This change updates the random number generation utility to use tf.uint64 for bitwise operations and constants. The seed generation function (next_seed_fn) has been rewritten to implement a proper 7-bit LFSR (PRBS7) using bitwise operations, replacing the previous power-based approximation.

tensorflow_model_optimization/python/core/internal/tensor_encoding/utils/tf_utils.py

@@ -153,33 +153,71 @@ def _cmwc_random_sequence(num_elements, seed):
 
   # Create constants needed for the algorithm. The constants and notation
   # follows from the above reference.
-  a = tf.tile(tf.constant([3636507990], tf.int64), [parallelism])
-  b = tf.tile(tf.constant([2**32], tf.int64), [parallelism])
-  logb_scalar = tf.constant(32, tf.int64)
+  a = tf.tile(tf.constant([3636507990], tf.uint64), [parallelism])
+  b = tf.tile(tf.constant([2**32], tf.uint64), [parallelism])
+  logb_scalar = tf.constant(32, tf.uint64)
   logb = tf.tile([logb_scalar], [parallelism])
-  f = tf.tile(tf.constant([0], dtype=tf.int64), [parallelism])
-  bits = tf.constant(0, dtype=tf.int64, name='bits')
+  f = tf.tile(tf.constant([0], dtype=tf.uint64), [parallelism])
+  bits = tf.constant(0, dtype=tf.uint64, name='bits')
 
   # TensorArray used in tf.while_loop for efficiency.
   values = tf.TensorArray(
       dtype=tf.float64, size=num_iters, element_shape=[parallelism])
   # Iteration counter.
   num = tf.constant(0, dtype=tf.int32, name='num')
   # TensorFlow constant to be used at multiple places.
-  val_53 = tf.constant(53, tf.int64, name='val_53')
+  val_53 = tf.constant(53, tf.uint64, name='val_53')
 
   # Construct initial sequence of seeds.
   # From a single input seed, we construct multiple starting seeds for the
   # sequences to be computed in parallel.
   def next_seed_fn(i, val, q):
-    val = val**7 + val**6 + 1  # PRBS7.
+    """Generates the next seed using a 7-bit LFSR.
+
+    This function implements a proper 7-bit Fibonacci LFSR with the polynomial
+    x^7 + x^6 + 1. It takes the lower 7 bits of `val` as the current state,
+    computes the next state, and writes it to the TensorArray `q`.
+
+    Args:
+      i: The current index in the while loop.
+      val: The current seed value (tf.uint64). The lower 7 bits are used as
+        the LFSR state.
+      q: The tf.TensorArray to write the generated seed into.
+
+    Returns:
+      A tuple of (i + 1, new_val, q), where `new_val` is the next state of the
+      LFSR.
+    """
+    state = tf.bitwise.bitwise_and(val, tf.constant(0x7F, tf.uint64))
+    # Avoid zero state, which is a trapping state for this LFSR polynomial.
+    state = tf.bitwise.bitwise_or(
+        state,
+        tf.cast(tf.equal(state, tf.constant(0, tf.uint64)), tf.uint64)
+    )
+    # Feedback bit = bit 7 (index 6) ^ bit 6 (index 5)
+    feedback = tf.bitwise.bitwise_and(
+        tf.bitwise.bitwise_xor(
+            tf.bitwise.right_shift(state, tf.constant(6, tf.uint64)),
+            tf.bitwise.right_shift(state, tf.constant(5, tf.uint64))
+        ),
+        tf.constant(1, tf.uint64)
+    )
+    # Shift left and insert feedback
+    val = tf.bitwise.bitwise_and(
+        tf.bitwise.bitwise_or(
+            tf.bitwise.left_shift(state, tf.constant(1, tf.uint64)),
+            feedback
+        ),
+        tf.constant(0x7F, tf.uint64)
+    )
     q = q.write(i, val)
     return i + 1, val, q
 
-  q = tf.TensorArray(dtype=tf.int64, size=parallelism, element_shape=())
+  q = tf.TensorArray(dtype=tf.uint64, size=parallelism, element_shape=())
+  seed_u64 = tf.cast(seed, tf.uint64)
   _, _, q = tf.while_loop(lambda i, _, __: i < parallelism,
                           next_seed_fn,
-                          [tf.constant(0), seed, q])
+                          [tf.constant(0), seed_u64, q])
   c = q = q.stack()
 
   # The random sequence generation code.
@@ -193,9 +231,10 @@ def cmwc_step(f, bits, q, c, num, values):
       f.set_shape((1,))  # Correct for failed shape inference.
     bits += logb_scalar
     def add_val(bits, f, values, num):
+      mask_53 = tf.constant(2**53 - 1, tf.uint64)
       new_val = tf.cast(
-          tf.bitwise.bitwise_and(f, (2**val_53 - 1)),
-          dtype=tf.float64) * (1 / 2**val_53)
+          tf.bitwise.bitwise_and(f, mask_53),
+          dtype=tf.float64) * (1.0 / 2.0**53)
       values = values.write(num, new_val)
       f += tf.bitwise.right_shift(f, val_53)
       bits -= val_53

tensorflow_model_optimization/python/core/internal/tensor_encoding/utils/tf_utils_test.py

@@ -188,6 +188,15 @@ def test_tf_int32_seed_raises(self):
     with self.assertRaisesRegex(TypeError, 'tf.int64 Tensor'):
       tf_utils._cmwc_random_sequence(10, tf.constant(123, tf.int32))
 
+  def test_reproduction_b511305971(self):
+    """Verifies that the PRNG does not produce negative states or bounds violations."""
+    # Reproduction steps from b/511305971
+    sequence = tf_utils._cmwc_random_sequence(
+        1000, tf.constant(12345, dtype=tf.int64))
+    sequence = self.evaluate(sequence)
+    self.assertAllGreaterEqual(sequence, 0.0)
+    self.assertAllLessEqual(sequence, 1.0)
+
 
 class RandomSignsCMWCTests(tf.test.TestCase, parameterized.TestCase):
   """Tests for `random_signs_cmwc` method."""