Ean Garvey

Probably the tuned LoRA configuration, I don't see this without the LoRA or without tuning. We probably have to make some tuning config changes, I'll narrow this down and disable...

It looks like this happens without LoRAs now, too... I've got a minimal reproducer with the following IR: [https://storage.googleapis.com/shark-public/ean/unet_lc/module_forward_dispatch_5.mlir](https://storage.googleapis.com/shark-public/ean/unet_lc/module_forward_dispatch_5.mlir) ``` %7 = linalg.matmul_transpose_b {compilation_info = #iree_codegen.compilation_info} ins(%3, %4 : tensor,...

@lum33n have you tried with a custom VAE? How are the results? Can we close this issue or does it need further attention?

Hey, please try after deleting the .vmfb -- if you still have problems let me know, I saw this a few times during testing

For ld.lld I got it working with `New-Item -ItemType SymbolicLink -Path "C:\\ROCm\5.5\bin\ld.lld" -Target "C:\\ROCm\5.5\bin\ld.lld.exe"` and `$env:Path += ';C:\AMD\ROCm\5.5\bin'` This should solve the issues posted above. This is just coming out...

several *hours*? this is a known driver issue that usually resolves in 8-20mins. Can you share more info about hardware/versions so we can try to recreate this? Thanks

This message happens to be the last one that prints before the generated image is returned. Does it ever generate an image? Also, would you mind trying another model to...

.ckpt should get picked up. If it's a custom VAE or LORA those go in the respective subfolders. If you want the weights finetuned for 768x768 from the "non-base" SD2.1...

FWIW this also fails with vector distribution explicitly disabled:, same dispatch: ``` :0: error: LLVM Translation failed for operation: builtin.unrealized_conversion_cast /home/eagarvey/SHARK/models/stabilityai_sdxl_turbo_1_64_1024x1024_fp16_gfx942/EulerDiscrete_unet_4_dispatches/configured_compiled_scheduled_unet_run_forward$async_dispatch_50.mlir:2:2: error: failed to translate the MLIR LLVM dialect to...

I was able to minimize further with the following torch IR, which includes a few operations on the norm input tensor that appear in the model: ``` module @module {...