Add vGPU support (#52)

* Add vGPU support

* Add logging, safer orphan cleanup

* Add test

* Fix performance issue with looking up profile types

* Run as root

* Automatically install headers

* Update inference test to use DKMS

* Address review comments

* Performance optimized /resources endpoint for mdev scanning

* Fix remaining calculation

* Fix availability math

Author: sjmiller609

.air.toml

@@ -5,14 +5,14 @@ tmp_dir = "tmp"
 [build]
   args_bin = []
   bin = "./tmp/main"
-  cmd = "go build -tags containers_image_openpgp -o ./tmp/main ./cmd/api && sudo setcap cap_net_admin,cap_net_bind_service=+eip ./tmp/main"
+  cmd = "go build -tags containers_image_openpgp -o ./tmp/main ./cmd/api"
   delay = 1000
   exclude_dir = ["assets", "tmp", "vendor", "testdata", "bin", "scripts", "data", "kernel"]
   exclude_file = []
   exclude_regex = ["_test.go"]
   exclude_unchanged = false
   follow_symlink = false
-  full_bin = ""
+  full_bin = "sudo ./tmp/main"
   include_dir = []
   include_ext = ["go", "tpl", "tmpl", "html", "yaml"]
   include_file = []

cmd/api/api/instances.go

@@ -190,6 +190,14 @@ func (s *ApiService) CreateInstance(ctx context.Context, request oapi.CreateInst
 		hvType = hypervisor.Type(*request.Body.Hypervisor)
 	}
 
+	// Parse GPU configuration (vGPU mode)
+	var gpuConfig *instances.GPUConfig
+	if request.Body.Gpu != nil && request.Body.Gpu.Profile != nil && *request.Body.Gpu.Profile != "" {
+		gpuConfig = &instances.GPUConfig{
+			Profile: *request.Body.Gpu.Profile,
+		}
+	}
+
 	// Calculate default resource limits when not specified (0 = auto)
 	// Uses proportional allocation based on CPU: (vcpus / cpuCapacity) * resourceCapacity
 	if diskIOBps == 0 {
@@ -220,6 +228,7 @@ func (s *ApiService) CreateInstance(ctx context.Context, request oapi.CreateInst
 		Devices:                  deviceRefs,
 		Volumes:                  volumes,
 		Hypervisor:               hvType,
+		GPU:                      gpuConfig,
 	}
 
 	inst, err := s.InstanceManager.CreateInstance(ctx, domainReq)
@@ -685,5 +694,17 @@ func instanceToOAPI(inst instances.Instance) oapi.Instance {
 		oapiInst.Volumes = &oapiVolumes
 	}
 
+	// Convert GPU info
+	if inst.GPUProfile != "" {
+		gpu := &oapi.InstanceGPU{
+			Profile: lo.ToPtr(inst.GPUProfile),
+		}
+		// Only set MdevUuid when non-empty to avoid "mdev_uuid": "" in output
+		if inst.GPUMdevUUID != "" {
+			gpu.MdevUuid = lo.ToPtr(inst.GPUMdevUUID)
+		}
+		oapiInst.Gpu = gpu
+	}
+
 	return oapiInst
 }

cmd/api/api/resources.go

@@ -56,6 +56,12 @@ func (s *ApiService) GetResources(ctx context.Context, _ oapi.GetResourcesReques
 		})
 	}
 
+	// Add GPU status if available
+	if status.GPU != nil {
+		gpuStatus := convertGPUResourceStatus(status.GPU)
+		resp.Gpu = &gpuStatus
+	}
+
 	return oapi.GetResources200JSONResponse(resp), nil
 }
 
@@ -75,3 +81,38 @@ func convertResourceStatus(rs resources.ResourceStatus) oapi.ResourceStatus {
 		Source:         source,
 	}
 }
+
+func convertGPUResourceStatus(gs *resources.GPUResourceStatus) oapi.GPUResourceStatus {
+	result := oapi.GPUResourceStatus{
+		Mode:       oapi.GPUResourceStatusMode(gs.Mode),
+		TotalSlots: gs.TotalSlots,
+		UsedSlots:  gs.UsedSlots,
+	}
+
+	// Convert profiles (vGPU mode)
+	if len(gs.Profiles) > 0 {
+		profiles := make([]oapi.GPUProfile, len(gs.Profiles))
+		for i, p := range gs.Profiles {
+			profiles[i] = oapi.GPUProfile{
+				Name:          p.Name,
+				FramebufferMb: p.FramebufferMB,
+				Available:     p.Available,
+			}
+		}
+		result.Profiles = &profiles
+	}
+
+	// Convert devices (passthrough mode)
+	if len(gs.Devices) > 0 {
+		devices := make([]oapi.PassthroughDevice, len(gs.Devices))
+		for i, d := range gs.Devices {
+			devices[i] = oapi.PassthroughDevice{
+				Name:      d.Name,
+				Available: d.Available,
+			}
+		}
+		result.Devices = &devices
+	}
+
+	return result
+}

cmd/api/main.go

@@ -22,6 +22,7 @@ import (
 	"github.com/kernel/hypeman"
 	"github.com/kernel/hypeman/cmd/api/api"
 	"github.com/kernel/hypeman/cmd/api/config"
+	"github.com/kernel/hypeman/lib/devices"
 	"github.com/kernel/hypeman/lib/guest"
 	"github.com/kernel/hypeman/lib/hypervisor/qemu"
 	"github.com/kernel/hypeman/lib/instances"
@@ -200,6 +201,26 @@ func run() error {
 		return fmt.Errorf("reconcile device state: %w", err)
 	}
 
+	// Reconcile mdev devices (clears orphaned vGPUs from crashed VMs)
+	// Build mdev info from instances - only destroys mdevs tracked by hypeman
+	logger.Info("Reconciling mdev devices...")
+	var mdevInfos []devices.MdevReconcileInfo
+	if allInstances != nil {
+		for _, inst := range allInstances {
+			if inst.GPUMdevUUID != "" {
+				mdevInfos = append(mdevInfos, devices.MdevReconcileInfo{
+					InstanceID: inst.Id,
+					MdevUUID:   inst.GPUMdevUUID,
+					IsRunning:  inst.State == instances.StateRunning || inst.State == instances.StateUnknown,
+				})
+			}
+		}
+	}
+	if err := devices.ReconcileMdevs(app.Ctx, mdevInfos); err != nil {
+		// Log but don't fail - mdev cleanup is best-effort
+		logger.Warn("failed to reconcile mdev devices", "error", err)
+	}
+
 	// Initialize ingress manager (starts Caddy daemon and DNS server for dynamic upstreams)
 	logger.Info("Initializing ingress manager...")
 	if err := app.IngressManager.Initialize(app.Ctx); err != nil {