f32 wrapper

LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False

LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False

RMSNorm as in the microGPT script: normalize by root mean square, no learned parameters.

RMSNorm as in the microGPT script: normalize by root mean square, no learned parameters.

Guard that ensures trace files are written when dropped. This is an opaque wrapper around the optional ChromeLayer guard.

MLP classifier for authorization decisions with power-of-two dimensions. Architecture follows article classification pattern: Input -> Li

Transformer block: RMSNorm + Attention + RMSNorm + MLP, with residual connections.

Transformer block: RMSNorm + Attention + RMSNorm + MLP, with residual connections.

Multi-head causal self-attention, following the microGPT script structure.

Multi-head causal self-attention, following the microGPT script structure.

Fixed batch size version that outputs [16, 16] instead of [1, 16, 16]

Implements only the first 3 steps of LayerNorm: Step 1: Calculate mean Step 2: Center the data Step 3: Calculate variance Ou

Feed-forward MLP with ReLU activation, matching the microGPT script.

Feed-forward MLP with ReLU activation, matching the microGPT script.

Smaller MLP classifier with power-of-two dimensions for SNARK compatibility. Architecture: Input -> Linear(512->256) -> ReLU -> Linear(256->1

Minimal GPT model following the microGPT script architecture. Differences from standard GPT-2: - RMSNorm instead of LayerNorm (as in t

Minimal GPT model following the microGPT script architecture. Differences from standard GPT-2: - RMSNorm instead of LayerNorm (as in t

Multi-layer perceptron matching the original perceptron_2.onnx architecture. Architecture: Input [1, 4] -> Linear(4, 30) -> ReLU -> Linear(30

Wrapper for prefix polynomial evaluations, used for type safety in prefix operations.

Read-only and mutable access layer shared by prover and verifier helpers. Generic helper for recovering existing opening claims.

Prover state for element-wise addition sumcheck protocol. Maintains the equality polynomial and operand polynomials needed to generate sumcheck messa

Verifier for element-wise addition sumcheck protocol. Verifies that the prover's sumcheck messages are consistent with the claimed addition operation

Prefix component for bitwise AND lookup table decomposition.

Bitwise AND suffix