Ean Garvey
Ean Garvey
I tried a few different configurations, and found a potentially useful runtime error when using inlined weights with SDXL: ``` Assertion failed: !!(iree_hal_resource_is(base_value, &iree_hal_rocm_buffer_vtable)), file C:\V\iree\experimental\rocm\rocm_buffer.c, line 25 ``` That...
I was a bit confused by this as well, but this was for sure run with HIP driver. Will validate with cli
OK, so if I switch from my local build, configured with: ``` cmake -GNinja -B ../iree-build --log-level=VERBOSE -DIREE_BUILD_PYTHON_BINDINGS=ON -DIREE_BUILD_COMPILER=ON -DPython3_EXECUTABLE=C:\\V\SHARK-Turbine\turb.env\Scripts\python.exe -DCMAKE_BUILD_TYPE=Release -DIREE_HAL_DRIVER_VULKAN=ON -DIREE_HAL_DRIVER_CUDA=OFF -DIREE_EXTERNAL_HAL_DRIVERS="rocm" -DIREE_ENABLE_CPUINFO=ON -DIREE_HAL_DRIVER_ROCM=ON -DIREE_ENABLE_LLD=ON -DIREE_ENABLE_RUNTIME_TRACING=OFF -DIREE_ENABLE_ASSERTIONS=ON -DIREE_ENABLE_SPLIT_DWARF=ON...
Are the pip releases built with assertions disabled? It could explain this, if the driver is still casting the wrong pointer.
The provided IR with external weights has an attention op. This was unintended. This IR has the attention op decomposed: https://sharkpublic.blob.core.windows.net/sharkpublic/ean/vae_f32_num/stable_diffusion_xl_base_1_0_bs1_512x512_fp32_vae_decode_decomp.mlir
> [This](https://gist.github.com/hanhanW/3797206187b96304baa99cfb0226524e) is the script that I used to compare the results. I don't see numeric issues with bf16 demotion only. They are close with `(atol=0.1, rtol=0.05)` config. This should...
It would also be helpful to know which of these flags are dependent on vector distribution, or to have some useful error messages for any flags that are not mutually...