[jenerestain]

Summary

Add experimental Linux discrete AMD RDNA3/RDNA4 support using the existing gpu.Probe contract. APUs/UMA, other AMD generations, and AMD Windows are explicitly excluded.

• Add an amdgpu sysfs/DRM adapter without ROCm, amd-smi, CGO, inference contexts, or new module dependencies.
• Join card/process observations by normalized PCI identity, retain DRM card numbering, support legacy/modern resident counters, deduplicate clients, and conservatively leave shared/unobservable bytes in the driver baseline.
• Add --gpu-backend auto|nvidia|amd to daemon/service selection and the swap benchmark. Preserve NVIDIA's existing probes and optional streaming; refuse ambiguous automatic selection.
• Guard AMD admissions against stale/failed samples and insufficient driver free bytes. Keep unreadable PIDs explicit, preserve sampled process identity, and require driver baseline confirmation before release/park completion.
• Add fixture/broker/service regressions, a repeatable opt-in bounded allocation/handoff check, and milestone/scope documentation in contrib/amd/README.md.

Validation

The unchanged starting tree passed the same baseline checks. Final implementation checked with Go 1.27.1 and the repository's exact linter version:

• go build ./... && test -z "$(gofmt -l .)" && go vet ./... && go test ./...
• go run github.com/golangci/golangci-lint/v2/cmd/golangci-lint@v2.14.0 run ./... — 0 issues
• GOOS=windows go build ./... && GOOS=windows go vet ./...
• go test -race ./internal/gpu ./internal/broker ./internal/cli
• Opt-in read-only probe against a W7900, kernel 6.12.107+deb13-amd64, including existing ROCm and Vulkan clients
• Isolated W7900 GEM allocation -> measured PID attribution -> higher-priority drain -> driver-confirmed release -> next allocation, repeated after final changes

The live test used sequential 64 MiB buffers, its own daemon/socket/config, and no installed service. Initial and final driver usage matched exactly at 41,657,159,680 bytes. Existing workload process identities and measured memory were unchanged; the 780M was excluded. The event trace confirmed lease.freed for the first child before lease.granted for the second.

Explicit limits / follow-up gates

Integration with current main

Merged upstream main through edc6cd9 without rebasing/force-pushing. Preserved its documentation split and VFIO hold/probe-suppression behavior, and added a guarded-AMD VFIO recovery regression. The combined tree passed the isolated W7900 allocation/drain/handoff check again. In this later run, existing workloads continued with the same process identities while one workload's allocations changed; the unattributed driver baseline returned to its original value and both test children exited. The unchanged-total figure above refers to the earlier stable-background runs, not a promise that other workloads stop allocating.

The existing VFIO command remains NVIDIA-specific; AMD VFIO CLI support and physical passthrough validation are not included.

All local checks above passed again on the integrated tree. GitHub's CI and docs workflows currently report action_required with no jobs started for this fork PR; maintainer approval is needed before upstream CI, including native Windows tests, can be confirmed green. No merge is requested until that gate passes.

• RDNA4 GC 12.0.0/12.0.1 is fixture-tested, not physically tested here. Run the documented read-only and opt-in allocation checks on a real RDNA4 card before claiming hardware validation.
• No NVIDIA GPU was available for live regression testing; NVIDIA tests and Windows build/vet pass. Native Windows execution remains a CI gate.
• The live allocation test validates the driver/lease path, not model correctness or engine-specific ROCm park/resume support. Park safety has broker regression coverage; actual engine/runtime capabilities still matter.
• Shared/unobservable background allocations can conservatively delay release confirmation. Whole-card leases retain report-only contested-grant semantics; no reservations or engine device isolation are claimed.
• No unrelated fixes, package/version changes, production workload restarts, or APU/UMA scheduling changes.

[jenerestain]

Closing at the author's request so APU support and validation on a Radeon 780M can be completed before resubmitting AMD support.