Vint Lee, Minh Nguyen, Leena Elzeiny, Chun Deng, Pieter Abbeel, John Wawrzynek

Paper

Use conda environment found in environment.yaml

conda env create -f environment.yaml
conda activate chipdiffusion

Training and evaluation experiments will log data to Weights & Biases by default. Set your name and W&B project in the config files using the logger.wandb_entity and logger.wandb_project options before running the commands below. Turn off W&B logging by appending logger.wandb=False to the commands below.

For running evaluations that require clustering, download shmetis and hmetis and place in the repo's root directory.

• diffusion contains code for training, fine-tuning, and evaluating models
• data-gen contains code for generating synthetic datasets
• data-gen/outputs will be used to store the generated datasets
• datasets is used for other datasets (like IBM and ISPD benchmarks).
• notebooks has useful scripts and functions for evaluating placements (measuring congestion in particular) and inspecting benchmark files.
• parsing has scripts for converting and clustering benchmarks in the DEF/LEF format (such as IBM).

Generate v0, v1, and v2 datasets for training:

PYTHONPATH=. python data-gen/generate_parallel.py versions@_global_=v0


PYTHONPATH=. python data-gen/generate_parallel.py versions@_global_=v1


PYTHONPATH=. python data-gen/generate_parallel.py versions@_global_=v2 num_train_samples=5000 num_val_samples=2500

Configs are also provided for running dataset design experiments. Since we only use these for evaluation, not for training, we only need to generate a few circuits:

PYTHONPATH=. python data-gen/generate_parallel.py versions@_global_=vertex-0.7x num_train_samples=0 num_val_samples=200


PYTHONPATH=. python data-gen/generate_parallel.py versions@_global_=distribution-linear num_train_samples=0 num_val_samples=200

For experiments on scale factor, the scale factor has to be specified by including gen_params.edge_dist.dist_params.scale=<SCALE_FACTOR>. For example:

PYTHONPATH=. python data-gen/generate_parallel.py versions@_global_=scale gen_params.edge_dist.dist_params.scale=0.8 num_train_samples=0 num_val_samples=200

For easier debugging, use data-gen/generate.py.

After generating data, we train models on the v1 dataset:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/train_graph.py method=train_large task=v1.61

We can train smaller models using:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/train_graph.py method=train_medium task=v1.61 model/size@model.backbone_params=medium


CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/train_graph.py method=train_small task=v1.61 model/size@model.backbone_params=small

Once the models have been trained, we can fine-tune them on v2:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/train_graph.py method=finetune_large task=v2.61 mode@_global_=finetune from_checkpoint=v1.61.train_large.61/step_3000000.ckpt

Evaluating on v1 dataset without guidance:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/eval.py task=v1.61 method=eval from_checkpoint=v2.61.finetune_large.61/step_250000.ckpt legalizer@_global_=none guidance@_global_=none num_output_samples=128

Evaluating zero-shot on clustered IBM benchmark with guidance:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/eval.py method=eval_guided task=ibm.cluster512.v1 from_checkpoint=v2.61.finetune_large.61/step_250000.ckpt num_output_samples=18

Macro-only evaluation for IBM and ISPD benchmarks:

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/eval.py method=eval_macro_only task=ibm.cluster512.v1 from_checkpoint=v2.61.finetune_large.61/step_250000.ckpt legalizer@_global_=opt-adam num_output_samples=18 model.grad_descent_steps=20 model.hpwl_guidance_weight=16e-4 legalization.alpha_lr=8e-3 legalization.hpwl_weight=12e-5 legalization.legality_potential_target=0 legalization.grad_descent_steps=20000 macros_only=True


CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python diffusion/eval.py method=eval_macro_only task=ispd2005 from_checkpoint=v2.61.finetune_large.61/step_250000.ckpt legalizer@_global_=opt-adam guidance@_global_=opt num_output_samples=8 model.grad_descent_steps=20 model.hpwl_guidance_weight=16e-4 legalization.alpha_lr=8e-3 legalization.hpwl_weight=12e-5 legalization.legality_potential_target=0 legalization.grad_descent_steps=20000 macros_only=True

Examples of generated placements, for both clustered and macro-only settings, can be found here.

Input netlist is stored using PyTorch-Geometric's Data object.

Input placements for training are stored as numpy arrays.

Placement outputs are saved as pickle files containing a single numpy array with (x, y) coordinates for each object.

To obtain the IBM dataset, download the benchmark in DEF/LEF format to benchmarks/ibm and run the code in parsing:

PYTHONPATH=. python parsing/cluster.py

PYTHONPATH=. python parsing/cluster.py num_clusters=0

The code will parse the DEF/LEFs, cluster the netlists if num_clusters is non-zero, then output the dataset as pickle files to datasets/clustered directory.

To obtain the ISPD dataset for running evaluations, download the ISPD benchmark in bookshelf format to benchmarks/ispd2005, then use this notebook.

Once the benchmark files have been generated, copy this config into the benchmark directory, and change val_samples as needed.

For convenience, we provide the training checkpoint for the Large+v2 model at this link. To use it, copy the large-v2 directory into your logs directory and specify from_checkpoint accordingly when running the commands above.

Note that if the checkpoint loads correctly, the code will print successfully loaded state dict for model before running training or evaluation; otherwise, successfully loaded model will be printed instead, and the code will default to random model weights. Hyperparameter mismatch is a common cause of failure, and we provide the training config used for reference.

If you found our work useful, please cite:

@inproceedings{
      lee2025chipdiffusion,
      title={Chip Placement with Diffusion Models},
      author={Vint Lee and Minh Nguyen and Leena Elzeiny and Chun Deng and Pieter Abbeel and John Wawrzynek},
      booktitle={Forty-second International Conference on Machine Learning},
      year={2025},
      url={https://arxiv.org/abs/2407.12282}
}