diff --git a/docs/content/features/GPU-acceleration.md b/docs/content/features/GPU-acceleration.md
index fb939a47a..085617c24 100644
--- a/docs/content/features/GPU-acceleration.md
+++ b/docs/content/features/GPU-acceleration.md
@@ -458,6 +458,8 @@ watch -n1 rocm-smi
 rocm-smi --showallinfo
 ```
 
+LocalAI reads AMD GPU memory for `/api/resources` and the model fit check from `rocm-smi`. ROCm 7 and later images no longer ship `rocm-smi`; there LocalAI reads the same figures from the amdgpu driver in sysfs (`/sys/class/drm/card*/device/mem_info_*`). On APUs with a small VRAM carve-out and a large GTT pool (for example Strix Halo), the GTT pool counts toward GPU memory in both cases.
+
 ### Intel
 
 ```bash
diff --git a/pkg/xsysinfo/amd_sysfs_meminfo_test.go b/pkg/xsysinfo/amd_sysfs_meminfo_test.go
new file mode 100644
index 000000000..5e75a3781
--- /dev/null
+++ b/pkg/xsysinfo/amd_sysfs_meminfo_test.go
@@ -0,0 +1,87 @@
+package xsysinfo
+
+import (
+	"os"
+	"path/filepath"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("readAMDSysfsGPUMemory", func() {
+	var root string
+
+	BeforeEach(func() {
+		root = filepath.Join(GinkgoT().TempDir(), "class", "drm")
+		Expect(os.MkdirAll(root, 0o755)).To(Succeed())
+	})
+
+	// Values read inside a ROCm 7+ container on Strix Halo (gfx1151),
+	// where rocm-smi is no longer shipped.
+	strixHalo := map[string]string{
+		"vendor":              "0x1002\n",
+		"mem_info_vram_total": "536870912\n",
+		"mem_info_vram_used":  "240300032\n",
+		"mem_info_gtt_total":  "128849018880\n",
+		"mem_info_gtt_used":   "33055911936\n",
+	}
+
+	It("reports a shared-memory APU with GTT folded into totals", func() {
+		writeSysfsCard(root, "card0", "0000:c5:00.0", strixHalo)
+
+		got := readAMDSysfsGPUMemory(root)
+		Expect(got).To(HaveLen(1))
+		Expect(got[0].Vendor).To(Equal(VendorAMD))
+		Expect(got[0].Index).To(Equal(0))
+		Expect(got[0].TotalVRAM).To(Equal(uint64(536870912 + 128849018880)))
+		Expect(got[0].UsedVRAM).To(Equal(uint64(240300032 + 33055911936)))
+		Expect(got[0].FreeVRAM).To(Equal(got[0].TotalVRAM - got[0].UsedVRAM))
+		Expect(got[0].UsagePercent).To(BeNumerically(">", 0))
+	})
+
+	It("keeps a discrete card on VRAM only", func() {
+		writeSysfsCard(root, "card1", "0000:03:00.0", map[string]string{
+			"vendor":              "0x1002\n",
+			"mem_info_vram_total": "25769803776\n",
+			"mem_info_vram_used":  "1073741824\n",
+			"mem_info_gtt_total":  "34359738368\n",
+			"mem_info_gtt_used":   "1048576\n",
+		})
+
+		got := readAMDSysfsGPUMemory(root)
+		Expect(got).To(HaveLen(1))
+		Expect(got[0].Index).To(Equal(1))
+		Expect(got[0].TotalVRAM).To(Equal(uint64(25769803776)))
+		Expect(got[0].UsedVRAM).To(Equal(uint64(1073741824)))
+	})
+
+	It("ignores non-AMD cards and connector entries", func() {
+		writeSysfsCard(root, "card0", "0000:c5:00.0", strixHalo)
+		// A connector hanging off the same device must not count twice.
+		Expect(os.MkdirAll(filepath.Join(root, "card0-DP-1"), 0o755)).To(Succeed())
+		Expect(os.Symlink(filepath.Join(root, "..", "devices", "pci", "0000:c5:00.0"),
+			filepath.Join(root, "card0-DP-1", "device"))).To(Succeed())
+		// An NVIDIA card exposing the same attribute names is not ours.
+		nv := map[string]string{}
+		for k, v := range strixHalo {
+			nv[k] = v
+		}
+		nv["vendor"] = "0x10de\n"
+		writeSysfsCard(root, "card1", "0000:01:00.0", nv)
+
+		got := readAMDSysfsGPUMemory(root)
+		Expect(got).To(HaveLen(1))
+		Expect(got[0].Index).To(Equal(0))
+	})
+
+	It("reports nothing for an AMD card without mem_info attributes", func() {
+		writeSysfsCard(root, "card0", "0000:c5:00.0", map[string]string{
+			"vendor": "0x1002\n",
+		})
+		Expect(readAMDSysfsGPUMemory(root)).To(BeEmpty())
+	})
+
+	It("reports nothing when the DRM tree is missing", func() {
+		Expect(readAMDSysfsGPUMemory(filepath.Join(root, "nope"))).To(BeEmpty())
+	})
+})
diff --git a/pkg/xsysinfo/gpu.go b/pkg/xsysinfo/gpu.go
index e017096b2..ab167ab68 100644
--- a/pkg/xsysinfo/gpu.go
+++ b/pkg/xsysinfo/gpu.go
@@ -614,10 +614,20 @@ func getNVIDIAGPUMemory() []GPUMemoryInfo {
 	return gpus
 }
 
