Skip to content

v0.2.2

Latest

Choose a tag to compare

@Zachary-wW Zachary-wW released this 05 Nov 08:32
· 11 commits to master since this release
Immutable release. Only release title and notes can be modified.
042a681

Highlights

  • New embodied models and recipes: LoongForge-VLDA, LingBot-VLA-v2, GR00T-N1.7, and Wall-OSS-0.5, plus updated Pi0.5, xVLA, FastWAM, and DreamZero training and evaluation paths.
  • New LLM / VLM support: Kimi-K3 for both LLMs and VLMs, GLM-5.3-flash, and Qwen3.8-27B.
  • More capable data pipeline: Ego2Robot converts first-person human manipulation videos into LeRobot v3.0 data for 16 robot morphologies.
  • Training-system improvements: tensor-wise Muon optimization, FP8 gradient communication, TorchAO FP8 FlexAttention, and variable-length ChunkPipe.
  • Production readiness: stronger async/DCP checkpoint handling, broader GPU regression, sensitive-data scanning, and an operator-controlled release workflow.

What's Changed

Embodied models and training

  • Added LoongForge-VLDA with two-stage training, checkpointing, and closed-loop LIBERO evaluation. The released recipe reports 98.5% on LIBERO and 84.5% zero-shot on LIBERO-Plus. See the cookbook and #285.
  • Added LingBot-VLA-v2 integration on a tensor-wise optimizer with per-parameter precision, overlapped publish, and Muon support (#211), with validated RoboTwin / LIBERO defaults (#306).
  • Completed GR00T-N1.7 inference, LIBERO / SimplerEnv training and evaluation recipes, custom operators, and CUDA-graph performance work (#178, #191, #195, #245, #300).
  • Added Wall-OSS-0.5 training with FSDP2 + DMuon (#192, #217) and fused RoPE / RMSNorm / SwiGLU / MoE operators (#195).
  • Expanded Pi0.5 with validated LIBERO / RoboTwin fine-tuning and evaluation recipes (#313), FlexAttention and activation-checkpoint optimizations (#179), and TorchAO FP8 FlexAttention support (#248). The published Pi0.5 recipe reaches 3.00x training speedup.
  • Expanded FastWAM with inference / evaluation integration (#222), fused optimizer and compiled-block acceleration (#171), multi-dataset training (#269), and a LIBERO evaluation protocol aligned with the official harness (#274, #279, #319).
  • Expanded DreamZero with inference and evaluation integration (#188), an optimized Wan2.2-5B FSDP recipe (#213), Delta-FP8 parameter AllGather (#220), and a compiled Wan2.1-14B FSDP recipe (#323).
  • Added FlashAttention-2 support for xVLA (#177).

LLMs and VLMs

  • Added Kimi-K3 LLM and VLM training and checkpoint conversion (#215).
  • Added training support for GLM-5.3-flash (#230) and Qwen3.8-27B (#176).
  • Enabled fused DSA and ChunkPipe on A-series GPUs (#233) and variable-length chunks in ChunkPipe (#273).
  • Added FP8 linear conversion with Transformer Engine and TorchAO (#218), plus an FP8 DDP gradient communication hook with error feedback and selective-precision exemptions (#223, #231).
  • Fixed MLA recompute isolation for non-MLA components (#266) and V-dimension padding for MLA FlashAttention (#309).

Data, checkpoints, and tooling

  • Added the Ego2Robot pipeline for first-person video-to-robot data conversion across 16 robot morphologies, outputting LeRobot v3.0 (#229, #237).
  • Improved LeRobot-v2 decoding for multi-frame observations (#262) and FastWAM image-padding / dataset defaults (#240).
  • Hardened checkpoint resume and save paths: early async checkpoint markers and DCP resume (#261), isolated async-checkpoint Gloo collectives (#283), alias/tied-tensor deduplication (#317), and asynchronous ZeRO optimizer-state saving (#320).
  • Fixed Hugging Face safetensors shard loading to preserve filename zero-padding (#326).
  • Restructured data and checkpoint tools into self-contained directories (#263) and moved dependency patches under third_party/ (#290).

Reliability, docs, and release engineering

  • Rebuilt CI around static checks, sensitive-data scanning, GPU regression, stale-run invalidation, and an operator-controlled release workflow (#151, #181, #182, #275, #291).
  • Added Docker and DreamZero quick starts, reorganized embodied tutorials, refreshed the benchmark and project pages, and added AtomGit mirror links.
  • Reduced the repo-wide Ruff backlog from 1,055 findings to zero and added CI checks for workflow YAML, documentation links, licensing, secrets, and image policy.

Container image

The v0.2.2 NVIDIA GPU image is available on Docker Hub:

docker pull loongforge/loongforge:0.2.2

For reproducible deployments, use the versioned tag. The image includes the unified LLM / VLM / embodied / diffusion runtime.

New Contributors

Full Changelog: v0.2.0...v0.2.2