Optimum RBLN

🤗 Optimum RBLN provides an interface between HuggingFace libraries (Transformers, Diffusers) and RBLN NPUs, including ATOM and Rebel100.

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-compiler library, which is required for running optimum-rbln, is only available for approved users. Please refer to the installation guide for instructions on accessing and installing rebel-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

• Install uv (refer to this link for detailed commands)

The below commands create a virtual environment in .venv and install optimum-rbln in editable mode along with its dependencies. See CONTRIBUTING.md for the full development setup.

git clone https://github.com/rbln-sw/optimum-rbln.git
cd optimum-rbln
uv sync

Need Help?

• Join discussions and get answers in our Developer Community
• Contact maintainers at support@rebellions.ai