Reading the Kokoro checkpoint in PowerShell

src/runspace/Torch.Checkpoint.psm1 reads a torch.save v1.6+ checkpoint — a zip holding one pickle and raw storages — with no Python and no native dependency. Only the pickle opcodes torch emits are implemented; anything else throws rather than guessing.

Result

kokoro-v1_0.pth, 327,212,226 bytes, 510 zip entries:

tensors = 548        parse = ~1.9 s
bert=25  bert_encoder=2  predictor=122  decoder=375  text_encoder=24
total parameters = 81,763,410   (311.9 MB fp32)

Shapes resolve correctly, e.g.

predictor.module.F0.1.conv1.weight_v   float32 [256,512,3]
predictor.module.F0.1.conv1.weight_g   float32 [256,1,1]
predictor.module.F0.1.pool.weight_v    float32 [512,1,3]
predictor.module.F0.1.conv1x1.weight_v float32 [256,512,1]

Validation

Against the R009 specimen bundle, which was produced from this same checkpoint and has been executed correctly on V73. Fusing weight_norm as w = g * v / ||v|| over dim 0 and transposing the PyTorch (out, in, k) layout to the specimen's (1, k, in, out):

elements=393,216   maxAbsDiff=8.196E-008   over 1e-6 = 0

That is fp32 rounding. It validates storage offsets, element order, dtype, and the weight_norm fusion convention in one comparison.

A weaker check was tried first and discarded: ||weight_v|| does not equal weight_g, because after training both are free parameters and only the direction of v matters.

Notes for anyone extending it

Three PowerShell behaviours bit this implementation and are commented in the source: GetNewClosure() snapshots variables, so a cursor helper advanced its own copy; variables carry type constraints across a scope, so reusing $k/$v as pickle keys failed against earlier [int]/[long] declarations; and @(...) flattens, so the args tuple absorbed the nested size tuple until it was built through a List[object]. Variable names are also case-insensitive, so $O/$o are one variable.

What this unblocks

Weights no longer require Python. Static int8 quantization needs only per-output-channel ranges from these values, with no calibration. The same reader feeds a reference evaluator, which is what activation ranges and a relative accuracy gate both depend on.