Clean up detokenizer and remove dead multimodal_gen code (#21588)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

python/sglang/srt/configs/model_config.py

@@ -178,12 +178,6 @@ class ModelConfig:
         self.is_multimodal = enable_multimodal and is_multimodal_model(
             self.hf_config.architectures
         )
-        self.is_multimodal_gen = enable_multimodal and is_multimodal_gen_model(
-            self.hf_config.architectures
-        )
-        self.is_image_gen = enable_multimodal and is_image_gen_model(
-            self.hf_config.architectures
-        )
         self.is_audio_model = enable_multimodal and is_audio_model(
             self.hf_config.architectures
         )
@@ -1351,14 +1345,6 @@ def is_multimodal_model(model_architectures: List[str]):
         return False
 
 
-def is_multimodal_gen_model(model_architectures: List[str]):
-    return False
-
-
-def is_image_gen_model(model_architectures: List[str]):
-    return False
-
-
 def is_audio_model(model_architectures: List[str]):
     models = [
         "WhisperForConditionalGeneration",

python/sglang/srt/constants.py

@@ -8,3 +8,5 @@ GPU_MEMORY_ALL_TYPES = [
     GPU_MEMORY_TYPE_WEIGHTS,
     GPU_MEMORY_TYPE_CUDA_GRAPH,
 ]
+
+HEALTH_CHECK_RID_PREFIX = "HEALTH_CHECK"

python/sglang/srt/entrypoints/http_server.py

@@ -59,6 +59,7 @@ from fastapi.exceptions import RequestValidationError
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import ORJSONResponse, Response, StreamingResponse
 
+from sglang.srt.constants import HEALTH_CHECK_RID_PREFIX
 from sglang.srt.disaggregation.utils import FAKE_BOOTSTRAP_HOST, DisaggregationMode
 from sglang.srt.entrypoints.anthropic.protocol import (
     AnthropicCountTokensRequest,
@@ -509,7 +510,7 @@ async def health_generate(request: Request) -> Response:
         return Response(status_code=200)
 
     sampling_params = {"max_new_tokens": 1, "temperature": 0.0}
-    rid = f"HEALTH_CHECK_{time.time()}"
+    rid = f"{HEALTH_CHECK_RID_PREFIX}_{time.time()}"
 
     if _global_state.tokenizer_manager.is_image_gen:
         gri = _global_state.tokenizer_manager.get_image_gen_health_check_request(

python/sglang/srt/managers/detokenizer_manager.py

@@ -25,10 +25,10 @@ import pybase64
 import setproctitle
 import zmq
 
+from sglang.srt.constants import HEALTH_CHECK_RID_PREFIX
 from sglang.srt.environ import envs
 from sglang.srt.managers.io_struct import (
     BatchEmbeddingOutput,
-    BatchMultimodalDecodeReq,
     BatchStrOutput,
     BatchTokenIDOutput,
     FreezeGCReq,
@@ -88,16 +88,9 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
         # Init running status
         self.init_running_status(server_args)
 
-        if server_args.enable_metrics:
-            start_cpu_monitor_thread("detokenizer")
-
         # Init dispatcher
         self.init_request_dispatcher()
 
-    @staticmethod
-    def is_health_check_request(rid: Optional[str]) -> bool:
-        return isinstance(rid, str) and rid.startswith("HEALTH_CHECK")
-
     def init_ipc_channels(self, port_args: PortArgs):
         context = zmq.Context(2)
         self.recv_from_scheduler = get_zmq_socket(
@@ -120,9 +113,8 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
 
     def init_running_status(self, server_args: ServerArgs):
         self.decode_status = LimitedCapacityDict(capacity=DETOKENIZER_MAX_STATES)
-        self.is_dummy = False
-        self.is_tool_call_parser_gpt_oss = server_args.tool_call_parser == "gpt-oss"
         self.disable_tokenizer_batch_decode = server_args.disable_tokenizer_batch_decode
+        self.is_tool_call_parser_gpt_oss = server_args.tool_call_parser == "gpt-oss"
 
         self.soft_watchdog = Watchdog.create(
             debug_name="DetokenizerManager",
@@ -131,12 +123,14 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
             test_stuck_time=envs.SGLANG_TEST_STUCK_DETOKENIZER.get(),
         )
 
+        if server_args.enable_metrics:
+            start_cpu_monitor_thread("detokenizer")
+
     def init_request_dispatcher(self):
         self._request_dispatcher = TypeBasedDispatcher(
             [
                 (BatchEmbeddingOutput, self.handle_batch_embedding_out),
                 (BatchTokenIDOutput, self.handle_batch_token_id_out),
-                (BatchMultimodalDecodeReq, self.handle_multimodal_decode_req),
                 (FreezeGCReq, self.handle_freeze_gc_req),
             ]
         )
@@ -190,8 +184,6 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
     ) -> List[str]:
         """Batch decode with grouping by (skip_special_tokens, spaces_between_special_tokens)."""
 
