verl-project/verl v0.7.1

on GitHub

Highlight

Model Engine

Megatron

• Support R3 router replay with vllm and sglang #4840 #4986 #5185
• Support MTP training in SFT/RL #4981 #4936
• LoRA training enhancement with megatron-bridge: actor/ref share, LoRA adapter only refit, etc #4673 #4632
• Support Qwen3.5 series training with mbridge #5381

VeOmni

• New veomni training backend with FSDP+SP+EP #4882

torchtitan

• New torchtitan training backend with FSDP+TP+PP+CP+EP, roadmap: #5306

Rollout Engine

vLLM

• Separate model runner from training process and refit weights by cuda ipc #4280
• FP8 rollout enhancement
• Upgrade to vllm==0.17.0

SGLang

• Support router replay
• FP8 rollout enhancement
• Upgrade to sglang==0.5.9

TensorRT-LLM

• New tensorrt-llm rollout backend, roadmap: #5042

Checkpoint Engine

• Add checkpoint engine manager #5031
• Add hccl, kimi checkpoint engine backend #4885 #4954

Trainer

• one-step-off/fully async trainer refactor with verl-core
  • Unify checkpoint engine #5029
  • Unify partial rollout agent loop with auto resume #5487
  • Ascend NPU support for one-step-off/fully async

