compile_unet error for `StableDiffusionLatentUpscalePipeline`

[kunc01]

Hi! Thanks for your amazing work.

I've tested on a V100 GPU, stable-fast works perfect for StableDiffusionControlNetInpaintPipeline, but for latent-upscaler and corresponding StableDiffusionLatentUpscalePipeline, sfast.compilers.diffusion_pipeline_compiler.compile_unet leads to

, while vae and text_encoder still work well.

I've tried options like enable_xformers, enable_triton, enable_cuda_graph, prefer_lowp_gemm and enable_fused_linear_geglu, but the error persists.

Looking forward to your reply, thank you!

[chengzeyi]

@kunc01 This problem has been fixed in the latest nightly release. Please try it.

[kunc01]

@chengzeyi Thanks for your reply，I tried 1.0.2.dev20240127+torch211cu118 on a V100 and 1.0.2.dev20240127+torch211cu121 on a A10，but found new errors.

[chengzeyi]

@chengzeyi Thanks for your reply，I tried 1.0.2.dev20240127+torch211cu118 on a V100 and 1.0.2.dev20240127+torch211cu121 on a A10，but found new errors.

Just fixed it. Please retry.

[kunc01]

@chengzeyi

[chengzeyi]

@chengzeyi

I guess this is related to Triton. can you try running with disabling Triton? Or you can share a minimal reproducing script so that I can reproduce it quickly.

[kunc01]

@chengzeyi Inference succeeded with enable_triton = False

upsampler = StableDiffusionLatentUpscalePipeline.from_pretrained("stabilityai/sd-x2-latent-upscaler", torch_dtype=torch.float16).to("cuda") config = CompilationConfig.Default() config.enable_xformers = True config.enable_triton = False config.enable_cuda_graph = True upsampler = compile(upsampler, config)