Unify over-length errors into the PayloadTooLargeError 413 format

Over-long inputs produced two different client errors depending on
which bound rejected them: the TokenizerManager pre-check (raw
context_len) returned 413 PayloadTooLargeError ('The input (N tokens)
is longer than the model's context length (M tokens).'), while inputs
between that and the scheduler's stricter effective limit hit
validate_input_length and returned 400 BAD_REQUEST with different
wording (and a confusing 'X exceeds X' message since the check is >=).

Unify on the 413 format end to end:
- validate_input_length wording now matches the TokenizerManager
  message, reporting the effective per-request limit.
- set_finish_with_abort takes status_code/err_type; the scheduler
  length-rejection sites abort with REQUEST_ENTITY_TOO_LARGE +
  PayloadTooLargeError. The batch handler previously queued the
  over-long request WITHOUT marking it aborted (it proceeded to
  prefill) — also fixed.
- Non-streaming aborts with 413 raise PayloadTooLargeError (now a
  ValueError subclass so raw /generate-style endpoints that only
  catch ValueError still respond; the OpenAI layer's except clause
  is reordered to win and emit the 413 format).
- Streaming abort responses prefer the scheduler-provided err_type
  over the HTTPStatus name.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

python/sglang/srt/entrypoints/openai/serving_base.py

@@ -130,18 +130,19 @@ class OpenAIServingBase(ABC):
             return self.create_error_response(
                 message=e.detail, err_type=str(e.status_code), status_code=e.status_code
             )
+        except PayloadTooLargeError as e:
+            # Must precede ValueError: PayloadTooLargeError subclasses it.
+            return self.create_error_response(
+                message=str(e),
+                err_type="PayloadTooLargeError",
+                status_code=413,
+            )
         except ValueError as e:
             return self.create_error_response(
                 message=str(e),
                 err_type="BadRequest",
                 status_code=400,
             )
-        except PayloadTooLargeError as e:
-            return self.create_error_response(
-                message=str(e),
-                err_type="PayloadTooLargeError",
-                status_code=413,
-            )
         except DS32EncodingError as e:
             logger.info(f"DS32EncodingError: {e}")
             return self.create_error_response(

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -970,6 +970,9 @@ class OpenAIServingChat(OpenAIServingBase):
                         err_type, status_code = (
                             self._streaming_http_error_type_and_status(code)
                         )
+                        # Prefer the scheduler-provided error type (e.g.
+                        # PayloadTooLargeError) over the HTTPStatus name.
+                        err_type = finish_reason.get("err_type") or err_type
                         error = self.create_streaming_error_response(
                             finish_reason.get("message", "Generation aborted."),
                             err_type,

python/sglang/srt/entrypoints/openai/serving_completions.py

@@ -281,6 +281,9 @@ class OpenAIServingCompletion(OpenAIServingBase):
                     err_type, status_code = self._streaming_http_error_type_and_status(
                         code
                     )
+                    # Prefer the scheduler-provided error type (e.g.
+                    # PayloadTooLargeError) over the HTTPStatus name.
+                    err_type = finish_reason.get("err_type") or err_type
                     error = self.create_streaming_error_response(
                         finish_reason.get("message", "Generation aborted."),
                         err_type,

python/sglang/srt/managers/schedule_batch.py

@@ -1203,7 +1203,12 @@ class Req(ReqDllmMixin):
             self.extend_input_len,
         )
 