What's Changed

• [ci] feat: add npu unit test by @yyyy2000 in #4626
• [recipe] fix: workaround for making the one-step off-policy recipe compatible with IPv6 environments on Ascend NPU by @ji-huazhong in #4782
• [fsdp] feat: integrate PrefixGrouper for GRPO training acceleration by @kevssim in #4368
• [rollout] fix: use model_dump() for proper Pydantic serialization in token2text by @yurekami in #4706
• [doc] chore: Change the name of npu unit test workflow by @yyyy2000 in #4800
• [model] feat: support per sample temperature in trainer by @vermouth1992 in #4787
• [tool] fix: add tools in single_turn_agent by @Junxiao-Zhao in #4798
• [recipe] feat: migrate recipe to the dedicated repo verl-recipe as a submodule by @tongyx361 in #4795
• [model] fix: fix temp dtype by @vermouth1992 in #4813
• [vllm, sglang, rollout] fix: Fix a mistake when running run_qwen3_vl-30b-megatron.sh with latest verl and vllm0.12 by @cboss6 in #4810
• [ckpt] feat: add checkpoint-engine abstraction by @wuxibin89 in #4775
• [doc, ci] fix: Update Ascend doc and fix e2e_ascend CI by @FightingZhen in #4816
• [trainer] feat: VeOmniEngine supports qwen3_vl ulysses by @A1waysBeenHere in #4806
• [doc] chore: fix checkpoint engine image link by @wuxibin89 in #4821
• [hardware] fix: automatically set device for SFT case by @A1waysBeenHere in #4828
• [data] feat: TransferQueue - Update TransferQueue version and docs by @0oshowero0 in #4829
• [doc] Update docs about fully_async_policy by @jsfanfanfan in #4826
• [ckpt] fix: FSDP save ckpt after validation by @wdl339 in #4799
• [perf] feat: Add MFU for Qwen3-VL dense by @zhihaofang1017 in #4753
• [tool] fix: avoid nested ToolResponse in SandboxFusionTool by @Winston-Yuan in #4833
• [vllm] fix: fix error in vllm patch for diff vllm version and add ci for moe with fp8 rollout by @Agoniii in #4824
• [algo] feat: add optimal token baseline and variance proxy by @jiawei415 in #4678
• [megatron] fix: Fix error in megatron workers by @zhihaofang1017 in #4832
• [misc] feat: delete unnecessary base class in agent loop worker and vLLMHttpServer by @PeterSH6 in #4838
• [misc] feat: consolidate tensordict before dispatch by @vermouth1992 in #4830
• [training_utils] fix: json encode error in filelogger by @zhuangqh in #4811
• [ckpt] chore: skip saving hf_checkpoint during megatron+lora training & add a separate lora merge script by @Junxiao-Zhao in #4839
• [rollout, vllm] fix: accuracy issue in verl serve mode + vllm-ascend + dp + ep + tp scenarios by @leo-pony in #4783
• [fsdp] feat: add validate process on trainer node when use_trainer_do_validate=True by @chenjiaoAngel in #4683
• [misc] fix: recipe submodule accidentally been removed by @wuxibin89 in #4843
• [worker, training_utils] fix: Engine Metric Aggregation by @JacobHelwig in #4778
• [rollout] fix: configurable agent loop + multimodal data for fully-async by @XChen-Zero in #4842
• [ci] test: switch the vlm rl test case in the npu environment to use the model engine by @ji-huazhong in #4844
• [ckpt] fix: Megatron save ckpt after validation by @wdl339 in #4841
• [megatron] feat: Share actor and ref in LoRA by @HollowMan6 in #4673
• [fsdp, megatron] fix: Engine Rollout Worker LoRA Parameter Update by @JacobHelwig in #4836
• [algo, rollout, sglang] feat: Support router replay with sglang by @moehanabi in #4840
• [perf] feat: Add MFU for Qwen3-VL MoE by @zhihaofang1017 in #4859
• [misc] fix: fix 3d position_ids for train_mini_batch by @wdl339 in #4860
• fix(sft_trainer): Fix global_tokens and total_tokens metrics always showing 0.0 by @khazic in #4854
• [rollout,vllm] feat: support vllm scheduling policy config and generate setting priority by @RobotGF in #4874
• [ckpt] fix: prevent data loss when max_ckpt_to_keep=1 by @jreiml in #4873
• [worker] feat: New engine share actor and ref for LoRA by @HollowMan6 in #4867
• [worker] fix: new engine saves megatron LoRA adapters checkpoints by @HollowMan6 in #4866
• [ckpt] fix: properly handle optimizer offloading for HybridDeviceOptimizer by @jreiml in #4870
• [doc] chore: update verl meetup by @wuxibin89 in #4884
• [vllm] Fix CLI argument serialization for list types by @jreiml in #4869
• [data] fix: build_messages for multi-modal data by @ccilery in #4864
• [rollout] fix: wrong display about Prometheus when using SGLang. by @jsfanfanfan in #4858
• [vllm] fix: pad data_hp to be multiples of block_size by @Agoniii in #4835
• [ci] feat: add ci to automatically submit PR request if precommit fails by @vermouth1992 in #4878
• [doc] chore: async README backticks by @JacobHelwig in #4898
• [doc] fix: correct typo in script comment by @Prozac614 in #4900
• [veomni] refactor: minor refactoring to ensure veomni engine compatibility with forward_only mode by @ji-huazhong in #4889
• [sglang] fix: sglang TP+DP support / port bug by @hustmf in #4715
• Either remove + prefix: 'actor_rollout_ref.model.enable_activation_of… by @Tomsawyerhu in #4910
• [vllm] fix: vllm_config arg gets removed in newer WorkerWrapperBase by @HollowMan6 in #4915
• Correct Attention FLOPS estimation in flops_counter.py by @HaochenYuan in #4929
• [algo, doc] feat: trust region sequence masking - (1) k3 KL avg and (2) veto for max criterion by @szrlee in #4544
• [rollout] feat: use rollout and validate parallel process by @chenjiaoAngel in #4863
• [model] feat: Add qwen3_vl_moe in VL_TYPE2INDEX for image_mask and vedio_mask computation by @A1waysBeenHere in #4923
• [data] feat: TransferQueue - fix rm_score error of TransferQueue by @baymax591 in #4928
• [vllm] fix: Update get_encoding import for vllm versions 0.13.0 and above by @xhx1022 in #4934
• [ci] fix: Add hydra-core to pre-commit installation by @vermouth1992 in #4892
• [data] feat: TransferQueue - Unify the return of reward by @walterchenchn in #4902
• Revert "Correct Attention FLOPS estimation in flops_counter.py" by @vermouth1992 in #4937
• [misc] fix: Correct Docstring arg in main() (PPO trainer) by @rfy48 in #4943
• [misc] fix: resolve pre-commit hook execution errors by @ji-huazhong in #4941
• [model] fix: qwen3-vl-30b npu_patch fix by @bjf-frz in #4888
• [veomni] feat: support offloading/loading the veomni model/optimizer by @ji-huazhong in #4916
• [rollout,mbridge] feat: add metrics for rollout num preempted and fix mbrideg freeze moe by @RobotGF in #4956
• [single_controller] fix: pass max_colocate_count and detached params when merging RayResourcePool by @wdl339 in #4949
• [single_controller] feat: Support dispatch/collect nested tensors with 3 or more dimensions by @JacobHelwig in #4940
• [env] fix: upgrade torch, cudnn and deps versions in vllm image to fix performance issue by @Begunner in #4960
• [training_utils] fix: correctly _resolve_device when not specified by @HollowMan6 in #4961
• [trainer] fix: pass scores device type to group_mean_std call by @HollowMan6 in #4962
• [training_utils] fix: Correct Attention TFLOPS estimation & fix CI by @HaochenYuan in #4959
• [training_utils] A bug that caused device selection in group statistics to fail has been covered by tests. by @JohnConnor123 in #4967
• [doc, data] fix: resolve broken documentation hyperlinks by @aphrodite1028 in #4970
• [sglang, rollout] feat: support sglang as rollout engine in fully async policy by @AniZpZ in #4191
• [megatron] feat: Using MTP in RL Training and Inference by @ArronHZG in #4936
• [megatron] fix: fix megatron sync_weights oom on user_trainer_do_validate mode by @chenjiaoAngel in #4944
• [rollout,vllm] fix: num_preempted metrics fix and typo correction in vllm async server by @RobotGF in #4976
• [ray,rollout,trtllm] feat: Adding tensorrt_llm as new rollout engine by @joyang-nv in #4665
• [data] fix: use lazy import for qwen_vl_utils in vision_utils.py by @Wheeeeeeeeels in #4991
• [recipe,tool] feat: make GSM8K multiturn tool quickstart actually work by @letsgetai in #4998
• [perf] feat: verl profiler system support Agent Loop scenario and integrate torch.profiler by @mengchengTang in #4320
• [vllm] fix: vllm TP+DP suuport bug by @ccilery in #4969
• [fsdp] fix: use module instead of function for fully_shard_module by @moaead in #5002
• [misc] fix: update version in the main branch by @yyDing1 in #5006
• [ckpt] feat: add Hccl ckpt engine backend by @hanhan-networking in #4885
• [reward] fix: conditionally include reward_extra_keys in meta_info based on rm_scores presence by @none0663 in #5005
• [rollout, vllm, sglang] fix: set default max_model_len by @ji-huazhong in #5018
• [ci] fix: fix ci by @vermouth1992 in #5022
• [ckpt] fix: npu load checkpoint by @Li-Yongwen in #4938
• [megatron] fix: patch mcore for MLA support with flash_attn by @HollowMan6 in #4931
• [BREAKING][worker, rollout, vllm] feat: implement vLLM colocated training-inference rollout with process separation by @jianjunzhong in #4280
• [megatron] feat: LoRA adapter only refit (TensorLoRARequest) by @HollowMan6 in #4632
• [veomni] refactor: no long check the attn_implementation/moe_implementation in VeOmniEngineConfig by @A1waysBeenHere in #5019
• [veomni] feat: support model resharding between veomni and rollout engine by @ji-huazhong in #5033
• [trtllm] fix: Fixes for TRTLLM rollout by @hchings in #5032
• [ci] fix: docker transformers==4.57.6 by @yyyy2000 in #5053
• [rollout] feat: set max_model_len by max_model_len or use max_position_embedding by @RobotGF in #5052
• [ckpt] feat: add CheckpointEngineManager by @wuxibin89 in #5031
• [doc] feat: add npu gspo practice by @wucong25 in #4988
• [ci] chore: move to verl-project by @wuxibin89 in #5059
• [vllm, sglang] feat: opt for FP8 rollout memory by @Agoniii in #4997
• [model] feat: add API to support automatically support engine backend by @vermouth1992 in #5050
• [megatron, training_utils] fix: Patch MoEAlltoAllTokenDispatcher.preprocess for router replay by @HollowMan6 in #4986
• [rollout, perf, cfg] fix: Add global step info and support more profile control params for rollout profiling (sglang backend) by @bithighrr in #5025
• [fsdp, megatron] feat: Support fully-async training on Ascend NPU by @acat-rw in #5043
• [doc, trainer] fix: shoudn't use rollout routing replay data for R2 by @HollowMan6 in #4973
• [doc] feat: add dapo multi model optimization practice by @ChibiQuest in #5044
• [ci] chore: fix ci failure by @wuxibin89 in #5068
• [ci] chore: fix npu ci failure by @wucong25 in #5064
• [sglang,ci,doc] feat: Update Ascend Dockerfile and docker build workflow to 8.3.RC1 version for VeRL + Sglang by @xiazhahe in #5065
• [megatron, training_utils] fix: router replay R3 align router replay data with global layer indices by @HollowMan6 in #5037
• [trainer] fix: resolve dataset config in agent loop by @yyDing1 in #5034
• [ckpt,rollout] fix: sleep_replicas before save_ckpt to avoid OOM by @RobotGF in #5079
• [reward, ci] fix: colocate reward model ci break by @yyDing1 in #5084
• [reward] fix: fix reward computation in _validate when use_reward_loop=True and reward_model.enable=True by @none0663 in #5054
• [rollout] fix: fix cpu allocation error in tensorrt_llm rollout manager by @SchumiDing in #5085
• Revert "[reward] fix: fix reward computation in _validate when use_reward_loop=True and reward_model.enable=True" by @wuxibin89 in #5091
• [trtllm] fix: minor fixes to trtllm rollout by @hchings in #5095
• [sglang] feat: add NPU GRPO training scripts for Qwen2.5-32B (FSDP/SGLang backends) by @xiazhahe in #5062
• [doc] chore: update arch image by @wuxibin89 in #5106
• [rollout] feat: automatically resume generation on abort by @wuxibin89 in #5071
• [sglang, doc] feat: add NPU GRPO training scripts for Qwen3-30B (Megaton/SGLang backends) and update doc by @hustmf in #5060
• [megatron] fix: megatron async save ckpt fix by @Leem-Li in #5016
• [model] feat: add NPU GRPO training scripts for Qwen2.5-32B/Qwen3-30B (Megaton/vLLM backends) by @psyloy in #4984
• [data] feat: Add support for Llama3.2-11-b-vision by @SchumiDing in #5112
• [vllm] feat: revert to the default behavior of cudagraph_mode by @vermouth1992 in #5109
• [fsdp] fix: Handle different transformers versions for Vision2Seq models in FSDP model merger by @liangxuZhang in #5108
• [megatron] feat: Support MTP training in SFT by @arvyanh in #4981
• [sglang] fix: update wiki to support speculative decode rollout by @ArronHZG in #5116
• [training_utils] fix: add upcasting for seq_len_effective to avoid potential overflow in calculate_workload by @albertcity in #5110
• [ci] feat: add npu workflow，e2e_sft_llm&model&reward_model_vllm by @yyyy2000 in #5039
• [doc] chore: update readme for SPEAR algorithm by @yuleiqin in #5124
• [vllm] feat: add shared memory support for weight transfer and IPC support checks by @jianjunzhong in #5089
• Revert "[rollout] feat: automatically resume generation on abort" by @PeterSH6 in #5127
• [sglang] fix: skip MoE router layers for FP8 quantization by @eternally-z in #5122
• [vllm] feat: get gpt-oss encoding on demand by @vermouth1992 in #5131
• [rollout,vllm] feat: revert default value of max_num_seqs by @RobotGF in #5139
• [misc] feat: Modify transformers dependency version in requirements by @vermouth1992 in #5141
• [ci] fix: fix ci by downgrade transformers to <5 by @vermouth1992 in #5143
• [misc] chore: rename huggingface-cli to hf to favor transformers v5 by @vermouth1992 in #5145
• [worker,rollout] refactor: remove set_expandable_segments calls in vllm separation mode by @RobotGF in #5144
• [megatron] fix: patched_routing accepts arbitrary args by @HollowMan6 in #5155
• Revert "[worker,rollout] refactor: remove set_expandable_segments calls in vllm separation mode" by @vermouth1992 in #5156
• [model, algo] feat: implement SAC algorithm and support Pi0.5 model by @Miical in #5118
• [training_utils] fix: Resolved bugs and conflicts in the fully async caused by multiple PRs by @ZLiao097 in #5100
• [reward] feat: split reward loop manager and agent loop manager by @yyDing1 in #5134
• [Doc] feat: update README to add new awesome project RuleReasoner by @jacklanda in #5157
• [trainer] feat: move save_ckpt before update_weights and validate by @RobotGF in #5137
• [ci,doc] feat: Add ascend_ci_guide by @yyyy2000 in #5163
• [rollout] fix: remove dtype cast by @vermouth1992 in #5117
• [hardware] fix: update architecture check and CANN toolkit path retrieval in device.py by @jianjunzhong in #5142
• [vllm] fix: build_app() missing 1 required positional argument: 'supported_tasks' by @HollowMan6 in #5093
• [perf] feat: clear megatron global buffer memory by @wuxibin89 in #5173
• [vllm, rollout] fix: Use different seeds for vllm by @victordion in #5179
• [ci] fix: fully async ci break by @yyDing1 in #5166
• [vllm] feat: make seed configurable and different among replicas by @vermouth1992 in #5181
• [data] fix: keyword video_metadata by @sophiayyya in #5177
• [megatron] fix: checkpoints uses fully_reshardable by default when supported by @HollowMan6 in #5154
• [trtllm, rollout] test: add unittest by @hchings in #5102
• [reward] refactor: migrate all reward managers to the new asynchronous reward manager by @yyDing1 in #5189
• [vllm] fix: handle multimodal inputs correctly in full async mode by @Silas-11 in #5160
• [megatron] feat: fused kernel suppport for new model engine by @HollowMan6 in #5191
• [fsdp] feat: Merge lora in fsdp training to speed up rollout by @amzfang in #5115
• [megatron] Add Megatron-Bridge support in fully async policy by @eternally-z in #5196
• [perf] fix: infer server profiler args fix by @mengchengTang in #5121
• [doc, perf] feat: add perf_tuning_on_ascend by @tardis-key in #5104
• [ci] feat: add three npu workflow yml test by @daikang6 in #4978
• [vllm] fix: ignore MoE router layers for FP8 quantization by @zpqiu in #5107
• [worker, training_utils] fix: Metric Aggregation Across DP Ranks by @JacobHelwig in #5203
• [megatron] fix: add protections for logits_processor_args.pop("loss_mask"), which may cause the forward_fn of value net collapse by @albertcity in #5204
• [trtllm] fix: reduce peak mem usage during update_weight() by @hchings in #5212
• [algo] feat: support rollout router replay in MegatronEngine by @xhx1022 in #5185
• [trtllm] fix: add synchronization before resume kv_cache to prevent oom in non-leader ranks by @shuyixiong in #5208
• [perf] feat: add images_seqlens on mfu calculation for engine_worker by @alwaysyiyu in #5207
• [reward] fix: Add assert to prevent reward NaN caused by overlong_cfg.len=0 by @ZLiao097 in #5216
• [recipe] refactor: refactor ray trainer for separate recipe use. (fully async / one step off) by @ArronHZG in #5184
• [BREAKING][reward] refactor: remove reward model worker code and invocation by @yyDing1 in #5194
• [fsdp] fix: Support trust_remote_code during FSDP HugginFace checkpoint save by @thvasilo in #5200
• [worker] feat: Avoid redundant base weight sync when engine doesn't sleep by @JohnConnor123 in #5147
• [ci] chore: fix npu ci by @wucong25 in #5218
• [vllm] fix: apply moe weight loader patch for standard wight loading by @zjchenn in #5234
• [ci] chore: fix npu ci setuptools by @yyyy2000 in #5238
• [ci] chore: fix npu ci setuptools, keep update pip and packaging by @yyyy2000 in #5239
• [reward] fix: preserve input non_tensor_batch in AgentLoopManager when reward_loop_worker_handles is None by @none0663 in #5195
• [perf] fix: fix npu profiling scripts by @tongtong0613 in #5226
• [megatron] feat: use yaml to manage mbridge args by @Kite0011 in #4584
• [algo] feat: reduce routed expert padding via NestedTensor and uint8 dtype by @xhx1022 in #5240
• [ray,trainer] feat: add master port range configuration for port range by @RobotGF in #5201
• [BREAKING][reward] refactor: deprecate batch reward manager by @yyDing1 in #5237
• [fsdp] feat: add script for qwen3next training on npu platform by @zjchenn in #5236
• [doc] fix: Update ascend_sglang_best_practices.rst by @hustmf in #5261
• [vllm, rollout] feat: update abort function with vllm internal pause_generation api by @PeterSH6 in #5253
• [veomni, trainer] feat: add rl support for veomni backend by @ji-huazhong in #4882
• [vllm] fix: run post-load weight processing once after async IPC sync by @zjchenn in #5235
• [doc] chore: version of dapo_multi_model_optimization_practice by @ChibiQuest in #5263
• [rollout] feat: make more rollout flags configurable to trtllm backend by @Superjomn in #5258
• [doc] refactor: update reward documents by @yyDing1 in #5272
• [doc] chore: Ascend retool practice doc by @LeoYao123 in #5266
• [vllm, rollout] fix: auto-downgrade cudagraph_mode to PIECEWISE when DCP is enabled by @Siritao in #5262
• [fsdp, veomni, trainer] fix: restrict npu-patch scope to avoid veomni backend interference by @ji-huazhong in #5268
• [BREAKING][reward] refactor: the full reward configuration by @yyDing1 in #5255
• [ci] chore: delete redundant npu ci by @yyyy2000 in #5259
• [fsdp, megatron] refactor: Refactor Fully Async Implementation via Engine Workers by @ZLiao097 in #5269
• [megatron, model] chore: add example of nemotron nano v3 by @ISEEKYAN in #5284
• [misc] chore: fix veomni_trainer.yaml by @wuxibin89 in #5285
• [megatron] fix: fallback to moe_router_padding_for_fp8 in router replay patch by @xhx1022 in #5283
• [reward] fix: backward compatibility with old reward config by @yyDing1 in #5287
• [reward] fix: reward model args and reward_kwargs bug by @yyDing1 in #5289
• [doc] chore: gspo update config and add version with npu by @chengminhua in #5279
• [fsdp,veomni] fix: remove FSDPUlyssesShardingManager to make eval_mode/train_mode reentrant by @wuxibin89 in #5305
• [veomni] refactor: Modify dp related parameters to align with FSDP backend and remove temporarily unsupported TP/PP/CP parameters by @ChengQianqian in #5303
• [trtllm] feat: use max utilization scheduler by default by @tongyuantongyu in #5302
• [worker, tool] fix: stabilize agent loop extra fields schema by @denismegerle in #5301
• [algo] feat: add NPU SAPO training script for Qwen3-8B (FSDP/vLLM backends) by @Vvictorrrr in #5257
• [fsdp, vllm] feat: add NPU GRPO training scripts for Qwen3-VL-8B (FSDP/VLLM backends) by @zhihaofang1017 in #5250
• [fsdp, vllm] feat: add NPU GRPO training scripts for Qwen3-VL-30B (FSDP/VLLM backends) by @alwaysyiyu in #5260
• [model,cfg] fix: type annotation for Lora target_modules by @thvasilo in #5223
• [megatron] feat: Support LoRA training with FP16 using Megatron-Bridge. by @xichengpro in #4648
• [ci] fix: main pre-commit by @pengwu22 in #5318
• [misc] refactor: delete remaining batch-mode code in single controller by @ji-huazhong in #5319
• [rollout] fix: make skip rollout compatible with async mode by @ChengQianqian in #5320
• [veomni, trainer] fix: padding pixel value with padding_scale for vl model by @A1waysBeenHere in #5322
• [fsdp,algo] feat: add NVFP4 QAT (Quantization-Aware Training) support by @zhangyimi in #5190
• [docs] Add new awesome work using Verl by @MING-ZCH in #5328
• [vllm] feat: remove workers from vLLMHttpServer by @tongyx361 in #5330
• Revert "[vllm] feat: remove workers from vLLMHttpServer" by @PeterSH6 in #5333
• [misc] refactor: remove deprecated codes by @ji-huazhong in #5336
• [misc] fix: include config files for experimental entrypoints in package data by @guillemgt in #5343
• [ci] chore: set torch-npu to 2.7.1.post2 in ascend dockerfile by @ji-huazhong in #5345
• Revert "[ci] chore: set torch-npu to 2.7.1.post2 in ascend dockerfile" by @ji-huazhong in #5353
• [reward] fix: empty class_dict for standalone reward model resource pool by @yyDing1 in #5348
• [trainer] feat: Add Torchtitan as alternative training engine by @acisseJZhong in #5051
• [training_utils] fix: mask out-of-bounds vocab entries fused kernel LCE logsumexp by @EricMarcus-ai in #5349
• [rollout] fix: Include routed_experts in ToolAgentLoop return value to support R3 router replay by @mirrorboat in #5368
• [misc] fix: pass torch dtype when init random model by @HollowMan6 in #5370
• [ci] chore: pin version cupy-cuda12x==13.6.0 by @wuxibin89 in #5377
• [doc] chore: ascend add performance analysis guide and update some version info by @chengminhua in #5324
• [trainer] feat: Support RL trainer with TorchtitanEngine by @acisseJZhong in #5356
• [algo] feat: Exception for agg_loss when dp_size > 1 but global information is absent & fix: correct & consistent loss aggregation for "seq-mean-token-sum-norm" by @tongyx361 in #5366
• [rollout] fix: make run_uvicorn behavior more reliable by @tongyuantongyu in #5383
• [doc] feat: update documentation for The Optimal Token Baseline and Rollout Correction by @jiawei415 in #5380
• [trainer] refactor: remove fsdp_sft_trainer.py by @wuxibin89 in #5382
• [ci] fix: occasional CI failures caused by sglang server port conflicts by @pengwu22 in #5310
• [fsdp] fix: add aggressive_empty_cache at end of init_model to prevent vLLM OOM by @EricMarcus-ai in #5384
• [doc, worker] feat: Enable Megatron-Bridge for MTP by @HollowMan6 in #5323
• [ckpt] feat: add kimi ckpt engine backend by @kip-cxj in #4954
• [misc] feat: ignore pyrightconfig.json to allow users to customize pyrightconfig to fix breaks by @tongyx361 in #5385
• [ci] chore: update triton-ascend and fix npu ut by @yyyy2000 in #5396
• [fsdp, megatron] feat: refactor fully-async and one-step-off training to support multiple checkpoint engine backends by @Shangwei-Li in #5029
• [doc] feat: add fully async and one step off to PR Checklist by @ArronHZG in #5404
• [doc] chore: ascend update gspo optimization practice document by @chengminhua in #5408
• [algo] feat: add DPPO with binary TV or binary KL implementation by @QPHutu in #5397
• [doc] chore: npu best practice doc by @hustmf in #5415
• [algo] fix: seq mean and default scale factor loss_mask.shape[-1] as in seq-mean-token-sum-norm by @tongyx361 in #5417
• [megatron] fix: missing model offload to CPU for forward_only mode by @xhx1022 in #5406
• [megatron] feat: enhance model offloading and loading for frozen parameters by @RobotGF in #5412
• [perf] fix: the overwritten of Torch_profile with multi steps. by @Rhetee in #5395
• [trainer] feat: add padding for tensor alignment in preprocess_thd_no_padding function by @RobotGF in #5410
• [tool] fix: handle empty image inputs in ToolAgentLoop by @denismegerle in #5420
• [rollout, data] fix: honor train_max_samples/val_max_samples in fully async rollouter by @denismegerle in #5359
• [tool] refactor: remove tool schema plumbing from SingleTurnAgentLoop by @denismegerle in #5425
• [misc] feat: Add code for data grouping in no-padding scenario by @Kite0011 in #5424
• [doc] add Dr. MAS to awesome work by @langfengQ in #5427
• [BREAKING][rollout,cfg] refactor: get rid of actor_rollout_ref config from rollout by @wuxibin89 in #5418
• [ci] chore: bump the version of vllm-ascend to v0.11.0 in the ascend dockerfile by @ji-huazhong in #5431
• [doc] chore: fix npu docs by @wucong25 in #5428
• [doc] fix: fix npu retool docs by @LeoYao123 in #5449
• [data] refactor: TransferQueue - retire legacy integration codes by @0oshowero0 in #5454
• [ci] fix: failed trtllm_unit_tests with attribute error by @HollowMan6 in #5446
• [megatron] fix: pass dp_group to rearrange_micro_batches to fix DeepEP timeout by @xhx1022 in #5451
• [rollout] fix: remove unexpected concurrency bound at 1000 by @tongyuantongyu in #5402
• [data] fix: accept jsonl dataset files by @zqzten in #5456
