[model-gateway]: add gRPC router embeddings endpoint implementation (#15273)

python/sglang/srt/entrypoints/grpc_server.py

@@ -455,10 +455,22 @@ class SGLangSchedulerServicer(sglang_scheduler_pb2_grpc.SglangSchedulerServicer)
         input_text = grpc_req.tokenized.original_text
         input_ids = list(grpc_req.tokenized.input_ids)
 
+        # Convert sampling params
+        sampling_params = self._convert_sampling_params(grpc_req.sampling_params)
+
+        # For embedding requests, max_new_tokens should be 0.
+        # The scheduler logic expects an integer, not None.
+        sampling_params.max_new_tokens = 0
+
+        sampling_params.normalize(tokenizer=None)
+
         return TokenizedEmbeddingReqInput(
             rid=grpc_req.request_id,
             input_text=input_text,
             input_ids=input_ids,
+            image_inputs={"mm_items": []},
+            token_type_ids=list(grpc_req.token_type_ids),
+            sampling_params=sampling_params,
         )
 
     def _convert_sampling_params(

python/sglang/srt/grpc/grpc_request_manager.py

@@ -688,7 +688,9 @@ class GrpcRequestManager:
                     batch_out.prompt_tokens[i] if batch_out.prompt_tokens else 0
                 ),
                 "finish_reason": (
-                    batch_out.finish_reason[i] if batch_out.finish_reason else None
+                    batch_out.finished_reasons[i]
+                    if batch_out.finished_reasons
+                    else None
                 ),
             }
 

python/sglang/srt/managers/scheduler.py

@@ -1710,9 +1710,14 @@ class Scheduler(
         if recv_req.image_inputs is not None:
             image_inputs = self._get_multimodal_inputs(recv_req.image_inputs)
             # Expand a single image token into multiple dummy tokens for receiving image embeddings
-            req.origin_input_ids = self.pad_input_ids_func(
-                req.origin_input_ids, image_inputs
-            )
+            # The `pad_input_ids_func` is model-specific and may be None for
+            # embedding models or models not requiring special padding.
+            # If None, `req.origin_input_ids` is expected to be correctly populated already.
+            if self.pad_input_ids_func:
+                req.origin_input_ids = self.pad_input_ids_func(
+                    req.origin_input_ids, image_inputs
+                )
+
             req.extend_image_inputs(image_inputs)
 
             if len(req.origin_input_ids) >= self.max_req_input_len: