Skip to content

CUDA Dependency Updates Tracker

This page tracks workarounds and pinned dependencies that depend on external packages improving their CUDA 13 support. As package maintainers release updated versions with better CUDA compatibility, the items below should be revisited and updated.

Current Workarounds

FAISS GPU Wheel for CUDA 13

Status: ⚠️ Requires Action

Current Workaround: Use faiss-cpu in requirements-cuda13.txt

File: requirements-cuda13.txt

Issue:

  • FAISS GPU wheels are unavailable for CUDA 13 + Python 3.13 due to limited build coverage from Meta
  • CPU variant is a functional fallback but lacks GPU acceleration for vector similarity search

Action Required When Fixed:

  1. Monitor FAISS releases for CUDA 13 + Python 3.13 wheel availability
  2. Test faiss-gpu with CUDA 13.x in CI environment
  3. Replace faiss-cpu with faiss-gpu in requirements-cuda13.txt
  4. Update documentation in docs/getting-started/installation.md and docs/getting-started/environments.md to remove the fallback note

Related Issue: Meta/FAISS issue tracker for CUDA 13 wheel builds

ONNX Runtime GPU Version Pinning

Status: 📌 Version Pinned

Current Pin: onnxruntime-gpu==1.24.1 in requirements-cuda.txt

File: requirements-cuda.txt

Reason for Pinning:

  • Version 1.24.1 is known to work with CUDA 12 user-space library linkage (libcublasLt.so.12, etc.) across both CUDA 12.x and 13.x environments
  • Later versions may have changed their CUDA 12 binary compatibility or introduced stricter version requirements

Action Required When Fixed:

  1. Monitor ONNX Runtime releases for CUDA 12–13 compatibility improvements
  2. Test newer versions (1.25.x, 1.26.x+) with:
  3. CUDA 12.x environments
  4. CUDA 13.x environments (with conda CUDA 12 runtime libs for linkage)
  5. CPU provider fallback behavior
  6. If newer versions offer better compatibility or performance, update to onnxruntime-gpu>=1.24.1,<2.0 (or specific newer pin)
  7. Update CI/CD testing to verify linkage across versions
  8. Document the upgrade path in the changelog

Related Issue: Check ONNX Runtime issues for "CUDA 13" and "CUDA compatibility"

CuPy Prerelease on CUDA 13

Status: 🔧 Prerelease Required

Current Configuration: Uses --pre flag and https://pip.cupy.dev/pre in requirements-cuda13.txt

File: requirements-cuda13.txt

Reason for Prerelease:

  • CUDA 13 stable cupy-cuda13x wheels may not be available from the main PyPI index
  • Prerelease wheels from CuPy's development server ensure CUDA 13 support

Action Required When Fixed:

  1. Monitor CuPy releases for CUDA 13 stable wheel availability
  2. When stable wheels are published:
  3. Remove --pre flag from requirements-cuda13.txt
  4. Remove custom index URL https://pip.cupy.dev/pre
  5. Update to pinned stable version (e.g., cupy-cuda13x==X.Y.Z)
  6. Test in CI with stable index to confirm compatibility
  7. Update docs/getting-started/installation.md to note that CUDA 13 installation is now fully stable

Related Issue: Monitor CuPy GitHub Releases for CUDA 13 stable tag

TensorRT Version-Specific Pins

Status: ✅ Versioned, May Improve

Current Configuration: Separate tensorrt-cu12, tensorrt-cu13 (unversioned) in version-specific files

Files: requirements-cuda12.txt, requirements-cuda13.txt

Considerations:

  • If NVIDIA publishes unified TensorRT CPU wheels with better version compatibility, we may simplify to a single tensorrt package
  • CUDA 13 support may improve with new releases; consider stricter pinning if compatibility issues emerge

Action Required When Fixed:

  1. Monitor NVIDIA TensorRT releases for unified CUDA version support
  2. If NVIDIA publishes CPU/GPU-agnostic wheels:
  3. Move TensorRT to requirements-cuda.txt (shared base)
  4. Remove version-specific TensorRT packages
  5. Simplify requirements structure
  6. If CUDA 13 TensorRT stability improves, consider pinning to a specific version range

Related Issue: NVIDIA TensorRT issue tracker

PyTorch Index URLs

Status: 📦 Version-Specific, Monitor

Current Configuration:

  • CUDA 12: --extra-index-url https://download.pytorch.org/whl/cu128
  • CUDA 13: --extra-index-url https://download.pytorch.org/whl/cu130

Files: requirements-cuda12.txt, requirements-cuda13.txt

Considerations:

  • PyTorch uses a custom index URL distribution strategy; check for future changes to their build/distribution infrastructure
  • CUDA 13 might be integrated into the main PyPI wheels in future major versions, eliminating the need for custom indexes

Action Required When Fixed:

  1. Monitor PyTorch installation docs for index URL changes
  2. If PyTorch integrates CUDA variants into standard PyPI:
  3. Remove custom index URLs from both CUDA requirement files
  4. Simplify to pinned PyTorch versions (e.g., torch>=2.1.0)
  5. Update CI to verify installation without custom indexes

Monitoring Checklist

When updating dependencies:

  • Check FAISS releases for CUDA 13 wheel availability
  • Review ONNX Runtime release notes and test compatibility
  • Monitor CuPy stable CUDA 13 wheel status
  • Evaluate TensorRT unified wheel roadmap
  • Check PyTorch for index URL or distribution strategy changes
  • Run full CI/CD suite with new dependencies
  • Update this document with resolved items

Summary Table

Package Current Fix File Priority Check Frequency
FAISS faiss-cpu fallback requirements-cuda13.txt High Monthly
ONNX Runtime Version pin 1.24.1 requirements-cuda.txt High Quarterly
CuPy Prerelease CUDA 13 requirements-cuda13.txt Medium Monthly
TensorRT Version-specific import requirements-cudaX.txt Medium Quarterly
PyTorch Custom index URLs requirements-cudaX.txt Low Quarterly

How to Report Resolved Items

When any of these items is resolved:

  1. Update the relevant requirement file
  2. Update this document, marking the item as ✅ Resolved
  3. Update related documentation (installation.md, environments.md)
  4. Add a note to CHANGELOG.md under the new release
  5. Remove the resolved item from the monitoring checklist