Add GPU/Device Support and Fix Symlink Deduplication Issues

[akshaver]

Fixes-102, Fixes-306, Fixes-401, Fixes-579

What

This MR introduces four patches that add GPU support to docker-slim and fix critical symlink processing bugs:

1. Implement docker device request in CLI - Adds --cro-device-request and --cro-runtime CLI flags to enable GPU access during container profiling. Allows passing device requests as JSON (e.g., --cro-device-request '{"Count":-1, "Capabilities":[["gpu"]]}' --cro-runtime nvidia).

2. Fix issues with duplicate symlink processing - Resolves a bug where files accessed through multiple symlinked paths (e.g., /usr/local/cuda/ vs /usr/local/cuda-12.9/) would be copied multiple times, with later copies randomly overwriting with 0-byte content. Implements inode-based deduplication to ensure only one canonical copy is kept.

3. Example use case with nvidia runtime - Adds documentation and example scripts demonstrating GPU workloads:

   • test_nvidia_smi.sh - Slims ubuntu to run nvidia-smi
   • test_nvidia_pytorch.sh - Slims nvidia-pytorch to run CUDA tests

4. Non-trivial example of slimming vllm - Adds a comprehensive example for slimming VLLM (LLM inference) containers with full API test suite validation.

Why

• GPU Support (docker-slim removing cuda which is required for GPU computation #102, please help me how to slim gpu docker #401): Many modern workloads require GPU access, especially in ML/AI contexts. Without --cro-device-request and --cro-runtime, it was impossible to profile containers that required GPU access during execution, making docker-slim unusable for CUDA-based images.

• Symlink Bug (Incomplete Image Nvidia/PyTorch #306, Non usable python-gunicorn docker slimed image #579): The symlink deduplication bug caused slimmed images to have corrupted (0-byte) library files, particularly in NVIDIA CUDA containers where /usr/local/cuda symlinks to versioned directories like /usr/local/cuda-12.9. This made slimmed images non-functional, as critical .so files would be empty.

How Tested

• Unit Tests: Added comprehensive test coverage for:

  • Device request JSON parsing and CLI flag handling (clifvgetter_test.go, config_test.go, device_request_test.go)
  • Symlink deduplication logic (dedup_test.go)
  • Ptrace return status handling (ptrace_test.go)
  • File copy operations (fsutil_test.go)

• Integration Tests:

  • test_nvidia_smi.sh - Verifies slimmed ubuntu image can run nvidia-smi with identical output to original
  • test_nvidia_pytorch.sh - Verifies slimmed pytorch image passes CUDA tests
  • test_nvidia_vllm.sh - Comprehensive VLLM API test suite comparing original vs slimmed image behavior (15 endpoint tests)

• Manual Testing: Successfully slimmed NVIDIA NIM LLM containers with GPU workloads and verified functional parity.

[kcq]

thanks for the PR @akshaver ! Do you mind moving it here , so it'll be easier to review and integrate.

[akshaver]

thanks for the PR @akshaver ! Do you mind moving it here , so it'll be easier to review and integrate.

Will do.

[akshaver]

@kcq Done - See minttoolkit-179