Support updating weights at once by stopping all requests (#6698)

Signed-off-by: Tianyu Zhou <albert.zty@antgroup.com>
Co-authored-by: Zilin Zhu <zhuzilinallen@gmail.com>
This commit is contained in:
Albert
2025-07-02 22:26:06 -07:00
committed by GitHub
co-authored by Zilin Zhu
parent b044400dd3
commit d3c275b117
7 changed files with 190 additions and 13 deletions
+9 -1
View File
@@ -740,6 +740,8 @@ class UpdateWeightFromDiskReqInput:
model_path: str
# The format to load the weights
load_format: Optional[str] = None
# Whether to abort all requests before updating weights
abort_all_requests: bool = False
@dataclass
@@ -759,6 +761,8 @@ class UpdateWeightsFromDistributedReqInput:
group_name: str = "weight_update_group"
# Whether to flush the cache after updating weights
flush_cache: bool = True
# Whether to abort all requests before updating weights
abort_all_requests: bool = False
@dataclass
@@ -780,6 +784,8 @@ class UpdateWeightsFromTensorReqInput:
load_format: Optional[str] = None
# Whether to flush the cache after updating weights
flush_cache: bool = True
# Whether to abort all requests before updating weights
abort_all_requests: bool = False
@dataclass
@@ -858,7 +864,9 @@ class SlowDownReqOutput:
@dataclass
class AbortReq:
# The request id
rid: str
rid: str = ""
# Whether to abort all requests
abort_all: bool = False
@dataclass