Can TurboDiffusion be applied to Autoregressive generation in a training-free manner?

Great work on the project. I am interested in applying your acceleration techniques (SageSLA/SageAttention) to Autoregressive (AR) video generation models.

I have two specific questions regarding the adaptation:

1. Training-free vs. Fine-tuning Can the current SageSLA or SageAttention directly replace the attention mechanism in a pre-trained AR model without any training? Or is fine-tuning mandatory (e.g., to align the Linear Attention branch or adapt to the quantization)?

2. KV Cache Support For AR decoding, KV Cache is essential. Does your current INT8 implementation support INT8 KV Cache management to save memory during step-by-step generation?

I noticed in the roadmap/documentation that there are plans to extend TurboDiffusion to Autoregressive (AR) models. Could you share a rough estimate of when we might expect AR support to land?

Start with the project roadmap/documentation and the existing SageSLA/SageAttention INT8 implementation. Investigate whether the current attention replacements and KV-cache handling cover autoregressive decoding, then document the training requirements, INT8 KV-cache support, and expected scope or timing for AR support.