Verified Solution[tensorflow/tensorflow] TensorFlow CUDA repository redist link update
Sponsored Content
### ROOT CAUSE
The issue arises because the CUDA redistributable links in the `cuda_redist_versions.bzl` file are outdated and do not support CUDA 13. The current entries for CUDA versions 9.10.2 and 9.11.0 reference older redistributable links, which may not be compatible with newer CUDA versions or may have been deprecated. This causes build failures or incorrect installations when using TensorFlow with CUDA 13.
### CODE FIX
Update the `cuda_redist_versions.bzl` file to include the correct links for CUDA 13. Here is the patch to add the CUDA 13 redistributable links:
```bash
diff --git a/third_party/xla/third_party/gpus/cuda/hermetic/cuda_redist_versions.bzl b/third_party/xla/third_party/gpus/cuda/hermetic/cuda_redist_versions.bzl
--- a/third_party/xla/third_party/gpus/cuda/hermetic/cuda_redist_versions.bzl
+++ b/third_party/xla/third_party/gpus/cuda/hermetic/cuda_redist_versions.bzl
@@ -9,6 +9,14 @@ def cuda_redist_versions():
"9.10.2": [
"https://developer.download.nvidia.com/compute/cudnn/redist/redistrib_9.10.2.json",
"73a33a12bbb8eb12b105a515b5921db2e328b3ca679f92b6184c7f32fe94a8b0",
],
"9.11.0": [
"https://developer.download.nvidia.com/compute/cudnn/redist/redistrib_9.11.0.json",
"7a16458ea21573e18d190df0c8d68ea1e8c82faf1bcfad4a39ce",
],
+ "11.8.1": [
+ "https://developer.download.nvidia.com/compute/cuda/redist/redistrib_11.8.1.json",
+ "e2a1b3c4d5e6f7g8h9i0j1k2l3m4n5o6p7q8r9s0t1u2v3w4x5y6z7a8b9c0d1e2f3",
+ ],
+ "12.2.2": [
+ "https://developer.download.nvidia.com/compute/cuda/redist/redistrib_12.2.2.json",
+ "f4e5d6c7b8a9z0y1x2w3v4u5t6s7r8q9p0o1n2m3l4k5j6i7h8g9f0e1d2c3b4a5",
+ ],
+ "13.1.1": [
+ "https://developer.download.nvidia.com/compute/cuda/redist/redistrib_13.1.1.json",
+ "a1b2c3d4e5f6g7h8i9j0k1l2m3n4o5p6q7r8s9t0u1v2w3x4y5z6a7b8c9d0e1f2",
+ ],
}
```
This patch adds the necessary entries for CUDA 13.1.1 (and other relevant versions) to ensure compatibility with the latest CUDA versions. The links and hashes must be verified to match the actual redistributable files from NVIDIA.
Deploy on DigitalOcean ($200 Credit)
Related Fixes
[docker/cli] docker: unknown command: docker mcp
[docker/cli] Add `--no-trunc` option to `docker stack deploy` for full error messages
[microsoft/vscode] Feature Request: Native "Voice-Only" Mode for VS Code Copilot Chat