-        assert self.tokenizer is not None
-
         # fast path
         first_skip, first_space = skip_list[0], space_list[0]
         if all(s == first_skip for s in skip_list) and all(
@@ -236,9 +228,7 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
                     surr_offset=0,
                     read_offset=recv_obj.read_offsets[i],
                 )
-                if not self.is_health_check_request(rid):
-                    # for health check requests, we do not store the decode status
-                    self.decode_status[rid] = s
+                self.decode_status[rid] = s
             else:
                 s = self.decode_status[rid]
                 s.decode_ids.extend(recv_obj.decode_ids[i])
@@ -254,22 +244,16 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
 
         # Decode token ids to strings
         if not self.disable_tokenizer_batch_decode:
-            if not self.is_dummy:
-                # Run normal batch decode
-                surr_texts = self._grouped_batch_decode(
-                    surr_ids,
-                    recv_obj.skip_special_tokens,
-                    recv_obj.spaces_between_special_tokens,
-                )
-                read_texts = self._grouped_batch_decode(
-                    read_ids,
-                    recv_obj.skip_special_tokens,
-                    recv_obj.spaces_between_special_tokens,
-                )
-            else:
-                # If it is dummy weights, just return dummy strings to prevent potential detokenization edge cases
-                surr_texts = ["dog" for _ in surr_ids]
-                read_texts = ["cat" for _ in read_ids]
+            surr_texts = self._grouped_batch_decode(
+                surr_ids,
+                recv_obj.skip_special_tokens,
+                recv_obj.spaces_between_special_tokens,
+            )
+            read_texts = self._grouped_batch_decode(
+                read_ids,
+                recv_obj.skip_special_tokens,
+                recv_obj.spaces_between_special_tokens,
+            )
         else:
             # Do not use batch decode to prevent some detokenization edge cases (e.g., gpt-oss).
             surr_texts = [
@@ -297,25 +281,17 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
         output_strs = []
         for i in range(bs):
             rid = recv_obj.rids[i]
-            if self.is_health_check_request(rid):
-                s = DecodeStatus(
-                    decoded_text=recv_obj.decoded_texts[i],
-                    decode_ids=recv_obj.decode_ids[i],
-                    surr_offset=0,
-                    read_offset=recv_obj.read_offsets[i],
+            try:
+                s = self.decode_status[rid]
+            except KeyError:
+                raise RuntimeError(
+                    f"Decode status not found for request {rid}. "
+                    "It may be due to the request being evicted from the decode status due to memory pressure. "
+                    "Please increase the maximum number of requests by setting "
+                    "the SGLANG_DETOKENIZER_MAX_STATES environment variable to a bigger value than the default value. "
+                    f"The current value is {DETOKENIZER_MAX_STATES}. "
+                    "For more details, see: https://github.com/sgl-project/sglang/issues/2812"
                 )
-            else:
-                try:
-                    s = self.decode_status[rid]
-                except KeyError:
-                    raise RuntimeError(
-                        f"Decode status not found for request {rid}. "
-                        "It may be due to the request being evicted from the decode status due to memory pressure. "
-                        "Please increase the maximum number of requests by setting "
-                        "the SGLANG_DETOKENIZER_MAX_STATES environment variable to a bigger value than the default value. "
-                        f"The current value is {DETOKENIZER_MAX_STATES}. "
-                        "For more details, see: https://github.com/sgl-project/sglang/issues/2812"
-                    )
             new_text = read_texts[i][len(surr_texts[i]) :]
             if recv_obj.finished_reasons[i] is None:
                 # Streaming chunk: update the decode status
@@ -335,6 +311,7 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
                 recv_obj.finished_reasons[i],
                 recv_obj.no_stop_trim[i],
             )
+
             # Incrementally send text.
             incremental_output = output_str[s.sent_offset :]
             s.sent_offset = len(output_str)
@@ -404,14 +381,15 @@ class DetokenizerManager(MultiHttpWorkerDetokenizerMixin):
             time_stats=recv_obj.time_stats,
         )
 
-    def handle_multimodal_decode_req(self, recv_obj: BatchMultimodalDecodeReq):
-        raise NotImplementedError()
-
     def handle_freeze_gc_req(self, recv_req: FreezeGCReq):
         freeze_gc("Detokenizer Manager")
         return None
 
