feat(system): report per-model DRM VRAM

Expose optional resident device memory for local backend process trees.
Deduplicate DRM clients and omit unsupported or incomplete readings.
Document accounting limits and preserve a measured zero in JSON.

Closes #11970.

Assisted-by: Codex:gpt-6
This commit is contained in:
localai-org-maint-bot committed 2026-09-13 12:12:43 +00:00
1 parent d463316caf
commit da2342a359
11 files changed
+401 -1

No files matched your search

+11
View File
@@ -1,10 +1,13 @@
package localai
import (
"strconv"
"github.com/labstack/echo/v4"
"github.com/mudler/LocalAI/core/config"
"github.com/mudler/LocalAI/core/schema"
"github.com/mudler/LocalAI/pkg/model"
"github.com/mudler/LocalAI/pkg/xsysinfo"
)
// SystemInformations returns the system informations
@@ -32,6 +35,14 @@ func SystemInformations(cl *config.ModelConfigLoader, ml *model.ModelLoader, app
if cfg, ok := cl.GetModelConfig(m.ID); ok {
entry.Backend = cfg.Backend
}
if process := m.Process(); process != nil {
pid, err := strconv.Atoi(process.CurrentPID())
if err == nil {
if used, ok := xsysinfo.ProcessVRAM(pid); ok {
entry.SizeVRAM = &used
}
}
}
sysmodels = append(sysmodels, entry)
}
return c.JSON(200,
@@ -0,0 +1,49 @@
// SPDX-License-Identifier: MIT
package localai_test
import (
"encoding/json"
"net/http"
"net/http/httptest"
"os"
"path/filepath"
"github.com/labstack/echo/v4"
"github.com/mudler/LocalAI/core/config"
"github.com/mudler/LocalAI/core/http/endpoints/localai"
"github.com/mudler/LocalAI/pkg/model"
"github.com/mudler/LocalAI/pkg/system"
process "github.com/mudler/go-processmanager"
. "github.com/onsi/ginkgo/v2"
. "github.com/onsi/gomega"
)
var _ = Describe("SystemInformations memory", func() {
It("keeps model metadata and omits VRAM for remote or stopped backends", func() {
path, err := os.MkdirTemp("", "system-info-")
Expect(err).NotTo(HaveOccurred())
DeferCleanup(os.RemoveAll, path)
configFile := filepath.Join(path, "remote.yaml")
Expect(os.WriteFile(configFile, []byte("name: remote\nbackend: llama-cpp\n"), 0600)).To(Succeed())
cl := config.NewModelConfigLoader(path)
Expect(cl.ReadModelConfig(configFile)).To(Succeed())
ml := model.NewModelLoader(&system.SystemState{})
store := model.NewInMemoryModelStore()
store.Set("remote", model.NewModel("remote", "worker:50051", nil))
store.Set("stopped", model.NewModel("stopped", "", &process.Process{}))
ml.SetModelStore(store)
app := echo.New()
app.GET("/system", localai.SystemInformations(cl, ml, &config.ApplicationConfig{}))
rec := httptest.NewRecorder()
app.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/system", nil))
Expect(rec.Code).To(Equal(http.StatusOK))
var response struct {
Models []map[string]any `json:"loaded_models"`
}
Expect(json.Unmarshal(rec.Body.Bytes(), &response)).To(Succeed())
Expect(response.Models).To(ConsistOf(
map[string]any{"id": "remote", "backend": "llama-cpp"},
map[string]any{"id": "stopped"},
))
})
})
+3
View File
@@ -204,6 +204,9 @@ type SysInfoModel struct {
// so it is resolved from the model's config; empty when the model was
// loaded without one (a loose file, or a config since removed).
Backend string `json:"backend,omitempty"`
// SizeVRAM is DRM-accounted resident device memory in bytes. Nil means
// the backend process tree has no complete supported reading.
