feat(node-info): report Linux AMD GPU utilization

services/nvpair-node-info/README.md

@@ -7,6 +7,19 @@ SPDX-License-Identifier: Apache-2.0
 
 A Go service that exposes this machine's hardware inventory (GPUs, CPU, physical memory) over a small HTTP API at `/v1/node-info`. It advertises nothing over mDNS itself — its parent (the broker) registers its `ni` port with the `nvpair-node-scanner` discovery daemon, which folds it into this node's single `_nvpair-node` record and fetches `/v1/node-info` to enrich the node for peers.
 
+## Linux AMD telemetry
+
+AMD adapters use PCI identity and the amdgpu `gpu_busy_percent` sysfs counter.
+This requires neither ROCm nor CUDA. Missing, unreadable, malformed, or out of
+range samples are unavailable. Valid samples, including idle samples, refresh
+the existing `telemetryValid` and `msSince` freshness contract. The existing
+wire format omits a zero utilization value, so inspect freshness as well.
+
+AMD memory fields remain omitted. On Strix Halo, reserved framebuffer VRAM,
+GTT capacity, the HSA GPU-visible pool, and system MemAvailable are distinct
+and overlapping measurements. This change does not label system RAM or GTT
+as dedicated VRAM. NVIDIA collection remains enabled on mixed-vendor hosts.
+
 ## Communication
 
 Two surfaces:

services/nvpair-node-info/gpu_linux.go

@@ -8,7 +8,9 @@ package main
 import (
 	"context"
 	"log"
+	"os"
 	"os/exec"
+	"path/filepath"
 	"strconv"
 	"strings"
 	"time"
@@ -22,13 +24,15 @@ import (
 // collector from blocking indefinitely.
 const nvidiaSmiTimeout = 3 * time.Second
 
+const amdPCIRoot = "/sys/bus/pci/devices"
+
 // detectGPUs enumerates GPUs on Linux. It prefers nvidia-smi, which yields the
 // marketing name, total VRAM, and a stable per-GPU UUID we reuse as the join
 // key (statsKey) against the dynamic stats collector's snapshot. When
 // nvidia-smi is absent — no NVIDIA driver, or an AMD/Intel-only host — it falls
-// back to ghw, which reports adapter names but no VRAM and no join key, so
-// those hosts list their GPUs without dynamic VRAM/utilization (matching the
-// pre-existing non-Windows behavior).
+// back to ghw. AMD adapters receive a PCI-derived join key for dynamic
+// utilization, while other adapters list their names without dynamic VRAM or
+// utilization (matching the pre-existing non-Windows behavior).
 //
 // On unified-memory architectures (UMA, e.g. Grace-Blackwell / DGX Spark)
 // nvidia-smi reports [N/A] for memory.total because the GPU shares system
@@ -45,6 +49,12 @@ func detectGPUs() []GPUInfo {
 					}
 				}
 			}
+			// Preserve NVIDIA records and include AMD adapters on mixed hosts.
+			for _, gpu := range detectGPUsGHW() {
+				if strings.HasPrefix(gpu.statsKey, "amd:") {
+					gpus = append(gpus, gpu)
+				}
+			}
 			return gpus
 		}
 	}
@@ -53,7 +63,8 @@ func detectGPUs() []GPUInfo {
 
 // detectGPUsGHW is the ghw-based fallback, identical in spirit to the
 // non-Windows/non-Linux path in gpu_other.go: enumerate display adapters and
-// return names only (VramBytes stays 0, statsKey stays empty).
+// return adapter names. AMD adapters also receive PCI-derived stats keys for
+// dynamic utilization; VramBytes stays 0 for every ghw-derived adapter.
 func detectGPUsGHW() []GPUInfo {
 	gpu, err := ghw.GPU()
 	if err != nil {
@@ -66,11 +77,56 @@ func detectGPUsGHW() []GPUInfo {
 		if card.DeviceInfo != nil && card.DeviceInfo.Product != nil {
 			name = card.DeviceInfo.Product.Name
 		}
-		gpus = append(gpus, GPUInfo{Name: name})
+		gpus = append(gpus, GPUInfo{Name: name, statsKey: amdStatsKey(amdPCIRoot, card.Address)})
 	}
 	return gpus
 }
 
