Spaces:
Running on Zero
Running on Zero
Commit History
Today
-
perf: overlap frame processing with GPU inference Running
Verified signatureDecode and preprocess the next frame, and finish the previous one, while the GPU runs the current frame. Only the GPU step depends on the temporal prior, so every computed value is unchanged. Load the compiled FP16 package once per GPU worker. ZeroGPU runs each request in a fresh copy of the worker context, so the ContextVar cache reloaded the package on every request. Log the package load time. Validated with 28 tests and lint/format checks. Local eager 692-frame matting: 79.6s -> 71.8s with byte-identical mattes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>2ed6676
-
perf: enable compiled FP16 inference on ZeroGPU
Unverified signatureCompile for the allocated GPU at startup, validate consecutive alpha mattes, and reuse a self-contained AOT package in request workers. Retain eager FP16 when a ZeroGPU compilation worker fails and expose a startup rollback variable. Embed compiler constants and release integer buffers safely to handle this model's incompatibilities with the pinned Spaces AOT helpers. Keep packages private at runtime and log complete request stages. Validated with 26 tests, lint and format checks, full 692-frame GPU comparison, fresh-process loading and exact temporal reset. Local adapter inference: 55.1 seconds; alpha MAE: 0.000008839.aa831fd