SizeVRAM *uint64 `json:"size_vram,omitempty"`
}
type SystemInformationResponse struct {
+25
View File
@@ -0,0 +1,25 @@
// SPDX-License-Identifier: MIT
package schema_test
import (
"encoding/json"
"github.com/mudler/LocalAI/core/schema"
. "github.com/onsi/ginkgo/v2"
. "github.com/onsi/gomega"
)
var _ = Describe("SysInfoModel memory", func() {
It("omits unavailable VRAM while preserving a measured zero", func() {
entry := schema.SysInfoModel{ID: "model"}
encoded, err := json.Marshal(entry)
Expect(err).NotTo(HaveOccurred())
Expect(string(encoded)).To(MatchJSON(`{"id":"model"}`))
zero := uint64(0)
entry.SizeVRAM = &zero
encoded, err = json.Marshal(entry)
Expect(err).NotTo(HaveOccurred())
Expect(string(encoded)).To(MatchJSON(`{"id":"model","size_vram":0}`))
})
})
+27 -1
View File
@@ -21,6 +21,30 @@ Returns available backends and currently loaded models.
| `backends` | `array` | List of available backend names (strings) |
| `loaded_models` | `array` | List of currently loaded models |
| `loaded_models[].id` | `string` | Model identifier |
| `loaded_models[].backend` | `string` | Backend name, when a model configuration is available |
| `loaded_models[].size_vram` | `integer` | Optional DRM-accounted resident device memory, in bytes |
### Per-model VRAM
On Linux, `size_vram` reports resident device memory for the local backend
process and its child processes. LocalAI reads `drm-resident-local*` and
`drm-resident-vram*` from `/proc` and counts each DRM client once per GPU.
Host-memory regions are excluded. The reading includes buffers attributed
to the backend, without separating weights, KV cache, and other allocations.
See the [kernel DRM accounting specification](https://docs.kernel.org/gpu/drm-usage-stats.html)
for these counters.
The field is omitted when accounting is unavailable or incomplete. This
includes external and distributed backends, macOS, proprietary NVIDIA
drivers, primary DRM nodes (`/dev/dri/card*`), missing resident counters,
and unreadable process information.
A present value of `0` means the supported counters report zero bytes.
Treat an absent field as unknown.
This is a snapshot of driver accounting, not a memory reservation. Shared
buffers can appear in different clients' counters, and allocations can change
during collection. Do not treat the sum across models as exclusive physical
GPU usage. These readings do not replace capacity checks when scheduling work.
### Usage
@@ -40,7 +64,9 @@ curl http://localhost:8080/system
],
"loaded_models": [
{
"id": "my-llama-model"
"id": "my-llama-model",
"backend": "llama-cpp",
"size_vram": 5368709120
},
{
"id": "whisper-1"
+159
View File
@@ -0,0 +1,159 @@
//go:build linux
// SPDX-License-Identifier: MIT
package xsysinfo
import (
"bufio"
"bytes"
"math"
"os"
"path/filepath"
"strconv"
"strings"
)
// ProcessVRAM reports device-local resident bytes accounted to a process tree
// by DRM. Unsupported or incomplete accounting returns false, not a measured zero.
func ProcessVRAM(pid int) (uint64, bool) {
return processVRAM("/proc", pid)
}
func processVRAM(procRoot string, pid int) (uint64, bool) {
if pid <= 0 {
return 0, false
}
clients := map[string]uint64{}
seen := map[int]bool{}
pending := []int{pid}
for len(pending) > 0 {
current := pending[len(pending)-1]
pending = pending[:len(pending)-1]
if seen[current] {
continue
}
seen[current] = true
base := filepath.Join(procRoot, strconv.Itoa(current))
fds, err := os.ReadDir(filepath.Join(base, "fd"))
if err != nil {
return 0, false
}
for _, fd := range fds {
target, err := os.Readlink(filepath.Join(base, "fd", fd.Name()))
if err != nil {
return 0, false
}
// A mixed DRM/NVIDIA tree cannot provide a complete DRM reading.
if strings.HasPrefix(target, "/dev/nvidia") {
return 0, false
}
if !strings.HasPrefix(target, "/dev/dri/render") {
// Primary nodes can also own allocations. Until their device
// identity is resolved, omitting them would undercount the tree.
if strings.HasPrefix(target, "/dev/dri/") {
return 0, false
}
continue
}
data, err := os.ReadFile(filepath.Join(base, "fdinfo", fd.Name()))
if err != nil {
return 0, false
}
client, used, ok := drmResidentClient(data)
if !ok {
return 0, false
}
key := target + ":" + client
// dup() and fork() can expose the same client more than once. The
// snapshot is not atomic; retain its largest observed reading.
