sglang 0.3.6.post2__py3-none-any.whl → 0.3.6.post3__py3-none-any.whl

sglang/srt/server.py

@@ -23,6 +23,7 @@ import json
 import logging
 import multiprocessing as mp
 import os
+import signal
 import threading
 import time
 from http import HTTPStatus
@@ -79,7 +80,7 @@ from sglang.srt.utils import (
     configure_logger,
     delete_directory,
     is_port_available,
-    kill_child_process,
+    kill_process_tree,
     maybe_set_triton_cache_manager,
     prepare_model_and_tokenizer,
     set_prometheus_multiproc_dir,
@@ -92,7 +93,7 @@ logger = logging.getLogger(__name__)
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 
-
+# Fast API
 app = FastAPI()
 app.add_middleware(
     CORSMiddleware,
@@ -103,7 +104,7 @@ app.add_middleware(
 )
 
 tokenizer_manager: TokenizerManager = None
-_max_total_num_tokens = None
+scheduler_info: Dict = None
 
 ##### Native API endpoints #####
 
@@ -171,7 +172,7 @@ async def flush_cache():
 
 @app.get("/start_profile")
 @app.post("/start_profile")
-async def start_profile():
+async def start_profile_async():
     """Start profiling."""
     tokenizer_manager.start_profile()
     return Response(
@@ -182,7 +183,7 @@ async def start_profile():
 
 @app.get("/stop_profile")
 @app.post("/stop_profile")
-async def stop_profile():
+async def stop_profile_async():
     """Stop profiling."""
     tokenizer_manager.stop_profile()
     return Response(
@@ -233,6 +234,8 @@ async def close_session(obj: CloseSessionReqInput, request: Request):
         )
 
 
+# fastapi implicitly converts json in the request to obj (dataclass)
+@app.api_route("/generate", methods=["POST", "PUT"])
 @time_func_latency
 async def generate_request(obj: GenerateReqInput, request: Request):
     """Handle a generate request."""
@@ -266,11 +269,7 @@ async def generate_request(obj: GenerateReqInput, request: Request):
             )
 
 
-# fastapi implicitly converts json in the request to obj (dataclass)
-app.post("/generate")(generate_request)
-app.put("/generate")(generate_request)
-
-
+@app.api_route("/encode", methods=["POST", "PUT"])
 @time_func_latency
 async def encode_request(obj: EmbeddingReqInput, request: Request):
     """Handle an embedding request."""
@@ -283,10 +282,7 @@ async def encode_request(obj: EmbeddingReqInput, request: Request):
         )
 
 
-app.post("/encode")(encode_request)
-app.put("/encode")(encode_request)
-
-
+@app.api_route("/encode", methods=["POST", "PUT"])
 @time_func_latency
 async def classify_request(obj: EmbeddingReqInput, request: Request):
     """Handle a reward model request. Now the arguments and return values are the same as embedding models."""
@@ -299,10 +295,6 @@ async def classify_request(obj: EmbeddingReqInput, request: Request):
         )
 
 
-app.post("/classify")(classify_request)
-app.put("/classify")(classify_request)
-
-
 ##### OpenAI-compatible API endpoints #####
 
 
@@ -380,11 +372,11 @@ def launch_engine(
     server_args: ServerArgs,
 ):
     """
-    Launch the Tokenizer Manager in the main process, the Scheduler in a subprocess, and the Detokenizer Manager in another subprocess.
+    Launch the TokenizerManager in the main process, the Scheduler in a subprocess, and the DetokenizerManager in another subprocess.
     """
 
     global tokenizer_manager
-    global _max_total_num_tokens
+    global scheduler_info
 
     # Configure global environment
     configure_logger(server_args)
@@ -450,8 +442,8 @@ def launch_engine(
     if server_args.chat_template:
         load_chat_template_for_openai_api(tokenizer_manager, server_args.chat_template)
 
-    # Wait for model to finish loading & get max token nums
-    scheduler_info = []
+    # Wait for model to finish loading
+    scheduler_infos = []
     for i in range(len(scheduler_pipe_readers)):
         data = scheduler_pipe_readers[i].recv()
 
@@ -459,10 +451,10 @@ def launch_engine(
             raise RuntimeError(
                 "Initialization failed. Please see the error messages above."
             )
-        scheduler_info.append(data)
+        scheduler_infos.append(data)
 
