<h2 id="t5xxl-google-flan-from-fp32-f2osen59s">T5xxl Google FLAN from FP32</h2>NOTE: It appears as when the decoder blocks are pruned from the model it looses the ability to interact with FLUX or SD 3.5 - When I merged the models with the full 42GB file this did not happen.Trimmed Sizes:<ol><li>FP32 18GB (42GB with Decoder Weights)</li><li>FP16/BF16 9GB</li><li>FP8 4.5GB (I do not recommend the FP8 version it seems to have lost to much precision)</li></ol>I created a <a target="_blank" rel="ugc" href="https://civarchive.com/articles/8570/a-simple-script-for-extracting-t5-or-t5-flan">tool</a> to extract the FLAN T5.<ul><li>I have a seed to seed <a target="_blank" rel="ugc" href="https://civarchive.com/articles/8629">image comparison</a> on FLAN T5</li><li>Speed increase may be from reduced VRAM Load and or reduced file size decreasing time to load in the case of CPU offloading.</li><li>All Models built of the FULL FP32 FLAN model</li><li>I have had issues with GGUF quantization using either the T5 header or T5 decoder header, and have been unable to quantize the model using GGUF</li></ul>Per the <a target="_blank" rel="ugc" href="https://www.apache.org/licenses/LICENSE-2.0">Apache 2.0 license</a> FLAN is attributed to Google

T5xxl Google FLAN from FP32 (CLIP ONLY)

Files with SHA256: 0815bcf3bf178dc3eca6d6f7c03dc96a68565f70b352b880434b9d7bed2061a4