diff --git a/.ci/docker/ci_commit_pins/pytorch.txt b/.ci/docker/ci_commit_pins/pytorch.txt
index 759062dec6f..13b827111b3 100644
--- a/.ci/docker/ci_commit_pins/pytorch.txt
+++ b/.ci/docker/ci_commit_pins/pytorch.txt
@@ -1 +1 @@
-7cda4017ddda554752e89069ae205be5e8388f59
+90f1e7bed15ca5e48c61c5b6dc5ad4810524f82f
diff --git a/backends/xnnpack/test/ops/test_slice_copy.py b/backends/xnnpack/test/ops/test_slice_copy.py
index 857c78480ad..f8189ab9862 100644
--- a/backends/xnnpack/test/ops/test_slice_copy.py
+++ b/backends/xnnpack/test/ops/test_slice_copy.py
@@ -67,7 +67,7 @@ def forward(self, x):
 
         inputs = (torch.randn(1, 1, 3, 3),)
         # Note that two of the slices are optimized away as they are identity.
-        self._test_slice_copy(ConvSlice(), inputs, 4, 2)
+        self._test_slice_copy(ConvSlice(), inputs, 2, 2)
 
     def test_fp32_slice_copy_default_start(self):
         """
@@ -95,7 +95,7 @@ def forward(self, x):
         (
             Tester(module, inputs)
             .export()
-            .check_count({"torch.ops.aten.slice.Tensor": 3})
+            .check_count({"torch.ops.aten.slice.Tensor": 1})
             .to_edge_transform_and_lower()
             .check_not(["torch.ops.higher_order.executorch_call_delegate"])
         )
diff --git a/devtools/inspector/tests/inspector_utils_test.py b/devtools/inspector/tests/inspector_utils_test.py
index 2416fd89838..29f6f7ce4a8 100644
--- a/devtools/inspector/tests/inspector_utils_test.py
+++ b/devtools/inspector/tests/inspector_utils_test.py
@@ -610,9 +610,6 @@ def test_compare_intermediate_outputs_sequence_and_non_sequence(self):
         with self.assertRaises(ValueError):
             compare_intermediate_outputs(a, b, L1Comparator())
 
-    @unittest.skip(
-        "TODO: enable the test after required feature has been built in pytorch core nightly version"
-    )
     def test_equip_debug_handle_to_export_program_success(self):
         """Test that propagate_back_debug_handle returns True and properly equips debug handles."""
         # Create a test model
@@ -679,9 +676,6 @@ def test_equip_debug_handle_to_export_program_failure(self):
         # Check that it returns False due to mismatch
         self.assertFalse(result)
 
-    @unittest.skip(
-        "TODO: enable the test after required feature has been built in pytorch core nightly version"
-    )
     def test_equip_debug_handle_to_export_program_op_to_be_removed_in_to_edge(self):
         """Test that propagate_back_debug_handle returns True and properly equips debug handles when an op is removed in to_edge"""
 
diff --git a/install_requirements.py b/install_requirements.py
index 43cc3b2c76d..7570ff3cfc5 100644
--- a/install_requirements.py
+++ b/install_requirements.py
@@ -71,7 +71,7 @@ def python_is_compatible():
 #
 # NOTE: If you're changing, make the corresponding change in .ci/docker/ci_commit_pins/pytorch.txt
 # by picking the hash from the same date in https://hud.pytorch.org/hud/pytorch/pytorch/nightly/
-NIGHTLY_VERSION = "dev20250706"
+NIGHTLY_VERSION = "dev20250714"
 
 
 def install_requirements(use_pytorch_nightly):
diff --git a/runtime/core/portable_type/c10/c10/util/BFloat16.h b/runtime/core/portable_type/c10/c10/util/BFloat16.h
index 0f7cecda46b..06236df1fc8 100644
--- a/runtime/core/portable_type/c10/c10/util/BFloat16.h
+++ b/runtime/core/portable_type/c10/c10/util/BFloat16.h
@@ -4,6 +4,7 @@
 // 1 bit for the sign, 8 bits for the exponent and 7 bits for the mantissa.
 
 #include <c10/macros/Macros.h>
+#include <c10/util/bit_cast.h>
 #include <cmath>
 #include <cstdint>
 #include <cstring>
@@ -67,13 +68,7 @@ inline C10_HOST_DEVICE uint16_t round_to_nearest_even(float src) {
 #endif
     return UINT16_C(0x7FC0);
   } else {
-    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
-    union {
-      uint32_t U32; // NOLINT(facebook-hte-BadMemberName)
-      float F32; // NOLINT(facebook-hte-BadMemberName)
-    };
-
-    F32 = src;
+    const uint32_t U32 = c10::bit_cast<uint32_t>(src);
     uint32_t rounding_bias = ((U32 >> 16) & 1) + UINT32_C(0x7FFF);
     return static_cast<uint16_t>((U32 + rounding_bias) >> 16);
   }
@@ -111,9 +106,7 @@ struct alignas(2) BFloat16 {
 #endif
 };
 
-C10_API inline std::ostream& operator<<(
-    std::ostream& out,
-    const BFloat16& value) {
+inline std::ostream& operator<<(std::ostream& out, const BFloat16& value) {
   out << (float)value;
   return out;
 }
diff --git a/runtime/core/portable_type/c10/c10/util/Half.h b/runtime/core/portable_type/c10/c10/util/Half.h
index 373881f21e5..bdcf7458145 100644
--- a/runtime/core/portable_type/c10/c10/util/Half.h
+++ b/runtime/core/portable_type/c10/c10/util/Half.h
@@ -414,7 +414,7 @@ struct alignas(2) Half {
 #endif
 };
 
-C10_API inline std::ostream& operator<<(std::ostream& out, const Half& value) {
+inline std::ostream& operator<<(std::ostream& out, const Half& value) {
   out << (float)value;
   return out;
 }
diff --git a/runtime/core/portable_type/c10/c10/util/bit_cast.h b/runtime/core/portable_type/c10/c10/util/bit_cast.h
index 380cfa7db1c..d7d2aa8dd39 100644
--- a/runtime/core/portable_type/c10/c10/util/bit_cast.h
+++ b/runtime/core/portable_type/c10/c10/util/bit_cast.h
@@ -3,6 +3,8 @@
 #include <cstring>
 #include <type_traits>
 
+#include <c10/macros/Macros.h>
+
 #if __has_include(<bit>) && (defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L)
 #include <bit>
 #define C10_HAVE_STD_BIT_CAST 1
@@ -23,7 +25,7 @@ using std::bit_cast;
 // See https://en.cppreference.com/w/cpp/numeric/bit_cast for more
 // information as well as the source of our implementations.
 template <class To, class From>
-std::enable_if_t<
+C10_HOST_DEVICE std::enable_if_t<
     sizeof(To) == sizeof(From) && std::is_trivially_copyable_v<From> &&
         std::is_trivially_copyable_v<To>,
     To>