clients[key] = max(clients[key], used)
}
// A worker may be spawned by any thread, not just the thread leader.
tasks, err := os.ReadDir(filepath.Join(base, "task"))
if err != nil || len(tasks) == 0 {
return 0, false
}
for _, task := range tasks {
data, err := os.ReadFile(filepath.Join(base, "task", task.Name(), "children"))
if err != nil {
return 0, false
}
for _, raw := range strings.Fields(string(data)) {
child, err := strconv.Atoi(raw)
if err != nil || child <= 0 {
return 0, false
}
pending = append(pending, child)
}
}
}
var total uint64
for _, used := range clients {
if used > math.MaxUint64-total {
return 0, false
}
total += used
}
return total, len(clients) > 0
}
func drmResidentClient(data []byte) (string, uint64, bool) {
var client string
var total uint64
found := false
scanner := bufio.NewScanner(bytes.NewReader(data))
for scanner.Scan() {
key, value, ok := strings.Cut(scanner.Text(), ":")
if !ok {
continue
}
if key == "drm-client-id" {
id, err := strconv.ParseUint(strings.TrimSpace(value), 10, 64)
if err != nil {
return "", 0, false
}
client = strconv.FormatUint(id, 10)
}
region, resident := strings.CutPrefix(key, "drm-resident-")
if !resident || !isVRAMRegion(region) {
continue
}
used, ok := drmResidentBytes(value)
if !ok || used > math.MaxUint64-total {
return "", 0, false
}
total += used
found = true
}
return client, total, scanner.Err() == nil && client != "" && found
}
func drmResidentBytes(value string) (uint64, bool) {
fields := strings.Fields(value)
if len(fields) == 0 || len(fields) > 2 {
return 0, false
}
n, err := strconv.ParseUint(fields[0], 10, 64)
if err != nil {
return 0, false
}
unit := uint64(1)
if len(fields) == 2 {
switch strings.ToLower(fields[1]) {
case "b":
case "kib":
unit = 1 << 10
case "mib":
unit = 1 << 20
case "gib":
unit = 1 << 30
default:
return 0, false
}
}
if n > math.MaxUint64/unit {
return 0, false
}
return n * unit, true
}
+105
View File
@@ -0,0 +1,105 @@
//go:build linux
// SPDX-License-Identifier: MIT
package xsysinfo
import (
"os"
"path/filepath"
"strconv"
. "github.com/onsi/ginkgo/v2"
. "github.com/onsi/gomega"
)
var _ = Describe("ProcessVRAM", func() {
var root string
write := func(path, contents string) {
Expect(os.MkdirAll(filepath.Dir(path), 0750)).To(Succeed())
Expect(os.WriteFile(path, []byte(contents), 0600)).To(Succeed())
}
addProcess := func(pid int, children string) {
base := filepath.Join(root, strconv.Itoa(pid))
Expect(os.MkdirAll(filepath.Join(base, "fd"), 0750)).To(Succeed())
write(filepath.Join(base, "task", strconv.Itoa(pid), "children"), children)
}
addFD := func(pid, fd int, render, info string) {
base := filepath.Join(root, strconv.Itoa(pid))
name := strconv.Itoa(fd)
Expect(os.Symlink("/dev/dri/"+render, filepath.Join(base, "fd", name))).To(Succeed())
write(filepath.Join(base, "fdinfo", name), info)
}
BeforeEach(func() {
var err error
root, err = os.MkdirTemp("", "process-vram-")
Expect(err).NotTo(HaveOccurred())
DeferCleanup(os.RemoveAll, root)
addProcess(100, "")
})
It("sums resident device memory across GPUs and child processes without duplicate clients", func() {
write(filepath.Join(root, "100/task/101/children"), "200")
addProcess(200, "")
info := "drm-client-id: 7\ndrm-total-local0: 900 MiB\ndrm-resident-local0: 128 MiB\ndrm-resident-system0: 4 GiB\n"
addFD(100, 3, "renderD128", info)
