evo-jax-compute

evo-jax-compute

All JAX computation logic for the five problem tasks, dispatching by problem ID.

Quick Start

import sys
sys.path.insert(0, '/app/environment/skills/evo-jax-compute/scripts')
from compute_utils import dispatch_task, compute_basic_reduce, compute_scan_rnn

# Dispatch by task ID
result = dispatch_task('basic_reduce', 'compute mean of each row', jax_data)

# Or call directly
result = compute_basic_reduce(x_array)

Key Functions

compute_basic_reduce(x)

Row-wise mean: jnp.mean(x, axis=1). Input (M,N) -> Output (M,).

compute_map_square(x)

Element-wise square: jnp.square(x). Same shape in/out.

compute_grad_logistic(data)

Gradient of logistic loss w.r.t. weights.

• Loss: L(w) = mean(logaddexp(0, -y * (X @ w)))
• Uses jax.grad(loss_fn, argnums=0) to differentiate w.r.t. w
• jnp.logaddexp(0, z) for numerical stability (scalar 0 auto-promotes)
• Input: dict with keys 'x'/'X' (N,D), 'y' (N,), 'w' (D,)
• Output: gradient array (D,)

compute_scan_rnn(data)

RNN forward pass using jax.lax.scan.

• CRITICAL: Uses weight-RIGHT-multiply: h_new = tanh(x_t @ Wx + h @ Wh + b)
• x_t is on LEFT of Wx, h is on LEFT of Wh (standard doc convention)
• Weight matrices are closed over, NOT passed through carry
• Carry dtype must match init dtype exactly
• Input: dict with 'seq' (T,I), 'init' (H,), 'Wx', 'Wh', 'b'
• Output: stacked hidden states (T, H)

compute_jit_mlp(data)

JIT-compiled 2-layer MLP.

• Layer 1: h = relu(X @ W1 + b1) using jax.nn.relu
• Layer 2: out = h @ W2 + b2
• Uses @jax.jit decorator
• Input: dict with 'X' (B,F), 'W1' (F,H), 'b1' (H,), 'W2' (H,O), 'b2' (O,)
• Output: (B, O)

dispatch_task(task_id, description, data)

Dispatch to correct compute function by exact ID match or fuzzy keyword matching.

Domain Knowledge

• RNN uses RIGHT-multiply (x_t @ Wx), verified against reference outputs
• jnp.logaddexp(0, x) safely computes log(1+exp(x)) without overflow
• jax.grad requires scalar return; use jnp.mean() to aggregate
• jax.lax.scan carry must maintain identical dtype/shape across iterations
• JIT may cause ~1e-7 float differences due to XLA operation reordering
• jax.nn.relu is canonical; derivative at 0 is defined as 0