Setting Up the Google Coral Edge TPU on Ubuntu 24.04 (2026)

TL;DR: Four kernel API changes break gasket-dkms 1.0 on kernels ≥ 6.8. Patch the four lines in /usr/src/gasket-1.0/, rebuild with DKMS, reload udev rules so the device re-enumerates to 18d1:9302, then test inference inside a Debian/Ubuntu container with python3-pycoral. The hardware runs fine once the plumbing is sorted.

What is the Google Coral Edge TPU?

The Google Coral USB Accelerator is a small USB 3.0 dongle containing Google’s Edge TPU ASIC — a purpose-built chip for running quantized (int8) TensorFlow Lite models. At around $30–$60 it delivers 4 TOPS (tera-operations per second) of inference throughput in a form factor smaller than a thumb drive.

It is not a general-purpose GPU. It only runs TFLite int8 models that have been compiled with the Edge TPU compiler, and each model must fit within its 8 MB of on-chip SRAM (larger models get partially cached — the “cold” portion runs on CPU). But for the right workloads — image classification, object detection, audio classification, person detection — it is genuinely fast.

The Coral product line also includes M.2 and mini-PCIe versions (which use the same apex PCIe driver), but this post focuses on the USB variant.

The problem: gasket-dkms fails on modern kernels

When you install gasket-dkms (the kernel module package that ships the apex PCIe driver for Coral M.2 boards) on Ubuntu 24.04 running kernel 6.8–6.17, apt leaves the package in a broken state:

iF  gasket-dkms    1.0-18    all    DKMS source for the gasket driver

The i means the package is installed, F means it failed to configure. Running sudo apt dist-upgrade or sudo dpkg --configure -a loops forever trying and failing to build the kernel module.

Even if you only have the USB Coral and never plan to use the PCIe driver, gasket-dkms is a dependency of libedgetpu1-std, so fixing it unblocks the whole stack.

The build log at /var/lib/dkms/gasket/1.0/build/make.log shows four distinct errors:

gasket_interrupt.c:161: error: too many arguments to function 'eventfd_signal'
gasket_page_table.c:57: error: expected ',' or ';' before 'DMA_BUF'
gasket_core.c:1375: error: 'no_llseek' undeclared; did you mean 'noop_llseek'?
gasket_core.c:1841: error: too many arguments to function 'class_create'

These are all upstream kernel API changes that accumulated since the last time the gasket-dkms package was updated.

Patching the source

The source lives at /usr/src/gasket-1.0/. DKMS re-reads this directory every time it builds, so patching it directly persists across kernel upgrades.

1. eventfd_signal lost its n argument (kernel ≥ 6.8)

// gasket_interrupt.c:161 — before
eventfd_signal(ctx, 1);

// after
eventfd_signal(ctx);

sudo sed -i 's/eventfd_signal(ctx, 1);/eventfd_signal(ctx);/' \
    /usr/src/gasket-1.0/gasket_interrupt.c

2. MODULE_IMPORT_NS now takes a string literal (kernel ≥ 6.13)

Before kernel 6.13, MODULE_IMPORT_NS(DMA_BUF) worked because the macro called __stringify() internally. From 6.13 onward the macro expects a pre-quoted string. Kernels < 6.13 on the same machine break if you use the new form, so we need a version guard:

// gasket_page_table.c — replace the bare macro with:
#include <linux/version.h>
#if LINUX_VERSION_CODE >= KERNEL_VERSION(6, 13, 0)
MODULE_IMPORT_NS("DMA_BUF");
#else
MODULE_IMPORT_NS(DMA_BUF);
#endif

Apply with Python so the quotes survive the shell:

sudo python3 -c "
with open('/usr/src/gasket-1.0/gasket_page_table.c', 'r') as f:
    c = f.read()
old = 'MODULE_IMPORT_NS(\"DMA_BUF\");'
new = '''#include <linux/version.h>
#if LINUX_VERSION_CODE >= KERNEL_VERSION(6, 13, 0)
MODULE_IMPORT_NS(\"DMA_BUF\");
#else
MODULE_IMPORT_NS(DMA_BUF);
#endif'''
with open('/usr/src/gasket-1.0/gasket_page_table.c', 'w') as f:
    f.write(c.replace(old, new))
print('done')
"

Note: You need to run this after the initial sed pass that changed MODULE_IMPORT_NS(DMA_BUF) to MODULE_IMPORT_NS("DMA_BUF"). If you are starting fresh, just insert the version-guarded block directly.

