FlashRT VLA Residual Gates

This package provides FlashRT joint residual/gate kernels for Hugging Face Kernel Hub.

It is intended for VLA/video model blocks that maintain separate video, action, and und token groups but update them in the same block.

Kernels

• bias_residual_bf16: fused residual + x + bias for one BF16 token segment.
• joint3_bias_gate_residual_bf16: fused video/action/und residual updates where video and action both use bias and gate.
• joint3_bias_gate_residual_action_nobias_bf16: fused video/action/und residual updates where video uses bias+gate, action uses gate, and und uses a plain residual add.
• joint3_bias_fp8_gate_residual_bf16: the same joint update with a statically scaled FP8 E4M3 video gate.
• joint3_bias_fp8_gate_residual_action_nobias_bf16: FP8 video gate with no action bias.

Hardware

• CUDA 12.8+
• BF16-capable NVIDIA GPUs

Current local source validation is on RTX 5090. Broader hardware rows should be added after installed-artifact validation.

Upstream

The serving source of truth remains FlashRT:

https://github.com/LiangSu8899/FlashRT