Repository Issue Activity (beta)

#2804 Agentic rollout: shared miles-side layer and Harbor in-process follow-ups

#3133 [Bug] SGLang router fails to start when Ray's node address is a hostname: launch_router --host <hostname> raises "Invalid address: invalid socket address syntax"

#3123 master

#1837 Tracking issue for refactoring and enhancements

#3113 TITO strict gate false-positives on qwen3 multi-turn history (empty think-skeleton re-render)

#3111 sandbox credentials: make the key file the only supply path (env supply masks broken file relays)

#3110 examples: launcher recipe boilerplate is copied across every agentic leg

#2705 [AMD][LoRA] AMD LoRA RL support tracker

#329 [Feature Request] MLFlow

#2856 p2p weight-sync crashes on Qwen3-235B-A22B (MoE) with MagicMock process-group error

#3088 verifiers: lift the renderers<0.1.11 cap once sglang relaxes its openai pin

#2853 [Roadmap] 2026 H2 Roadmap

#615 [Roadmap] Blackwell MXFP8 and NVFP4 RL training

#2362 [BUG] Current OPD TopK reward cannot distill teacher but only maximize student's entropy

#2800 Agentic rollout: drop infra-aborted samples per sample (refill) instead of per group

#2207 [RFC] Load compute kernels from the Hugging Face Hub

#2254 [rollout] Define lifecycle ownership and recovery for fully async rollout

#1360 [refactor] weight update refactor

#2722 [DeepSeek-V4-Flash-0731] Define and validate the SGLang/Megatron train–rollout numerical parity

#2258 [RFC] Multi-LoRA backend compatible with the Tinker API

#1334 Mid-flight `/generate` retry can double-generate: `_post` retries any exception and there is no rid to abort the orphaned request

#2689 fix(rollout): handle malformed and non-object tool arguments before execution

#2679 Weight sync fails for bridge-built models (Qwen3.5/3.6/3.8 GDN, Qwen3-VL): --spec is ignored in bridge mode but megatron_to_hf expects the spec's names

#2592 openenv agent metrics do not report why an episode ended (no end_reason / truncated flag)

#2668 [Bug] LoRA native checkpoints duplicate adapter shards across DP ranks