tests: e2e convergence with sgd

tests/conftest.py

@@ -1,7 +1,9 @@
 from __future__ import annotations
 
+import random
 from typing import TYPE_CHECKING
 
+import numpy as np
 import pytest
 
 from skinnygrad import runtime
@@ -22,3 +24,9 @@ def engine(request: pytest.FixtureRequest) -> object:
     if request.param == runtime.NumPyEngine:
         return runtime.NumPyEngine()
     raise ValueError("invalid internal test config")
+
+
+@pytest.fixture(autouse=True)
+def set_random_seeds(seed: float = 42):
+    np.random.seed(seed)
+    random.seed(seed)

tests/test_convergence.py

@@ -0,0 +1,22 @@
+from skinnygrad import optim, tensors
+
+
+def test_sgd_training_loop():
+    a = tensors.Tensor.random_normal(1, 5, mean=5, var=2, requires_grad=False)
+    b = tensors.Tensor.random_normal(5, 1, mean=5, var=2, requires_grad=True)
+    opt = optim.SGD((a, b), lr=0.001)
+
+    initial_loss, loss_val = None, None
+    for _ in range(1000):  # Run for 100 iterations
+        with opt:
+            loss = ((a @ b) - 10) ** 2
+            loss.backprop()
+
+        loss_val = loss.sum().realize()
+        assert isinstance(loss_val, float)
+        if initial_loss is None:
+            initial_loss = loss_val
+
+    assert initial_loss is not None
+    assert loss_val is not None
+    assert 0 <= loss_val < 0.01 < initial_loss  # Ensure loss decreases and is not zero