-// getAMDGPUMemory queries AMD GPUs using rocm-smi.
-// Prefers --showmeminfo all so APU GTT pools are visible (#12058); falls
-// back to vram-only on older rocm-smi builds that reject "all".
+// getAMDGPUMemory queries AMD GPUs using rocm-smi, falling back to the
+// amdgpu sysfs attributes when rocm-smi is absent (ROCm 7+ images ship
+// only amd-smi) or reports nothing.
 func getAMDGPUMemory() []GPUMemoryInfo {
+	if gpus := getAMDRocmSMIMemory(); len(gpus) > 0 {
+		return gpus
+	}
+	return readAMDSysfsGPUMemory(defaultSysfsDRMPath)
+}
+
+// getAMDRocmSMIMemory prefers --showmeminfo all so APU GTT pools are
+// visible (#12058); falls back to vram-only on older rocm-smi builds that
+// reject "all".
+func getAMDRocmSMIMemory() []GPUMemoryInfo {
 	if _, err := exec.LookPath("rocm-smi"); err != nil {
 		return nil
 	}
@@ -758,38 +768,45 @@ func parseRocmSMIMemInfoCSV(output string) []GPUMemoryInfo {
 			gttTotal, gttUsed = scaleRocmMemIfMB(gttTotal, gttUsed)
 		}
 
-		totalBytes, usedBytes := vramTotal, vramUsed
-		if amdSharedMemoryAPU(vramTotal, gttTotal) {
-			totalBytes = vramTotal + gttTotal
-			usedBytes = vramUsed + gttUsed
-			xlog.Debug("AMD shared-memory APU detected; folding GTT into VRAM totals",
-				"vram_total", vramTotal, "gtt_total", gttTotal, "combined_total", totalBytes)
-		}
-
-		freeBytes := uint64(0)
-		if totalBytes > usedBytes {
-			freeBytes = totalBytes - usedBytes
-		}
-
-		usagePercent := 0.0
-		if totalBytes > 0 {
-			usagePercent = float64(usedBytes) / float64(totalBytes) * 100
-		}
-
-		gpus = append(gpus, GPUMemoryInfo{
-			Index:        parseAMDDeviceIndex(parts[0]),
-			Name:         "AMD GPU",
-			Vendor:       VendorAMD,
-			TotalVRAM:    totalBytes,
-			UsedVRAM:     usedBytes,
-			FreeVRAM:     freeBytes,
-			UsagePercent: usagePercent,
-		})
+		gpus = append(gpus, amdGPUMemoryInfo(parseAMDDeviceIndex(parts[0]),
+			vramTotal, vramUsed, gttTotal, gttUsed))
 	}
 
