Dev Roadmap 2026 Q3

[SolitaryThinker]

Inference

• Add DGX Spark support [infra] Add DGX Spark and multi-architecture CUDA support #1447
• Streamline and improve documentation for DGX Spark
• Improve nvfp4/fp8 inference API and example scripts @kevin314 [misc] Clean up QAD 5090 example inference scripts #1496
• batching support [feat] Add opt-in dynamic generation batching #1453 @macthecadillac
• device specific optimal example scripts (eg 5090, 3090, GB200, etc)
• UI frontend for inference

Model support

Tracking issue: #1407

Image Models

• Flux 1 [feat] Fix FLUX.1-dev port: native RoPE, parity tests, SSIM reference #1321
• Flux 2 [model] Flux2 Klein Port #1349
• GLM-Image [feat] Add GLM-Image inference support #1030

Video Models

• Kandinsky-5 T2V/I2V @leffff [feat] Add Kandinsky-5 T2V/I2V pipeline support #1471
• Cosmos3 @SolitaryThinker
• Add Lingbot-Video @Davids048 [new-model] Add LingBot-Video TI2V inference #1600

AR models

• DreamX-world 5B @Suckl [feat] Add DreamX-World 5B Cam and AR pipelines #1538
• Add Lingbot-World @Davids048 [new-model] Port LingBot-World-v2 #1579
• Stable-Video-Infinity @H1yori233 [feat]: Stable-Video-Infinity inference on Wan 2.1 I2V 14B 480P #1344

Dreamverse

• Dreamverse i2v Support @SolitaryThinker
• Support Cosmos 2.5 Predict 2B
• Add 5090 support

Training

• Add Muon optimizer @alexzms [feat] train: add Muon optimizer (FSDP2/DTensor-aware) alongside AdamW #1519
• Add Kandinsky FT @leffff
• Add LTX2.3 FT @SolitaryThinker
• webui frontend for FT

Distillation

• anyflow [feat] Add AnyFlow any-step video distillation (pretrain + on-policy) #1371
• Teacher forcing and causal consistency distillation @H1yori233 [feat] Add Clean-history Teacher Forcing and Causal Consistency Distillation #1505
• ReROPE @H1yori233 [feat] Relativistic RoPE re-indexing for long causal rollouts #1454
• Add TDM distillation https://arxiv.org/pdf/2503.06674
• i2v distillation @alexzms @Satyam-53
• kandinsky QAD support @leffff [feat] Add Kandinsky5 QAD training pipeline: data preprocessing, QAT finetune, QAT-aware DMD distillation #1601
• LTX2.3 QAD support @SolitaryThinker
• Optimize QAD MFU (in progress)

RL

• Add DiffusionNFT Wan Video RL training @Abecid [feat] Add DiffusionNFT Wan video RL training #1476
• PromptRL @Davids048

Kernels

• Flashinfer Attention backend @SolitaryThinker
• Optimize nvpf4 QAT kernel for DGX Spark
• Kernel fusion support for popular models
  • Wan
  • LTX
  • Cosmos
  • Kandinsky

CI / Perf Optimization

Tracking Issue: #1374

• Land Performance CI phase 1 @Satyam-53
• Baseline and coverage for Wan, LTX, and Cosmos model @Satyam-53
• Spin up the perf dashboard on a separate CPU instance @Satyam-53

[Satyam-53]

Performance Dashboard is now live.
More models and gpu support coming soon.