Release v0.1.4

What's Changed

Major Updates

• Added Universal Context Parallel and THD CP support, including routing replay, VLM handling, and updated recipes by @hijkzzz in #27
• Added standalone checkpoint evaluation with --eval.eval_only by @hijkzzz in adfada1
• Added independent retention policies for DCP and HF checkpoints by @hijkzzz in 6302bd6

Rollout & Training

• Exposed NCCL fallback for vLLM tensor-parallel all-reduce by @k21993 in #25
• Vectorized default REINFORCE returns by @morluto in #29
• Made Ray actor constructors synchronous by @andrewwhitecdw in #30
• Ensured ChatAgent.prompt_text is always a string by @andrewwhitecdw in #31
• Simplified ChatAgent prompt extraction by @hijkzzz in #32
• Let the kernel allocate router listening ports by @hijkzzz in #33
• Improved vLLM weight-update verification and KV-cache invalidation by @hijkzzz in #36
• Unified rollout sampling arguments across training and eval-only modes by @hijkzzz in 5023eee

Fixes

• Skipped redundant tied lm_head weight updates by @k21993 in #26
• Fixed exported HF dtype metadata for multimodal models by @hijkzzz in 54b0a87
• Fixed Quack and CUTLASS DSL compatibility and added build-time import checks by @hijkzzz in #34
• Fixed the FlashInfer cubin and Python version mismatch by @hijkzzz in #35
• Removed duplicate dependency-pin generation and redundant Docker installation steps by @hijkzzz in #37
• Corrected the --actor.entropy_coef help text by @k21993 in #24

New Contributors

• @morluto made their first contribution in #29
• @andrewwhitecdw made their first contribution in #30

Full Changelog: v0.1.3...v0.1.4