 	return gpus
 }
 
+// amdGPUMemoryInfo builds the report for one AMD GPU, folding GTT into
+// the totals when the sizes match a shared-memory APU (#12058).
+func amdGPUMemoryInfo(index int, vramTotal, vramUsed, gttTotal, gttUsed uint64) GPUMemoryInfo {
+	totalBytes, usedBytes := vramTotal, vramUsed
+	if amdSharedMemoryAPU(vramTotal, gttTotal) {
+		totalBytes = vramTotal + gttTotal
+		usedBytes = vramUsed + gttUsed
+		xlog.Debug("AMD shared-memory APU detected; folding GTT into VRAM totals",
+			"vram_total", vramTotal, "gtt_total", gttTotal, "combined_total", totalBytes)
+	}
+
+	freeBytes := uint64(0)
+	if totalBytes > usedBytes {
+		freeBytes = totalBytes - usedBytes
+	}
+
+	usagePercent := 0.0
+	if totalBytes > 0 {
+		usagePercent = float64(usedBytes) / float64(totalBytes) * 100
+	}
+
+	return GPUMemoryInfo{
+		Index:        index,
+		Name:         "AMD GPU",
+		Vendor:       VendorAMD,
+		TotalVRAM:    totalBytes,
+		UsedVRAM:     usedBytes,
+		FreeVRAM:     freeBytes,
+		UsagePercent: usagePercent,
+	}
+}
+
 func looksLikeRocmMemInfoHeader(line string) bool {
 	lower := strings.ToLower(line)
 	return strings.Contains(lower, "device") && strings.Contains(lower, "memory")
diff --git a/pkg/xsysinfo/sysfsgpu.go b/pkg/xsysinfo/sysfsgpu.go
index e32686a2a..170f5f716 100644
--- a/pkg/xsysinfo/sysfsgpu.go
+++ b/pkg/xsysinfo/sysfsgpu.go
@@ -200,3 +200,44 @@ func ghwHasVendor(cards []*gpu.GraphicsCard, vendor string) bool {
 	}
 	return false
 }
+
+// readAMDSysfsGPUMemory reports AMD GPU memory from the amdgpu driver's
+// mem_info_* attributes under the DRM sysfs tree.
+//
+// It is the fallback for images without rocm-smi: ROCm 7+ images ship
+// only amd-smi, so a working Strix Halo reported "No GPU found" and
+// gpu_count 0 while inference ran on the GPU. The kernel exposes the
+// same VRAM and GTT figures rocm-smi reads, so the shared-memory APU
+// rule (#12058) applies unchanged.
+func readAMDSysfsGPUMemory(root string) []GPUMemoryInfo {
+	var gpus []GPUMemoryInfo
+	for _, card := range scanSysfsGPUs(root) {
+		if card.Vendor != VendorAMD {
+			continue
+		}
+		deviceDir := filepath.Join(root, card.Card, "device")
+		vramTotal, ok := readSysfsDecUint(filepath.Join(deviceDir, "mem_info_vram_total"))
+		if !ok {
+			continue
+		}
+		vramUsed, _ := readSysfsDecUint(filepath.Join(deviceDir, "mem_info_vram_used"))
+		gttTotal, _ := readSysfsDecUint(filepath.Join(deviceDir, "mem_info_gtt_total"))
+		gttUsed, _ := readSysfsDecUint(filepath.Join(deviceDir, "mem_info_gtt_used"))
+		gpus = append(gpus, amdGPUMemoryInfo(parseAMDDeviceIndex(card.Card),
+			vramTotal, vramUsed, gttTotal, gttUsed))
+	}
+	return gpus
+}
+
+func readSysfsDecUint(path string) (uint64, bool) {
+	// #nosec G304 -- same provenance as readSysfsHexUint.
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		return 0, false
+	}
+	value, err := strconv.ParseUint(strings.TrimSpace(string(raw)), 10, 64)
+	if err != nil {
+		return 0, false
+	}
+	return value, true
+}