feat: Add model version tracking with API endpoints and response metadata (#8795)

python/sglang/srt/managers/io_struct.py

@@ -798,6 +798,8 @@ class UpdateWeightFromDiskReqInput:
     load_format: Optional[str] = None
     # Whether to abort all requests before updating weights
     abort_all_requests: bool = False
+    # Optional: Update weight version along with weights
+    weight_version: Optional[str] = None
 
 
 @dataclass
@@ -819,6 +821,8 @@ class UpdateWeightsFromDistributedReqInput:
     flush_cache: bool = True
     # Whether to abort all requests before updating weights
     abort_all_requests: bool = False
+    # Optional: Update weight version along with weights
+    weight_version: Optional[str] = None
 
 
 @dataclass
@@ -842,6 +846,8 @@ class UpdateWeightsFromTensorReqInput:
     flush_cache: bool = True
     # Whether to abort all requests before updating weights
     abort_all_requests: bool = False
+    # Optional: Update weight version along with weights
+    weight_version: Optional[str] = None
 
 
 @dataclass
@@ -872,6 +878,14 @@ class InitWeightsUpdateGroupReqOutput:
     message: str
 
 
+@dataclass
+class UpdateWeightVersionReqInput:
+    # The new weight version
+    new_version: str
+    # Whether to abort all running requests before updating
+    abort_all_requests: bool = True
+
+
 @dataclass
 class GetWeightsByNameReqInput:
     name: str

python/sglang/srt/managers/tokenizer_manager.py

@@ -1529,6 +1529,7 @@ class TokenizerManager:
                 "id": rid,
                 "finish_reason": recv_obj.finished_reasons[i],
                 "prompt_tokens": recv_obj.prompt_tokens[i],
+                "weight_version": self.server_args.weight_version,
             }
 
             if getattr(state.obj, "return_logprob", False):