Release Notes - 0.5.0

FuriosaAI SDK 0.5.0 release includes approximately 87 bug fixes, added functionalities, and improvements. The following are some of the key changes:

Compiler Improvement

0.5.0 release adds NPU acceleration support for the following operators. You can find the entire list of acceleration-supported operators at List of Supported Operators for Warboy Acceleration.

  • BatchNormalization

  • ConvTranspose

  • LeakyRelu

  • Pow

  • Sqrt

  • Sub

Additionally, we have improved the compiler with this release by adding support for Opset 13 operator of Onnx.

Session API Improvement

API Improvement 1 (commit b1d2b74): you can now designate NPU device in Session API. With 0.5.0, you can designate NPU device when generating a session, whereas previously you could only designate device to be used through the environment variable NPU_DEVNAME. If not explicitly stated, designated device in the environment variable NPU_DEVNAME will be used, as was done before.

from furiosa.runtime import session

sess1 = session.create("model1.onnx", device="npu0pe0")
sess2 = session.create("model2.onnx", device="npu0pe1")

# Asynchronous API
async_sess, queue = session.create_async("model2.onnx", device="npu1pe2")

API Improvement 2 (commit 4f1f114) is support for tensor name. The existing API was limited in that it could only identify the order of the tensor built into the model, and pass it as an input parameter to From 0.5.0, you can explicitly designate name of input and output tensors as shown below.

np1 = np.random.randint(0, 255, session_input.shape, dtype=np.uint8)
np2 = np.random.randint(0, 255, session_input.shape, dtype=np.uint8)
sess1.run_with(outputs=["output1"], inputs={
  "input2": np2,
  "input1": np1,

Error Diagnosis Message & Error Handling Improvements

In case of an error, version information for debugging and compiler log are output independently. This enables easier bug reporting. You can find more information about reporting issues and customer service at Bug Report.

>>> from furiosa.runtime import session
>>> session.create("mnist-8.onnx")
Saving the compilation log into /home/furiosa/.local/state/furiosa/logs/compile-20211121223028-l5w4g6.log
2021-11-22T06:30:28.423371Z ERROR fail to compile the model: the static shape of tensor 'input' contains an unsupported dimension value: Some(DimParam("batch_size"))
Information Dump
- Python version: 3.8.10 (default, Sep 28 2021, 16:10:42)  [GCC 9.3.0]
- furiosa-libnux path:
- furiosa-libnux version: 0.5.0 (rev: 407c0c51f built at 2021-11-18 22:32:34)
- furiosa-compiler version: 0.5.0 (rev: 407c0c51f built at 2021-11-18 22:32:34)
- furiosa-sdk-runtime version: Furiosa SDK Runtime  (libnux 0.5.0 407c0c51f 2021-11-18 22:32:34)

Please check the compiler log at /home/furiosa/.local/state/furiosa/logs/compile-20211121223028-l5w4g6.log.
If you have a problem, please report the log file to with the information dumped above.

Improvements such as error handling for the following issues are also included.

  • Error fix when using duplicate devices (commit 01aaa40)

  • Added timeout of CompletionQueue & error handling fix for session connection termination (commit 21cba85)

  • Hanging issue fix for interruption during compiling (commit a0f4bd7)

Introducing Furiosa Server (serving framework)

0.5.0 includes Furiosa Server, a serving framework that supports GRPC and REST API. You can easily install it by running pip install furiosa-sdk[server]. By running it with the command below, the model can be served immediately with the NPU. You can find more detailed instructions and functions at Model Server (Serving Framework).

furiosa server \
--model-path MNISTnet_uint8_quant_without_softmax.tflite \
--model-name mnist

Introducing Furiosa Model package

From 0.5.0, the optimized model for the FuriosaAI NPU can be used directly as a Python package. You can easily install it with the command pip install furiosa-sdk[models], and can immediately be used in Session API as shown in the following example.

import asyncio

from furiosa.registry import Model
from import MLCommonsResNet50
from furiosa.runtime import session

resnet50: Model =
sess = session.create(resnet50.model, device='npu0pe0')

Command line NPU management tool: furiosactl

0.5.0 includes furiosactl, a command line NPU management tool. You can install it with apt install furiosa-toolkit. You can use this tool to check NPU device status, as well as identify idle NPUs. You can find apt server configuration instructions at APT server configuration.

$ furiosactl info

| NPU  | Name             | Temp. | Power   | PCI-BDF      | PCI-DEV |
| npu0 | FuriosaAI Warboy |  34°C | 12.92 W | 0000:01:00.0 | 510:0   |

$ furiosactl list
| NPU  | DEVNAME   | Type      | Status |
| npu0 | npu0      | All PE(s) | Ready  |
|      | npu0pe0   | Single PE | Ready  |
|      | npu0pe1   | Single PE | Ready  |
|      | npu0pe0-1 | PE Fusion | Ready  |

Kubernetes support

0.5.0 includes NPU support for Kubernetes. You can install the NPU device plugin and node labeller with the command below, and have the NPU be scheduled together when deploying pods. More details can be found at Kubernetes Support.

kubectl apply -f
kubectl apply -f