Default package

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: RequestAuthInfo, ResponseAuthInfo

• string username = 1

• bytes public_key = 2

• string expiration_time = 3

• bytes signature = 4

• MessageCode code = 1

  in case of a protocol violation, this will be the error message

• bytes group_id = 2

  a unique group identifier, same as in MessageFromLeader

• bytes peer_id = 3

  sender's rpc peer_id, used for coordination

• optional Tensor tensor_part = 4

  either peer's local tensor part (rpc input) or group average of this part (rpc output)

• double weight = 5

  peers will be averaged in proportion to these weights

Used in: Tensor

• NONE = 0

• MEANSTD_16BIT = 1

• FLOAT16 = 2

• QUANTILE_8BIT = 3

• UNIFORM_8BIT = 4

• BLOCKWISE_8BIT = 5

• bytes metadata = 1

• optional Tensor tensor_part = 2

(message has no fields)

• bytes serialized_info = 1

• string uid = 1

• repeated Tensor tensors = 2

• bytes metadata = 3

• repeated Tensor tensors = 2

• bytes metadata = 3

• string uid = 1

• optional RequestAuthInfo auth = 1

• repeated bytes keys = 2

  a list of DHTID search keys encoded as bytes

• optional NodeInfo peer = 3

  optional, same behavior as in DHT.ping

• optional ResponseAuthInfo auth = 1

• repeated FindResult results = 2

  for each item, return value/expiration (if found) and nearest peers

• optional NodeInfo peer = 3

  respondent's node id, for you to update routing table

Used in: FindResponse

• ResultType type = 1

  NONE | REGULAR | DICTIONARY

• bytes value = 2

  n/a | serialized value | serialized DictionaryDHTValue with serialized fields

• double expiration_time = 3

  n/a | expiration time | DictionaryDHTValue.latest_expiration_time

• repeated bytes nearest_node_ids = 4

  two aligned arrays: DHTIDs and PeerIDs for nearest peers (sorted by XOR distance)

  DHTIDs of the nearest peers serialized with node_id.to_bytes()

• repeated bytes nearest_peer_ids = 5

  libp2p PeerIDs of the nearest peers

• bytes schema_hash = 2

  A hash that describes follower's tensors (shapes, num tensors, etc)

• double expiration = 3

  Follower would like to **begin** all_reduce by this point in time

• bytes gather = 4

  optional metadata that is gathered from all peers (e.g. batch size or current loss)

• bool client_mode = 5

  if True, the incoming averager is a client with no capacity for averaging

• string group_key = 6

  group key identifying an All-Reduce bucket, e.g my_averager.0b011011101

Used in: AveragingData, MessageFromLeader

• NO_CODE = 0

  Default value that should not be used explicitly

• REQUEST_JOIN = 1

  "Dear maybe leader, will you have me in your group as a follower?"

• ACCEPTED = 2

  "I accept you in my group, you now commit to responding to me"

• BEGIN_ALLREDUCE = 3

  "We can begin allreduce now. These are your peers."

• PART_FOR_AVERAGING = 4

  "I am running allreduce with you, here's a part of my tensor that you should aggregate"

• AVERAGED_PART = 5

  "I aggregated your part with others and here's the average for that part"

• NOT_DECLARED = 6

  "I have not declared my group id yet, how the heck did you even find me? Go away."

• NOT_A_LEADER = 7

  "I am not a group a leader. Go ask my leader instead."

• BAD_EXPIRATION_TIME = 8

  "I will not accept you. I cannot guarantee that we begin before you expire."

• BAD_SCHEMA_HASH = 9

  "I will not accept you. I am not averaging the samy type of tensors as you."

• BAD_GROUP_ID = 10

  "I will not accept your request, your group id does not match with any groups i'm in."

• DUPLICATE_PEER_ID = 11

  "I will not accept you, i already have exactly the same peer id in my current group."

• GROUP_IS_FULL = 12

  "I will not accept you, my group already contains too many peers."

• NOT_LOOKING_FOR_GROUP = 13

  "I'm not available at the moment. Please, get lost."

• PROTOCOL_VIOLATION = 14

  "You did something so unspeakable that i don't have a special code for that."

• INTERNAL_ERROR = 15

  "I messed up, we will have to stop allreduce because of that."

• CANCELLED = 16

  "[from peer during allreduce] I no longer want to participate in AllReduce."

• GROUP_DISBANDED = 17

  "[from leader] The group is closed. Go find another group."

• BAD_GROUP_KEY = 18

  "I will not accept you. My current group key differs (maybe you used my older key)."

• MessageCode code = 1

• bytes group_id = 2

  a unique identifier of this group, only valid until allreduce is finished/failed

• bytes suggested_leader = 3

  if peer is already in a group, it'll provide us with a peer id of its leader

• repeated bytes ordered_peer_ids = 4

  a sequence of peers, each responsible for one shard during averaging

• repeated bytes gathered = 5

  metadata (gather) from all groupmates in the same order as their peer ids

Used in: FindRequest, FindResponse, PingRequest, PingResponse, StoreRequest, StoreResponse

• bytes node_id = 1

  note: both node_id and port are optional: if specified, ask peer to add you to its routing table; if either node_id or port is absent, simply request recipient info (for client-only mode)

  sender's own node id serialized with DHTID.to_bytes()

• optional RequestAuthInfo auth = 1

• optional NodeInfo peer = 2

  (optional) sender's own node info, same behavior as in DHT.rpc_ping

• bool validate = 3

  set to True if sender wants to validate that he is accessible and synchronized

• optional ResponseAuthInfo auth = 1

• optional NodeInfo peer = 2

  respondent's node id, for you to update routing table

• double dht_time = 4

  recipient's local DHT time - used to soft-synchronize peers

• bool available = 5

  if validate = True, this flag asserts that the sender is available for ping

Used in: FindRequest, PingRequest, StoreRequest

• optional AccessToken client_access_token = 1

• bytes service_public_key = 2

• double time = 3

• bytes nonce = 4

• bytes signature = 5

Used in: FindResponse, PingResponse, StoreResponse

• optional AccessToken service_access_token = 1

• bytes nonce = 2

• bytes signature = 3

Used in: FindResult

• NOT_FOUND = 0

• FOUND_REGULAR = 1

• FOUND_DICTIONARY = 2

• optional RequestAuthInfo auth = 1

• repeated bytes keys = 2

  three lists of the same length representing dht keys, dht values and expiration

  keys in the form of DHTID.generate(raw_key).to_bytes()

• repeated bytes subkeys = 3

  serialized subkeys for DictionaryDHTValue type. None means no subkey

• repeated bytes values = 4

  serialized value for i-th key

• repeated double expiration_time = 5

  expirations for i-th key (type = DHTExpiration)

• repeated bool in_cache = 6

  if in_cache[i], store i-th key in cache, else store normally

• optional NodeInfo peer = 7

  (optional) sender's own node info, same behavior as in DHT.rpc_ping

• optional ResponseAuthInfo auth = 1

• repeated bool store_ok = 2

  for every key, True means store accepted, False means store rejected/failed

• optional NodeInfo peer = 3

  respondent's node id, for you to update routing table

Used in: AveragingData, DownloadData, ExpertRequest, ExpertResponse

• bytes buffer = 1

• repeated uint32 size = 2

• bool requires_grad = 3

• string dtype = 4

• CompressionType compression = 5

• int32 chunks = 6

• int32 number = 1

• int32 number = 1