Configuration for replicating results

[q8888620002]

Hi

Thank you for developing and sharing this excellent library! I was wondering if it would be possible to provide the exact configurations used to replicate the results from the MolGPS paper, particularly the comparisons between MPNN++, GPS++, and Transformer.

It seems that some configurations in expts/neurips2023_configs may be outdated. Could you clarify if the latest configurations are available?

Also, I wonder if sharing an example yaml for multi-GPU training is possible. I've tried the following trainer configs

accelerator:
  type: gpu  # cpu or ipu or gpu
  config_override:
    datamodule:
      args:
        batch_size_training: 64
        batch_size_inference: 256
    trainer:
      trainer:
        precision: 32
trainer:
    precision: 32
    max_epochs: *max_epochs
    min_epochs: 1
    accumulate_grad_batches: 2
    check_val_every_n_epoch: 10
    devices: 4
    strategy: ddp

However, the following error shows up
RuntimeError: CUDA error: device-side assert triggered CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.

Thanks in advance!

[q8888620002]

Below is my script for test runs by following example in Multitask Molecular Modelling with Graphium on the IPU

def main():
    """Multi-Task pretraining with largemix dataset"""
    filename = 

    cfg = load_yaml_config("config_gps_10M_pcqm4m.yaml")

    cfg, accelerator_type = load_accelerator(cfg)

    datamodule = MultitaskFromSmilesDataModule(
        **cfg["datamodule"]["args"],
    )
    datamodule.prepare_data(True)
    datamodule.setup("fit")

    # Initialize the network
    model_class, model_kwargs = load_architecture(
        cfg,
        in_dims=datamodule.in_dims,
    )
    metrics = load_metrics(cfg)
    logger.info(metrics)

    predictor = load_predictor(
        cfg,
        model_class,
        model_kwargs,
        metrics,
        datamodule.get_task_levels(),
        accelerator_type,
        datamodule.featurization,
        datamodule.task_norms,
    )

    logger.info(predictor.model)
    logger.info(ModelSummary(predictor, max_depth=4))
    date_time_suffix = datetime.now().strftime("%d.%m.%Y_%H.%M.%S")
    trainer = load_trainer(cfg, accelerator_type, date_time_suffix)

    logger.info("About to set the max nodes etc.")
    predictor.set_max_nodes_edges_per_graph(datamodule, stages=["train", "val"])

    with SafeRun(
        name="TRAINING", raise_error=cfg["constants"]["raise_train_error"], verbose=True
    ):
        trainer.fit(model=predictor, datamodule=datamodule)

    predictor.set_max_nodes_edges_per_graph(datamodule, stages=["test"])

    # Run the model validation
    with SafeRun(
        name="TESTING", raise_error=cfg["constants"]["raise_train_error"], verbose=True
    ):
        trainer.test(
            model=predictor,
            datamodule=datamodule,
        )