Add kolors compile

[lixiang007666]

This PR is done:

• Add kolors compile&readme.

[lixiang007666]

当前支持的 kolors diffusers 在：
huggingface/diffusers#8812

目前还没有合并。

[lixiang007666]

动态 shape 日志：

oneflow：

oneflow backend compile...
Starting warmup...
Warmup complete.
Warmup time: 39.30 seconds
Generated image saved to kolors_oneflow_compile.png in 3.50 seconds.
Max used CUDA memory : 20.627GiB
Test run with multiple resolutions...
Running at resolution: 1024x1024
Inference time: 4.17 seconds
Running at resolution: 1024x768
Inference time: 2.86 seconds
Running at resolution: 1024x576
Inference time: 2.40 seconds
Running at resolution: 1024x512
Inference time: 2.20 seconds
Running at resolution: 1024x256
Inference time: 1.30 seconds
Running at resolution: 768x1024
Inference time: 2.89 seconds
Running at resolution: 768x768
Inference time: 2.42 seconds
Running at resolution: 768x576
Inference time: 2.10 seconds
Running at resolution: 768x512
Inference time: 1.68 seconds
Running at resolution: 768x256
Inference time: 1.03 seconds
Running at resolution: 576x1024
Inference time: 2.43 seconds
Running at resolution: 576x768
Inference time: 2.10 seconds
Running at resolution: 576x576
Inference time: 1.53 seconds
Running at resolution: 576x512
Inference time: 1.48 seconds
Running at resolution: 576x256
Inference time: 0.93 seconds
Running at resolution: 512x1024
Inference time: 2.22 seconds
Running at resolution: 512x768
Inference time: 1.72 seconds
Running at resolution: 512x576
Inference time: 1.47 seconds
Running at resolution: 512x512
Inference time: 1.34 seconds
Running at resolution: 512x256
Inference time: 0.86 seconds
Running at resolution: 256x1024
Inference time: 1.33 seconds
Running at resolution: 256x768
Inference time: 1.04 seconds
Running at resolution: 256x576
Inference time: 0.93 seconds
Running at resolution: 256x512
Inference time: 0.86 seconds
Running at resolution: 256x256
Inference time: 0.78 seconds

nexfort：

nexfort backend compile...
Starting warmup...
Warmup complete.
Warmup time: 314.58 seconds
Generated image saved to kolors_nexfort_compile.png in 2.31 seconds.
Max used CUDA memory : 19.435GiB
Test run with multiple resolutions...
Running at resolution: 1024x1024
Inference time: 5.95 seconds
Running at resolution: 1024x768
Inference time: 4.04 seconds
Running at resolution: 1024x576
Inference time: 3.48 seconds
Running at resolution: 1024x512
Inference time: 3.13 seconds
Running at resolution: 1024x256
Inference time: 2.39 seconds
Running at resolution: 768x1024
Inference time: 4.09 seconds
Running at resolution: 768x768
Inference time: 3.50 seconds
Running at resolution: 768x576
Inference time: 3.07 seconds
Running at resolution: 768x512
Inference time: 2.45 seconds
Running at resolution: 768x256
Inference time: 2.40 seconds
Running at resolution: 576x1024
Inference time: 3.57 seconds
Running at resolution: 576x768
Inference time: 3.07 seconds
Running at resolution: 576x576
Inference time: 2.49 seconds
Running at resolution: 576x512
Inference time: 2.48 seconds
Running at resolution: 576x256
Inference time: 2.46 seconds
Running at resolution: 512x1024
Inference time: 3.18 seconds
Running at resolution: 512x768
Inference time: 2.51 seconds
Running at resolution: 512x576
Inference time: 2.49 seconds
Running at resolution: 512x512
Inference time: 2.49 seconds
Running at resolution: 512x256
Inference time: 2.46 seconds
Running at resolution: 256x1024
Inference time: 2.49 seconds
Running at resolution: 256x768
Inference time: 2.50 seconds
Running at resolution: 256x576
Inference time: 2.48 seconds
Running at resolution: 256x512
Inference time: 2.47 seconds
Running at resolution: 256x256
Inference time: 2.49 seconds

[lixiang007666]

发现 nexfort 最近的更新有些优化导致速度变慢了很多，我定位下是哪个提交。

[lixiang007666]

发现 nexfort 最近的更新有些优化导致速度变慢了很多，我定位下是哪个提交。

pipe = compile_pipe(
pipe, backend="nexfort", options=options, ignores=['text_encoder'], fuse_qkv_projections=True
)

加了 ignores=['text_encoder'] 后，性能变差很多，不太合理。

于是定位到 https://github.com/siliconflow/nexfort/pull/91 ，发现这个 commit 之前不会出现这种情况。

不过这个问题对这个 PR 的合并没影响。

[lixiang007666]

TODO:

• a100 测速。
• 质量报告。

[strint]

加了 ignores=['text_encoder'] 后，性能变差很多，不太合理。

因为 kolors 用的 chatglm，印象中模型会比之前 sdxl 的 text encoder 大，这块开销应该是明显的