diff --git a/CMakeLists.txt b/CMakeLists.txt
index e261bf1ac8..64de3baa9f 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -222,7 +222,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # `+PTX` in TORCH_CUDA_ARCH_LIST is not preserved here. If a kernel really
   # needs PTX, add `+PTX` to that kernel's component-specific arch list below.
   #
-  clear_cuda_arches(CUDA_ARCH_FLAGS)
+  clear_cuda_gencode_flags(CUDA_ARCH_FLAGS)
+  warn_if_ptx_arch_requested("${CUDA_ARCH_FLAGS}")
   extract_unique_cuda_archs_ascending(CUDA_ARCHS "${CUDA_ARCH_FLAGS}")
   message(STATUS "CUDA target architectures: ${CUDA_ARCHS}")
   # Filter the target architectures by the supported supported archs
diff --git a/cmake/utils.cmake b/cmake/utils.cmake
index bbae89c1f5..6ded936213 100644
--- a/cmake/utils.cmake
+++ b/cmake/utils.cmake
@@ -220,11 +220,11 @@ endmacro()
 #
 # Example:
 #   CMAKE_CUDA_FLAGS="-Wall -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75"
-#   clear_cuda_arches(CUDA_ARCH_FLAGS)
+#   clear_cuda_gencode_flags(CUDA_ARCH_FLAGS)
 #   CUDA_ARCH_FLAGS="-gencode arch=compute_70,code=sm_70;-gencode arch=compute_75,code=sm_75"
 #   CMAKE_CUDA_FLAGS="-Wall"
 #
-macro(clear_cuda_arches CUDA_ARCH_FLAGS)
+macro(clear_cuda_gencode_flags CUDA_ARCH_FLAGS)
     # Extract all `-gencode` flags from `CMAKE_CUDA_FLAGS`
     string(REGEX MATCHALL "-gencode arch=[^ ]+" CUDA_ARCH_FLAGS
       ${CMAKE_CUDA_FLAGS})
@@ -235,6 +235,26 @@ macro(clear_cuda_arches CUDA_ARCH_FLAGS)
       ${CMAKE_CUDA_FLAGS})
 endmacro()
 
+#
+# Warn when a caller requested PTX code generation through global CUDA arch
+# flags. vLLM removes those flags and reapplies per-source gencode flags, so the
+# user's global PTX request will not be preserved.
+#
+function(warn_if_ptx_arch_requested CUDA_ARCH_FLAGS)
+  foreach(_ARCH_FLAG ${CUDA_ARCH_FLAGS})
+    if(_ARCH_FLAG MATCHES "code=.*compute_[0-9]+[af]?")
+      message(WARNING
+        "PTX code generation requested in CUDA architecture flags "
+        "(${_ARCH_FLAG}), but vLLM does not preserve global PTX requests "
+        "when normalizing per-source CUDA architectures. Remove '+PTX' from "
+        "TORCH_CUDA_ARCH_LIST or rely on vLLM's built-in per-kernel PTX "
+        "selection.")
+      return()
+    endif()
+  endforeach()
+endfunction()
+
+
 #
 # Extract unique CUDA architectures from a list of compute capabilities codes in 
 # the form `<major><minor>[<letter>]`, convert them to the form sort 
diff --git a/docs/getting_started/installation/gpu.cuda.inc.md b/docs/getting_started/installation/gpu.cuda.inc.md
index 4b82178a4d..a417bf4af7 100644
--- a/docs/getting_started/installation/gpu.cuda.inc.md
+++ b/docs/getting_started/installation/gpu.cuda.inc.md
@@ -156,6 +156,9 @@ cd vllm
 uv pip install -e . --torch-backend=auto
 ```
 
+!!! note "CUDA Architecture & PTX Flags"
+    vLLM normalizes CUDA architectures on a per-source basis to optimize build times and wheel sizes. Global `+PTX` requests in `TORCH_CUDA_ARCH_LIST` (e.g., `TORCH_CUDA_ARCH_LIST="8.0+PTX"`) are ignored for general extension targets; vLLM generates PTX only for specific internal kernels that require it.
+
 !!! tip
     Building from source requires a lot of compilation. If you are building from source repeatedly, it's more efficient to cache the compilation results.
 
diff --git a/tests/test_cmake_utils.py b/tests/test_cmake_utils.py
index d0673bc462..dd0c6261d3 100644
--- a/tests/test_cmake_utils.py
+++ b/tests/test_cmake_utils.py
@@ -1,10 +1,22 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import shutil
 import subprocess
+import sys
 from pathlib import Path
 
 
+def _get_cmake_bin() -> str:
+    cmake = shutil.which("cmake")
+    if cmake:
+        return cmake
+    venv_cmake = Path(sys.executable).parent / "cmake"
+    if venv_cmake.is_file():
+        return str(venv_cmake)
+    return "cmake"
+
+
 def test_exact_family_arch_precedes_generic_family_fallback(tmp_path: Path):
     repo_root = Path(__file__).parents[1]
     script = tmp_path / "test_cuda_archs.cmake"
@@ -20,7 +32,7 @@ endif()
 """
     )
 
-    subprocess.run(["cmake", "-P", script], check=True)
+    subprocess.run([_get_cmake_bin(), "-P", script], check=True)
 
 
 def test_extract_archs_prefers_sass_target_over_corrupted_virtual_arch(
@@ -45,4 +57,26 @@ endif()
 """
     )
 
-    subprocess.run(["cmake", "-P", script], check=True)
+    subprocess.run([_get_cmake_bin(), "-P", script], check=True)
+
+
+def test_clear_cuda_gencode_flags(tmp_path: Path):
+    repo_root = Path(__file__).parents[1]
+    script = tmp_path / "test_clear_flags.cmake"
+    script.write_text(
+        f"""
+cmake_minimum_required(VERSION 3.26)
+include("{repo_root / "cmake" / "utils.cmake"}")
+set(CMAKE_CUDA_FLAGS "-Wall -gencode arch=compute_80,code=sm_80")
+clear_cuda_gencode_flags(CUDA_ARCH_FLAGS)
+if(NOT "${{CMAKE_CUDA_FLAGS}}" STREQUAL "-Wall ")
+  message(FATAL_ERROR "Expected '-Wall ', got '${{CMAKE_CUDA_FLAGS}}'")
+endif()
+if(NOT "${{CUDA_ARCH_FLAGS}}" STREQUAL "-gencode arch=compute_80,code=sm_80")
+  message(FATAL_ERROR "Expected '-gencode arch=compute_80,code=sm_80', "
+    "got '${{CUDA_ARCH_FLAGS}}'")
+endif()
+"""
+    )
+
+    subprocess.run([_get_cmake_bin(), "-P", script], check=True)