addFD(100, 4, "renderD128", info)
addFD(200, 3, "renderD128", info)
addFD(200, 4, "renderD129", "drm-client-id: 7\ndrm-resident-vram0: 256 MiB\n")
used, ok := processVRAM(root, 100)
Expect(ok).To(BeTrue())
Expect(used).To(Equal(uint64(384 * 1024 * 1024)))
})
It("distinguishes a measured zero from unavailable accounting", func() {
addFD(100, 3, "renderD128", "drm-client-id: 7\ndrm-resident-local0: 0 B\n")
used, ok := processVRAM(root, 100)
Expect(ok).To(BeTrue())
Expect(used).To(BeZero())
})
DescribeTable("does not invent readings from unsupported or invalid accounting",
func(info string) {
addFD(100, 3, "renderD128", info)
_, ok := processVRAM(root, 100)
Expect(ok).To(BeFalse())
},
Entry("no resident keys", "drm-client-id: 7\ndrm-total-vram0: 128 MiB\n"),
Entry("host memory only", "drm-client-id: 7\ndrm-resident-system0: 128 MiB\n"),
Entry("no client identity", "drm-resident-vram0: 128 MiB\n"),
Entry("malformed size", "drm-client-id: 7\ndrm-resident-vram0: unknown KiB\n"),
Entry("unknown unit", "drm-client-id: 7\ndrm-resident-vram0: 128 widgets\n"),
Entry("overflow", "drm-client-id: 7\ndrm-resident-vram0: 18446744073709551615 GiB\n"),
)
It("omits a partial reading if a child cannot be inspected", func() {
addFD(100, 3, "renderD128", "drm-client-id: 7\ndrm-resident-vram0: 128 MiB\n")
write(filepath.Join(root, "100/task/100/children"), "200")
_, ok := processVRAM(root, 100)
Expect(ok).To(BeFalse())
})
It("omits a partial reading if another DRM client lacks accounting", func() {
addFD(100, 3, "renderD128", "drm-client-id: 7\ndrm-resident-vram0: 128 MiB\n")
addFD(100, 4, "renderD129", "drm-client-id: 8\n")
_, ok := processVRAM(root, 100)
Expect(ok).To(BeFalse())
})
DescribeTable("omits mixed readings with unsupported GPU descriptors",
func(target string) {
addFD(100, 3, "renderD128", "drm-client-id: 7\ndrm-resident-vram0: 128 MiB\n")
Expect(os.Symlink(target, filepath.Join(root, "100/fd/4"))).To(Succeed())
_, ok := processVRAM(root, 100)
Expect(ok).To(BeFalse())
},
Entry("primary DRM node", "/dev/dri/card0"),
Entry("NVIDIA device", "/dev/nvidia0"),
)
It("returns unavailable for missing processes or no DRM descriptors", func() {
for _, pid := range []int{-1, 0, 100, 999} {
_, ok := processVRAM(root, pid)
Expect(ok).To(BeFalse())
}
})
})
+9
View File
@@ -0,0 +1,9 @@
//go:build !linux
// SPDX-License-Identifier: MIT
package xsysinfo
// ProcessVRAM is unavailable on platforms without Linux DRM fdinfo accounting.
func ProcessVRAM(pid int) (uint64, bool) {
return 0, false
}
+4
View File
@@ -7736,6 +7736,10 @@ const docTemplate = `{
},
"id": {
"type": "string"
},
"size_vram": {
"description": "SizeVRAM is DRM-accounted resident device memory in bytes. Nil means\nthe backend process tree has no complete supported reading.",
"type": "integer"
}
}
},
+4
View File
@@ -7733,6 +7733,10 @@
},
"id": {
"type": "string"
},
"size_vram": {
"description": "SizeVRAM is DRM-accounted resident device memory in bytes. Nil means\nthe backend process tree has no complete supported reading.",
"type": "integer"
}
}
},
+5
View File
@@ -2616,6 +2616,11 @@ definitions:
type: string
id:
type: string
size_vram:
description: |-
SizeVRAM is DRM-accounted resident device memory in bytes. Nil means
the backend process tree has no complete supported reading.
type: integer
type: object
schema.SystemInformationResponse:
properties: