yqzhishen
5c14181606
Pre-release audit fixes
C1: Fix max_beta config value silently ignored in DDPM linear schedule
H1: Fix load_ckpt regression when key_in_ckpt=None with prefix_in_ckpt set
H3: Update base.yaml to reflect defaults
H4: Fix diff_accelerator doc default (dpm-solver -> ddim)
M2: Fix in-place mutation in SwiGLU (gate *= ratio -> gate = gate * ratio)
M3: ChainedOptimizer init now raises ValueError for unmatched params
M5: Update stale 'PyTorch 1.13 temporary workaround' comments
M6: Align variance backbone ONNX tracing to use consistent float dtype
M8: Fix schema doc defaults (predict_tension/predict_voicing -> false, spec_min -> [-12])
M9: Replace bare except with except (IndexError, TypeError)
M10: Replace ambiguous condition cache comments with clear explanations
L1: Replace unused loop variable with underscore
L3: Fix mutable default arguments in muon.py
L5: Fix 'optimzier' typo in chained_optimizer.py
L6: Remove unused Iterable import
L7: Optimize stretch embed GRU flatten_parameters (once per forward, not every call)
L8: Add comments explaining variance scaling factor values
L10: Replace np.sqrt with math.sqrt in attention scaling
L13: Fix 'dynamix_axes' typo in acoustic_exporter.py
L14: Update onnx_helper.py docstring (onnxsim -> onnxslim)
P8: Fix f-strings without interpolated variables
Misc: Remove unused XavierUniformInitLinear import from modules/toplevel.py