaSky-AI/SkyRL#2393 · opened on Oct 2, 2026
[skycap] Follow-ups after the Harbor integration
NovaSky-AI/SkyRL#2318 · opened on Sep 28, 2026
[skycap] standalone tito/trace capture proxy for RL and post-training
NovaSky-AI/SkyRL#2306 · opened on Sep 25, 2026
3
Support For Parallel Environment Steps
NovaSky-AI/SkyRL#2301 · opened on Sep 24, 2026
2
Add a dedicated documentation page for all SkyRL metrics
NovaSky-AI/SkyRL#2300 · opened on Sep 24, 2026
naive_dapo.compute_score: tiemout=True typo makes the \pi branch always raise (swallowed by bare except:), so correct numeric answers to \pi questions score 0.0
NovaSky-AI/SkyRL#2289 · opened on Sep 24, 2026
[tinker] Support newer Tinker SDKs: sampling fails with SDK 0.30.1
NovaSky-AI/SkyRL#2287 · opened on Sep 24, 2026
pad_training_input_batch pads TensorList fields with tensor[0].clone() — loss-masked pad rows carry a real row's payload (ragged-consumer crash; wrong pixel_values/image_grid_thw on pad rows)
NovaSky-AI/SkyRL#2263 · opened on Sep 22, 2026
1
fully_async (non-colocated): training actor emits all-NaN logits on every microbatch of the first step after the first NCCL weight sync; weights/optimizer/exports all finite
NovaSky-AI/SkyRL#2247 · opened on Sep 19, 2026
3
fully_async: non-colocated weight sync ignores clear_kv_cache_on_weight_sync — stale prefix-cache KV trained on after each sync
NovaSky-AI/SkyRL#2246 · opened on Sep 19, 2026
[bug][fsdp] FSDP2 manual offload/load corrupts DTensor unshard state under colocate_all=True, causing silent non-deterministic logit errors
NovaSky-AI/SkyRL#2215 · opened on Sep 16, 2026
1
[feature] Honor max_tokens_per_microbatch in the loss-path forward (forward with loss_fn)
NovaSky-AI/SkyRL#2211 · opened on Sep 15, 2026
[weight sync] Improve offload/backload semantics while training with LoRA with colocate_all=True
NovaSky-AI/SkyRL#2206 · opened on Sep 14, 2026
Support full-distribution comparison to validate bitwise consistency between rollout and training
NovaSky-AI/SkyRL#2201 · opened on Sep 12, 2026
1
[tinker] Retain only the K most recent checkpoints
NovaSky-AI/SkyRL#2116 · opened on Sep 1, 2026
[train] FSDP2 cannot train glm4v VLMs: per-nn.Embedding fully_shard leaves vision position_embedding sharded at raw .weight read (grid_sample DTensor crash)
NovaSky-AI/SkyRL#2087 · opened on Aug 22, 2026
Multi-turn VLM generator corrupts trajectories for all thinking models: history-editing chat templates break the deferred obs-token offsets
NovaSky-AI/SkyRL#2075 · opened on Aug 21, 2026
[deps] Migrate custom vLLM-router fork on top of 0.1.15 release of vllm-router
NovaSky-AI/SkyRL#2070 · opened on Aug 19, 2026
1
[vllm] Remove old inference worker wrap code after 0.26.0 update
NovaSky-AI/SkyRL#2020 · opened on Aug 12, 2026
[examples] Add example startup scripts for tinker engine
NovaSky-AI/SkyRL#1983 · opened on Aug 4, 2026
1
Qwen3.6-27B MTP is unsafe with prefix caching enabled
NovaSky-AI/SkyRL#1981 · opened on Aug 4, 2026

Navigation Menu

Repository navigation

Issues

Issues

[model] GLM-5.3-Flash Tracker

[tinker] SkyRL-Train create_model leaks the placement group when the first model fails to build

[tinker] Per-architecture definition of LoRA groups for train_attn / train_mlp on FSDP

[skyrl-gym][text2sql] 44% of SkyRL-SQL-653 tasks give reward 1.0 to any query that returns no rows

[tracker] SkyRL VLM training

[skycap] Follow-ups after the Harbor integration

[skycap] standalone tito/trace capture proxy for RL and post-training

Support For Parallel Environment Steps

Add a dedicated documentation page for all SkyRL metrics

naive_dapo.compute_score: `tiemout=True` typo makes the `\pi` branch always raise (swallowed by bare `except:`), so correct numeric answers to `\pi` questions score 0.0

[tinker] Support newer Tinker SDKs: sampling fails with SDK 0.30.1

pad_training_input_batch pads TensorList fields with tensor[0].clone() — loss-masked pad rows carry a real row's payload (ragged-consumer crash; wrong pixel_values/image_grid_thw on pad rows)

fully_async (non-colocated): training actor emits all-NaN logits on every microbatch of the first step after the first NCCL weight sync; weights/optimizer/exports all finite

fully_async: non-colocated weight sync ignores clear_kv_cache_on_weight_sync — stale prefix-cache KV trained on after each sync

[bug][fsdp] FSDP2 manual offload/load corrupts DTensor unshard state under `colocate_all=True`, causing silent non-deterministic logit errors

[feature] Honor max_tokens_per_microbatch in the loss-path forward (forward with loss_fn)

[weight sync] Improve offload/backload semantics while training with LoRA with `colocate_all=True`

Support full-distribution comparison to validate bitwise consistency between rollout and training

[tinker] Retain only the K most recent checkpoints

[train] FSDP2 cannot train glm4v VLMs: per-nn.Embedding fully_shard leaves vision position_embedding sharded at raw .weight read (grid_sample DTensor crash)

Multi-turn VLM generator corrupts trajectories for all thinking models: history-editing chat templates break the deferred obs-token offsets

[deps] Migrate custom vLLM-router fork on top of 0.1.15 release of vllm-router

[vllm] Remove old inference worker wrap code after 0.26.0 update

[examples] Add example startup scripts for tinker engine

Qwen3.6-27B MTP is unsafe with prefix caching enabled