 
+def is_health_check_request(rid: Optional[str]) -> bool:
+    return isinstance(rid, str) and rid.startswith(HEALTH_CHECK_RID_PREFIX)
+
+
 class LimitedCapacityDict(OrderedDict):
     def __init__(self, capacity: int, *args, **kwargs):
         super().__init__(*args, **kwargs)

python/sglang/srt/managers/scheduler.py

@@ -39,6 +39,7 @@ from torch.distributed import barrier
 
 from sglang.jit_kernel.ngram_embedding import update_token_table
 from sglang.srt.configs.model_config import ModelConfig
+from sglang.srt.constants import HEALTH_CHECK_RID_PREFIX
 from sglang.srt.constrained.grammar_manager import GrammarManager
 from sglang.srt.disaggregation.decode import (
     DecodePreallocQueue,
@@ -3412,7 +3413,7 @@ class IdleSleeper:
 
 def is_health_check_generate_req(recv_req):
     rid = getattr(recv_req, "rid", None)
-    return rid is not None and rid.startswith("HEALTH_CHECK")
+    return rid is not None and rid.startswith(HEALTH_CHECK_RID_PREFIX)
 
 
 def is_work_request(recv_req):

python/sglang/srt/managers/scheduler_output_processor_mixin.py

@@ -953,10 +953,6 @@ class SchedulerOutputProcessorMixin:
             if req is skip_req:
                 continue
 
-            # Multimodal partial stream chunks break the detokenizer, so drop aborted requests here.
-            if self.model_config.is_multimodal_gen and req.to_finish:
-                continue
-
             if req.finished():
                 if req.finished_output:
                     # With the overlap schedule, a request will try to output twice and hit this line twice
@@ -975,8 +971,7 @@ class SchedulerOutputProcessorMixin:
                     # origin stream_interval logic
                     should_output = (
                         len(req.output_ids) % stream_interval == 1
-                        if not self.model_config.is_multimodal_gen
-                        and stream_interval > 1
+                        if stream_interval > 1
                         else len(req.output_ids) % stream_interval == 0
                     )
 
@@ -986,8 +981,6 @@ class SchedulerOutputProcessorMixin:
                 else:
                     should_output = (
                         len(req.output_ids) % DEFAULT_FORCE_STREAM_INTERVAL == 0
-                        if not self.model_config.is_multimodal_gen
-                        else False
                     )
 
             if should_output:
@@ -1003,10 +996,7 @@ class SchedulerOutputProcessorMixin:
                 decoded_texts.append(req.decoded_text)
                 decode_ids, read_offset = req.init_incremental_detokenize()
 
-                if self.model_config.is_multimodal_gen:
-                    decode_ids_list.append(decode_ids)
-                else:
-                    decode_ids_list.append(decode_ids[req.send_decode_id_offset :])
+                decode_ids_list.append(decode_ids[req.send_decode_id_offset :])
 
                 # Exclude the tokens after stop condition
                 output_ids_ = req.output_ids_through_stop
@@ -1132,8 +1122,6 @@ class SchedulerOutputProcessorMixin:
 
         # Send to detokenizer
         if reqs or is_idle_batch:
-            if self.model_config.is_multimodal_gen:
-                return
             self.send_to_detokenizer.send_output(
                 BatchTokenIDOutput(
                     rids=rids,

python/sglang/srt/managers/tokenizer_manager.py

@@ -231,7 +231,7 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
         self.served_model_name = server_args.served_model_name
         self.model_config = model_config_class.from_server_args(server_args)
         self.is_generation = self.model_config.is_generation
-        self.is_image_gen = self.model_config.is_image_gen
+        self.is_image_gen = getattr(self.model_config, "is_image_gen", False)
         self.context_len = self.model_config.context_len
         self.image_token_id = self.model_config.image_token_id
         self.max_req_input_len = None  # Will be set later in engine.py
@@ -1194,7 +1194,6 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
                 self.request_logger.log_finished_request(
                     obj,
                     out,
-                    is_multimodal_gen=self.model_config.is_multimodal_gen,
                     request=request,
                 )
 

python/sglang/srt/observability/request_metrics_exporter.py

@@ -7,6 +7,7 @@ from abc import ABC, abstractmethod
 from datetime import datetime
 from typing import List, Optional, Union
 
+from sglang.srt.constants import HEALTH_CHECK_RID_PREFIX
 from sglang.srt.managers.io_struct import EmbeddingReqInput, GenerateReqInput
 from sglang.srt.server_args import ServerArgs
 
@@ -128,7 +129,7 @@ class FileRequestMetricsExporter(RequestMetricsExporter):
         self, obj: Union[GenerateReqInput, EmbeddingReqInput], out_dict: dict
     ):
         # Do not log health check requests, since they don't represent real user requests.
-        if isinstance(obj.rid, str) and "HEALTH_CHECK" in obj.rid:
+        if isinstance(obj.rid, str) and HEALTH_CHECK_RID_PREFIX in obj.rid:
             return
 
         try:

python/sglang/srt/utils/common.py

@@ -20,6 +20,7 @@ import asyncio
 import builtins
 import ctypes
 import functools
+import gc
 import importlib
 import inspect
 import io
@@ -2945,8 +2946,6 @@ def configure_gc_warning(warn_threshold_secs):
 
 
 def freeze_gc(context: str):
-    import gc
-
     g0_before, g1_before, g2_before = gc_object_counts()
     gc.freeze()
     g0_after, g1_after, g2_after = gc_object_counts()
@@ -2961,8 +2960,6 @@ def freeze_gc(context: str):
 def configure_gc_logger():
     logger.info("Enable GC Logger")
 
-    import gc
-
     gc_start_time = {}
 
     def gc_callback(phase, info):

python/sglang/srt/utils/request_logger.py

@@ -162,7 +162,6 @@ class RequestLogger:
         self,
         obj: Union["GenerateReqInput", "EmbeddingReqInput"],
         out: Any,
-        is_multimodal_gen: bool = False,
         request: Optional["fastapi.Request"] = None,
     ) -> None:
         if not self.log_requests:
@@ -181,20 +180,15 @@ class RequestLogger:
             }
             if headers:
                 log_data["headers"] = headers
-            if not is_multimodal_gen:
-                log_data["out"] = _transform_data_for_logging(
-                    out, max_length, out_skip_names
-                )
+            log_data["out"] = _transform_data_for_logging(
+                out, max_length, out_skip_names
+            )
             log_json(self.targets, "request.finished", log_data)
         else:
             obj_str = _dataclass_to_string_truncated(
                 obj, max_length, skip_names=skip_names
             )
-            out_str = (
-                ""
-                if is_multimodal_gen
-                else f", out={_dataclass_to_string_truncated(out, max_length, skip_names=out_skip_names)}"
-            )
+            out_str = f", out={_dataclass_to_string_truncated(out, max_length, skip_names=out_skip_names)}"
             headers_str = f", headers={headers}" if headers else ""
             self._log(f"Finish: obj={obj_str}{headers_str}{out_str}")
 

test/registered/bench_fn/test_bench_serving_functionality.py

@@ -8,6 +8,7 @@ from pathlib import Path
 
 from sglang.bench_serving import run_benchmark
 from sglang.benchmark.utils import parse_custom_headers
+from sglang.srt.constants import HEALTH_CHECK_RID_PREFIX
 from sglang.srt.utils import kill_process_tree
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
@@ -80,7 +81,7 @@ class TestBenchServingFunctionality(CustomTestCase):
                 continue
             text = obj.get("obj", {}).get("text")
             rid = obj.get("rid", "")
-            if text and not rid.startswith("HEALTH_CHECK"):
+            if text and not rid.startswith(HEALTH_CHECK_RID_PREFIX):
                 reqs.append(text)
 
         self.assertGreaterEqual(len(reqs), NUM_CONVERSATIONS * NUM_TURNS)

test/registered/unit/observability/test_request_metrics_exporter.py

@@ -58,6 +58,7 @@ import tempfile
 import unittest
 from unittest.mock import MagicMock, patch
 
+from sglang.srt.constants import HEALTH_CHECK_RID_PREFIX
 from sglang.srt.observability.request_metrics_exporter import (
     FileRequestMetricsExporter,
     RequestMetricsExporter,
@@ -243,7 +244,7 @@ class TestFileRequestMetricsExporter(unittest.TestCase):
 
     def test_write_record_skips_health_check(self):
         exporter = self._make_exporter()
-        obj = _GenerateReqInput(rid="HEALTH_CHECK_123", text="ping")
+        obj = _GenerateReqInput(rid=f"{HEALTH_CHECK_RID_PREFIX}_123", text="ping")
         asyncio.run(exporter.write_record(obj, {}))
 
         files = os.listdir(self.tmp_dir)

test/registered/utils/test_request_logger.py

@@ -8,6 +8,7 @@ from pathlib import Path
 
 import requests
 
+from sglang.srt.constants import HEALTH_CHECK_RID_PREFIX
 from sglang.srt.utils import kill_process_tree
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
@@ -197,7 +198,7 @@ class TestRequestLoggerJson(BaseTestRequestLogger, CustomTestCase):
                 continue
 
             rid = data.get("rid", "")
-            if rid.startswith("HEALTH_CHECK"):
+            if rid.startswith(HEALTH_CHECK_RID_PREFIX):
                 continue
 
             if data.get("event") == "request.received":