3. no_llseek renamed to noop_llseek (kernel ≥ 6.x)

sudo sed -i 's/\.llseek = no_llseek,/.llseek = noop_llseek,/' \
    /usr/src/gasket-1.0/gasket_core.c

4. class_create lost its module argument (kernel ≥ 6.4)

// gasket_core.c:1841 — before
class_create(driver_desc->module, driver_desc->name);

// after
class_create(driver_desc->name);

sudo sed -i 's/class_create(driver_desc->module, driver_desc->name)/class_create(driver_desc->name)/' \
    /usr/src/gasket-1.0/gasket_core.c

Rebuilding with DKMS

Remove the broken DKMS state and add the patched source back:

sudo dkms remove gasket/1.0 --all
sudo dkms add gasket/1.0

Build for the running kernel first (sanity check):

sudo dkms build gasket/1.0 -k $(uname -r)

If that succeeds, install it:

sudo dkms install gasket/1.0 -k $(uname -r)

Then let dpkg finish the configuration (it will build for all remaining kernels that have headers installed):

sudo dpkg --configure gasket-dkms

You should see all builds succeed and the final dpkg -l gasket-dkms output show ii (installed, configured):

ii  gasket-dkms    1.0-18    all    DKMS source for the gasket driver

USB device: udev rules and re-enumeration

The Coral USB Accelerator ships in two USB states:

The Coral apt repo installs a udev rule at /lib/udev/rules.d/60-libedgetpu1-std.rules:

SUBSYSTEM=="usb",ATTRS{idVendor}=="1a6e",ATTRS{idProduct}=="089a",GROUP="plugdev"
SUBSYSTEM=="usb",ATTRS{idVendor}=="18d1",ATTRS{idProduct}=="9302",GROUP="plugdev"

If the device was plugged in before the rules were reloaded it will show root:root permissions. Reload and re-trigger:

sudo udevadm control --reload-rules
sudo udevadm trigger --subsystem-match=usb

After a second or two, check:

lsusb | grep -E '1a6e|18d1'
# Bus 002 Device 003: ID 18d1:9302 Google Inc.

The device should now show as 18d1:9302 (firmware loaded) with root:plugdev permissions:

ls -la /dev/bus/usb/002/003
# crw-rw---- 1 root plugdev 189, 130 Jun 24 09:28 /dev/bus/usb/002/003

Make sure your user is in the plugdev group:

sudo usermod -a -G plugdev $USER
# Log out and back in, or: newgrp plugdev

Testing: why you need a container

Here is the annoying part. libedgetpu1-std 16.0 (the current Coral apt package) was compiled against TFLite ~2.5–2.13. On Ubuntu 24.04 the only available Python is 3.12, and the Coral apt packages (python3-pycoral, python3-tflite-runtime) declare Depends: python3 (< 3.10).

The tflite-runtime and ai-edge-litert packages on PyPI are versions 2.14+ — too new to be ABI-compatible with the old libedgetpu delegate interface at inference time. The delegate loads but fails with:

RuntimeError: Encountered an unresolved custom op.
Node number 1 (EdgeTpuDelegateForCustomOp) failed to invoke.

The cleanest workaround is a Docker container running Ubuntu 20.04 (focal), which supports Python 3.8 and has working python3-pycoral packages. Pass your USB device into the container with --device:

# Find the USB device path first
lsusb | grep '18d1:9302'
# Bus 002 Device 003: ID 18d1:9302 Google Inc.
# → /dev/bus/usb/002/003

docker run --rm \
  --device /dev/bus/usb/002/003 \
  -v /tmp/coral_test:/coral_test \
  ubuntu:20.04 \
  bash -c "
    export DEBIAN_FRONTEND=noninteractive
    apt-get update -qq
    apt-get install -y -qq gnupg curl python3 python3-numpy python3-pil
    curl -fsSL https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key add -
    echo 'deb https://packages.cloud.google.com/apt coral-edgetpu-stable main' \
        > /etc/apt/sources.list.d/coral.list
    apt-get update -qq
    apt-get install -y -qq libedgetpu1-std python3-pycoral
    python3 /coral_test/your_test.py
  "

A minimal classification test

Prepare test assets:

mkdir -p /tmp/coral_test && cd /tmp/coral_test

# Quantized MobileNetV2 compiled for Edge TPU
wget 'https://github.com/google-coral/test_data/raw/master/mobilenet_v2_1.0_224_quant_edgetpu.tflite'
# Same model without Edge TPU ops (CPU baseline)
wget 'https://github.com/google-coral/test_data/raw/master/mobilenet_v2_1.0_224_quant.tflite'
# ImageNet labels
wget 'https://raw.githubusercontent.com/google-coral/test_data/master/imagenet_labels.txt'
# Test image (Grace Hopper portrait)
wget 'https://github.com/google-coral/test_data/raw/master/grace_hopper.bmp'

