Remove the unused jax enable_x64. (keras-team#21737)

* Remove the unused jax x64 context.

* Fix ops.trace dtype.

Author: james77777778

keras/src/backend/common/variables_test.py

@@ -971,36 +971,17 @@ def test_mul(self, dtypes):
     def test_truediv(self, dtypes):
         import jax.numpy as jnp
 
-        try:
-            # JAX v0.8.0 and newer
-            from jax import enable_x64
-        except ImportError:
-            # JAX v0.7.2 and older
-            from jax.experimental import enable_x64
-
-        # We have to disable x64 for jax since jnp.true_divide doesn't respect
-        # JAX_DEFAULT_DTYPE_BITS=32 in `./conftest.py`. We also need to downcast
-        # the expected dtype from 64 bit to 32 bit when using jax backend.
-        with enable_x64(False):
-            dtype1, dtype2 = dtypes
-            x1 = backend.Variable(
-                "ones", shape=(1,), dtype=dtype1, trainable=False
-            )
-            x2 = backend.Variable(
-                "ones", shape=(1,), dtype=dtype2, trainable=False
-            )
-            x1_jax = jnp.ones((1,), dtype=dtype1)
-            x2_jax = jnp.ones((1,), dtype=dtype2)
-            expected_dtype = standardize_dtype(
-                jnp.true_divide(x1_jax, x2_jax).dtype
-            )
-            if "float64" in (dtype1, dtype2):
-                expected_dtype = "float64"
-            if backend.backend() == "jax":
-                expected_dtype = expected_dtype.replace("64", "32")
+        dtype1, dtype2 = dtypes
+        x1 = backend.Variable("ones", shape=(1,), dtype=dtype1, trainable=False)
+        x2 = backend.Variable("ones", shape=(1,), dtype=dtype2, trainable=False)
+        x1_jax = jnp.ones((1,), dtype=dtype1)
+        x2_jax = jnp.ones((1,), dtype=dtype2)
+        expected_dtype = standardize_dtype(
+            jnp.true_divide(x1_jax, x2_jax).dtype
+        )
 
-            self.assertDType(x1 / x2, expected_dtype)
-            self.assertDType(x1.__rtruediv__(x2), expected_dtype)
+        self.assertDType(x1 / x2, expected_dtype)
+        self.assertDType(x1.__rtruediv__(x2), expected_dtype)
 
     @parameterized.named_parameters(
         named_product(dtypes=itertools.combinations(NON_COMPLEX_DTYPES, 2))

keras/src/backend/jax/numpy.py

@@ -1240,14 +1240,7 @@ def tile(x, repeats):
 
 def trace(x, offset=0, axis1=0, axis2=1):
     x = convert_to_tensor(x)
-    dtype = None
-    # TODO: Remove the condition of uint8 and uint16 once we have jax>=0.4.27
-    # for both CPU & GPU environments.
-    # uint8 and uint16 will be casted to uint32 when jax>=0.4.27 but to int32
-    # otherwise.
-    if standardize_dtype(x.dtype) in ("bool", "uint8", "uint16"):
-        dtype = "int32"
-    return jnp.trace(x, offset=offset, axis1=axis1, axis2=axis2, dtype=dtype)
+    return jnp.trace(x, offset=offset, axis1=axis1, axis2=axis2)
 
 
 def tri(N, M=None, k=0, dtype=None):

keras/src/backend/numpy/numpy.py

@@ -1177,8 +1177,10 @@ def trace(x, offset=0, axis1=0, axis2=1):
     axis2 = standardize_axis_for_numpy(axis2)
     x = convert_to_tensor(x)
     dtype = standardize_dtype(x.dtype)
-    if dtype not in ("int64", "uint32", "uint64"):
-        dtype = dtypes.result_type(dtype, "int32")
+    if dtype in ("bool", "int8", "int16"):
+        dtype = "int32"
+    elif dtype in ("uint8", "uint16"):
+        dtype = "uint32"
     return np.trace(x, offset=offset, axis1=axis1, axis2=axis2, dtype=dtype)
 
 

keras/src/backend/tensorflow/numpy.py

@@ -2701,8 +2701,11 @@ def tile(x, repeats):
 def trace(x, offset=0, axis1=0, axis2=1):
     x = convert_to_tensor(x)
     dtype = standardize_dtype(x.dtype)
-    if dtype not in ("int64", "uint32", "uint64"):
-        dtype = dtypes.result_type(dtype, "int32")
+    if dtype in ("bool", "int8", "int16"):
+        dtype = "int32"
+    elif dtype in ("uint8", "uint16"):
+        dtype = "uint32"
+    x = tf.cast(x, dtype)
     x_shape = tf.shape(x)
     x = moveaxis(x, (axis1, axis2), (-2, -1))
     # Mask out the diagonal and reduce.
@@ -2711,10 +2714,7 @@ def trace(x, offset=0, axis1=0, axis2=1):
         x,
         tf.zeros_like(x),
     )
-    # The output dtype is set to "int32" if the input dtype is "bool"
-    if standardize_dtype(x.dtype) == "bool":
-        x = tf.cast(x, "int32")
-    return tf.cast(tf.reduce_sum(x, axis=(-2, -1)), dtype)
+    return tf.reduce_sum(x, axis=(-2, -1))
 
 
 def tri(N, M=None, k=0, dtype=None):

keras/src/backend/torch/numpy.py

@@ -1646,8 +1646,9 @@ def tile(x, repeats):
 def trace(x, offset=0, axis1=0, axis2=1):
     x = convert_to_tensor(x)
     dtype = standardize_dtype(x.dtype)
-    if dtype != "int64":
-        dtype = dtypes.result_type(dtype, "int32")
+    if dtype in ("bool", "int8", "int16", "uint8"):
+        # Torch backend doesn't support uint32 dtype.
+        dtype = "int32"
     return torch.sum(
         torch.diagonal(x, offset, axis1, axis2),
         dim=-1,

keras/src/ops/numpy.py

@@ -6365,8 +6365,13 @@ def compute_output_spec(self, x):
         x_shape[self.axis2] = -1
         output_shape = list(filter((-1).__ne__, x_shape))
         output_dtype = backend.standardize_dtype(x.dtype)
-        if output_dtype not in ("int64", "uint32", "uint64"):
-            output_dtype = dtypes.result_type(output_dtype, "int32")
+        if output_dtype in ("bool", "int8", "int16"):
+            output_dtype = "int32"
+        elif output_dtype in ("uint8", "uint16"):
+            output_dtype = "uint32"
+        if output_dtype == "uint32" and backend.backend() == "torch":
+            # Torch backend doesn't support uint32 dtype.
+            output_dtype = "int32"
         return KerasTensor(output_shape, dtype=output_dtype)