v0.8.0

What's Changed

• fix(policies): respect skip_normalization_weights on checkpoint load by @shuheng-liu in #425
• Add Cosmos3-Super 64B action-generation benchmark (policy mode) by @WilliamYue37 in #426
• docs(models): document all policies on the models page by @shuheng-liu in #427
• feat(train): optional torch.compile of flow-matching module for pi05/pi07 by @WilliamYue37 in #428
• feat(deepspeed): validate ZeRO-3 for pi05 full fine-tune + ZeRO-2 benchmark by @WilliamYue37 in #430
• feat(robocasa): per-task official horizon for sim eval by @shuheng-liu in #431
• feat(eval): make eval simulation seeding reproducible and configurable by @shuheng-liu in #433
• Adding learnable embeddings for modality specific by @akshay18iitg in #432
• feat(policies): add cosmos3 (Qwen3-VL-32B reasoner + flow-matching expert) by @WilliamYue37 in #429
• feat(policies): optional knowledge_insulation + skip zero-weight loss terms by @shuheng-liu in #437
• feat(pi05_mem): RLDX-1 STSS motion module video encoder by @akshay18iitg in #438
• feat(grpc): optional x-tuner-api-key authentication by @cenxiaoxinxin-dotcom in #439
• docs(cosmos3): add tensor-flow ASCII diagram to model docs by @WilliamYue37 in #434
• chore: bump version to v0.8.0 by @WilliamYue37 in #440

New Contributors

• @cenxiaoxinxin-dotcom made their first contribution in #439

Full Changelog: v0.7.0...v0.8.0