     # Assume all schedulers have same max_total_num_tokens
-    _max_total_num_tokens = scheduler_info[0]["max_total_num_tokens"]
+    scheduler_info = scheduler_infos[0]
 
 
 def launch_server(
@@ -476,12 +468,12 @@ def launch_server(
 
     1. HTTP server: A FastAPI server that routes requests to the engine.
     2. SRT engine:
-        1. Tokenizer Manager: Tokenizes the requests and sends them to the scheduler.
+        1. TokenizerManager: Tokenizes the requests and sends them to the scheduler.
         2. Scheduler (subprocess): Receives requests from the Tokenizer Manager, schedules batches, forwards them, and sends the output tokens to the Detokenizer Manager.
-        3. Detokenizer Manager (subprocess): Detokenizes the output tokens and sends the result back to the Tokenizer Manager.
+        3. DetokenizerManager (subprocess): Detokenizes the output tokens and sends the result back to the Tokenizer Manager.
 
     Note:
-    1. The HTTP server and Tokenizer Manager both run in the main process.
+    1. The HTTP server and TokenizerManager both run in the main process.
     2. Inter-process communication is done through ICP (each process uses a different port) via the ZMQ library.
     """
     launch_engine(server_args=server_args)
@@ -490,7 +482,7 @@ def launch_server(
     if server_args.api_key:
         add_api_key_middleware(app, server_args.api_key)
 
-    # add prometheus middleware
+    # Add prometheus middleware
     if server_args.enable_metrics:
         add_prometheus_middleware(app)
         enable_func_timer()
@@ -502,7 +494,7 @@ def launch_server(
     t.start()
 
     try:
-        # Listen for HTTP requests
+        # Update logging configs
         LOGGING_CONFIG["formatters"]["default"][
             "fmt"
         ] = "[%(asctime)s] %(levelprefix)s %(message)s"
@@ -511,6 +503,8 @@ def launch_server(
             "fmt"
         ] = '[%(asctime)s] %(levelprefix)s %(client_addr)s - "%(request_line)s" %(status_code)s'
         LOGGING_CONFIG["formatters"]["access"]["datefmt"] = "%Y-%m-%d %H:%M:%S"
+
+        # Listen for HTTP requests
         uvicorn.run(
             app,
             host=server_args.host,
@@ -526,8 +520,7 @@ def launch_server(
 async def _get_server_info():
     return {
         **dataclasses.asdict(tokenizer_manager.server_args),  # server args
-        "memory_pool_size": await tokenizer_manager.get_memory_pool_size(),  # memory pool size
-        "max_total_num_tokens": _max_total_num_tokens,  # max total num tokens
+        **scheduler_info,
         "version": __version__,
     }
 
@@ -562,6 +555,15 @@ def _set_envs_and_config(server_args: ServerArgs):
             "at https://docs.flashinfer.ai/installation.html.",
         )
 
+    # Register the signal handler.
+    # The child processes will send SIGQUIT to this process when any error happens
+    # This process then clean up the whole process tree
+    def sigquit_handler(signum, frame):
+        kill_process_tree(os.getpid())
+
+    signal.signal(signal.SIGQUIT, sigquit_handler)
+
+    # Set mp start method
     mp.set_start_method("spawn", force=True)
 
 
@@ -588,7 +590,7 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
         if pipe_finish_writer is not None:
             pipe_finish_writer.send(last_traceback)
         logger.error(f"Initialization failed. warmup error: {last_traceback}")
-        kill_child_process(include_self=True)
+        kill_process_tree(os.getpid())
         return
 
     model_info = res.json()
@@ -621,9 +623,10 @@ def _wait_and_warmup(server_args, pipe_finish_writer):
         if pipe_finish_writer is not None:
             pipe_finish_writer.send(last_traceback)
         logger.error(f"Initialization failed. warmup error: {last_traceback}")
-        kill_child_process(include_self=True)
+        kill_process_tree(os.getpid())
         return
 
+    # Debug print
     # logger.info(f"{res.json()=}")
 
     logger.info("The server is fired up and ready to roll!")
@@ -690,7 +693,7 @@ class Runtime:
 
     def shutdown(self):
         if self.pid is not None:
-            kill_child_process(self.pid, include_self=True)
+            kill_process_tree(self.pid)
             self.pid = None
 
     def cache_prefix(self, prefix: str):
@@ -800,18 +803,11 @@ class Engine:
     launching the HTTP server adds unnecessary complexity or overhead,
     """
 
-    def __init__(self, *args, **kwargs):
-
+    def __init__(self, log_level: str = "error", *args, **kwargs):
         # before python program terminates, call shutdown implicitly. Therefore, users don't have to explicitly call .shutdown()
         atexit.register(self.shutdown)
 
-        # runtime server default log level is log
-        # offline engine works in scripts, so we set it to error
-
-        if "log_level" not in kwargs:
-            kwargs["log_level"] = "error"
-
-        server_args = ServerArgs(*args, **kwargs)
+        server_args = ServerArgs(*args, log_level=log_level, **kwargs)
         launch_engine(server_args=server_args)
 
     def generate(
@@ -914,7 +910,7 @@ class Engine:
             return ret
 
     def shutdown(self):
-        kill_child_process()
+        kill_process_tree(os.getpid(), include_parent=False)
 
     def get_tokenizer(self):
         global tokenizer_manager
@@ -934,5 +930,11 @@ class Engine:
         loop = asyncio.get_event_loop()
         return loop.run_until_complete(encode_request(obj, None))
 
+    def start_profile(self):
+        tokenizer_manager.start_profile()
+
+    def stop_profile(self):
+        tokenizer_manager.stop_profile()
+
     async def get_server_info(self):
         return await _get_server_info()

sglang/srt/server_args.py

@@ -144,7 +144,7 @@ class ServerArgs:
         if self.served_model_name is None:
             self.served_model_name = self.model_path
 
-        if self.chunked_prefill_size <= 0:
+        if self.chunked_prefill_size is not None and self.chunked_prefill_size <= 0:
             # Disable chunked prefill
             self.chunked_prefill_size = None
 

sglang/srt/utils.py

@@ -443,26 +443,14 @@ def assert_pkg_version(pkg: str, min_version: str, message: str):
         )
 
 
-def kill_parent_process():
-    """Kill the parent process and all children of the parent process."""
-    current_process = psutil.Process()
-    parent_process = current_process.parent()
-    kill_child_process(
-        parent_process.pid, include_self=True, skip_pid=current_process.pid
-    )
-    try:
-        current_process.kill()
-    except psutil.NoSuchProcess:
-        pass
-
-
-def kill_child_process(pid=None, include_self=False, skip_pid=None):
-    """Kill the process and all its children process."""
-    if pid is None:
-        pid = os.getpid()
+def kill_process_tree(parent_pid, include_parent: bool = True, skip_pid: int = None):
+    """Kill the process and all its child processes."""
+    if parent_pid is None:
+        parent_pid = os.getpid()
+        include_parent = False
 
     try:
-        itself = psutil.Process(pid)
+        itself = psutil.Process(parent_pid)
     except psutil.NoSuchProcess:
         return
 
@@ -475,13 +463,13 @@ def kill_child_process(pid=None, include_self=False, skip_pid=None):
         except psutil.NoSuchProcess:
             pass
 
-    if include_self:
+    if include_parent:
         try:
             itself.kill()
 
             # Sometime processes cannot be killed with SIGKILL (e.g, PID=1 launched by kubernetes),
             # so we send an additional signal to kill them.
-            itself.send_signal(signal.SIGINT)
+            itself.send_signal(signal.SIGQUIT)
         except psutil.NoSuchProcess:
             pass
 

sglang/test/test_utils.py

@@ -22,7 +22,7 @@ from sglang.bench_serving import run_benchmark
 from sglang.global_config import global_config
 from sglang.lang.backend.openai import OpenAI
 from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
-from sglang.srt.utils import get_bool_env_var, kill_child_process
+from sglang.srt.utils import get_bool_env_var, kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.utils import get_exception_traceback
 
@@ -504,7 +504,7 @@ def run_unittest_files(files: List[str], timeout_per_file: float):
             )
             assert ret_code == 0
         except TimeoutError:
-            kill_child_process(process.pid, include_self=True)
+            kill_process_tree(process.pid)
             time.sleep(5)
             print(
                 f"\nTimeout after {timeout_per_file} seconds when running {filename}\n",
@@ -578,7 +578,7 @@ def run_bench_serving(
             run_benchmark(warmup_args)
         res = run_benchmark(args)
     finally:
-        kill_child_process(process.pid, include_self=True)
+        kill_process_tree(process.pid)
 
     assert res["completed"] == num_prompts
     return res
@@ -611,7 +611,7 @@ def run_bench_one_batch(model, other_args):
         lastline = output.split("\n")[-3]
         output_throughput = float(lastline.split(" ")[-2])
     finally:
-        kill_child_process(process.pid, include_self=True)
+        kill_process_tree(process.pid)
 
     return output_throughput
 
@@ -677,8 +677,14 @@ def run_and_check_memory_leak(
     enable_mixed_chunk,
     disable_overlap,
     chunked_prefill_size,
+    assert_has_abort,
 ):
-    other_args = ["--chunked-prefill-size", str(chunked_prefill_size)]
+    other_args = [
+        "--chunked-prefill-size",
+        str(chunked_prefill_size),
+        "--log-level",
+        "debug",
+    ]
     if disable_radix_cache:
         other_args += ["--disable-radix-cache"]
     if enable_mixed_chunk:
@@ -710,8 +716,8 @@ def run_and_check_memory_leak(
     workload_func(base_url, model)
 
     # Clean up everything
-    kill_child_process(process.pid, include_self=True)
-    kill_child_process(process.pid, include_self=True)
+    kill_process_tree(process.pid)
+    kill_process_tree(process.pid)
     stdout.close()
     stderr.close()
     if os.path.exists(STDOUT_FILENAME):
@@ -723,14 +729,19 @@ def run_and_check_memory_leak(
     # Assert success
     has_new_server = False
     has_leak = False
+    has_abort = False
     for line in output_lines:
         if "The server is fired" in line:
             has_new_server = True
         if "leak" in line:
             has_leak = True
+        if "Abort" in line:
+            has_abort = True
 
     assert has_new_server
     assert not has_leak
+    if assert_has_abort:
+        assert has_abort
 
 
 def run_mmlu_test(
@@ -761,6 +772,7 @@ def run_mmlu_test(
         enable_mixed_chunk,
         disable_overlap,
         chunked_prefill_size,
+        assert_has_abort=False,
     )
 
 
@@ -800,4 +812,5 @@ def run_mulit_request_test(
         enable_mixed_chunk,
         enable_overlap,
         chunked_prefill_size,
+        assert_has_abort=False,
     )

sglang/utils.py

@@ -348,9 +348,9 @@ def wait_for_server(base_url: str, timeout: int = None) -> None:
 
 
 def terminate_process(process):
-    from sglang.srt.utils import kill_child_process
+    from sglang.srt.utils import kill_process_tree
 
-    kill_child_process(process.pid, include_self=True)
+    kill_process_tree(process.pid)
 
 
 def print_highlight(html_content: str):

sglang/version.py

@@ -1 +1 @@
-__version__ = "0.3.6.post2"
+__version__ = "0.3.6.post3"

{sglang-0.3.6.post2.dist-info → sglang-0.3.6.post3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.3.6.post2
+Version: 0.3.6.post3
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -241,6 +241,7 @@ Requires-Dist: sglang[runtime_common]; extra == "srt"
 Requires-Dist: torch; extra == "srt"
 Requires-Dist: vllm>=0.6.3.post1; extra == "srt"
 Requires-Dist: cuda-python; extra == "srt"
+Requires-Dist: flashinfer>=0.1.6; extra == "srt"
 Provides-Extra: srt-hip
 Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
 Requires-Dist: torch; extra == "srt-hip"

