Is there an existing issue for this?
Contact Details
No response
What should this feature add?
Flashpack is a high-throughput Diffusers-compatible weight storage format. It improves load times by between 3x and 20x versus Safetensors. It currently supports quantized Diffusers models and may add a single-file offline distribution option in the future (fal-ai/flashpack#32).
You may wish to consider defaulting to Flashpack in version 7 and providing Flashpack versions of the various Starter Models, with SDNQ for quantization as required. This would potentially improve Invoke's time-to-first-generation speed considerably, particularly on lower-end systems where such enhancements are most important.
Alternatives
No response
Additional Content
No response
Is there an existing issue for this?
Contact Details
No response
What should this feature add?
Flashpack is a high-throughput Diffusers-compatible weight storage format. It improves load times by between 3x and 20x versus Safetensors. It currently supports quantized Diffusers models and may add a single-file offline distribution option in the future (fal-ai/flashpack#32).
You may wish to consider defaulting to Flashpack in version 7 and providing Flashpack versions of the various Starter Models, with SDNQ for quantization as required. This would potentially improve Invoke's time-to-first-generation speed considerably, particularly on lower-end systems where such enhancements are most important.
Alternatives
No response
Additional Content
No response