diff --git a/.readthedocs.yml b/.readthedocs.yml
index 4b829fc279..9631c71e80 100644
--- a/.readthedocs.yml
+++ b/.readthedocs.yml
@@ -7,7 +7,8 @@ build:
     post_create_environment:
       - pip install uv
     post_install:
-      - VIRTUAL_ENV=$READTHEDOCS_VIRTUALENV_PATH uv pip install -r doc/requirements.txt
+      # RTD renders documentation without executing custom PyTorch operators.
+      - DP_ENABLE_PYTORCH=0 VIRTUAL_ENV=$READTHEDOCS_VIRTUALENV_PATH uv pip install -r doc/requirements.txt
   apt_packages:
     - inkscape
 sphinx:
diff --git a/source/CMakeLists.txt b/source/CMakeLists.txt
index 1161f59ee7..4ccc18542c 100644
--- a/source/CMakeLists.txt
+++ b/source/CMakeLists.txt
@@ -404,7 +404,10 @@ if(ENABLE_PYTORCH AND NOT DEEPMD_C_ROOT)
   else()
     set(DEEPMD_TORCH_HAS_CUDA FALSE)
   endif()
-  if(NOT Torch_VERSION VERSION_LESS "2.1.0")
+  # PyTorch 2.14 raised the minimum C++ standard from C++17 to C++20.
+  if(NOT Torch_VERSION VERSION_LESS "2.14.0")
+    set_if_higher(CMAKE_CXX_STANDARD 20)
+  elseif(NOT Torch_VERSION VERSION_LESS "2.1.0")
     set_if_higher(CMAKE_CXX_STANDARD 17)
   elseif(NOT Torch_VERSION VERSION_LESS "1.5.0")
     set_if_higher(CMAKE_CXX_STANDARD 14)
diff --git a/source/lmp/tests/dpa4_spin_harness.py b/source/lmp/tests/dpa4_spin_harness.py
index c48eb5f600..26a4cd3f95 100644
--- a/source/lmp/tests/dpa4_spin_harness.py
+++ b/source/lmp/tests/dpa4_spin_harness.py
@@ -176,7 +176,8 @@ def compute_expected(
     proc = sp.run([sys.executable, "-c", script], capture_output=True, text=True)
     if proc.returncode != 0:
         raise RuntimeError(f"Failed to compute expected values:\n{proc.stderr}")
-    result = json.loads(proc.stdout.strip())
+    # Warp and other backends may log to stdout before the final JSON payload.
+    result = json.loads(proc.stdout.strip().splitlines()[-1])
 
     # Raw DeepEval force_mag (dE/dspin), scaled by LAMMPS's own
     # spin_norm / hbar unit convention (see HBAR_METAL) before comparison.
diff --git a/source/lmp/tests/test_lammps_dpa4_chg_spin_deepspin_pt2.py b/source/lmp/tests/test_lammps_dpa4_chg_spin_deepspin_pt2.py
index 3d76b17aac..775725c703 100644
--- a/source/lmp/tests/test_lammps_dpa4_chg_spin_deepspin_pt2.py
+++ b/source/lmp/tests/test_lammps_dpa4_chg_spin_deepspin_pt2.py
@@ -180,7 +180,8 @@ def _compute_expected() -> None:
     proc = sp.run([sys.executable, "-c", script], capture_output=True, text=True)
     if proc.returncode != 0:
         raise RuntimeError(f"Failed to compute expected values:\n{proc.stderr}")
-    result = json.loads(proc.stdout.strip())
+    # Warp and other backends may log to stdout before the final JSON payload.
+    result = json.loads(proc.stdout.strip().splitlines()[-1])
 
     # Raw DeepEval force_mag (dE/dspin), scaled by LAMMPS's own
     # spin_norm / hbar unit convention (see ``_HBAR_METAL`` above) before
diff --git a/source/lmp/tests/test_lammps_dpa4_zbl_pt2.py b/source/lmp/tests/test_lammps_dpa4_zbl_pt2.py
index 099e236294..32f20257df 100644
--- a/source/lmp/tests/test_lammps_dpa4_zbl_pt2.py
+++ b/source/lmp/tests/test_lammps_dpa4_zbl_pt2.py
@@ -188,7 +188,8 @@ def _compute_expected() -> None:
     proc = sp.run([sys.executable, "-c", script], capture_output=True, text=True)
     if proc.returncode != 0:
         raise RuntimeError(f"Failed to compute expected values:\n{proc.stderr}")
-    result = json.loads(proc.stdout.strip())
+    # Warp and other backends may log to stdout before the final JSON payload.
+    result = json.loads(proc.stdout.strip().splitlines()[-1])
 
     expected_e = result["e"]
     expected_ae = np.array(result["ae"])
diff --git a/source/op/pt/CMakeLists.txt b/source/op/pt/CMakeLists.txt
index 6948f50335..ddfe38253b 100644
--- a/source/op/pt/CMakeLists.txt
+++ b/source/op/pt/CMakeLists.txt
@@ -212,9 +212,18 @@ if(USE_CUDA_TOOLKIT AND DEEPMD_TORCH_HAS_CUDA)
                                               ${DEEPMD_PT_CUDA_ARCHITECTURES})
   endif()
   target_link_libraries(deepmd_op_pt PRIVATE CUDA::cublas CUDA::cublasLt)
-  # libtorch headers require C++17; the CUDA sources must match.
-  set_target_properties(deepmd_op_pt PROPERTIES CUDA_STANDARD 17
-                                                CUDA_STANDARD_REQUIRED ON)
+  # libtorch headers require at least C++17, and PyTorch 2.14+ requires C++20.
+  # Preserve an explicitly higher CUDA standard requested by the user.
+  set(DEEPMD_PT_CUDA_STANDARD 17)
+  if(NOT Torch_VERSION VERSION_LESS "2.14.0")
+    set_if_higher(DEEPMD_PT_CUDA_STANDARD 20)
+  endif()
+  if(CMAKE_CUDA_STANDARD)
+    set_if_higher(DEEPMD_PT_CUDA_STANDARD ${CMAKE_CUDA_STANDARD})
+  endif()
+  set_target_properties(
+    deepmd_op_pt PROPERTIES CUDA_STANDARD ${DEEPMD_PT_CUDA_STANDARD}
+                            CUDA_STANDARD_REQUIRED ON)
   if(DEEPMD_ENABLE_DPA1_HIGH_LMAX)
     target_compile_definitions(deepmd_op_pt
                                PRIVATE DEEPMD_ENABLE_DPA1_HIGH_LMAX=1)