Test the case when max_new_tokens is very large (#1038)

python/sglang/srt/managers/detokenizer_manager.py

@@ -32,7 +32,7 @@ from sglang.srt.managers.io_struct import (
 )
 from sglang.srt.managers.schedule_batch import FINISH_MATCHED_STR
 from sglang.srt.server_args import PortArgs, ServerArgs
-from sglang.utils import find_printable_text, get_exception_traceback, graceful_registry
+from sglang.utils import find_printable_text, get_exception_traceback
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 
@@ -164,8 +164,6 @@ def start_detokenizer_process(
     port_args: PortArgs,
     pipe_writer,
 ):
-    graceful_registry(inspect.currentframe().f_code.co_name)
-
     try:
         manager = DetokenizerManager(server_args, port_args)
     except Exception:

python/sglang/srt/managers/policy_scheduler.py

@@ -15,6 +15,7 @@ limitations under the License.
 
 """Request policy scheduler"""
 
+import os
 import random
 from collections import defaultdict
 from contextlib import contextmanager
@@ -24,9 +25,11 @@ from sglang.srt.managers.schedule_batch import Req, ScheduleBatch
 from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.radix_cache import TreeNode
 
-# Clip the max new tokens for the request whose max_new_tokens is very large.
+# Clip the estimation of max_new_tokens for the request whose max_new_tokens is very large.
 # This can prevent the server from being too conservative.
-CLIP_MAX_NEW_TOKENS = 4096
+# Note that this only clips the estimation in the scheduler but does not change the stop
+# condition. The request can still generate tokens until it hits the unclipped max_new_tokens.
+CLIP_MAX_NEW_TOKENS = int(os.environ.get("SGLANG_CLIP_MAX_NEW_TOKENS", "4096"))
 
 
 class PolicyScheduler:

python/sglang/srt/openai_api/adapter.py

@@ -77,7 +77,7 @@ class FileMetadata:
 batch_storage: Dict[str, BatchResponse] = {}
 file_id_request: Dict[str, FileMetadata] = {}
 file_id_response: Dict[str, FileResponse] = {}
-# map file id to file path in SGlang backend
+# map file id to file path in SGLang backend
 file_id_storage: Dict[str, str] = {}
 
 
@@ -335,7 +335,7 @@ async def process_batch(tokenizer_manager, batch_id: str, batch_request: BatchRe
         }
 
     except Exception as e:
-        print("error in SGlang:", e)
+        print("error in SGLang:", e)
         # Update batch status to "failed"
         retrieve_batch = batch_storage[batch_id]
         retrieve_batch.status = "failed"

python/sglang/srt/server_args.py

@@ -64,7 +64,7 @@ class ServerArgs:
 
     # Other
     api_key: Optional[str] = None
-    file_storage_pth: str = "SGlang_storage"
+    file_storage_pth: str = "SGLang_storage"
 
     # Data parallelism
     dp_size: int = 1

python/sglang/test/test_utils.py

@@ -398,6 +398,8 @@ def popen_launch_server(
     timeout: float,
     api_key: Optional[str] = None,
     other_args: tuple = (),
+    env: Optional[dict] = None,
+    return_stdout_stderr: bool = False,
 ):
     _, host, port = base_url.split(":")
     host = host[2:]
@@ -417,7 +419,16 @@ def popen_launch_server(
     if api_key:
         command += ["--api-key", api_key]
 
-    process = subprocess.Popen(command, stdout=None, stderr=None)
+    if return_stdout_stderr:
+        process = subprocess.Popen(
+            command,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            env=env,
+            text=True,
+        )
+    else:
+        process = subprocess.Popen(command, stdout=None, stderr=None, env=env)
 
     start_time = time.time()
     while time.time() - start_time < timeout: