feat(yolo): add YOLO26 models

[jamjamjon]

No description provided.

[darkSuperman]

It seems YOLOE-26 isn't included? Could you add it? Thank

[jamjamjon]

@darkSuperman
Recently I’ve had limited time. I need to optimize the cudarc code first, and then review the Ultralytics code to see whether it’s possible to export an ONNX model architecture similar to YOLOE. As a result, support for YOLOE-26 will be added a bit later.

[darkSuperman]

@jamjamjon Uh oh, it seems that exporting to YOLOE-26 turns it into a regular YOLO-26 model and no longer supports prompts. ultralytics/ultralytics#23250 (comment)

[jamjamjon]

@darkSuperman For algorithms like YOLOE and YOLO-World, to improve deployment portability, the text prompt is often directly fused into the model, turning it into a lightweight YOLO-seg model. This means the model loses its ability to accept text prompts.

In practice, there are generally two approaches:

1. Prompt-free: replace text prompts with 4,000+ predefined classes;
2. Prompt-based: select a specific set of target classes of interest and then export the model to ONNX.

In usls, YOLOE provides both approaches: prompt-free and prompt-based.
The prompt-based approach exports two ONNX models—text-encoder and vision-encoder (SAVPE)—to retain both text and vision prompt capabilities.

Therefore, support for YOLOE-26 still requires reviewing the corresponding implementation in detail.

[jamjamjon]

@darkSuperman Both prompt-free and prompt-based YOLOE-26 are now supported. Code refinements for improved performance are coming next.