How wuhy68/chatedav1 works
Project Type and Purpose
• ChatEDAv1 is a specialized autonomous agent built for Electronic Design Automation. It integrates a fine-tuned large language model (named AutoMage) with a suite of EDA tools. The system is responsible for managing the complex design flow—from Register-Transfer Level (RTL) description to final GDSII outputs—by planning tasks, generating scripts, and executing them through defined interfaces.
• The README describes how ChatEDA leverages a set of EDA tools (with an example being the OpenRoad API) and also provides data sets – one for training (ChatEDA-train-example.json) and another benchmark for evaluation (ChatEDA-Bench.txt).
File Structure
• Top-level directories/files include:
• LICENSE and README.md for licensing and project overview.
• api_doc/: Contains API documentation and implementation files (openroad_api.py, openroad_api_impl.py, and parse_mk_config.py). This part is crucial as it defines the interface layer between the ChatEDA controller and the underlying EDA tools.
• data/: Includes train and test subdirectories. These hold the training datasets for fine-tuning AutoMage as well as benchmark examples to evaluate the system.
• The structure shows separation between API (for interfacing with EDA tool functionalities) and data (for training and evaluation).
Key Components
a. LLM Component – AutoMage Model:
• Represents the autonomous intelligence driving the decision-making and script generation.
• This component receives high-level design tasks and produces actionable instructions.
b. EDA Tool Chain:
• Represented by the API documented in api_doc (e.g., the OpenRoad API and its implementation). These tools act as executors that carry out the design tasks on the RTL/GDSII flow.
• This layer might include additional EDA tools if integrated further.
c. Data Repositories:
• Training Data: Found under data/train (e.g., ChatEDA-train-example.json) that is used to fine-tune the AutoMage model.
• Evaluation Benchmark: Found under data/test (e.g., ChatEDA-Bench.txt) for comprehensive performance evaluations.
d. Orchestrator/Controller:
• While not explicitly a separate file, the design flow implies a component (the autonomous agent "ChatEDA") that controls interactions between the LLM and the EDA tools. This can be represented as a middleware layer coordinating the overall process.
Establish Relationships and Interactions
• Data Flow:
• Training Data → Fine-tuning process → AutoMage (LLM).
• User Requirements/Design Specifications/input tasks (implied) are processed by the AutoMage model.
• AutoMage then generates a series of instructions which are passed to the EDA Tool Chain via the OpenRoad API.
• The EDA tools execute the tasks, and their outputs (design artifacts) may be used for further evaluation.
• Integration Flow:
Architectural Principles
• Separation of concerns – clearly divide the system into data, intelligence, middleware, and execution modules.
• That build or orchestration scripts (as implied by the file tree) are represented if they further detail how components interact.
• That the system’s design enables modularity, allowing potential updates or expansions (e.g., adding new EDA tools or modifying the LLM).