This release is a pre-release and may not be stable for production use.
Optimum RBLN
🤗 Optimum RBLN provides an interface between HuggingFace libraries (Transformers, Diffusers) and RBLN NPUs, including ATOM and REBEL.
This library enables seamless integration between the HuggingFace ecosystem and RBLN NPUs through a comprehensive toolkit for model loading and inference across single and multi-NPU environments. While we maintain a list of officially validated models and tasks, users can easily adapt other models and tasks with minimal modifications.
Key Features
🚀 High Performance Inference
- Optimized model execution on RBLN NPUs through RBLN SDK compilation
- Support for both single and multi-NPU inference
- Integrated with RBLN Runtime for optimal performance
🔧 Easy Integration
- Seamless compatibility with HuggingFace Model Hub
- Drop-in replacement for existing HuggingFace pipelines
- Minimal code changes required for NPU acceleration
Seamless Replacement for Existing HuggingFace Code
Swap the HuggingFace class for its RBLN counterpart. Passing a HuggingFace model id compiles the
model for the NPU on the first run; passing a directory of previously compiled artifacts loads them
directly, so compilation is skipped.
- from diffusers import StableDiffusionXLPipeline
+ from optimum.rbln import RBLNStableDiffusionXLPipeline
# Load model
model_id = "stabilityai/stable-diffusion-xl-base-1.0"
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
- pipe = StableDiffusionXLPipeline.from_pretrained(model_id)
+ pipe = RBLNStableDiffusionXLPipeline.from_pretrained(model_id)
# Generate image
image = pipe(prompt).images[0]
# Save image result
image.save("image.png")
+ # (Optional) Save compiled artifacts to skip the compilation step in future runs
+ pipe.save_pretrained("compiled_sdxl")
Compile Ahead of Time with the CLI
Instead of compiling inside your script, you can compile a model up front with the optimum-rbln-cli
command and load the resulting artifacts later:
# Compile and save the artifacts to ./compiled_qwen3
optimum-rbln-cli --model-id Qwen/Qwen3-4B -o ./compiled_qwen3 \
--max_seq_len 8192 --batch_size 1 --num_devices 4
from optimum.rbln import RBLNQwen3ForCausalLM
# Load the compiled artifacts (no recompilation)
model = RBLNQwen3ForCausalLM.from_pretrained("./compiled_qwen3")
Useful CLI helpers:
optimum-rbln-cli --list-classes # list available RBLN classes
optimum-rbln-cli --class RBLNQwen3ForCausalLM --show-rbln-config # show accepted rbln_config keys
optimum-rbln-cli --examples # show more usage examples
Documentation
Check out the documentation of Optimum RBLN for more advanced usage.
Getting Started
Note: The
rebel-compilerlibrary, which is required for runningoptimum-rbln, is only available for approved users. Please refer to the installation guide for instructions on accessing and installingrebel-compiler.
Install from PyPI
To install the latest release of this package:
pip install optimum-rbln --extra-index-url https://download.pytorch.org/whl/cpu
Install from source
Prerequisites
The below command installs optimum-rbln along with its dependencies.
git clone https://github.com/rbln-sw/optimum-rbln.git
cd optimum-rbln
./scripts/uv-sync.sh
Need Help?
- Join discussions and get answers in our Developer Community
- Contact maintainers at support@rebellions.ai
1 maintainer
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file optimum_rbln-0.11.3a2.tar.gz.
File metadata
- Download URL: optimum_rbln-0.11.3a2.tar.gz
- Upload date:
- Size: 627.8 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/6.2.0 CPython/3.12.12
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
87d7d87bdd47547a90a33a8b7ef3dcf698ff553303b5dc2087a91d7b8ef61fb1
|
|
| MD5 |
4c5f1632e6e52e519e9b1ca1fdf3621c
|
|
| BLAKE2b-256 |
bb64eb5af7a5d5d3991c2eefee6c097dd86529d5f9d570aa465c3c3707030a25
|
File details
Details for the file optimum_rbln-0.11.3a2-py3-none-any.whl.
File metadata
- Download URL: optimum_rbln-0.11.3a2-py3-none-any.whl
- Upload date:
- Size: 674.0 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via:
twine/6.2.0 CPython/3.12.12
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
51cdc7df892ef5d47075b79d22242f16da0c6779a529e4a18bac3a80f048ea86
|
|
| MD5 |
61173dc913652b16a309d05baa48fbdd
|
|
| BLAKE2b-256 |
7735f25d6ba2109583e150b00178f40b977a62b095d55b1644863c8c49482898
|
