amd-shark

/

sdxl-quant-int8

like 2

Follow

AMD SHARK 17

Updates the following:

• Add <weight|input>_zp_dtype to quant_param.json to differentiate between exported versions
• Update input/weight zero-points to be int8 (not uint8)
• Update the math model and tests to incorporate the above changes
• Remove SmoothQuant multipliers from layers that aren't quantized
• Upload new quant_param.json
• Upload new params.safetensors
• Upload new example output out.safetensors
• Confirm compliant FID of model (FID ∈ (23.0108, 23.9501)): 23.89
• Confirm compliant CLIP score of model (CLIP ∈ (31.686, 31.813)): 31.86

Strikethrough items were updated outside this PR.