{sglang-0.3.6.post2.dist-info → sglang-0.3.6.post3.dist-info}/RECORD

@@ -1,16 +1,16 @@
 sglang/__init__.py,sha256=3M0oz0ZA8fULhV5LwQ4hxh-MRdHsOJRD1D63C60pdG4,1616
 sglang/api.py,sha256=NdO6cYnklnEBQBKqQjlqI8-P1EownKQ71t5ibCGhEVo,6953
 sglang/bench_latency.py,sha256=oZjSAzX7dUiSu-zdz0dkyUPo-qAX_lsXFH1gf03akgI,76
-sglang/bench_offline_throughput.py,sha256=z6uA6Gxa_nFZa0cOXi7MJDuX82xcqk5WfqBMavd8a-s,10929
-sglang/bench_one_batch.py,sha256=AVMpCBWEsMI2TlMK55JPgPJu0kHg8DI0WV_Bhd4pJgc,15668
-sglang/bench_one_batch_server.py,sha256=hYc3r9JQOLrfqmKgKPOmP0Kr63Sya9wPV_dHzMRZ2Dw,5924
+sglang/bench_offline_throughput.py,sha256=3OrFI26PmoVTU3pQrBFC50AZI7HpKKuk4vYycbkDjhY,12428
+sglang/bench_one_batch.py,sha256=iSev0LruPdfJ49mVeCJNFREmgex2omDSpTgwHgRDNIo,15692
+sglang/bench_one_batch_server.py,sha256=-fV9FTLNNcSIy0pgYeggXedPVK0fVsXZqVQswT8OMOY,5945
 sglang/bench_serving.py,sha256=hI7FjaERyqKBrYtKewDU6E4rSufKxqsUPyUgtWtTKSI,52545
 sglang/check_env.py,sha256=rE4ZAG0e6M-Xd-qdHcKclN8Qav6b9gEh4yvlV_TbOg0,5450
 sglang/global_config.py,sha256=fnT0U9vlHdGaQFKN9tYTnUF4-eVW4HYQURd5zvPtrg0,1286
-sglang/launch_server.py,sha256=U17c44CbbpMBm2JQxVLaz1mfUKk7PgBDhTLAFNeJEvI,362
+sglang/launch_server.py,sha256=4y2QeSj0wVNB9MJQZeahD4ahTDU6gwqo7MPUytyFop0,403
 sglang/launch_server_llavavid.py,sha256=tGc17S1vUfLwbi1GB26oOdXxTWr7gjlqpTrPnrMRNO8,1007
-sglang/utils.py,sha256=eCvD3fZCALr-MuyZxJL7HAeeqqpxAxf4LJrf7OiCbco,11547
-sglang/version.py,sha256=_Aams_yVBpGe9-85k-kF3qpgcd3D_AsWkVfMFmCWh3c,28
+sglang/utils.py,sha256=r4Dw-xffcrTRposls-gqyoYxjgJNYhVduK_6bDN_Vj4,11526
+sglang/version.py,sha256=pyFIjLZBxCp2AwTbeLFaYhEL1dFXGzpYW00Vkg2755Y,28
 sglang/lang/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sglang/lang/chat_template.py,sha256=jprS3-In2FTUoedKwZg-HYvDwU8RTIYntOlf2zoN2sU,14814
 sglang/lang/choices.py,sha256=-W1DVw9N9ZliVpvmWrzIXG4cswAah8eMQrHWzkS3D8o,6234
@@ -29,9 +29,9 @@ sglang/srt/conversation.py,sha256=u9zFU8aMYzwHUbQRKU76B_T-jfLlPoxUcWG_nRbDM2I,21
 sglang/srt/hf_transformers_utils.py,sha256=sUUCpjbTHuYDMuwOaz00nH5fataXKjliD8gCxXU64sw,6712
 sglang/srt/mm_utils.py,sha256=1ScBunw_x4W8ebM_AcJ62-1T2mfT8NlMJqdAhkF1lb0,12367
 sglang/srt/model_parallel.py,sha256=QR-Alqo0sElDXPJ79N1PhUHHKiEHPQn3dyXduMP-SHQ,3664
-sglang/srt/server.py,sha256=tH_22tnksy3bbhYu_njjx5L59pb9lJ7tU40Z2BLoiaI,30894
-sglang/srt/server_args.py,sha256=CfmpU6_EDnxJzpJiRx2n6AhOPCtrHPOf-7wEtTF__L0,30834
-sglang/srt/utils.py,sha256=QXc01TOB7abpL6p3KzfP7u2xFZohQ-ThbI5DAJGoHeI,33894
+sglang/srt/server.py,sha256=1A_RdzTgeVPKcoZvsLs0dH9U3ZOY2MWjS6X3EUmwzPs,31011
+sglang/srt/server_args.py,sha256=PbkhdNkr46Ngv3_JPplo5jLw78pebRxNVTiIb-9uPVA,30876
