v0.3.8

What's Changed

• fix: refactor of weight mapper no need unsplit map specification by @lfengad in #436
• Remove unsupported fields from Rollout parallelism config. by @foreverlms in #437
• fix: add support for reasoning vla / avla usage by @lfengad in #439
• Disable DeepEP for architectures older than Hopper by @bastefaniak in #441
• Fix n_local_experts computation in DeepseekV3 and Qwen3 MoE by @bastefaniak in #440
• feat: off policy sequence masking by @xlu451 in #431
• fix: slurm launch dp replica support by @lfengad in #443
• fix: fix deepep usage due to synchronize issue by @lfengad in #445

New Contributors

• @bastefaniak made their first contribution in #441

Full Changelog: v0.3.7...v0.3.8

v0.3.7

What's Changed

• fix dpsk hf convert by @yy-code-nv in #415
• Fix: pass norm_topk_prob for qwen3_vl_moe and intern_vl by @kane-vln in #414
• Fix named buffer init by @yy-code-nv in #411
• Fix moe implementation by @yy-code-nv in #416
• Customize build_model with extra hf_config_args by @kane-vln in #418
• RFC: refactor trainer for better customization by @foreverlms in #412
• rfc: colocated mode by @lfengad in #413
• feat: full custom case example with readme by @lfengad in #421
• Add hooks for SFT validation by @foreverlms in #419
• feat: Refine for custom example by @lfengad in #422
• Update deepseek weight mapping for GRPO(vllm >= 0.10.0) by @kane-vln in #410
• feat: update non-text rollout cases interface handling. by @lfengad in #425
• feat: unbiased kl estimate by @xlu451 in #423
• feat: control of weight version in DAPO case. by @lfengad in #426
• fix: data type specification fixed and refine by @lfengad in #429
• fix qwen3 moe weight exporting by @foreverlms in #428
• fix(controller): guard zero division when no policy replicas registered by @xlu451 in #435

Full Changelog: v0.3.6...v0.3.7

v0.3.6

What's Changed

• fix: More for training strategy consistency and metrics demonstration by @lfengad in #356
• Fix: qwen3_vl_moe encoder use FlashAttnMeta by @kane-vln in #358
• Support TP for HFModel by @kane-vln in #355
• fix: Fix regression for more metrics in validation case by @lfengad in #360
• feat: Add post process for rollout generation in data packer by @lfengad in #361
• feat: SFT training with DDP to load model at only master rank by @lfengad in #359
• Support video input for qwen3-vl/hf vlm datapacker by @kane-vln in #365
• Update tests for datapacker by @kane-vln in #367
• Enable local dataset loading and fetching for Policy and Rollout. by @foreverlms in #354
• feat: Decoupled loss for async RL by @lfengad in #368
• Remove prompt_idxs which is not needed now. by @foreverlms in #371
• Fix: add tp_slice_dim initialization in state dict conversion by @kane-vln in #372
• [FRC] Couple tokenizer with data packer by @heslami in #311
• Support Nemotron-Nano SFT by @kane-vln in #373
• Support sequence packing for HFModel by @kane-vln in #369
• feat: dapo case move rollout filter into rollout worker by @lfengad in #387
• Add expandable segmentation for pytorch allocator by @yy-code-nv in #388
• Add the deepep support for Qwen3-MoE models by @yufanhuangNV in #389
• Fix: resolve version incompatibility between FA3 and TE  by @kane-vln in #391
• Add sanity check for parallelism by @foreverlms in #390
• fix: Fix hf gradient checking by @lfengad in #394
• Enable FP4 dynamic quantization of linear layers for policy training by @yufanhuangNV in #374
• rfc: restructure of some common used logic in parallel map by @lfengad in #395
• Fix: fp4 compatible with python env by @lfengad in #398
• fix: qwen2.5 vl case execution fix by @lfengad in #399
• RFC: Refactor rollout worker part by @lfengad in #396
• fix: resume from ckpt of hf buffer handling by @lfengad in #401
• Fix qwen2-5 modeling by @yy-code-nv in #404
• fix: stop issue due to validation by @lfengad in #405
• Fix qwen3-moe and qwen3-vl-moe safetensors export by @foreverlms in #406
• Support allgather moe dispatcher by @kane-vln in #402
• Force FSDP warp to ensure consistent mix-precision training behavior, fix qwen3-moe deepep bug by @yy-code-nv in #409

