6cd0cd3bd99ec2df47bb571a18e0b70819134504,rllib/agents/impala/impala.py,OverrideDefaultResourceRequest,default_resource_request,#Any#Any#,101

Before Change

    def default_resource_request(cls, config):
        cf = dict(cls._default_config, **config)
        Trainer._validate_config(cf)
        return Resources(
            cpu=cf["num_cpus_for_driver"],
            gpu=cf["num_gpus"],
            memory=cf["memory"],
            object_store_memory=cf["object_store_memory"],
            extra_cpu=cf["num_cpus_per_worker"] * cf["num_workers"] +
            cf["num_aggregation_workers"],
            extra_gpu=cf["num_gpus_per_worker"] * cf["num_workers"],
            extra_memory=cf["memory_per_worker"] * cf["num_workers"],
            extra_object_store_memory=cf["object_store_memory_per_worker"] *
            cf["num_workers"])


def make_learner_thread(local_worker, config):
    if config["num_gpus"] > 1 or config["num_data_loader_buffers"] > 1:

After Change

        // Return PlacementGroupFactory containing all needed resources
        // (already properly defined as device bundles).
        return PlacementGroupFactory(
            bundles=[{
                // Driver + Aggregation Workers:
                // Force to be on same node to maximize data bandwidth
                // between aggregation workers and the learner (driver).
                // Aggregation workers tree-aggregate experiences collected
                // from RolloutWorkers (n rollout workers map to m
                // aggregation workers, where m < n) and always use 1 CPU
                // each.
                "CPU": cf["num_cpus_for_driver"] +
                cf["num_aggregation_workers"],
                "GPU": cf["num_gpus"]
            }] + [
                {
                    // RolloutWorkers.
                    "CPU": cf["num_cpus_per_worker"],

Instances