+sglang/srt/utils.py,sha256=TWeASu4TOqIbvb-rJ0CYvFcMyk67hPJxQZnvyqrKu8k,33585
 sglang/srt/configs/__init__.py,sha256=_usVIXHQjft4PAJ1Y-yGQOn2QNOv501GYMlQwpGXbns,208
 sglang/srt/configs/exaone.py,sha256=Duxd4yQoKy8GWEzZD_kCY_OzmN_67CTJL_Kgn0eXk3g,10731
 sglang/srt/configs/model_config.py,sha256=r5N_OO4w3_R3kZ80P-ZPECscXmspI41d1vc6uEE9ixM,9526
@@ -61,9 +61,6 @@ sglang/srt/layers/attention/triton_ops/decode_attention.py,sha256=BE63WhKiutSNkh
 sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py,sha256=1pSXfY3EEaM7iRN_uElHnAfsrJMhTFbu9fj8Z0O2PbE,21480
 sglang/srt/layers/attention/triton_ops/extend_attention.py,sha256=Gfct-0_l-S2ZrP4F-zkzNiFbmd3C3f7uJovacOuDxaA,11472
 sglang/srt/layers/attention/triton_ops/prefill_attention.py,sha256=lojFXRZMLWkzS2Y8uxaolnQhXaWKG19mCAWaF5KQeiI,6087
-sglang/srt/layers/fused_moe_grok/__init__.py,sha256=rj_JBzcP--eaaM6LGQ-u580uQvqLisp5JtGBAs1fVYc,80
-sglang/srt/layers/fused_moe_grok/fused_moe.py,sha256=bxRcjdALxeY3FDnKivGOoNr6Er1kh6CCPtlAp7pjz50,23844
-sglang/srt/layers/fused_moe_grok/layer.py,sha256=v-o5YHYEU2HIEZwouyuc3UyfNj7YQrEYOO_BXKELU7Y,23453
 sglang/srt/layers/fused_moe_triton/__init__.py,sha256=PHKFqd2hPOO-g9kSMseg2g76lpg9OGXQDThWU6bt9vs,902
 sglang/srt/layers/fused_moe_triton/fused_moe.py,sha256=qwfRBOeY5DT48Q6z71Eh9cjFehvs_K6eLIVWNL044Ug,28363
 sglang/srt/layers/fused_moe_triton/layer.py,sha256=URDkTt8xEqnqpO5tb_3L7JlhlO53VWfqDDNSRYEu-LY,21545
@@ -72,17 +69,17 @@ sglang/srt/layers/quantization/base_config.py,sha256=daK9p0aijMszLUm1W4Pc33FK87M
 sglang/srt/lora/lora.py,sha256=KhhO9aKCyFWvJnhI07lZKANIvNjtt882HrTYFNBZMv0,15065
 sglang/srt/lora/lora_config.py,sha256=a2fTQESlCbG1xLiBYy4ptZ6c0Burcqyg1_6V1XSok-Y,1506
 sglang/srt/lora/lora_manager.py,sha256=DHiqdl0_4wQ5PxZBZtlCpP14515mDV2_H9tzL3Rdss8,12886
-sglang/srt/managers/data_parallel_controller.py,sha256=JxRtJJTVn1FU2iD292rLZPftAsR4_8j4d3yF8j0dvBc,8327
-sglang/srt/managers/detokenizer_manager.py,sha256=oWquBe0yvSwILwllMBJFJUEgBt1NEM_3KluAc0T6Pnw,7333
-sglang/srt/managers/image_processor.py,sha256=foLv3QVW_A8IRjRcHOKn0_HC771JbPEz8ML1mGqYKYw,13685
-sglang/srt/managers/io_struct.py,sha256=WLXz-tyn0jR7zNO9feRBXgyjphVa8qR55OoEOUdzoVI,13751
-sglang/srt/managers/schedule_batch.py,sha256=jBABHbL7gyrKdrFrzScJ76MtvG2D9Y5HDx74qsclo80,44470
+sglang/srt/managers/data_parallel_controller.py,sha256=psI4FAuBGjtdnEuwagnGdtRqvqSSxOROfNKQqVDqlVA,8382
+sglang/srt/managers/detokenizer_manager.py,sha256=TtrtE37XT5XcJzk8-R5rHZ16NHTPd5XZi8hf3h-sB2A,7462
+sglang/srt/managers/image_processor.py,sha256=Y8RgyrzbJjJTpjbnZDa5qiiG5wWjZ68rOXUPDi6kkFo,13698
