From fad66b9f8857f26444bb2f9deb32d5cd25277668 Mon Sep 17 00:00:00 2001
From: tlopex <820958424@qq.com>
Date: Sat, 13 Jun 2026 15:05:09 -0400
Subject: [PATCH] [CI] Drop redundant cmake/ninja install from the Linux wheel
 CUDA sidecar

Follow-up to the manylinux_cuda image switch. The prebuilt
quay.io/manylinux_cuda image already ships cmake (4.3.2, well above the
required 3.18) and make, and the sidecar's cmake invocation uses the
default Makefiles generator (no -G Ninja), so Ninja was never used.
Remove the now-redundant `pip install -U pip cmake ninja` and the
unused python_bin variable, and update the comment to match: the image
provides the build tools, and PATH only needs the bundled CPython (for
the configure) plus the CUDA toolchain (for nvcc).

Verified by running the script in the image with no pip install: it
configures with the image cmake, builds via make, and produces a
libtvm_runtime_cuda.so byte-for-byte the same size as before, still
NEEDED libcudart.so.13 with RUNPATH $ORIGIN.

Also fix a stale comment in the Windows sidecar: it claimed the conda
13.0.2 pin matched the Linux hook, but the Linux side is now CUDA 13.1
(from the image). The win-64 nvidia channel still caps at 13.0.x, so
Windows stays on 13.0.2 -- slightly behind Linux but harmless, since the
sidecar has no device code and links the CUDA runtime by soname only.
The Windows pip install is kept: that build uses the Ninja generator and
runs on a runner without the prebuilt CUDA image.
---
 ci/scripts/package/README.md                             | 6 ++++--
 .../package/manylinux_build_libtvm_runtime_cuda.sh       | 8 ++++----
 ci/scripts/package/windows_build_libtvm_runtime_cuda.bat | 9 +++++----
 3 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/ci/scripts/package/README.md b/ci/scripts/package/README.md
index f123fd1383b8..972c3153098d 100644
--- a/ci/scripts/package/README.md
+++ b/ci/scripts/package/README.md
@@ -23,7 +23,9 @@ and `[tool.scikit-build]`). This directory holds the few helper scripts that flo
 invokes:
 
 - `manylinux_build_libtvm_runtime_cuda.sh` — run by the `build_cuda_runtime` CI
-  stage; builds the `libtvm_runtime_cuda.so` sidecar inside the manylinux container.
+  stage; builds the `libtvm_runtime_cuda.so` sidecar inside the prebuilt
+  `quay.io/manylinux_cuda` image (CUDA toolkit preinstalled).
 - `windows_build_libtvm_runtime_cuda.bat` — the Windows equivalent (run with
-  `shell: cmd`), building `tvm_runtime_cuda.dll`.
+  `shell: cmd`); installs the CUDA toolkit via conda and builds
+  `tvm_runtime_cuda.dll`.
 - `build-environment.yaml` — conda environment for building the wheel.
diff --git a/ci/scripts/package/manylinux_build_libtvm_runtime_cuda.sh b/ci/scripts/package/manylinux_build_libtvm_runtime_cuda.sh
index 66e3d787b426..ea9e563046f3 100755
--- a/ci/scripts/package/manylinux_build_libtvm_runtime_cuda.sh
+++ b/ci/scripts/package/manylinux_build_libtvm_runtime_cuda.sh
@@ -27,14 +27,14 @@ set -euxo pipefail
 
 repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)"
 build_dir="${repo_root}/build-wheel-cuda"
-python_bin="/opt/python/cp310-cp310/bin/python"
 parallel="$(getconf _NPROCESSORS_ONLN 2>/dev/null || echo 4)"
 
 # Build the CUDA runtime sidecar with CUDA on and LLVM off, so it does not need
-# the LLVM prefix; the main CPU wheel links LLVM statically. pip-install
-# cmake/ninja rather than relying on whatever the image ships.
+# the LLVM prefix; the main CPU wheel links LLVM statically. The manylinux CUDA
+# image already ships cmake and make, and the build uses the default Makefiles
+# generator (no Ninja), so no build tools are installed here. Put the bundled
+# CPython and CUDA toolchain on PATH for the CMake configure and nvcc.
 export PATH="/opt/python/cp310-cp310/bin:/usr/local/cuda/bin:${PATH}"
-"${python_bin}" -m pip install -U pip cmake ninja
 nvcc --version
 
 rm -rf "${build_dir}"
diff --git a/ci/scripts/package/windows_build_libtvm_runtime_cuda.bat b/ci/scripts/package/windows_build_libtvm_runtime_cuda.bat
index 20523394ee1b..4864c18a80d2 100644
--- a/ci/scripts/package/windows_build_libtvm_runtime_cuda.bat
+++ b/ci/scripts/package/windows_build_libtvm_runtime_cuda.bat
@@ -36,10 +36,11 @@ set "conda_exe=conda"
 where conda >nul 2>nul || set "conda_exe=%CONDA%\Scripts\conda.exe"
 
 rem Install the pinned CUDA toolkit via conda from the nvidia channel, mirroring the
-rem LLVM-via-conda install used elsewhere. The win-64 channel caps at 13.0.x, matching
-rem the Linux hook's CUDA 13.0.2. The nvidia CDN occasionally returns a transient
-rem HTTP 5xx, so retry once; a half-finished first attempt can leave the prefix
-rem partially populated, so wipe it before retrying.
+rem LLVM-via-conda install used elsewhere. The win-64 channel caps at 13.0.x, so this
+rem pins 13.0.2 -- slightly behind the Linux image's CUDA 13.1, which is harmless: the
+rem sidecar has no device code and links the CUDA runtime by soname only. The nvidia
+rem CDN occasionally returns a transient HTTP 5xx, so retry once; a half-finished first
+rem attempt can leave the prefix partially populated, so wipe it before retrying.
 if not exist "%cuda_prefix%\Library\bin\nvcc.exe" (
   call "%conda_exe%" create -q -p "%cuda_prefix%" -c nvidia/label/cuda-13.0.2 cuda-toolkit -y
   if errorlevel 1 (