fix: `skip-tokenizer-init` by default in sglang (#2595)

components/backends/sglang/src/dynamo/sglang/args.py

@@ -128,6 +128,13 @@ def parse_args(args: list[str]) -> Config:
 
     server_args = ServerArgs.from_cli_args(parsed_args)
 
+    if not server_args.skip_tokenizer_init:
+        logging.warning(
+            "When using the dynamo frontend (python3 -m dynamo.frontend), we perform tokenization and detokenization "
+            "in the frontend. Automatically setting --skip-tokenizer-init to True."
+        )
+        server_args.skip_tokenizer_init = True
+
     return Config(server_args, dynamo_args)
 
 

components/backends/sglang/src/dynamo/sglang/request_handlers/decode_handler.py

@@ -102,7 +102,14 @@ class DecodeWorkerHandler(BaseWorkerHandler):
             if finish_reason:
                 out = {"token_ids": [], "finish_reason": finish_reason["type"]}
             else:
-                next_total_toks = len(res["output_ids"])
+                try:
+                    next_total_toks = len(res["output_ids"])
+                except KeyError:
+                    raise ValueError(
+                        f"Missing 'output_ids' in response. This often happens when using skip_tokenizer_init=True. "
+                        f"If you're using ModelType.CHAT or custom model configurations, you may need to modify "
+                        f"the tokenization/detokenization logic in your handler. Response keys: {list(res.keys())}"
+                    )
                 out = {"token_ids": res["output_ids"][num_output_tokens_so_far:]}
                 num_output_tokens_so_far = next_total_toks