hlb_cifar free speed (#16938)

* free speed for hlb cifar

* mirror

examples/hlb_cifar10.py

@@ -314,6 +314,9 @@ def train_cifar():
   opt_bias     = optim.SGD(params_bias,     lr=0.01, momentum=hyp['opt']['momentum'], nesterov=True, weight_decay=hyp['opt']['bias_decay'])
   opt_non_bias = optim.SGD(params_non_bias, lr=0.01, momentum=hyp['opt']['momentum'], nesterov=True, weight_decay=hyp['opt']['non_bias_decay'])
 
+  # realize model params and optimizer state before JIT to avoid cache misses
+  Tensor.realize(*params_dict.values(), *opt_bias.b, *opt_non_bias.b)
+
   # NOTE taken from the hlb_CIFAR repository, might need to be tuned
   initial_div_factor = hyp['opt']['initial_div_factor']
   final_lr_ratio = hyp['opt']['final_lr_ratio']

tinygrad/nn/datasets.py

@@ -8,7 +8,7 @@ def mnist(device=None, fashion=False):
          _mnist("t10k-images-idx3-ubyte.gz")[0x10:].reshape(-1,1,28,28).to(device), _mnist("t10k-labels-idx1-ubyte.gz")[8:].to(device)
 
 def cifar(device=None):
-  tt = tar_extract(Tensor.from_url('https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz', gunzip=True))
+  tt = tar_extract(Tensor.from_url('https://data.brainchip.com/dataset-mirror/cifar10/cifar-10-binary.tar.gz', gunzip=True))
   train = Tensor.cat(*[tt[f"cifar-10-batches-bin/data_batch_{i}.bin"].reshape(-1, 3073).to(device) for i in range(1,6)])
   test = tt["cifar-10-batches-bin/test_batch.bin"].reshape(-1, 3073).to(device)
   return train[:, 1:].reshape(-1,3,32,32), train[:, 0], test[:, 1:].reshape(-1,3,32,32), test[:, 0]