For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python class

TransferReqData

TransferReqData

class max.pipelines.kv_cache.TransferReqData(*, dst_name, src_name, transfer_name, transfer_ids, src_idxs, dst_idxs, src_replica_idx, dst_replica_idx, is_read=False, tp_shard_count=0, local_shards_used=<factory>)

source

Bases: Struct

Metadata associated with a transfer request.

This is safe to send between threads/processes.

Parameters:

• dst_name (str)
• src_name (str)
• transfer_name (str)
• transfer_ids (list[int])
• src_idxs (list[int])
• dst_idxs (list[int])
• src_replica_idx (int)
• dst_replica_idx (int)
• is_read (bool)
• tp_shard_count (int)
• local_shards_used (list[int])

dst_idxs

dst_idxs: list[int]

source

Length of destination indices can differ from len(transfer_ids).

dst_name

dst_name: str

source

Base name of destination engine.

dst_replica_idx

dst_replica_idx: int

source

Index of the destination replica this transfer is to.

is_read

is_read: bool

source

True if this is a READ (pull) transfer initiated by the destination.

local_shards_used

local_shards_used: list[int]

source

Physical TP shard indices on the initiator that own this transfer’s handles. Empty means “all shards in the recorded replica” (pre-flatten behavior). Required to release/status-check transfers when a flattened group has picked a subset of shards.

src_idxs

src_idxs: list[int]

source

Length of source indices can differ from len(transfer_ids).

src_name

src_name: str

source

Base name of source engine.

src_replica_idx

src_replica_idx: int

source

Index of the source replica this transfer is from.

tp_shard_count

tp_shard_count: int

source

Number of TP shards participating. 0 = all shards (backwards compat).

transfer_ids

transfer_ids: list[int]

source

Transfer IDs (one per TP shard in the replica).

transfer_name

transfer_name: str

source

Transfer name.