nvml

frigate/stats.py

@@ -151,9 +151,11 @@ async def set_gpu_stats(
             nvidia_usage = get_nvidia_gpu_stats()
 
             if nvidia_usage:
-                name = nvidia_usage["name"]
+                for i in nvidia_usage:
-                del nvidia_usage["name"]
+                    stats[nvidia_usage[i]["name"]] = {
-                stats[name] = nvidia_usage
+                        "gpu": round(nvidia_usage[i]["gpu"],2),
+                        "mem": round(nvidia_usage[i]["mem"],2),
+                    }
             else:
                 stats["nvidia-gpu"] = {"gpu": -1, "mem": -1}
                 hwaccel_errors.append(args)

frigate/util.py

@@ -16,6 +16,7 @@ from collections import Counter
 from collections.abc import Mapping
 from multiprocessing import shared_memory
 from typing import Any, AnyStr, Optional, Tuple
+import py3nvml.py3nvml as nvml
 
 import cv2
 import numpy as np
@@ -915,46 +916,37 @@ def get_intel_gpu_stats() -> dict[str, str]:
         return results
 
 
+def try_get_info(f, h, default='N/A'):
+    try:
+        v = f(h)
+    except nvml.NVMLError_NotSupported:
+        v = default
+    return v
+
+
+
 def get_nvidia_gpu_stats() -> dict[str, str]:
-    """Get stats using nvidia-smi."""
+    nvml.nvmlInit()
-    nvidia_smi_command = [
+    deviceCount = nvml.nvmlDeviceGetCount()
-        "nvidia-smi",
+    results = {}
-        "--query-gpu=gpu_name,utilization.gpu,memory.used,memory.total",
+    for i in range(deviceCount):
-        "--format=csv",
+        handle = nvml.nvmlDeviceGetHandleByIndex(i)
-    ]
+        meminfo = nvml.nvmlDeviceGetMemoryInfo(handle)
-
+        util = try_get_info(nvml.nvmlDeviceGetUtilizationRates, handle)
-    if (
+        if util != 'N/A':
-        "CUDA_VISIBLE_DEVICES" in os.environ
+            gpu_util = util.gpu
-        and os.environ["CUDA_VISIBLE_DEVICES"].isdigit()
-    ):
-        nvidia_smi_command.extend(["--id", os.environ["CUDA_VISIBLE_DEVICES"]])
-    elif (
-        "NVIDIA_VISIBLE_DEVICES" in os.environ
-        and os.environ["NVIDIA_VISIBLE_DEVICES"].isdigit()
-    ):
-        nvidia_smi_command.extend(["--id", os.environ["NVIDIA_VISIBLE_DEVICES"]])
-
-    p = sp.run(
-        nvidia_smi_command,
-        encoding="ascii",
-        capture_output=True,
-    )
-
-    if p.returncode != 0:
-        logger.error(f"Unable to poll nvidia GPU stats: {p.stderr}")
-        return None
         else:
-        usages = p.stdout.split("\n")[1].strip().split(",")
+            gpu_util = 0
-        memory_percent = f"{round(float(usages[2].replace(' MiB', '').strip()) / float(usages[3].replace(' MiB', '').strip()) * 100, 1)} %"
+        results[i] = {
-        results: dict[str, str] = {
+            "name": nvml.nvmlDeviceGetName(handle),
-            "name": usages[0],
+            "gpu": gpu_util,
-            "gpu": usages[1].strip(),
+            "mem": meminfo.used / meminfo.total * 100
-            "mem": memory_percent,
         }
 
         return results
 
 
+
 def ffprobe_stream(path: str) -> sp.CompletedProcess:
     """Run ffprobe on stream."""
     clean_path = escape_special_characters(path)

requirements-wheels.txt

@@ -11,6 +11,7 @@ peewee == 3.15.*
 peewee_migrate == 1.7.*
 psutil == 5.9.*
 pydantic == 1.10.*
+git+https://github.com/fbcotter/py3nvml#egg=py3nvml
 PyYAML == 6.0
 pytz == 2023.3
 tzlocal == 4.3

web/src/routes/System.jsx

@@ -268,8 +268,8 @@ export default function System() {
                         </Thead>
                         <Tbody>
                           <Tr>
-                            <Td>{gpu_usages[gpu]['gpu']}</Td>
+                            <Td>{gpu_usages[gpu]['gpu']}%</Td>
-                            <Td>{gpu_usages[gpu]['mem']}</Td>
+                            <Td>{gpu_usages[gpu]['mem']}%</Td>
                           </Tr>
                         </Tbody>
                       </Table>