v0.0.5

• added new repeatable --override-config "inference.batch_size=2" in CLI for quick experimentation of config changes
• in cases where you only want one stem (e.g. piano-only in bs_roformer-fruit-sw), use the config inference.requested_stems=["piano"]. in this case, the unrelated mask estimators in the state dict are surgically removed so it never ends up inferencing unrelated stems. as a fallback, models that do not implement a StemSelectionPlan will use index_select effectively "delete" unused stems
• By default, audio read/chunking/stitching/normalisation runs on the GPU. To reduce memory pressure, you can force the non-model components on your host device (CPU) with inference.io_device="cpu". This breaks bit-for-bit equivalence in the final output, but has virtually zero difference.
• the --cpu flag in splifft run is removed, and replaced with inference.model_device="cpu"

🔪 Breaking Changes:

• InferenceEngine now also supports construction with a Python dictionary so it was renamed to config:

  engine = InferenceEngine.from_pretrained(
  -   config_path="/path/to/config.json",
  +   config="/path/to/config.json",  # or a python dict containing the config
      checkpoint_path="/path/to/checkpoint.pt",
  )

• On Windows the list of models is now %LOCALAPPDATA%\splifft\Cache\{model_id}\config.json
• Config has been updated. Run splifft pull {{model_id}} --force-overwrite-config for the following:
  • bs_roformer-fruit-sw: default batch_size from 4 to 8 and force_weights_dtype to float16
  • mdx23c-aufr33-drumsep: default batch_size from 1 to 8

Full Changelog: v0.0.4...v0.0.5