08888dcb3e
update id and resource uri support
c6c804c241
Document latent Gaussian noise injection in README.md
9ff22f3dd8
Implement latent Gaussian noise injection during training for DEQ stabilization
d13682243e
Slower epsilon decay
28d770c8a3
Document epsilon curriculum in README.md
7493946b93
Introduce dynamic epsilon curriculum progression starting at 0.1
8435490c6e
Document compilation flags in README.md
012d26834c
Default --compile to True on Linux, and add overriding --no-compile flag
61306c812d
Document Anderson Acceleration solver in README.md
6cc317d1e4
Implement MPS-compatible Tikhonov-regularized Anderson Acceleration solver for DEQ models
b1ec23b1a7
Document the project in README.md
19e019b15b
Add selectable core-type option, default to transformer, and enable TF32 matmul precision
be6d595626
Integrate pure PyTorch MambaBlock as core and add configurable Taylor-order backpropagation
a99bb1b728
Add orthogonal initialization, core spectral normalization, weight decay, and split core learning rates
c84c31152d
Add --n-fixed-core configurable depth parameter for the recurrent reasoning core
e9e78d07a2
Add --epsilon CLI parameter and log FP-Diff in train.py
eafd79f05a
Dynamically label print metric as FP-Iterations in DEQ mode
088a76919f
Update .python-version to 3.14
9191a353f7
Add numpy dependency
12d5ecb096
Support logging actual steps count in DEQ model mode
3844d6fd7e
DEQ tests
21ac79f0e2
Initialize project: Sandwich Universal Transformer with ACT in PyTorch
7a8019d4c1
Client: display calibre books added