qmmd!

General description

qmmd is the "qmassa metrics daemon" that provides a Prometheus endpoint over HTTP to export GPU usage metrics. It uses the same internal library as the qmassa TUI application. By default, qmmd doesn't scan the /proc filesystem and process DRM fdinfo files for reporting GPU engines usage and relies on qmlib implementations for each kernel driver.

Requirements

The minimum requirements to compile & run qmmd are:

• Compile-time: Rust v1.88 or later, pkg-config and libudev development packages
• Runtime: Linux kernel v6.8 or later to report most usage stats

How to install it

The recommendation is to install qmmd using cargo. If you want to install the latest release on crates.io using qmmd's lock file:

cargo install --locked qmmd

If you want to install the latest development version using qmmd's lock file:

cargo install --locked --git https://github.com/ulissesf/qmassa qmmd

How to use it

[!IMPORTANT] You have to run qmmd as root so it can get all GPU usage metrics. In order to properly daemonize and manage it, please run qmmd under a systemd service (check the sample service file).

Export metrics from a specific GPU. The device is specified by its PCI device slot name or its sysname (for non-PCI devices).

sudo qmmd -d 0000:03:00.0

Running for only 5 iterations (stats updates).

sudo qmmd -n 5

Changing the interval between stats updates to 1s (1000 ms).

sudo qmmd -m 1000

Changing the IP to register the HTTP endpoint listener to 192.168.86.32.

sudo qmmd -i 192.168.86.32

Changing the port to register the HTTP endpoint listener to 9090.

sudo qmmd -p 9090

Attempt to use all DRM fdinfo files in the system to calculate GPU engines usage. This option will use more CPU than relying on default qmlib drivers.

sudo qmmd -f

Using perf PMU to report freqs only for 0000:03:00.0 with the xe driver.

sudo qmmd -o xe=devslot=0000:03:00.0,freqs=pmu

Metrics description

The table below shows all the metric names (with units), description, type, and labels. They follow the convention commonly used by Prometheus.

Below is an example of the metrics exported by the Prometheus HTTP endpoint.

# TYPE qmmd_gpu_info counter
qmmd_gpu_info{device="0000:03:00.0",pci_id="8086:E212",vendor_name="Intel Corporation",device_name="Battlemage G21 [Arc Pro B50]",revision="00",driver_name="xe",dev_type="Discrete (PF)",dev_node="/dev/dri/card1"} 1
qmmd_gpu_info{device="0000:03:00.0",pci_id="8086:E212",vendor_name="Intel Corporation",device_name="Battlemage G21 [Arc Pro B50]",revision="00",driver_name="xe",dev_type="Discrete (PF)",dev_node="/dev/dri/renderD128"} 1

# TYPE qmmd_gpu_memory_used_bytes gauge
qmmd_gpu_memory_used_bytes{device="0000:03:00.0",mem_type="smem"} 11472896
qmmd_gpu_memory_used_bytes{device="0000:03:00.0",mem_type="vram"} 722874368

# TYPE qmmd_gpu_memory_total_bytes gauge
qmmd_gpu_memory_total_bytes{device="0000:03:00.0",mem_type="smem"} 67253100544
qmmd_gpu_memory_total_bytes{device="0000:03:00.0",mem_type="vram"} 17095983104

# TYPE qmmd_gpu_engine_utilization_ratio gauge
qmmd_gpu_engine_utilization_ratio{device="0000:03:00.0",engine="bcs"} 0
qmmd_gpu_engine_utilization_ratio{device="0000:03:00.0",engine="ccs"} 0.9698813172214144
qmmd_gpu_engine_utilization_ratio{device="0000:03:00.0",engine="rcs"} 0.021016510095973474
qmmd_gpu_engine_utilization_ratio{device="0000:03:00.0",engine="vcs"} 0
qmmd_gpu_engine_utilization_ratio{device="0000:03:00.0",engine="vecs"} 0

# TYPE qmmd_gpu_actual_frequency_hertz gauge
qmmd_gpu_actual_frequency_hertz{device="0000:03:00.0",freq_id="gt0"} 2500000000
qmmd_gpu_actual_frequency_hertz{device="0000:03:00.0",freq_id="gt1"} 1100000000

# TYPE qmmd_gpu_maximum_frequency_hertz gauge
qmmd_gpu_maximum_frequency_hertz{device="0000:03:00.0",freq_id="gt0"} 2600000000
qmmd_gpu_maximum_frequency_hertz{device="0000:03:00.0",freq_id="gt1"} 1500000000

# TYPE qmmd_gpu_power_watts gauge
qmmd_gpu_power_watts{device="0000:03:00.0",domain="gpu"} 33.96474249418945
qmmd_gpu_power_watts{device="0000:03:00.0",domain="package"} 47.56483612825638

# TYPE qmmd_gpu_temperature_celsius gauge
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="mctrl"} 60
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="pcie"} 58
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="pkg"} 59
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram"} 52
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_0"} 52
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_1"} 52
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_2"} 52
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_3"} 52
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_4"} 52
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_5"} 50
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_6"} 52
qmmd_gpu_temperature_celsius{device="0000:03:00.0",sensor="vram_ch_7"} 52

# TYPE qmmd_gpu_fan_speed_rpm gauge
qmmd_gpu_fan_speed_rpm{device="0000:03:00.0",fan_id="1"} 2119

The supported metrics depend on the qmlib drivers implementation. For supported drivers and features, available driver options as well as kernel driver limitations/gaps, please check the qmlib drivers information.

License

Copyright © 2026 Ulisses Furquim

This project is distributed under the terms of the Apache License (Version 2.0). See LICENSE for details.