Update NVIDIA --gpu requests using CDI to match AMD logic by elezar · Pull Request #52145 · moby/moby

elezar · 2026-03-05T14:28:02Z

- What I did

Refactored the handling of --gpus requests for NVIDIA devices to align with AMD device requests. See #52048.

- How I did it

Introduced a cdiCacheInjector and update the AMD-specific code to make use of this type. I then migrated the NVIDIA implementation to this.

- How to verify it

- Human readable description for the release notes

- A picture of a cute animal (not mandatory but encouraged)

shiv-tyagi · 2026-03-05T15:58:13Z

daemon/devices_linux.go

 // RegisterGPUDeviceDrivers registers GPU device drivers.
 // If the cdiCache is provided, it is used to detect presence of CDI specs for AMD GPUs.
 // For NVIDIA GPUs, presence of CDI specs is detected by checking for the nvidia-cdi-hook binary.


Suggested change

// RegisterGPUDeviceDrivers registers GPU device drivers.

// If the cdiCache is provided, it is used to detect presence of CDI specs for AMD GPUs.

// For NVIDIA GPUs, presence of CDI specs is detected by checking for the nvidia-cdi-hook binary.

// RegisterGPUDeviceDrivers registers GPU device drivers.

// If the cdiCache is provided, it is used to discover the vendor via available CDI specs

// and translate the GPU requests to CDI device requests.

@elezar ptal 🤗

Done. Sorry for the delay.

vvoland · 2026-03-17T10:52:53Z

daemon/devices_linux.go

 	}
 }
+
+// a cdiCacheInjector uses the specified CDI cache to inject device requests


Suggested change

// a cdiCacheInjector uses the specified CDI cache to inject device requests

// cdiCacheInjector uses the specified CDI cache to inject device requests

vvoland · 2026-03-17T10:59:47Z

daemon/devices_linux.go

+	if c.optionalVendor != "" && vendor != c.optionalVendor {
+		return fmt.Errorf("CDI specs for required vendor %v not found", c.optionalVendor)
+	}
+	injector := &cdiDeviceInjector{
+		defaultCDIDeviceKind: c.optionalVendor + "/gpu",
+	}


Shouldn't it be called requiredVendor (or just vendor)? Otherwise the injectDevices will error out.

Otherwise, if optionalVendor is empty, we end up with defaultCDIDeviceKind: "/gpu"

You're right. The logic here is not correct. Let me work through it and also add unit tests to capture the intent.

Signed-off-by: Evan Lezar <[email protected]>

This change adds a cdiCacheInjector type for handling --gpus requests as vendor-specific CDI requests. The handling of --gpus requests for AMD devices is updated to use this type. Signed-off-by: Evan Lezar <[email protected]>

This change aligns the handling of --gpus requests using CDI for NVIDIA device with those for AMD devices. Instead of checking for the existence of the nvidia-cdi-hook executable, available CDI vendors are checked. This makes the implementation more robust w.r.t changing implementation details in NVIDIA CDI specs. Signed-off-by: Evan Lezar <[email protected]>

thaJeztah · 2026-03-18T07:15:45Z

🙈 now a linter is failing

goanalysis_metalinter: 1m24.292750831s 
daemon/devices_nvidia_linux.go:223:29: ST1016: methods on the same type should have the same receiver name (seen 1x "c", 2x "i") (staticcheck)
func (i *cdiDeviceInjector) injectDevices(s *specs.Spec, dev *deviceInstance) error {
                            ^

Signed-off-by: Evan Lezar <[email protected]>

vvoland · 2026-03-18T09:01:31Z

daemon/devices_nvidia_linux.go

+	ListVendors() []string
+}
+
+// createCDIInjector creates a injector for CDI devices from the specified a set


Suggested change

// createCDIInjector creates a injector for CDI devices from the specified a set

// createCDIDeviceInjector creates an injector for CDI devices from the specified a set

vvoland · 2026-03-18T09:02:49Z

daemon/devices_nvidia_linux.go

+func (i *cdiDeviceInjector) resolveVendor() (string, error) {
+	vendor, err := i.validVendors.getFirstAvailableIn(i.availableVendors.ListVendors())
+	if err != nil {
+		return "", err
+	}
+	return vendor, nil


nit, can be simplified:

Suggested change

func (i *cdiDeviceInjector) resolveVendor() (string, error) {

vendor, err := i.validVendors.getFirstAvailableIn(i.availableVendors.ListVendors())

if err != nil {

return "", err

}

return vendor, nil

func (i *cdiDeviceInjector) resolveVendor() (string, error) {

return i.validVendors.getFirstAvailableIn(i.availableVendors.ListVendors())

elezar · 2026-03-18T09:03:00Z

daemon/devices_nvidia_linux.go

+	if cdiCache != nil {
+		// Create a device injector that handles --gpus device requests using
+		// nvidia.com CDI requests.
+		injector := createCDIDeviceInjector(cdiCache, "nvidia.com")
+		// Register a named device driver for this injector so that a user can
+		// explicitly request it.
 		// This has no capabilities associated to not inadvertently match requests.
-		cdiDeviceDriver := &deviceDriver{
-			updateSpec: (&cdiDeviceInjector{
-				defaultCDIDeviceKind: "nvidia.com/gpu",
-			}).injectDevices,
+		nvidiaDrivers["nvidia.cdi"] = &deviceDriver{
+			updateSpec: injector.injectDevices,


One issue with this approach is that this will ALWAYS register an NVIDIA driver and the AMD driver will not be used. I need to rethink this and also how we will test this.

I've marked this as draft for the time being.

github-actions bot added the area/daemon Core Engine label Mar 5, 2026

elezar changed the title ~~Cdi vendor detection~~ Update NVIDIA --gpu requests using CDI to match AMD logic Mar 5, 2026

elezar mentioned this pull request Mar 5, 2026

Use CDI for GPU injection for AMD devices for --gpus #52048

Merged

vvoland added the kind/refactor PR's that refactor, or clean-up code label Mar 5, 2026

vvoland added this to the 29.3.1 milestone Mar 5, 2026

shiv-tyagi reviewed Mar 5, 2026

View reviewed changes

elezar force-pushed the cdi-vendor-detection branch from c61a286 to 5a0ed51 Compare March 12, 2026 13:37

vvoland reviewed Mar 17, 2026

View reviewed changes

elezar added 3 commits March 17, 2026 16:17

chore: Fix comment for firstSuccessfulUpdater type

eac1b14

Signed-off-by: Evan Lezar <[email protected]>

elezar force-pushed the cdi-vendor-detection branch from b3a33e2 to dad2476 Compare March 17, 2026 15:17

Resolve vendors in cdiDeviceInjector instead of using default CDI kind

2fc090d

Signed-off-by: Evan Lezar <[email protected]>

elezar force-pushed the cdi-vendor-detection branch from dad2476 to 2fc090d Compare March 18, 2026 08:26

vvoland reviewed Mar 18, 2026

View reviewed changes

elezar commented Mar 18, 2026

View reviewed changes

elezar marked this pull request as draft March 18, 2026 11:25

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Update NVIDIA --gpu requests using CDI to match AMD logic#52145

Update NVIDIA --gpu requests using CDI to match AMD logic#52145
elezar wants to merge 4 commits intomoby:masterfrom
elezar:cdi-vendor-detection

elezar commented Mar 5, 2026 •

edited by vvoland

Loading

Uh oh!

shiv-tyagi Mar 5, 2026

Uh oh!

thaJeztah Mar 12, 2026

Uh oh!

elezar Mar 12, 2026

Uh oh!

vvoland Mar 17, 2026

Uh oh!

vvoland Mar 17, 2026

Uh oh!

elezar Mar 17, 2026

Uh oh!

thaJeztah commented Mar 18, 2026

Uh oh!

vvoland Mar 18, 2026

Uh oh!

vvoland Mar 18, 2026

Uh oh!

elezar Mar 18, 2026

Uh oh!

elezar Mar 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	// a cdiCacheInjector uses the specified CDI cache to inject device requests
	// cdiCacheInjector uses the specified CDI cache to inject device requests

	// createCDIInjector creates a injector for CDI devices from the specified a set
	// createCDIDeviceInjector creates an injector for CDI devices from the specified a set

Conversation

elezar commented Mar 5, 2026 • edited by vvoland Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

thaJeztah commented Mar 18, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

elezar commented Mar 5, 2026 •

edited by vvoland

Loading