From 44104b0b7f3f915c0a3a9ce4d2b78758d33b9a38 Mon Sep 17 00:00:00 2001
From: George Hotz <72895+geohot@users.noreply.github.com>
Date: Mon, 1 Dec 2025 18:27:32 -0800
Subject: [PATCH] mnist with grad acc + Adam on CPU (#13520)

* mnist with grad acc + Adam on CPU

* still broken, but closer

* works w/o jit

* this works without the jit
---
 examples/gradaccum_mnist.py | 73 +++++++++++++++++++++++++++++++++++++
 1 file changed, 73 insertions(+)
 create mode 100644 examples/gradaccum_mnist.py

diff --git a/examples/gradaccum_mnist.py b/examples/gradaccum_mnist.py
new file mode 100644
index 0000000000..c46cd0ced3
--- /dev/null
+++ b/examples/gradaccum_mnist.py
@@ -0,0 +1,73 @@
+import itertools
+from examples.beautiful_mnist import Model
+from tinygrad import nn, Tensor, dtypes, Device
+from tinygrad.helpers import getenv, trange, partition
+
+# TODO: refactor this into optim/onnx
+def functional_adam(g:Tensor, m:Tensor, v:Tensor, b1_t:Tensor, b2_t:Tensor, lr=0.001, b1=0.9, b2=0.999, eps=1e-6) -> Tensor:
+  b1_t *= b1
+  b2_t *= b2
+  m.assign(b1 * m + (1.0 - b1) * g)
+  v.assign(b2 * v + (1.0 - b2) * (g * g))
+  m_hat = m / (1.0 - b1_t)
+  v_hat = v / (1.0 - b2_t)
+  return lr * (m_hat / (v_hat.sqrt() + eps))
+
+if __name__ == "__main__":
+  BS = getenv("BS", 512)
+  ACC_STEPS = getenv("ACC_STEPS", 4)
+
+  X_train, Y_train, X_test, Y_test = nn.datasets.mnist()
+  model = Model()
+
+  params = nn.state.get_parameters(model)
+
+  # set requires grad on the ones we need gradients of
+  for x in params:
+    if x.requires_grad is None: x.requires_grad_()
+
+  # split params (with grads) and buffers (without)
+  params, buffers = partition(nn.state.get_parameters(model), lambda x: x.requires_grad)
+  print(f"params: {len(params)} buffers: {len(buffers)}")
+
+  # optim params
+  pos_params = list(itertools.accumulate(params, lambda x,y: x+y.numel(), initial=0))
+  adam_m = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
+  adam_v = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
+  adam_b1_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False)
+  adam_b2_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False)
+  adam_params = [adam_m, adam_v, adam_b1_t, adam_b2_t]
+
+  @Tensor.train()
+  def microbatch(loss: Tensor, grads: Tensor):
+    samples = Tensor.randint(BS // ACC_STEPS, high=X_train.shape[0])
+    for t in params: t.grad = None
+    # divide by ACC_STEPS at the loss
+    uloss = (model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]) / ACC_STEPS).backward()
+    # concat the grads and assign them
+    loss.assign(loss + uloss)
+    grads.assign(grads + Tensor.cat(*[t.grad.contiguous().flatten() for t in params], dim=0))
+    Tensor.realize(loss, grads)
+
+  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100
+
+  test_acc = float('nan')
+  for i in (t:=trange(getenv("STEPS", 70))):
+    # microbatch sets the gradients
+    loss = Tensor.zeros(tuple()).contiguous()
+    grads = Tensor.zeros(pos_params[-1]).contiguous()
+    for _ in range(ACC_STEPS): microbatch(loss, grads)
+
+    # run optimizer (on CPU, where adam params live)
+    delta = functional_adam(grads.to("CPU"), adam_m, adam_v, adam_b1_t, adam_b2_t)
+
+    # update the params, copying back the delta one at a time to avoid OOM
+    for j,tt in enumerate(params):
+      tt.assign(tt.detach() - delta[pos_params[j]:pos_params[j+1]].reshape(tt.shape).to(Device.DEFAULT))
+
+    # realize everything
+    Tensor.realize(*params, *buffers, *adam_params)
+
+    # eval
+    if i%10 == 9: test_acc = get_test_acc().item()
+    t.set_description(f"loss: {loss.item():6.2f} test_accuracy: {test_acc:5.2f}%")