Search Results
6/22/2025, 6:47:57 PM
>>105672284
for complex prompts and styles, the ROI is worth it. chroma can do in a single 60s gen what SDXL and SD3.5 models can't do in six 10s gens.
>>105672309
no, sadly they didn't explain how.
>you can enable the torch one with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
I use this command:
>TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention --bf16-vae
I have no idea if I'm getting a perf boost from this.
for complex prompts and styles, the ROI is worth it. chroma can do in a single 60s gen what SDXL and SD3.5 models can't do in six 10s gens.
>>105672309
no, sadly they didn't explain how.
>you can enable the torch one with TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
I use this command:
>TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention --bf16-vae
I have no idea if I'm getting a perf boost from this.
Page 1