Tutorial 01: Compute API Basics

Goal: get comfortable with low-level backend namespaces.

Step 1: Initialize compute

import numpy as np
import grilly

backend = grilly.Compute()

Step 2: Run FNN ops

x = np.random.randn(16, 64).astype(np.float32)
w = np.random.randn(64, 32).astype(np.float32)
b = np.zeros(32, dtype=np.float32)

y = backend.fnn.linear(x, w, b)
y = backend.fnn.activation_relu(y)

Step 3: Run attention ops

q = np.random.randn(2, 4, 16, 32).astype(np.float32)
k = np.random.randn(2, 4, 16, 32).astype(np.float32)
v = np.random.randn(2, 4, 16, 32).astype(np.float32)

out = backend.attention.flash_attention2(q, k, v)
print(out.shape)

Step 4: Run memory search

query = np.random.randn(1, 128).astype(np.float32)
database = np.random.randn(2000, 128).astype(np.float32)

distances = backend.faiss.compute_distances(query, database)
topk_values, topk_indices = backend.faiss.topk(distances, k=5)

Step 5: Cleanup

backend.cleanup()

You now know the main backend layout and how to call operation families directly.