Clarification on QAIRT / QNN Compilation for Older Hexagon Architectures ( like v73)

[Omar-afk801]

Hi @Qualcomm-AI-research.
So as the MobileWan artifacts get released, a few open source Android developers and enthusiasts are preparing to backport the model for older NPU generations, specifically the Hexagon v73 architecture family (covering devices like the Snapdragon 8 Gen 2 and mid-range chips with native INT4 support, paired with high-capacity 12GB RAM configurations, likely.)
Though I do have a couple of questions regarding the performance and compatibility of this model:

1- Does the model specifically have any sort of layers or architectural differences that are exclusive to the V75 and up Hexagon NPU architectures? Or can Developers compile it for older chips normally using QAIRT?

2- Given the recurrent chunking mechanism, what is the hard-minimum VRAM/RAM requirement for the compilation graph context when executing a standard 3-step generation pass on older mobile NPUs (especially older flagships or newer premium mid range chips)?

3- Would Neodragon, Qualcomm AI Research's previous video diffusion model, be a (theoretically) better alternative to MobileWan for older flagships or newer Mid-Range chips?

[habibian]

Hi @Omar-afk801 ,

1. No, its a typical DiT with standard transformer blocks, so no tricky operation / layer

2. We run the model under 10GB RAM

3. Both Neodragon and MobileWan should be runnable on older NPUs. MobileWan has a better quality at a higher computational cost so I think it pretty much depends on your application / goal. I suggest first trying out their pytorch models off-target to compare generation qualities. If Neodragon quality meets your expectations, it is a much cheaper / faster model