Fix attention rng mismatch between forward and reverse direction

PiperOrigin-RevId: 272707157

Author: T2T Team

tensor2tensor/trax/layers/attention.py

@@ -344,7 +344,8 @@ def new_params_and_state(self, input_shape, input_dtype, rng):
 class BaseCausalAttention(base.Layer):
   """Base class for variants of causal self-attention."""
 
-  def __init__(self):
+  def __init__(self, mode='train'):
+    del mode
     super(BaseCausalAttention, self).__init__(n_inputs=3)
 
   def forward(self, inputs, params=(), state=(), rng=None, **kwargs):

tensor2tensor/trax/layers/reversible.py

@@ -101,8 +101,8 @@ def reverse(self, output, params=(), state=(), **kwargs):
       rngs = backend.random.split(rng, self._n_layers)
 
     layer_val = output
-    for layer, p, s, rng in reversed(zip(self.sublayers,
-                                         params, state, rngs)):
+    for layer, p, s, rng in reversed(list(zip(self.sublayers,
+                                              params, state, rngs))):
       layer_val = layer.reverse(layer_val, p, s, rng=rng, **kwargs)
 
     return layer_val
@@ -116,8 +116,8 @@ def reverse_and_grad(self, output, ct, params=(), state=(), **kwargs):
     layer_val = output
     layer_ct = ct
     params_ct = []
-    for layer, p, s, rng in reversed(zip(self.sublayers,
-                                         params, state, rngs)):
+    for layer, p, s, rng in reversed(list(zip(self.sublayers,
+                                              params, state, rngs))):
       layer_val, layer_ct = layer.reverse_and_grad(
           layer_val, layer_ct, p, s, rng=rng, **kwargs)
       layer_ct, p_ct = layer_ct

tensor2tensor/trax/models/research/reformer.py

@@ -254,14 +254,18 @@ def __init__(self, attention):
     super(ApplyAttentionWrapper, self).__init__(attention, [], [])
     self.attention = attention
 
-  def forward_and_backward(self, inputs, ct, **kwargs):
+  def forward_and_backward(self, inputs, ct, rng=None, **kwargs):
     # Simultaneous forward pass and backprop through the attention mechanism.
     qkv = inputs[:3]
     passthrough = inputs[3:]
     out_ct = ct[0]
     passthrough_ct = ct[1:]
+    if rng is not None:
+      # Adjust RNG to match the forward pass.
+      rng = backend.random.split(rng, self._n_layers)[0]
 
-    out, qkv_ct = self.attention.forward_and_backward(qkv, out_ct, **kwargs)
+    out, qkv_ct = self.attention.forward_and_backward(
+        qkv, out_ct, rng=rng, **kwargs)
     return (out,) + passthrough, qkv_ct + passthrough_ct
 
 

tensor2tensor/trax/models/research/reformer_test.py

@@ -21,10 +21,43 @@
 
 from absl.testing import absltest
 from absl.testing import parameterized
+import jax
+import numpy as onp
+
+from tensor2tensor.trax import backend
 from tensor2tensor.trax import layers as tl
+from tensor2tensor.trax.backend import numpy as np
 from tensor2tensor.trax.models.research import reformer
 
 
+class PoisonOnRNGMismatchAttention(tl.BaseCausalAttention):
+  """Fills gradients with NaNs if reverse rng does not match forward rng."""
+
+  # pylint: disable=protected-access
+  def forward_and_backward(self, inputs, ct, rng=None, **kwargs):
+    assert backend.get_name() == 'jax', (
+        'JAX backend is required to use forward_and_backward.')
+
+    if ct is not None and tl.Layer._STASH_OUT is not None:
+      recovered_rng = tl.Layer._STASH_OUT.pop(self)
+      is_same = (rng[0] == recovered_rng[0]) & (rng[1] == recovered_rng[1])
+      is_same = is_same.astype(np.float32)
+      # Divides by zero if rngs are not the same, which results in NaNs.
+      inputs = (inputs[0] / is_same, inputs[1] / is_same, inputs[2] / is_same)
+
+    def _do_forward(x):  # pylint: disable=invalid-name
+      res, _ = self.forward(x, rng=rng, **kwargs)
+      return res
+    output, vjpfun = jax.vjp(_do_forward, inputs)
+    return output, vjpfun(ct)[0]
+
+  def forward(self, inputs, params=(), state=(), rng=None, **kwargs):
+    if tl.Layer._STASH_IN is not None:
+      tl.Layer._STASH_IN[self] = rng
+    return inputs[2], state
+  # pylint: enable=protected-access
+
+
 class ReformerTest(parameterized.TestCase):
 
   def test_reformer_lm_forward_shape(self):
@@ -39,6 +72,33 @@ def test_reformer_lm_forward_shape(self):
         model, tuple(input_shape), integer_inputs=True)
     self.assertEqual(((1, 8, 16), (1, 8, 16)), final_shape)
 
+  def test_reformer_rng_consistency(self):
+    with backend.use_backend('jax'):
+      vocab_size = 16
+      batch_size = 1
+      input_shape = ((batch_size, 8), (batch_size, 8))
+      model = reformer.ReformerLM(
+          vocab_size, d_model=32, d_ff=64,
+          d_attention_key=16, d_attention_value=16, n_layers=1, n_heads=2,
+          max_len=16, n_chunks=2, n_attention_chunks=1, mode='train',
+          attention_type=PoisonOnRNGMismatchAttention)
+
+      rng = backend.random.get_prng(0)
+      params, state = model.initialize_once(
+          input_shape, (np.int32, np.int32), rng)
+
+      def dummy_loss_fn(params):
+        inputs = (np.zeros(input_shape[0], dtype=np.int32),) * 2
+        output = model(inputs, params=params, state=state, rng=rng)
+        dummy_loss = backend.numpy.sum(output[0])
+        return dummy_loss
+
+      grad_fn = backend.grad(dummy_loss_fn)
+      grads = grad_fn(params)
+      # PoisonOnRNGMismatchAttention uses NaNs to signal an rng mismatch.
+      for grad in jax.tree_util.tree_leaves(grads):
+        assert onp.all(onp.isfinite(grad))
+
 
 if __name__ == '__main__':
   absltest.main()