AOTInductor graphs (built with aokit) for the per-frame transformer block of apple/FastFrames-L, with FP8 GEMMs. No weights are included: the graphs bind each block's live weights at load time.
- Built for torch 2.13.0+cu130 on an RTX PRO 6000 (sm120), the ZeroGPU GPU. Other torch builds or GPUs need a rebuild.
- Two graphs per block (
ff_pre,ff_post), split at the KV-cache write; the attention length is the only dynamic shape (1 to 21 latent frames). Code:ff_aoti.pyin the FastFrames-L Live Space. - Per DiT pass vs eager FP8 on the build GPU: 152 -> 110 ms (cache slot 5), 219 -> 185 ms (slot 17); eager-vs-compiled feature cosine >= 0.998 over a 21-frame rollout.
- Downloads last month
- 17
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support