Support SDXL and its distributed inference

[Zars19]

The idea of patch parallelism comes from the CVPR 2024 paper Distrifusion. In order to reduce the difficulty of implementation, all communications in the example are synchronous.

This can help SDXL achieve better performance, especially when the resolution is very high

A100, 50 steps, 2048x2048, SDXL

Framework	sync_mode	n_gpu	latency(s)	speed_up	memory(MiB)
Torch	-	1	25.25	1x	42147
TRT	-	1	21.98	1.15x	42895
DistrFusion(Torch)	split_batch	2	13.33	1.89x	40173
Ours	split_batch	2	11.69	2.16x	42675
DistrFusion(Torch)	corrected_async_gn	4	8.27	3.05x	49087
DistrFusion(Torch)	full_sync	4	8.64	2.92x	51943
Ours	full_sync	4	7.73	3.27x	43073

[juney-nvidia]

@Zars19 thanks for the contribution to TensorRT-LLM!

@nv-guomingz can you help take care of this? :)

Thanks
June

[nv-guomingz]

@Zars19 thanks for the contribution to TensorRT-LLM!

@nv-guomingz can you help take care of this? :)

Thanks June

Sure, I'll collobrate with @Zars19 for enabling SDXL with TRT-LLM.

[nv-guomingz]

Hi @Zars19 , could u please resolve the code conflicts firstly?

[Zars19]

Hi @Zars19 , could u please resolve the code conflicts firstly?

I have resolved the conflict :) @nv-guomingz

[nv-guomingz]

Hi @Zars19 thanks for your patience.
Could u please update this MR by updating/rebasing those two commit(including one merge commit) into one commit which make us easy to integrate and testing?

[Zars19]

@nv-guomingz I completed the git rebase

[lmxyy]

Any updates on the code review?

[Zars19]

Any updates on the code review?

After rebasing the code, I haven't received feedback for a while now