Support loading weights from remote instance (#8215)

Signed-off-by: Anqi Shen <amy.saq@antgroup.com>
Co-authored-by: Chayenne <74843776+zhaochenyang20@users.noreply.github.com>
This commit is contained in:
amysaq2023
2025-09-12 17:40:22 +08:00
committed by GitHub
parent 1b1701f1f7
commit 30d20ce84f
18 changed files with 1042 additions and 6 deletions

View File

@@ -1020,6 +1020,44 @@ class UpdateWeightsFromTensorReqOutput:
message: str
@dataclass
class InitWeightsSendGroupForRemoteInstanceReqInput:
# The master address
master_address: str
# The ports for each rank's communication group
ports: str
# The rank in the communication group
group_rank: int
# The world size
world_size: int
# The group name
group_name: str = "weight_send_group"
# The backend
backend: str = "nccl"
@dataclass
class InitWeightsSendGroupForRemoteInstanceReqOutput:
success: bool
message: str
@dataclass
class SendWeightsToRemoteInstanceReqInput:
# The master address
master_address: str
# The ports for each rank's communication group
ports: str
# The group name
group_name: str = "weight_send_group"
@dataclass
class SendWeightsToRemoteInstanceReqOutput:
success: bool
message: str
@dataclass
class InitWeightsUpdateGroupReqInput:
# The master address