<h2 id="these-are-not-int8-convrot.">These are not INT8 ConvRot.</h2><p></p><p>These models are INT8 Tensorwise. if you are on <span style="color:rgb(64, 192, 87)"><strong>AMD ROCm</strong></span>, these models will pack a real punch if you take the time to custom build ComfyUI with <a target="_blank" rel="ugc" href="https://github.com/ROCm/flash-attention">ROCm\FlashAttention</a> + AMD Triton / AITER... But they are expected to pack <span style="color:rgb(64, 192, 87)"><strong>a performance boost on Nvidia</strong></span> cards too.</p><div data-youtube-video><iframe width="640" height="480" allowfullscreen="true" autoplay="false" disablekbcontrols="false" enableiframeapi="false" endtime="0" ivloadpolicy="0" loop="false" modestbranding="false" origin playlist src="https://www.youtube.com/embed/sF-HF0d04kw" start="0"></iframe></div><h3 id="minimum-start-up-parameters-for-rocmflashattention-+-amd-triton:"><br />Minimum start up parameters for ROCm\FlashAttention + AMD Triton:</h3><p>IMPORTANT: Do not install flashattention offered by pip get official <a target="_blank" rel="ugc" href="https://github.com/ROCm/flash-attention">ROCm\flashattention from github.</a></p><pre><code>export FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE

python main.py \
  --use-flash-attention \
  --disable-xformers </code></pre><h2 id="benchmarks">Benchmarks</h2><p>You can find <a rel="ugc" href="https://huggingface.co/PuppetVision/sdxl_base_and_refiner_01_int8_comfyui/blob/main/benchmarks/benchmarks.csv">raw benchmarks here</a>.<br /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/50bec26a-a1d1-4a2b-9712-8addb48f214b/original=true/50bec26a-a1d1-4a2b-9712-8addb48f214b.jpeg" /><br /><br /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ebd84165-9b73-44ab-8a9d-3cffa2b234f8/original=true/ebd84165-9b73-44ab-8a9d-3cffa2b234f8.jpeg" /><br /></p>

SDXL Base and Refiner 01 - INT8

Files with SHA256: c685a974add4dce82596fa535126e469511d1e922a1ba684a0fdb94d7c0deab4