New Contributors

• @yy-code-nv made their first contribution in #388
• @yufanhuangNV made their first contribution in #389

Full Changelog: v0.3.5...v0.3.6

v0.3.5

What's Changed

• Pass video_kwargs for policy input by @Dinghow in #346
• Support the case where position_ids is None for qwen-vl series by @kane-vln in #325
• Using FA3 in policy for hopper platform. by @foreverlms in #313
• Support qwen3-vl-moe grpo by @kane-vln in #345
• Fix docker build that lack of psutil by @foreverlms in #350
• [Fix] remove error raise when importing te by @foreverlms in #352
• fix: Refine on-policy and metrics by @lfengad in #351

Full Changelog: v0.3.4...v0.3.5

v0.3.4

What's Changed

• Support OAI-GPT-OSS by @foreverlms in #202
• feat: support multi-turn rl and tool call by @jingxu9x in #197
• fix: prompts payload from dataset fixed for general cases. by @lfengad in #285
• refactor: use api_client replace request url by @jingxu9x in #265
• Add cosine similarity check into context parallel test by @foreverlms in #289
• fix: For pending weight sync cmds in rollout to do them all at once. by @lfengad in #291
• fix: more cases compatible for the RLpayload format update by @lfengad in #293
• fix: Refine command fetch filter in validation case. by @lfengad in #295
• fix: Relax packing sequence test check by @lfengad in #299
• Support cp for VLMs by @kane-vln in #296
• Activation offload in policy by @foreverlms in #294
• Support gpt-oss based internvl of SFT by @kane-vln in #290
• Set p2r nccl group size default to 1. by @foreverlms in #308
• Move reward calculation to rollout worker by @lfengad in #292
• feat: custom sampler support by @lfengad in #309
• feat: custom batch sampler support and batch data loader in RL. by @lfengad in #314
• fix: add warning log for dataloader batch setting by @lfengad in #316
• [Feat] Implement GSPO by @Bin-NV in #306
• Optimize hfmodel loading by @kane-vln in #323
• Support qwen3_vl_moe sft by @kane-vln in #315
• Supporting pure text prompt for qwen2.5-vl by @zekunhao1995 in #326
• Fix HF_HOME model path by @Dinghow in #328
• Adjust samples generation for on-policy. by @foreverlms in #324
• Cache model downloads in CI by @bddppq in #332
• feat: update reference and optimizers periodically by @lfengad in #327
• feat: More metrics collection during RL by @lfengad in #330
• feat: lora alpha pattern by @xlu451 in #331
• Fix: qwen3_vl_moe data packer renaming by @kane-vln in #333
• HF way support of Qwen3-VL. by @foreverlms in #329
• fix weight version calc in on-policy by @foreverlms in #339
• feat: Refine step logic with optimization batch control by @lfengad in #338
• Disable dual streams in act-offloading by @foreverlms in #341
• feat: Reduce peak memory in P2R check process by @lfengad in #340
• Fix CI failure by @foreverlms in #344
• fix: fix unmatched token numbers in calculating token-mean loss by @lfengad in #342

New Contributors

• @Bin-NV made their first contribution in #306
• @zekunhao1995 made their first contribution in #326

