[Build] Upgrade CUDA Dockerfiles from GCC 10 to GCC 12 for C++20 compatibility (#44923)

Signed-off-by: Richard Barnes <rbarnes@meta.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>

CMakeLists.txt

@@ -20,6 +20,14 @@ set(CMAKE_CUDA_STANDARD_REQUIRED ON)
 set(CMAKE_HIP_STANDARD 20)
 set(CMAKE_HIP_STANDARD_REQUIRED ON)
 
+# PyTorch headers require C++20; GCC < 11.3 has incomplete C++20 support.
+if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND CMAKE_CXX_COMPILER_VERSION VERSION_LESS "11.3")
+  message(FATAL_ERROR
+    "GCC >= 11.3 is required to build vLLM (found ${CMAKE_CXX_COMPILER_VERSION}). "
+    "PyTorch's C++20 headers require a compiler with full C++20 support. "
+    "See: https://github.com/pytorch/pytorch/pull/167929")
+endif()
+
 
 # CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
 set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")

docker/Dockerfile

@@ -148,11 +148,13 @@ RUN if [ "${BUILD_OS}" = "manylinux" ]; then \
             sudo \
             python3-pip \
             libibverbs-dev \
-            # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
-            # as it was causing spam when compiling the CUTLASS kernels
-            gcc-10 \
-            g++-10 \
-        && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10 \
+            # GCC 10 was previously pinned to suppress spurious -Wredundant-move warnings
+            # from CUTLASS (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519). That bug
+            # was fixed in GCC 11. GCC >= 11.3 is now required because PyTorch's C++20 headers
+            # (pytorch/pytorch#167929) are not compatible with GCC < 11.3.
+            gcc-11 \
+            g++-11 \
+        && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 110 --slave /usr/bin/g++ g++ /usr/bin/g++-11 \
         # Install python dev headers if available (needed for cmake FindPython on Ubuntu 24.04
         # which ships cmake 3.28 and requires Development.SABIModule; silently skipped on
         # Ubuntu 20.04/22.04 where python3.x-dev is not available without a PPA)

docker/Dockerfile.nightly_torch

@@ -42,10 +42,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Reference: https://github.com/astral-sh/uv/pull/1694
 ENV UV_HTTP_TIMEOUT=500
 
-# Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
-# as it was causing spam when compiling the CUTLASS kernels
-RUN apt-get install -y gcc-10 g++-10
-RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10
+# GCC >= 11.3 required for PyTorch C++20 headers (pytorch/pytorch#167929).
+RUN apt-get install -y gcc-11 g++-11
+RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 110 --slave /usr/bin/g++ g++ /usr/bin/g++-11
 RUN <<EOF
 gcc --version
 EOF

docs/getting_started/installation/gpu.cuda.inc.md

@@ -139,6 +139,15 @@ You can find more information about vLLM's wheels in [Install the latest code](#
 
 #### Full build (with compilation) {#full-build}
 
+!!! note "Compiler requirement"
+    Building from source requires GCC/G++ ≥ 11.3. PyTorch's C++20 headers are
+    not compatible with GCC 10 or GCC < 11.3. On Ubuntu 22.04:
+    ```bash
+    sudo apt-get install -y gcc-11 g++-11
+    sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-11 110 \
+        --slave /usr/bin/g++ g++ /usr/bin/g++-11
+    ```
+
 If you want to modify C++ or CUDA code, you'll need to build vLLM from source. This can take several minutes:
 
 ```bash