Implementation:Pytorch Serve Intel GPU Utils

Overview

Intel_GPU_Utils provides utility functions for Intel GPU discovery and metrics collection via the xpu-smi CLI tool. It wraps subprocess calls to xpu-smi for GPU counting, device information retrieval, and command validation, enabling TorchServe to collect hardware metrics from Intel discrete GPUs.

Field	Value
Page Type	Implementation
Implementation Type	API Doc
Domains	Hardware_Monitoring, Intel_GPU
Knowledge Sources	Pytorch_Serve
Workflow	Metrics_Collection
Last Updated	2026-02-13 18:52 GMT

Description

This module provides a lightweight wrapper around the xpu-smi command-line tool for Intel GPU hardware metrics. It is used by TorchServe's Intel Extension for PyTorch (IPEX) integration to discover available Intel GPUs and collect runtime metrics such as utilization, memory usage, and temperature.

Key Functions

• check_cmd(cmd) (lines 12-32): Validates and executes a xpu-smi command, returning parsed CSV output
• count_gpus() (lines 35-41): Returns the number of available Intel GPUs by querying xpu-smi discovery
• list_gpu_info(num_gpus) (lines 44-54): Collects detailed metrics for a specified number of GPUs

Dependencies

Library	Usage
csv	Parsing CSV output from xpu-smi
subprocess	Executing xpu-smi CLI commands

Code Reference

Source Location

File	Lines	Repository
examples/intel_extension_for_pytorch/intel_gpu.py	L1-54	pytorch/serve

Key Functions

def check_cmd(cmd):
    """
    Validate and execute a xpu-smi command.
    Lines 12-32.

    Runs the given xpu-smi command via subprocess, captures
    stdout, and parses the CSV-formatted output.

    Parameters:
        cmd (str): The xpu-smi command string to execute.

    Returns:
        list: Parsed CSV rows from command output.

    Raises:
        subprocess.CalledProcessError: If xpu-smi command fails.
    """
    ...


def count_gpus():
    """
    Count the number of available Intel GPUs.
    Lines 35-41.

    Queries xpu-smi discovery to enumerate available
    Intel discrete GPU devices.

    Returns:
        int: Number of detected Intel GPUs.
    """
    ...


def list_gpu_info(num_gpus):
    """
    Collect detailed metrics for Intel GPUs.
    Lines 44-54.

    Iterates over the specified number of GPUs and
    collects device-level metrics via xpu-smi.

    Parameters:
        num_gpus (int): Number of GPUs to query.

    Returns:
        list: Per-GPU metric dictionaries.
    """
    ...

Import

import csv
import subprocess

I/O Contract

Function	Input	Output	Notes
check_cmd(cmd)	cmd: str (xpu-smi command)	list of parsed CSV rows	Wraps subprocess call to xpu-smi
count_gpus()	None	int (GPU count)	Queries xpu-smi discovery
list_gpu_info(num_gpus)	num_gpus: int	list of per-GPU metric dicts	Collects device-level metrics

Usage Examples

Example 1: Counting Available GPUs

from intel_gpu import count_gpus

num_gpus = count_gpus()
print(f"Detected {num_gpus} Intel GPU(s)")

Example 2: Collecting GPU Metrics

from intel_gpu import count_gpus, list_gpu_info

num_gpus = count_gpus()
gpu_metrics = list_gpu_info(num_gpus)

for i, metrics in enumerate(gpu_metrics):
    print(f"GPU {i}: {metrics}")

Example 3: Validating xpu-smi Command

from intel_gpu import check_cmd

# Execute a specific xpu-smi command
result = check_cmd("xpu-smi discovery --dump 1")
# result contains parsed CSV output rows

Related Pages

• Principle:Pytorch_Serve_Hardware_Metrics_Collection -- The hardware metrics collection pattern this module implements
• Implementation:Pytorch_Serve_Metrics_API - TorchServe metrics API that consumes GPU metrics
• Environment:Pytorch_Serve_Python_PyTorch_Runtime - Core Python/PyTorch runtime