• [single_controller] refactor: use BatchData to simplify concat and chunk in single_controller by @zw0610 in #5450
• [megatron] feat: Support DSA indexer LoRA mappings by @HollowMan6 in #5462
• [doc] fix: fix typo in agentic rl doc by @KevinZeng08 in #5461
• [misc] chore: support transformers 5 by @HollowMan6 in #5445
• [doc] fix: fix dapo multi model practice by @ChibiQuest in #5453
• [trainer] feat: Update trainer API for TorchtitanEngine by @acisseJZhong in #5457
• [rollout] refactor: bucketed transfer utils by @pengwu22 in #5309
• [rollout] feat: update trtllm docker by @Superjomn in #5386
• [doc] fix: fix npu retool doc by @LeoYao123 in #5467
• [ckpt] feat: add mooncake backend by @x1314aq in #5176
• [doc] chore: add ascend backend feature by @wucong25 in #5466
• [megatron] fix: support hybrid dense/MoE models in router replay with PP/VPP by @xhx1022 in #5452
• [megatron] fix: patch support newer mcore version by @HollowMan6 in #5372
• [ci] fix: sanity issue related to Last updated string by @HollowMan6 in #5477
• [rollout] feat: support auto resume on abort in FullyAsyncLLMServerManager by @wuxibin89 in #5430
• [trainer] feat: Support EP with TorchtitanEngine by @acisseJZhong in #5469
• docs: fix typo in kl_penalty docstring by @ZHAOoops in #5481
• [megatron] fix: add FP8 block quantization padding for EngineWorker by @zpqiu in #5440
• [ckpt, model] fix: preserve lora_alpha in model_merger via training meta by @Yatogaii in #5326
• [fsdp,algo] feat: Support QAT (NVFP4) in FSDPEngine for the unified engine_workers architecture by @zhangyimi in #5411
• [doc] feat: add mtp spec log by @ArronHZG in #5491
• [reward] feat: add example scripts for reward model usage by @yyDing1 in #5486
• [BREAKING][trtllm] feat: Add FP8 refit support for trtllm rollout by @shuyixiong in #5374
• [veomni,ci] fix: Modify default setting in veomni test scripts to prevent misunderstanding by @0oshowero0 in #5484
• [ckpt] fix: test issues of kimi and mooncake backend by @x1314aq in #5500
• [doc] chore: update FP8 guide with E2E training section and reorganization by @zpqiu in #5502
• [model,doc] feat: add qwen3 32B megatron 1k to 256k by @ChibiQuest in #5497
• [doc] chore: npu docker support vllm013 by @yyyy2000 in #5471
• [doc] fix: update recipe link to fix 404 not found by @tardis-key in #5286
• [ci] feat: add npu nightly ci by @daikang6 in #5225
• [data] fix: use %-style format placeholders in logger.warning() by @cavities12 in #5512
• [rollout] feat: global request-level load balancer single source routing by @aoshen524 in #5399
• [rollout] feat: Fix partial load problem, Add vlm support for trtllm rollout by @SchumiDing in #5149
• Revert "[rollout] feat: Fix partial load problem, Add vlm support for trtllm rollout" by @wuxibin89 in #5525
• [ckpt] fix: Fix checkpoint engine backend unset error by @ZLiao097 in #5473
• [rollout] feat: Fix partial load problem, Add vlm support for trtllm rollout by @SchumiDing in #5528
• [Megatron] feat: Support routing replay on NPU with performance and compatibility enhancements by @755651978 in #5298
• [rollout] fix: update checkpoint_engine bucket size parameter for Ascend compatibility by @nuerxiati in #5539
• [misc] feat: support dynamic bsz using group size by @Kite0011 in #5438
• [fully_async, one_step_off] feat: support auto resume on abort when using fully_async by @ArronHZG in #5487
• [doc] chore: add note for kimi ckpt engine by @kip-cxj in #5546
• [perf, trainer, training_utils] fix: Try to montior with mlflow up to 3 times, and avoid duplicate key processing in each step. by @sheilaliuxl in #5548
• [trainer] fix: support nsys when using sft_trainer_ray.py by @arvyanh in #5489
• [rollout] fix: reintroduce NCCL_CUMEM_ENABLE for weight synchronization in async rollout environments by @RobotGF in #5522
• [ci] feat: npu nightly ci log is redirected to the specified directory by @daikang6 in #5557
• [ci] fix: sft_trainer_ray ci break by @wuxibin89 in #5562
• [fsdp] fix: wrap embed_tokens/lm_head by name for peft models by @cavities12 in #5516
• [ci] chore: update npu ci to vllm013 by @yyyy2000 in #5523
• [algo] feat: support router replay in MegatronEngine by @xhx1022 in #5219
• [docker] feat: update stable image to vllm==0.17.0, sglang==0.5.9 by @Begunner in #5542
• [megatron, model] feat: qwen3.5 example by @ISEEKYAN in #5381
• [algo] feat: add GDPO (Group reward-Decoupled Normalization Policy Optimization) algorithm by @Rhetee in #5503
• [megatron] feat: model engine support mtp by @ArronHZG in #5561
• [doc] fix: fix te pip install instructions by @TKONIY in #5501
• [rollout] fix: agent loop copy read-only routed_experts before torch conversion by @HollowMan6 in #5519
• [ci] chore: change machine for npu ci by @yyyy2000 in #5578
• [megatron] fix: apply override_transformer_config inside mindspeed engine to avoid confict with other training engine by @ChengQianqian in #5589
• [rollout] fix: fix some compatibility issue with qwen vl seris support of trtllm rollout by @SchumiDing in #5583
• [misc] chore: bump version to 0.7.1 by @wuxibin89 in #5602

