Failed to execute cutlass gemm: internal error on turning.

[Ph0rk0z]

When you enable cuda graphs on turning, the compile fails. It succeeds on 3090 but that doesn't need the speedup. Is there something incompatible? I built from source.

[Ph0rk0z]

I found out that compiling the C extension for multiple architectures causes import errors.

[Ph0rk0z]

Source of the error is:

def _modify_model(
    m,
    enable_cnn_optimization=True,
    enable_fused_linear_geglu=True,
    prefer_lowp_gemm=True,
    enable_triton=False,
    enable_triton_reshape=False,
    enable_triton_layer_norm=False,
    memory_format=None,
):

inside: https://github.com/chengzeyi/stable-fast/blob/main/src/sfast/compilers/diffusion_pipeline_compiler.py

enable_fused_linear_geglu=True,

Geglu always enables even on unsupported architecture. I changed it to false and off I go.

I do wonder why SDnext is faster than comfy though.. maybe it's due to diffusers?