Full Changelog: v0.3.3...v0.3.4

v0.3.3

What's Changed

• Fix DDP for SFTTrainer by @kane-vln in #239
• Unify validation config by @Dinghow in #255
• Support of passing custom arguments to custom dataset script. by @foreverlms in #253
• fix: heartbeat opt for heavy cpu case. by @lfengad in #257
• Refactor: reset named_buffers in load_hf_weights by @kane-vln in #252
• fix:Remove transformer-engine dependence in requirement by @lfengad in #258
• [cleanup] Remove the old Deepseek-V3 implemention by @heslami in #260
• fix: reward filter in dynamic sampling + rollout outdate pause generation by @lfengad in #263
• [Fix] Deepseek V3 GRPO bug fix by @heslami in #259
• feat: Support epoch level save frequency by save_freq_in_epoch by @lfengad in #264
• InternVL sft support by @kane-vln in #254
• Revert "Support of passing custom arguments to custom dataset script.… by @gekurian in #267
• Fix the bug introduced in PR #252 by @gekurian in #268
• Support of passing custom arguments to custom dataset script by @foreverlms in #275
• Fallback to hfmodel pass if build model fails by @kane-vln in #276
• Fix: sync named_buffer for hfmodel in grpo mode by @kane-vln in #277
• fix: cpu intensive situation aware for controller and reward by @lfengad in #279
• fix: Lepton cross-node job sync for host preparation before start. by @lfengad in #278
• feat: SFT validation dataset and packer specification support by @lfengad in #281

Full Changelog: v0.3.2...v0.3.3

v0.3.2

What's Changed

• Use AutoModel in case architecture doesn't exist by @kane-vln in #236
• [1/n] Support Deepseek V3 SFT by @heslami in #190
• [3/n] Support Deepseek V3 GRPO / Deepseek R1 by @heslami in #240
• fix: outdated import from leptonai by @xlu451 in #243
• trtllm-pytorch as the rollout backend. by @foreverlms in #161
• feat: lora for grpo by @xlu451 in #222
• Fix: sync_model_vocab corner case by @kane-vln in #242
• feat: Only do for trainable params in weight sync. by @lfengad in #238
• feat: custom logger support by @lfengad in #245
• fix: Refine logger to only specified in data packer script. by @lfengad in #246
• feat: data type control in weight transfer. by @lfengad in #247
• fix: Remove cosmos-rl dependency in launch_all.py in normal mode. by @lfengad in #248
• feat: sequence packing in training for optimization by @lfengad in #211
• fix: min_filter_prefix_tokens corner case by @jcao-ai in #251

New Contributors

• @heslami made their first contribution in #190

Full Changelog: v0.3.1...v0.3.2

v0.3.1

What's Changed

• Optimization: Quick dataloader advancing after checkpoint resumed by @jcao-ai in #234
• fix: GRPO ckpt resume by @jcao-ai in #235

Full Changelog: v0.3.0...v0.3.1

v0.3.0

What's Changed

• fix: Refine slurm script by @lfengad in #223
• fix: hostname setting problem for lepton case when cross node replica by @lfengad in #224
• refactor hfmodel gradient_checkpoint by @kane-vln in #229
• Liger kernel integration by @jcao-ai in #226
• LoRA For HFModel by @kane-vln in #230
• fix: step_** digit sorting during ckpt selection by @jcao-ai in #232

Full Changelog: v0.2.9...v0.3.0

v0.2.9

What's Changed

• [LoRA] Support modules_to_save by @jcao-ai in #213
• fix: exit issue when replica unregister to make total steps larger. by @lfengad in #214
• minor fix: only report val_score if validation is enabled by @jcao-ai in #215
• fix: use kaiming init for lora by @jcao-ai in #216
• chore: pass use_rslora config by @xlu451 in #217
• Refactor activation checkpointing by @jcao-ai in #218
• HF VLMs Support by @kane-vln in #165
• feat: add deterministic feature by @jcao-ai in #220
• fix: slurm launch code root found issue. by @lfengad in #221

Full Changelog: v0.2.8...v0.2.9