fa08cfd489
This contains all performance hints to the API. Currently there's only bytes_per_pack, which splits large batches into multiple packs allows overlapping communication and computation. Currently we can only pack if all Tensors in the batch have known shapes. PiperOrigin-RevId: 297269428 Change-Id: Iaf7d7d3adf7c6cad59aa6079fbcd36b31e92c4b5
34 lines
996 B
Plaintext
34 lines
996 B
Plaintext
path: "tensorflow.distribute.ReplicaContext"
|
|
tf_class {
|
|
is_instance: "<class \'tensorflow.python.distribute.distribute_lib.ReplicaContext\'>"
|
|
is_instance: "<type \'object\'>"
|
|
member {
|
|
name: "devices"
|
|
mtype: "<type \'property\'>"
|
|
}
|
|
member {
|
|
name: "num_replicas_in_sync"
|
|
mtype: "<type \'property\'>"
|
|
}
|
|
member {
|
|
name: "replica_id_in_sync_group"
|
|
mtype: "<type \'property\'>"
|
|
}
|
|
member {
|
|
name: "strategy"
|
|
mtype: "<type \'property\'>"
|
|
}
|
|
member_method {
|
|
name: "__init__"
|
|
argspec: "args=[\'self\', \'strategy\', \'replica_id_in_sync_group\'], varargs=None, keywords=None, defaults=None"
|
|
}
|
|
member_method {
|
|
name: "all_reduce"
|
|
argspec: "args=[\'self\', \'reduce_op\', \'value\', \'experimental_hints\'], varargs=None, keywords=None, defaults=[\'None\'], "
|
|
}
|
|
member_method {
|
|
name: "merge_call"
|
|
argspec: "args=[\'self\', \'merge_fn\', \'args\', \'kwargs\'], varargs=None, keywords=None, defaults=[\'()\', \'None\'], "
|
|
}
|
|
}
|