[chore] add _repeated_blocks for regional compilation support

[RuixiangMa]

Purpose

Add _repeated_blocks attribute to diffusion transformer models to enable torch.compile regional compilation optimization and fixes the warning "Regional compilation skipped because the model does not define _repeated_blocks "

Test Plan

Test Result

NVIDIA 4090(48G)

curl -X POST http://localhost:8004/v1/images/generations   -H "Content-Type: application/json"   -d '{
    "prompt": "a majestic dragon perched on the mountain ridge of Vermont, misty morning atmosphere, photorealistic style",
    "size": "1024x1024",
    "num_inference_steps": 50,
    "cfg_scale": 4.0,
    "guidance_scale": 0.5,
    "seed": 42
  }' | jq -r '.data[0].b64_json' | base64 -d > dragon.png

Same quality, latency shown only！

Metric	NO _repeated_blocks	_repeated_blocks
Longcat-image	36 s/img	29 s/img
GLM-Image	60 s/img	58 s/img

Current tests cover only glm-image and longcat; several existing models remain unadded—may be necessary to include.

[hsliuustc0106]

any more models can benefit from the regional compilation support?

[hsliuustc0106]

lgtm

[RuixiangMa]

any more models can benefit from the regional compilation support?

compile once, reuse across all identical layers, which is introduced in PR #684， I think it is benefits all DiT models with repeated Transformer blocks.

[hsliuustc0106]

@wtomin do we have a doc recording this feature?

[wtomin]

@wtomin do we have a doc recording this feature?

https://github.com/vllm-project/vllm-omni/blob/main/docs/contributing/model/adding_diffusion_model.md#torchcompile-support.

It is brief but explains how to set _repeated_blocks.

[wtomin]

LGTM.