I build and review research software where model quality depends on reliable training, inference, evaluation, and control code.
|
🤖 VLA & robotics PI0/PI0.5, flow matching, policy learning, simulation, and control |
🌍 World models Video prediction, planning rollouts, guidance, and model-based RL |
🎨 Diffusion Generative modeling, distillation, quantization, and video systems |
⚡ LLM systems Transformers, distributed inference, retrieval, and evaluation |
Recent merged work across learning algorithms, numerical stability, inference efficiency, and simulation infrastructure.
View all 27 merged projects
| Area | Project / PR | Contribution focus |
|---|---|---|
| VLA | DeepThinkVLA #9 | Reasoning and action-generation path |
| VLA | FluxVLA #108 | PI0/PI0.5 action-horizon and attention alignment |
| VLA | OpenTau #547 | VLA training and evaluation updates |
| VLA / flow matching | tau-0-vla #13 | Beta time sampling and device-safe flow statistics |
| WAM / diffusion | WAM-Flow #20 | Variable-length text training and stage-2 path |
| VLA / policy | pi-zero-pytorch #27 | Batched RTC guidance-weight handling |
| Diffusion | stable-diffusion.cpp #1936 | Sampler and inference improvements |
| Diffusion / quantization | LightX2V #1474 | Symmetric FP8/INT8 Triton quantization |
| Diffusion | denoising-diffusion-pytorch #369 | Forward XM classifier-free guidance compatibility |
| Vision / distillation | vit-pytorch #372 | Numerically stable DINO loss in low precision |
| Generative modeling | vector-quantize-pytorch #254 | Masked residual beam-loss aggregation |
| World model | BiWM #13 | Wan DMD guidance-anchor correction |
| World model | nano-world-model #20 | Multi-frame planning-rollout alignment |
| World model / RL | stable-worldmodel #322 | TD-MPC2 scaling frozen during validation |
| Robotics / geometry | pypose #408 | Autograd-safe Lie-group quaternion normalization |
| RL / environments | Gymnasium #1682 | Deterministic snapshots for mutable observations |
| Simulation | genesis-world #3284 | Embodied simulation infrastructure |
| Physics | chrono #821 | Physics and simulation infrastructure |
| Distributed inference | xDiT #774 | Distributed diffusion inference path |
| Distributed training | Psi0 #103 | Token-mixture synchronization across ranks |
| Transformers | x-transformers #357 | Self-attention KV-cache reordering |
| Retrieval | LightRAG #3782 | Retrieval and graph pipeline improvements |
| LLM education | LLMs-from-scratch #1077 | Training and implementation updates |
| World models | awesome-llm-apps #1134 | AI application and model references |
| AI systems | ai-agent-book #1043 | Agent and LLM application references |
| LLM systems | helm-www #2237 | Evaluation system improvements |
| Diffusion / transformer | sdnext #5076 | Diffusion UI and backend integration |
- Correctness first — make edge cases explicit and testable.
- Reproducibility — keep randomness, distributed behavior, and evaluation stable.
- Efficiency — account for memory, precision, and throughput in every implementation.
All listed contributions link to merged upstream pull requests authored by zjn20030811.