bump version

[feat] REAL Loss (Rewards as Labels) for GRPO Training (#8424)

[bugfix] fix gemma4 template (#9026)

[bugfix] fix grpo moe model patch (#9027)

[bugfix] filter more grpo keys for model forward (#9025)

[megatron] support qwen3.5 CP (#9022)

[megatron] support multimodal MTP (#8390)

[model] fix gemma4 model_id (#9020)

[megatron] fix vit_attn_impl megatron (compat mcore-bridge) (#9019)

[bugfix] fix megatron lora_llm & pp (#9016)

Fix NPU device mismatch in Megatron DPO/RLHF training (#9014)

[bugfix] Fix mm token type ids (transformers 5.5.0 infer); update json.dumps ensure_ascii (#9015)

[bugfix] fix gkd bridge (#8510)

[model] support gemma4 (#8508)

[docs] update docs (#8504)