Save this as /tmp/coral_test/classify.py:

import time, numpy as np
from PIL import Image
from pycoral.utils import edgetpu
import tflite_runtime.interpreter as tflite

print("Coral devices:", edgetpu.list_edge_tpus())

with open("imagenet_labels.txt") as f:
    labels = [l.strip() for l in f]

img = Image.open("grace_hopper.bmp").resize((224, 224))
x = np.expand_dims(np.array(img, dtype=np.uint8), 0)

# CPU run
cpu = tflite.Interpreter(model_path="mobilenet_v2_1.0_224_quant.tflite")
cpu.allocate_tensors()
cpu.set_tensor(cpu.get_input_details()[0]["index"], x)
cpu.invoke()  # warm up
t0 = time.perf_counter()
for _ in range(50): cpu.invoke()
cpu_ms = (time.perf_counter() - t0) / 50 * 1000
cpu_out = cpu.get_tensor(cpu.get_output_details()[0]["index"])[0]

# Edge TPU run
delegate = tflite.load_delegate("libedgetpu.so.1")
tpu = tflite.Interpreter(
    model_path="mobilenet_v2_1.0_224_quant_edgetpu.tflite",
    experimental_delegates=[delegate]
)
tpu.allocate_tensors()
tpu.set_tensor(tpu.get_input_details()[0]["index"], x)
tpu.invoke()  # warm up — loads firmware + model onto TPU
t0 = time.perf_counter()
for _ in range(50): tpu.invoke()
tpu_ms = (time.perf_counter() - t0) / 50 * 1000
tpu_out = tpu.get_tensor(tpu.get_output_details()[0]["index"])[0]

top = lambda o: [(labels[i], int(o[i])) for i in np.argsort(o)[-3:][::-1]]
print(f"\nCPU  {cpu_ms:.1f} ms — {top(cpu_out)}")
print(f"TPU  {tpu_ms:.1f} ms — {top(tpu_out)}")
print(f"\nSpeedup: {cpu_ms/tpu_ms:.1f}×")

Run it inside the container:

Coral devices: [{'type': 'usb', 'path': '/sys/bus/usb/devices/2-6'}]

CPU  16.9 ms — [('military uniform', 234), ('suit, suit of clothes', 3), ('mortarboard', 2)]
TPU   4.2 ms — [('military uniform', 233), ('suit, suit of clothes', 3), ('mortarboard', 2)]

Speedup: 4.0×

The top-1 prediction (military uniform) is correct — Grace Hopper was a U.S. Navy admiral.

Checklist summary

FAQ

Does this affect the PCIe / M.2 Coral too?
Yes. The gasket and apex kernel modules are used by both the USB and PCIe/M.2 variants. The USB Coral does not need the apex driver for normal use (it talks over USB), but gasket-dkms still needs to build cleanly or dpkg stays broken.

Will future kernel upgrades break it again?
Only the MODULE_IMPORT_NS patch is kernel-version-sensitive and we handled it with a #if LINUX_VERSION_CODE guard. The other three changes (eventfd_signal, noop_llseek, class_create) are already the current API in 6.8+ — they will not regress. The version guard covers kernels < 6.13 and ≥ 6.13 simultaneously, so new kernel header packages will just work.

Can I avoid Docker entirely?
If you install Python 3.10 (e.g. via the deadsnakes PPA) and create a virtualenv, tflite-runtime 2.14 installs cleanly — but the ABI mismatch with libedgetpu.so.1 16.0 causes EdgeTpuDelegateForCustomOp to fail at invoke time. Until Google releases a new libedgetpu built against a recent TFLite, Docker with Ubuntu 20.04 (Python 3.8 + python3-pycoral 2.0.0) is the most reliable path. Alternatively, compile libedgetpu from source against TFLite 2.14 — a significant undertaking.

What is libedgetpu1-max?
libedgetpu1-max runs the Edge TPU ASIC at a higher clock speed for ~35% more throughput but produces more heat (the dongle gets warm). libedgetpu1-std is the conservative clocked version. They are mutually exclusive — install one or the other. Both are at version 16.0 in the Coral apt repo.

The gasket-dkms package still shows errors after patching — what now?
Check whether kernel headers for all installed kernels are present: ls /lib/modules/. For kernels without matching headers, DKMS cannot build. Either install the headers (sudo apt install linux-headers-$(uname -r)) or remove the old kernel entirely (sudo apt autoremove --purge linux-image-X.Y.Z-generic).