v2.1.0

What's Changed

• chore(swe): remove unused SWE SFT dataset loader by @Le8r0nJames in #1467
• fix(docker): rename zeroclaw installer flag to --skip-quickstart by @sitabulaixizawaluduo in #1469
• fix(megatron): use dp_reshardable optimizer sharding and load optimizer state correctly by @Le8r0nJames in #1468
• fix(hermes): preserve singleton online reward signal by @jszzr in #1474
• fix(mcore): TP-shard GroupRMSNorm gate-norm weight for DCP checkpointing by @Le8r0nJames in #1470
• fix(rollout): stop controller-managed workers from dp-scaling staleness capacity by @Le8r0nJames in #1471
• fix(v2): export local workflow statistics by @jszzr in #1478
• docs: update NPU installation guide for ascend-v1.0.4 by @Adiactive in #1502
• Docs: Update README and cli docs by @sitabulaixizawaluduo in #1521
• fix(engine): route text-only VLM microbatches through the padded forward by @Adiactive in #1529
• fix: align CP metrics and TP grad norm metadata by @yulangz in #1497
• feat(infra): add HTTP-based Ray Scheduler by @HwVanICI in #1441
• fix(data_proxy): require admin key for /register_model to prevent SSRF by @sebastionoss in #1550
• test(inference-service): send admin key in data proxy register_model tests by @sitabulaixizawaluduo in #1556
• fix(inference): reject incomplete sampling evidence by @EazyReal in #1554
• fix(proxy): drop retry-orphan completions to prevent trajectory split by @yulangz in #1498
• fix(tests): assert all_gather_param passthrough by storage identity by @Le8r0nJames in #1562
• feat(rollout): add grouped reward normalization controls by @dingzhiqiang in #1516
• feat(vlm): add Qwen3.6 LoRA GRPO training support for 27B and 35B-A3B by @Lei00764 in #1444
• fix(engine): gate per-PP weight sync on SGLang backend by @koladefaj in #1564
• docs: allow ray scheduler in agent workflow compatibility notes by @wuallen57730 in #1567
• fix(rollout): attribute output tokens to the serving version by @Le8r0nJames in #1569
• fix(engine): engage Megatron deterministic mode before model build by @Le8r0nJames in #1544
• fix: emit PEFT-standard disk LoRA adapter keys so vLLM can load them (incl. #1577) by @sitabulaixizawaluduo in #1579
• feat(examples): add SGLang patch enabling Bailing hybrid radix cache by @Le8r0nJames in #1573
• fix(openai): align proxy tool schemas with sglang chat-completions rendering by @Le8r0nJames in #1499
• feat(mcore): apply fp32 lm head forward when enabled by @Le8r0nJames in #1574
• feat(ppo): report rejection-aware token and log-prob statistics by @Le8r0nJames in #1572
• fix(infra): fail fast when a local inference server dies during launch by @Le8r0nJames in #1545
• perf: reduce Megatron training memory peaks by @yulangz in #1555
• feat(colocate): support AWEX colocated actor-rollout training by @Le8r0nJames in #1500
• fix(scheduler): treat every terminal Slurm state as a dead worker by @Le8r0nJames in #1583
• feat: update awex to 0.8.0 by @sitabulaixizawaluduo in #1591
• fix: isolate AWEX control barriers from payload transfer by @pyq623 in #1596
• fix(infra): stabilize inference worker routing and caches by @pyq623 in #1589
• feat(scheduler): honor reservation, exclusive and user env overrides on Slurm by @Le8r0nJames in #1584
• feat(engine): support fixed warmup steps by @daihaowz in #1597
• chore: change codeowners by @sitabulaixizawaluduo in #1599
• docs(algorithms): describe metadata-driven group normalization by @EazyReal in #1602
• fix(models): support FP32 operands with chunked LM head by @yulangz in #1594
• feat(scheduler): support grouped colocation in the Ray scheduler by @Le8r0nJames in #1575
• fix(models): configure bridge provider for deterministic builds by @sitabulaixizawaluduo in #1603
• fix(infra): preserve LD_PRELOAD in local launchers by @YAO-001 in #1578
• docs(api): correct obsolete no_save_optim guidance by @koladefaj in #1609
• feat(engine): enable model-owned THD for Qwen3-VL by @Adiactive in #1608
• feat(engine): support Qwen3-VL with native AWEX colocate by @sitabulaixizawaluduo in #1605
• docs: update NPU installation guide for ascend-v1.0.5 by @Adiactive in #1617
• fix(infra): pause proxy workers during weight updates by @Adiactive in #1618
• feat(trainer): add flexible GAE lambda strategies by @RanranranQAQ in #1600
• feat: support qwen35 awex colocate weight update by @sitabulaixizawaluduo in #1620
• chore(deps): bump awex to 0.8.1 by @sitabulaixizawaluduo in #1621
• fix(lora): regression in megatron+lora support due to missing vllm lora aliases by @gursimar in #1601
• fix(megatron): release actor weights after async save by @Le8r0nJames in #1615
• fix: make rollout sampling deterministic by @pyq623 in #1625
• feat(awex): add separation AdamW delta weight transfer by @pyq623 in #1623
• docs: clarify terminology in the M2PO guides by @Wrestle7 in #1635
• fix(v2): forward retry-orphan cleanup by @Le8r0nJames in #1627
• fix(v2): balance offline session routing by @Le8r0nJames in #1624
• feat(v2): support SWE proxy compatibility by @Le8r0nJames in #1628
• feat(v2): support grouped reward normalization by @Le8r0nJames in #1619
• fix(examples): retune GSM8K GRPO learning rate for FP32 master weights by @harooos in #1634
• perf(megatron): avoid synthetic BSHD padding rows by @EazyReal in #1632
• fix(api): return float zero after async reward timeout by @harooos in #1541
• chore: bump v2.1.0 by @sitabulaixizawaluduo in #1639

New Contributors

• @jszzr made their first contribution in #1474
• @Lei00764 made their first contribution in #1444
• @wuallen57730 made their first contribution in #1567
• @YAO-001 made their first contribution in #1578
• @RanranranQAQ made their first contribution in #1600
• @Wrestle7 made their first contribution in #1635
• @harooos made their first contribution in #1634

Full Changelog: v2.0.0...v2.1.0