search

alibaba / BladeDISC

BladeDISC is an end-to-end DynamIc Shape Compiler project for machine learning workloads.
Apache License 2.0
791 stars 159 forks source link

Diffusion model benchmark #904

Open zzpmiracle opened 1 year ago

zzpmiracle commented 1 year ago

We have support diffusers in https://github.com/alibaba/BladeDISC/issues/867 . This issue tracks performance of all the diffuser pipelines. For the concern of performance, we use BlaDNN to tuning models during runtime. The following pipelines would be tested:

zzpmiracle commented 1 year ago
pipeine model img size unet-pytorch unet-disc e2e-pytorch e2e-disc
stableDiffusionPipeline stabilityai/stable-diffusion-2-1 768*768 119.81ms 60.08ms 6.21s 3.14s
StableDiffusionImg2ImgPipeline runwayml/stable-diffusion-v1-5 768*512 66.22ms 29.49ms 3.25s 1.30s
StableDiffusionDepth2ImgPipeline stabilityai/stable-diffusion-2-depth 480*640 44.33ms 19.89ms 2.08s 0.89s
StableDiffusionInpaintPipeline runwayml/stable-diffusion-inpainting 512*512 37.52ms 19.17ms 2.63s 1.10s
AltDiffusionPipeline BAAI/AltDiffusion 512*512 37.93ms 18.74ms 2.65s 1.08s
tanyokwok commented 1 year ago

How about the e2e latency run with TensorRT/AIT?

vivienfanghuagood commented 1 year ago

pipeine model img size unet-pytorch unet-disc e2e-pytorch e2e-disc stableDiffusionPipeline stabilityai/stable-diffusion-2-1 768768 119.81ms 60.08ms 6.21s 3.14s StableDiffusionImg2ImgPipeline runwayml/stable-diffusion-v1-5 768512 66.22ms 29.49ms 3.25s 1.30s StableDiffusionDepth2ImgPipeline stabilityai/stable-diffusion-2-depth 480640 44.33ms 19.89ms 2.08s 0.89s StableDiffusionInpaintPipeline runwayml/stable-diffusion-inpainting 512512 37.52ms 19.17ms 2.63s 1.10s AltDiffusionPipeline BAAI/AltDiffusion 512*512 37.93ms 18.74ms 2.65s 1.08s

Can you share the scripts to run these benchmarks?

bonlime commented 1 year ago

@zzpmiracle Where I can find scripts to run this benchmarks? i've tried using it on the latest diffusers version, and it can't trace the components...

zzpmiracle commented 1 year ago

@zzpmiracle Where I can find scripts to run this benchmarks? i've tried using it on the latest diffusers version, and it can't trace the components...

maybe we can use docker images in this zhihu article https://zhuanlan.zhihu.com/p/631461489

zwj536 commented 1 year ago

@zzpmiracle Where I can find scripts to run this benchmarks? i've tried using it on the latest diffusers version, and it can't trace the components...

maybe we can use docker images in this zhihu article https://zhuanlan.zhihu.com/p/631461489

@zzpmiracle hi, do you have an example of Hires.fix ?