Overview

Currently, all-smi reads ROCm version for AMD GPUs but does not extract the kernel driver version. The libamdgpu_top library provides get_drm_version_struct() method that returns driver version information (major.minor.patchlevel format).

Current State

Already Implemented ✅

• ROCm Version: Retrieved via libamdgpu_top::get_rocm_version()
  • Source: $ROCM_PATH/.info/version file
  • Currently added to detail HashMap as "ROCm Version"
  • Location: src/device/readers/amd.rs:219-221

Not Implemented ❌

• Driver Version: Available but not extracted
  • Method: DeviceHandle::get_drm_version_struct() returns drmVersion struct
  • Fields: version_major, version_minor, version_patchlevel
  • Reference: references/amdgpu_top/crates/libamdgpu_top/src/app.rs:386-388
  • Reference: references/amdgpu_top/crates/amdgpu_top_json/src/dump.rs (JSON export example)

Implementation Tasks

1. AMD GPU Reader Enhancement

File: src/device/readers/amd.rs

• Add driver version extraction in collect_gpu_info() method
• Call device_handle.get_drm_version_struct() to get drmVersion
• Format as "major.minor.patchlevel" string (e.g., "6.12.0")
• Add to detail HashMap as "Driver Version"

Code location: Around line 210-220 where ROCm version is currently added

2. API Mode (Auto-exposed)

File: src/api/metrics/gpu.rs

• Verify all_smi_gpu_info metric automatically includes new driver_version label
• The export_device_info() function already dynamically adds all detail HashMap fields as labels
• Labels will be sanitized: "Driver Version" → "driver_version", "ROCm Version" → "rocm_version"

Expected output:

all_smi_gpu_info{gpu="AMD Instinct MI300X",instance="hostname",uuid="...",index="0",type="GPU",driver_version="6.12.0",rocm_version="6.3.0"} 1

3. Mock Server Enhancement

File: src/mock/templates/amd_gpu.rs

Currently, the AMD GPU mock template does NOT generate all_smi_gpu_info metric at all.

• Add add_gpu_info_metric() function similar to NVIDIA mock
• Include driver_version and rocm_version labels
• Use constants from src/mock/constants.rs:
  • Add DEFAULT_AMD_DRIVER_VERSION: &str = "6.12.0" (Linux kernel AMDGPU driver version)
  • Add DEFAULT_AMD_ROCM_VERSION: &str = "6.3.0"
• Call from add_gpu_metrics() after basic GPU metrics
• Ensure it does NOT include NVIDIA-specific labels (cuda_version)

Code location: After line 104 in build_amd_template()

4. Documentation Updates

• Update API.md AMD GPU section to list driver_version in metric labels table
• Update README.md if necessary

Technical References

drmVersion Struct

From references/amdgpu_top/crates/libamdgpu_top/src/xdna/bindings.rs:412-417:

pub struct drm_version {
    pub version_major: ::std::os::raw::c_int,
    pub version_minor: ::std::os::raw::c_int,
    pub version_patchlevel: ::std::os::raw::c_int,
    pub name_len: __kernel_size_t,
    pub name: *mut ::std::os::raw::c_char,
    // ... other fields
}

Usage Example

From references/amdgpu_top/crates/amdgpu_top_json/src/dump.rs:

let drm = self.get_drm_version_struct().map_or(Value::Null, |drm| json!({
    "major": drm.version_major,
    "minor": drm.version_minor,
    "patchlevel": drm.version_patchlevel,
}));

Expected Benefits

1. Consistency: Match NVIDIA GPU functionality which already exposes driver_version and cuda_version
2. Debugging: Easier troubleshooting with driver version information
3. Monitoring: Better cluster management with version tracking
4. API Completeness: Full hardware information in Prometheus metrics

Platform Requirements

• Linux with AMD GPU and ROCm/AMDGPU drivers
• glibc builds only (not available in musl due to library dependencies)
• Requires sudo or membership in video and render groups

Related Files

• src/device/readers/amd.rs - AMD GPU reader implementation
• src/api/metrics/gpu.rs - API metrics generation
• src/mock/templates/amd_gpu.rs - AMD GPU mock template
• src/mock/constants.rs - Mock server constants
• references/amdgpu_top/ - libamdgpu_top library reference

Acceptance Criteria

• Driver version appears in detail HashMap for AMD GPUs
• all_smi_gpu_info metric includes driver_version and rocm_version labels in API mode
• Mock server generates all_smi_gpu_info with appropriate AMD-specific labels
• Labels are properly sanitized (spaces to underscores)
• Documentation is updated
• No impact on NVIDIA or other platform metrics