[jpascucci-nv]

Description

closes #3211

Detect the kernel’s GCC major version from /proc/version, install matching gcc/g++ packages, and export CC/CXX for the driver build. Preserve existing compiler selection on older kernels. This addresses the 580.65.06 DKMS failure on 6.8.0-147, where GCC 13 rejects the kernel’s GCC 14 compiler flags, while retaining older-driver coverage. Validated shell syntax, lint, and eight mocked regression cases. Live GPU validation remains pending CI.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[coderabbitai]

📝 Summary

Summary by CodeRabbit

• Bug Fixes
  • GPU driver installation now selects a compiler compatible with the running kernel when its version can be detected and differs from the default.
  • If the kernel version cannot be detected, or the compiler versions already match, installation continues with the existing compiler selection.
  • This helps avoid compiler mismatches during driver installation while preserving the existing behavior when no adjustment is needed.

host_install installs build-essential first. It compares the detected kernel GCC major version with the default, installs matching GCC and G++ packages when they differ, and sets CC and CXX before installing DKMS and kernel headers.

Changes: GPU driver build dependencies

Layer / File(s)	Summary
Kernel-aware compiler selection ci/tools/install_gpu_driver.sh	host_install separates build-essential installation from DKMS and kernel-header installation. When the detected kernel GCC major version differs from the default, it installs matching GCC and G++ packages and sets CC and CXX.

Assessment against linked issues

Objective	Addressed	Explanation
Restore successful scheduled optional-deps CI runs [#3211]	❓	The issue does not identify the failure cause, so the available context cannot confirm that this compiler change resolves it.

Suggested reviewers: leofang

Priority: Normal

Change: Bug fix

Merge Risk: Low · up to 1cfdc

A future kernel metadata format change could silently break GPU driver builds; the small fail-closed fix should be applied before or shortly after merging.

- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

Finishing Touches: 💡 1

🛠️ Fix failing CI checks: 💡

• Commit to this branch
• Create a new PR

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.

[coderabbitai]

Actionable comments posted: 1

---

Review info

Run configuration

• Source mode: Auto
• Repository: NVIDIA/cuda-python
• Branch: main
• File: .coderabbit.yaml
• Commit: 94625c1
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 6c0eb1dd-251d-41db-8b0e-968109019e13

Commits

Reviewing files that changed from the base of the PR and between e912fa5 and 97c7336.

Files selected for processing: 1

• ci/tools/install_gpu_driver.sh

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 10 remain after this review.

[coderabbitai]

Major · Stability & Availability · Quick win

Supported by static analysis

Script executed:

gh run view 40620830773 --attempt 1 --log |
  rg -ni -C 3 'linux-headers|dkms|gcc|apt-get'

Repository: NVIDIA/cuda-python

Length of output: 261

---

Script executed:

nl -ba ci/tools/install_gpu_driver.sh | sed -n '60,140p'
rg -n -F -e 'dkms' -e 'CC=' -e 'CXX' -e 'linux-headers' -- ci/ || test "$?" -eq 1

Repository: NVIDIA/cuda-python

Length of output: 4506

---

important: Install the compiler before the DKMS and kernel headers.

Line 95 installs dkms and linux-headers-${kernel_release} before the compiler selection on lines 98-106. Installing headers can run dkms autoinstall for registered modules through the dkms package's header hook. That build uses the default compiler, not the CC/CXX exported on line 104. The GCC 14 flags can then fail this step, the same way the PR reports. The gcc -dumpversion call on line 99 requires build-essential, so install that package first, select the compiler, and then install dkms and the headers.

🐛 Suggested fix

--- "a/ci/tools/install_gpu_driver.sh"
+++ "b/ci/tools/install_gpu_driver.sh"
@@ -92,18 +92,20 @@
 host_install() {
   local kernel_release kernel_gcc default_gcc
   kernel_release=$(uname -r)
-  apt-get -y install build-essential dkms "linux-headers-${kernel_release}" psmisc kmod
+  apt-get -y install build-essential
 
   # Match the kernel's GCC major only when the default compiler differs.
   kernel_gcc=$(sed -nE 's/.*gcc(-[0-9]+)? (\([^)]*\) )?([0-9]+)\..*/\3/p' /proc/version)
   default_gcc=$(gcc -dumpversion | cut -d. -f1)
   echo "Kernel ${kernel_release}: GCC major=${kernel_gcc:-unknown}, default GCC major=${default_gcc}" >&2
   if [ -n "$kernel_gcc" ] && [ "$kernel_gcc" != "$default_gcc" ]; then
     apt-get -y install "gcc-${kernel_gcc}" "g++-${kernel_gcc}"
     # Installing a versioned compiler does not change the default; select it for DKMS.
     export CC="gcc-${kernel_gcc}" CXX="g++-${kernel_gcc}"
     echo "Kernel ${kernel_release}: using CC=${CC} CXX=${CXX}" >&2
   fi
+
+  apt-get -y install dkms "linux-headers-${kernel_release}" psmisc kmod
 
   # Download and verify the installer before tearing down the existing
   # host driver so a download failure or SHA-256 mismatch leaves the

Source: Path instructions

[coderabbitai]

Actionable comments posted: 1

---

Review info

Run configuration

• Source mode: Auto
• Repository: NVIDIA/cuda-python
• Branch: main
• File: .coderabbit.yaml
• Commit: 94625c1
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: fc47ca62-08d2-471f-a9eb-570e5952e0d8

Commits

Reviewing files that changed from the base of the PR and between 97c7336 and 1cfdc45.

Files selected for processing: 1

• ci/tools/install_gpu_driver.sh

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 9 remain after this review.