New Contributors

• @yyyy2000 made their first contribution in #4626
• @Junxiao-Zhao made their first contribution in #4798
• @cboss6 made their first contribution in #4810
• @wdl339 made their first contribution in #4799
• @zhihaofang1017 made their first contribution in #4753
• @Winston-Yuan made their first contribution in #4833
• @jiawei415 made their first contribution in #4678
• @XChen-Zero made their first contribution in #4842
• @khazic made their first contribution in #4854
• @jreiml made their first contribution in #4873
• @Prozac614 made their first contribution in #4900
• @hustmf made their first contribution in #4715
• @Tomsawyerhu made their first contribution in #4910
• @xhx1022 made their first contribution in #4934
• @walterchenchn made their first contribution in #4902
• @rfy48 made their first contribution in #4943
• @bjf-frz made their first contribution in #4888
• @JohnConnor123 made their first contribution in #4967
• @aphrodite1028 made their first contribution in #4970
• @AniZpZ made their first contribution in #4191
• @joyang-nv made their first contribution in #4665
• @Wheeeeeeeeels made their first contribution in #4991
• @letsgetai made their first contribution in #4998
• @moaead made their first contribution in #5002
• @hanhan-networking made their first contribution in #4885
• @Li-Yongwen made their first contribution in #4938
• @jianjunzhong made their first contribution in #4280
• @hchings made their first contribution in #5032
• @bithighrr made their first contribution in #5025
• @ChibiQuest made their first contribution in #5044
• @xiazhahe made their first contribution in #5065
• @SchumiDing made their first contribution in #5085
• @psyloy made their first contribution in #4984
• @liangxuZhang made their first contribution in #5108
• @arvyanh made their first contribution in #4981
• @albertcity made their first contribution in #5110
• @yuleiqin made their first contribution in #5124
• @eternally-z made their first contribution in #5122
• @Miical made their first contribution in #5118
• @jacklanda made their first contribution in #5157
• @victordion made their first contribution in #5179
• @sophiayyya made their first contribution in #5177
• @Silas-11 made their first contribution in #5160
• @amzfang made their first contribution in #5115
• @daikang6 made their first contribution in #4978
• @shuyixiong made their first contribution in #5208
• @alwaysyiyu made their first contribution in #5207
• @thvasilo made their first contribution in #5200
• @Superjomn made their first contribution in #5258
• @LeoYao123 made their first contribution in #5266
• @Siritao made their first contribution in #5262
• @ChengQianqian made their first contribution in #5303
• @tongyuantongyu made their first contribution in #5302
• @denismegerle made their first contribution in #5301
• @Vvictorrrr made their first contribution in #5257
• @zhangyimi made their first contribution in #5190
• @MING-ZCH made their first contribution in #5328
• @guillemgt made their first contribution in #5343
• @acisseJZhong made their first contribution in #5051
• @mirrorboat made their first contribution in #5368
• @kip-cxj made their first contribution in #4954
• @QPHutu made their first contribution in #5397
• @Rhetee made their first contribution in #5395
• @zqzten made their first contribution in #5456
• @KevinZeng08 made their first contribution in #5461
• @x1314aq made their first contribution in #5176
• @ZHAOoops made their first contribution in #5481
• @Yatogaii made their first contribution in #5326
• @cavities12 made their first contribution in #5512
• @755651978 made their first contribution in #5298
• @sheilaliuxl made their first contribution in #5548
• @TKONIY made their first contribution in #5501

Full Changelog: v0.7.0...v0.7.1