-    def set_finish_with_abort(self, error_msg: str):
+    def set_finish_with_abort(
+        self,
+        error_msg: str,
+        status_code: HTTPStatus = HTTPStatus.BAD_REQUEST,
+        err_type: str = "BadRequestError",
+    ):
         if get_tensor_model_parallel_rank() == 0:
             logger.error(f"{error_msg}, {self.rid=}")
         self.multimodal_inputs = None
@@ -1211,9 +1216,7 @@ class Req(ReqDllmMixin):
         self.origin_input_ids = [0]  # set it to one token to skip the long prefill
         self.return_logprob = False
         self.logprob_start_len = -1
-        self.to_finish = FINISH_ABORT(
-            error_msg, HTTPStatus.BAD_REQUEST, "BadRequestError"
-        )
+        self.to_finish = FINISH_ABORT(error_msg, status_code, err_type)
 
     def __repr__(self):
         return (

python/sglang/srt/managers/scheduler.py

@@ -1976,7 +1976,11 @@ class Scheduler(
             self.server_args.allow_auto_truncate,
         )
         if error_msg:
-            req.set_finish_with_abort(error_msg)
+            req.set_finish_with_abort(
+                error_msg,
+                status_code=HTTPStatus.REQUEST_ENTITY_TOO_LARGE,
+                err_type="PayloadTooLargeError",
+            )
             self._add_request_to_queue(req)
             return
 
@@ -2225,6 +2229,13 @@ class Scheduler(
             self.server_args.allow_auto_truncate,
         )
         if error_msg:
+            # NOTE: this path previously queued the over-long request without
+            # marking it aborted, letting it proceed to prefill.
+            req.set_finish_with_abort(
+                error_msg,
+                status_code=HTTPStatus.REQUEST_ENTITY_TOO_LARGE,
+                err_type="PayloadTooLargeError",
+            )
             self._add_request_to_queue(req)
             return
 

python/sglang/srt/managers/tokenizer_manager.py

@@ -122,8 +122,13 @@ asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 _REQUEST_STATE_WAIT_TIMEOUT = envs.SGLANG_REQUEST_STATE_WAIT_TIMEOUT.get()
 
 
-class PayloadTooLargeError(Exception):
-    """Exception raised when a request payload exceeds the model context length."""
+class PayloadTooLargeError(ValueError):
+    """Exception raised when a request payload exceeds the model context length.
+
+    Subclasses ValueError so callers that only handle ValueError (e.g. the raw
+    /generate endpoint) still return an error response; the OpenAI serving
+    layer catches it first to produce the 413 PayloadTooLargeError format.
+    """
 
 
 logger = logging.getLogger(__name__)
@@ -1207,12 +1212,20 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
                     # Check if this was an abort/error created by scheduler
                     if isinstance(out["meta_info"].get("finish_reason"), dict):
                         finish_reason = out["meta_info"]["finish_reason"]
-                        if (
-                            finish_reason.get("type") == "abort"
-                            and finish_reason.get("status_code")
-                            == HTTPStatus.BAD_REQUEST
+                        if finish_reason.get("type") == "abort" and finish_reason.get(
+                            "status_code"
+                        ) in (
+                            HTTPStatus.BAD_REQUEST,
+                            HTTPStatus.REQUEST_ENTITY_TOO_LARGE,
                         ):
                             if not is_stream:
+                                if (
+                                    finish_reason.get("status_code")
+                                    == HTTPStatus.REQUEST_ENTITY_TOO_LARGE
+                                ):
+                                    raise PayloadTooLargeError(
+                                        finish_reason["message"]
+                                    )
                                 raise ValueError(finish_reason["message"])
                             else:
                                 yield out

python/sglang/srt/managers/utils.py

@@ -133,10 +133,13 @@ def validate_input_length(
             req.origin_input_ids = req.origin_input_ids[:max_req_input_len]
             return None
         else:
+            # Keep the wording identical to the TokenizerManager-side
+            # PayloadTooLargeError so clients see one over-length format
+            # regardless of which bound (raw context length there, effective
+            # per-request limit here) rejected the request.
             error_msg = (
-                f"Input length ({len(req.origin_input_ids)} tokens) exceeds "
-                f"the maximum allowed length ({max_req_input_len} tokens). "
-                f"Use a shorter input or enable --allow-auto-truncate."
+                f"The input ({len(req.origin_input_ids)} tokens) is longer "
+                f"than the model's context length ({max_req_input_len} tokens)."
             )
             return error_msg