OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Infrastructure & Serving

Acceltran

AccelTran is an open-source Python research tool that simulates accelerator design spaces for flexible and heterogeneous transformer architectures, with a focus on sparsity-aware and monolithic 3D designs. It accepts accelerator configuration files and transformer model dictionaries, and can report per-cycle accelerator state and plot utilization and metric curves. The repository also includes separate pruning and ha

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Acceltran — OpenViST project profile in AI Infrastructure & Serving.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use AccelTran when you want to explore accelerator configurations for transformer workloads in the FlexiBERT 2.0 design space, or when you need the repository's documented pruning and synthesis workflows.

Who it is for

  • Researchers studying transformer accelerator architecture trade-offs.
  • Hardware and ML researchers comparing sparsity-aware accelerator configurations against transformer models.
  • Users investigating the repository's documented pruning and synthesis workflows.

Documented uses

  • Simulate an accelerator configuration against a transformer model dictionary.
  • Inspect accelerator state across simulation cycles and plot utilization and metric curves.
  • Evaluate pruning effects on SST-2 performance with the documented DP-BERT workflow.
  • Generate area and power reports for supported hardware modules using Synopsys Design Compiler.
Before you choose

Requirements & limits

The repository is public and licensed under BSD-3-Clause. Its documented environment setup uses Conda; the README lists Python 3.6–3.8 and PyTorch 1.8.1.

  • Simulator inputs are split between accelerator configuration files and transformer model-dictionary files, which must match the specified design space.
  • The repository includes distinct simulator, pruning and synthesis workflows; synthesis uses Synopsys Design Compiler.
  • The documented SST-2 pruning example uses an already weight-pruned model; further weight pruning requires the --prune_weights flag.
  • The README states that the project connects with the separate FlexiBERT 2.0 design-space framework.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run the documented BERT-Tiny simulation

Simulate the provided BERT-Tiny model with its example accelerator configuration and generate utilization and metric plots.

Before you start

  • Clone the AccelTran repository and initialize its submodules.
  • Set up the Conda environment with the repository's env_setup.sh script.
  • Use the provided BERT-Tiny model dictionary and configuration files.

Steps

  1. Clone the repository, enter its directory, and initialize and update submodules.

  2. Run source env_setup.sh to create the Conda environment.

  3. From simulator/, run the command below.

    Documented command
    python3 run_simulator.py --model_dict_path ./model_dicts/bert_tiny.json --config_path ./config/config_tiny.yaml --plot_steps 1000 --debug

Expected outcome

The simulator outputs accelerator state for each cycle and plots utilization and metric curves every 1,000 cycles.

Guide limitations

  • The configuration must conform to simulator/design_space/design_space.yaml.
  • The repository describes the environment setup as Conda-based; it says a pip requirements file is being prepared.
  • This recipe follows the repository instructions and is not a report of an execution test.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
17 October 2026
Original website title
jha-lab/acceltran · Official homepage title · retained in its source language
How we review sourcesSuggest a correction