Is your feature request related to a problem? Please describe.
Diffusers cannot load or run AMD Quark-quantized diffusion models through the standard from_pretrained APIs Today the only way to use a Quark-quantized UNet/transformer with Diffusers is to monkey-patch Quark into the auto-quantizer mappings at runtime, which is fragile and breaks when Diffusers internals change.
Describe the solution you'd like.
We would like to add Quark to AUTO_QUANTIZER_MAPPING (like bitsandbytes, GGUF, torchao, quanto, ModelOpt).
This would mean QuantizationMethod.QUARK, QuarkConfig, and QuarkDiffusersQuantizer would be registered in AUTO_QUANTIZER_MAPPING / AUTO_QUANTIZATION_CONFIG_MAPPING. There would be two load paths -- prequantized reload as well as online quantization. This also adds the optional amd-quark dependancy.
Describe alternatives you've considered.
We have tried runtime monkey-patching from Quark-side (currently in place).
Additional context.
This mirrors https://huggingface.co/docs/transformers/quantization/quark (the existing Transformers Quark integration). Implementation is ready at #14077.
Is your feature request related to a problem? Please describe.
Diffusers cannot load or run AMD Quark-quantized diffusion models through the standard from_pretrained APIs Today the only way to use a Quark-quantized UNet/transformer with Diffusers is to monkey-patch Quark into the auto-quantizer mappings at runtime, which is fragile and breaks when Diffusers internals change.
Describe the solution you'd like.
We would like to add Quark to AUTO_QUANTIZER_MAPPING (like bitsandbytes, GGUF, torchao, quanto, ModelOpt).
This would mean QuantizationMethod.QUARK, QuarkConfig, and QuarkDiffusersQuantizer would be registered in AUTO_QUANTIZER_MAPPING / AUTO_QUANTIZATION_CONFIG_MAPPING. There would be two load paths -- prequantized reload as well as online quantization. This also adds the optional amd-quark dependancy.
Describe alternatives you've considered.
We have tried runtime monkey-patching from Quark-side (currently in place).
Additional context.
This mirrors https://huggingface.co/docs/transformers/quantization/quark (the existing Transformers Quark integration). Implementation is ready at #14077.