+// amdStatsKey identifies AMD adapters by PCI address, independently of DRM
+// card numbering. The root argument permits tests without real hardware.
+func amdStatsKey(root, address string) string {
+	if !strings.Contains(address, ":") || filepath.Base(address) != address {
+		return ""
+	}
+	if strings.Count(address, ":") == 1 {
+		address = "0000:" + address
+	}
+	data, err := os.ReadFile(filepath.Join(root, address, "vendor"))
+	if err != nil || strings.TrimSpace(string(data)) != "0x1002" {
+		return ""
+	}
+	return "amd:" + address
+}
+
+// decodeAMDUtilization reads the amdgpu driver counter without a ROCm library,
+// subprocess or elevated privilege. Missing or invalid counters stay absent.
+// Memory is deliberately not mapped to VRAM: Strix Halo's GTT, reserved VRAM,
+// HSA pool and MemAvailable have different semantics and overlap.
+func decodeAMDUtilization(root string, out map[string]gpuStat) bool {
+	entries, err := os.ReadDir(root)
+	if err != nil {
+		return false
+	}
+	valid := false
+	for _, entry := range entries {
+		key := amdStatsKey(root, entry.Name())
+		if key == "" {
+			continue
+		}
+		data, err := os.ReadFile(filepath.Join(root, entry.Name(), "gpu_busy_percent"))
+		if err != nil {
+			continue
+		}
+		pct, err := strconv.ParseUint(strings.TrimSpace(string(data)), 10, 32)
+		if err != nil || pct > 100 {
+			continue
+		}
+		out[key] = gpuStat{UtilizationPct: uint32(pct)}
+		valid = true
+	}
+	return valid
+}
+
 // nvidiaSmiCSV runs `nvidia-smi --query-gpu=<fields> --format=csv,noheader,nounits`
 // and returns raw stdout. The caller parses the comma-separated rows. A missing
 // binary (not on PATH) surfaces as an exec error, which callers treat as "no

services/nvpair-node-info/gpu_linux_test.go

@@ -0,0 +1,57 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+//go:build linux
+
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestAMDUtilizationAvailability(t *testing.T) {
+	root := t.TempDir()
+	dir := filepath.Join(root, "0000:c5:00.0")
+	if err := os.MkdirAll(dir, 0700); err != nil {
+		t.Fatal(err)
+	}
+	write := func(name, value string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(value), 0600); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("vendor", "0x1002\n")
+	if got := amdStatsKey(root, "c5:00.0"); got != "amd:0000:c5:00.0" {
+		t.Fatal(got)
+	}
+	for _, value := range []string{"0", "83", "100", "101", "-1", "N/A"} {
+		write("gpu_busy_percent", value)
+		out := map[string]gpuStat{"NVIDIA-existing": {UtilizationPct: 7}}
+		valid := decodeAMDUtilization(root, out)
+		want := value == "0" || value == "83" || value == "100"
+		if valid != want {
+			t.Fatalf("%q valid=%v", value, valid)
+		}
+		_, exists := out["amd:0000:c5:00.0"]
+		if exists != want {
+			t.Fatalf("%q fabricated or lost sample", value)
+		}
+		if out["NVIDIA-existing"].UtilizationPct != 7 {
+			t.Fatal("NVIDIA overwritten")
+		}
+	}
+	if err := os.Remove(filepath.Join(dir, "gpu_busy_percent")); err != nil {
+		t.Fatal(err)
+	}
+	if decodeAMDUtilization(root, map[string]gpuStat{}) {
+		t.Fatal("missing counter marked valid")
+	}
+	write("vendor", "0x8086")
+	write("gpu_busy_percent", "70")
+	if decodeAMDUtilization(root, map[string]gpuStat{}) {
+		t.Fatal("non-AMD adapter accepted")
+	}
+}

services/nvpair-node-info/stats_linux.go

@@ -143,7 +143,9 @@ func (c *statsCollector) decodeSnapshot() *statsSnapshot {
 
 	gpu := make(map[string]gpuStat)
 	sampledAt := time.Time{}
-	if c.decodeGPU(gpu) {
+	nvidiaValid := c.decodeGPU(gpu)
+	amdValid := decodeAMDUtilization(amdPCIRoot, gpu)
+	if nvidiaValid || amdValid {
 		sampledAt = time.Now()
 	}
 	applyGPUStats(previous, snap, gpu, sampledAt)
@@ -161,7 +163,7 @@ func (c *statsCollector) decodeGPU(out map[string]gpuStat) bool {
 	csv, err := nvidiaSmiCSV("uuid,utilization.gpu,memory.used")
 	if err != nil {
 		if c.nvidiaUnavailable.CompareAndSwap(false, true) {
-			slog.Warn("nvidia-smi unavailable; GPU utilization / dedicated VRAM-used will not be reported",
+			slog.Debug("nvidia-smi unavailable; NVIDIA metrics unavailable; other GPU sources remain enabled",
 				"err", err)
 		}
 		return false

services/versions.json

@@ -5,7 +5,7 @@
   "components": {
     "ollama-proxy": "0.26.2",
     "lmstudio-proxy": "0.16.2",
-    "nvpair-node-info": "0.13.3",
+    "nvpair-node-info": "0.14.0",
     "nvpair-node-scanner": "0.20.3",
     "nvpair-manual-nodes": "0.11.1",
     "nvpair-workload-manager": "0.13.3",