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