+sglang/srt/managers/io_struct.py,sha256=bvhHIRSBpaCXFQqRBTpxy-hjvNtDxSfoDJ5XWCHoy6g,13646
+sglang/srt/managers/schedule_batch.py,sha256=dTeM0U1xvUq_GlHy8SQft6-pP76cSubPKzCClsQ9MgM,44801
 sglang/srt/managers/schedule_policy.py,sha256=ayFz4iPLIlG8mx5i1glTCAMHJPGpFedMP9UgRtqkNhA,12526
-sglang/srt/managers/scheduler.py,sha256=JVxV3Y5AU0OOOfePVM5dVPuuN_Kd9nwV3p3vH3CHQps,56059
-sglang/srt/managers/session_controller.py,sha256=hajOnkNZ_JpP4E-GKMVGzyJSK4sc9uF9t229uFuxkVs,2874
-sglang/srt/managers/tokenizer_manager.py,sha256=zYbKEKNuM1B3PXzA7jnDpxew-0rZXSX-7dHmVLWG3e4,26477
+sglang/srt/managers/scheduler.py,sha256=aEU-6_0w-HbpFTMSoiDqf3mj_UfedjugCUvnQBmhgoU,56571
+sglang/srt/managers/session_controller.py,sha256=Yp-IV3rXczACZxZXmF-QxW9CWICGy8KHQ9ttBGJ8WXA,2800
+sglang/srt/managers/tokenizer_manager.py,sha256=jLzoEIhQWzZX7rcLZ290vfnPY2ghxWdYhf7YJQtUC3s,25339
 sglang/srt/managers/tp_worker.py,sha256=1SQJ60iKS9e5vGY555fT1iZ4OtLumXzeWfB08fSWKbk,6176
-sglang/srt/managers/tp_worker_overlap_thread.py,sha256=7vhPebaOS4JamaS08CGf_hwxnUO7Gy_SXZXEPwNHKoY,7621
+sglang/srt/managers/tp_worker_overlap_thread.py,sha256=pLQOHj-nFrqHyVFP-JvrU--tjh1X1yET_NJIFHp0H0I,7990
 sglang/srt/mem_cache/base_prefix_cache.py,sha256=qEQwEkG4E5rab2ZoTqcesf5pR_J4nV2jBxIHsBJHtIM,924
 sglang/srt/mem_cache/chunk_cache.py,sha256=VcCpyrf5FOQ5xoKeOouCI5ZQLkZo_pgY1SPbDDkagGg,2492
 sglang/srt/mem_cache/flush_cache.py,sha256=GYcxmNXh4hsMpFfNOuCTpKilW7guZwTtAg_usVeM3J0,979
@@ -105,14 +102,14 @@ sglang/srt/models/gemma2.py,sha256=lbfQhQpUhf1MAEB_00Uo6rp20k4Hr353UbPKKuMsxec,1
 sglang/srt/models/gemma2_reward.py,sha256=cQawatbsfBuWQTueivYHl_17ZoQUHEelI1sr1y5pvfY,2556
 sglang/srt/models/gpt2.py,sha256=Th7_Dnkw82GFBOuMOTrHtA44JBPHRUtY3Qd73rQwzMc,9741
 sglang/srt/models/gpt_bigcode.py,sha256=lYo4ajy49VvvPkaduaFtOaCRT_ItqyNUE158S-BI5QA,10136
-sglang/srt/models/grok.py,sha256=rDIH_SFzauuEHcL_vCOSrYLjdBC3i3o_AcceL3amsJw,14927
+sglang/srt/models/grok.py,sha256=d6qvj_i_Pam4dV_WInUvw2cIH3s6hCj-skbgyvLld1E,13923
 sglang/srt/models/internlm2.py,sha256=DxbA15d9QR0tLOczpC6DkB8QyNHXJRdZatY6Nskwv1k,12170
 sglang/srt/models/internlm2_reward.py,sha256=Lr-JA0vfTQJt9q5oDMiopGuoXAevyEv5PAoDe2rsTJk,2425
 sglang/srt/models/llama.py,sha256=FSGuM3BamhuT5h2jedh5cSFwFYduOJwkAZJJ672awRw,16423
 sglang/srt/models/llama_classification.py,sha256=c8WZ1ADa3f6s2IJVoP10ouVgeCwv_ndns_qMgLrC6QI,3413
 sglang/srt/models/llama_embedding.py,sha256=2ex2jrz31osaAd9V8sJeN0qyxmk-L5NgOBkXL1puGhI,3166
 sglang/srt/models/llama_reward.py,sha256=prhHDPpf1k6tlQtGE6zq5gx0uSZAD3W5v7W28bdgy4U,4619
-sglang/srt/models/llava.py,sha256=HjC2TDLngpaN8HMYyGp5doEK32HeQN8iT2tYE_Slrtg,25130
+sglang/srt/models/llava.py,sha256=G6EcAJ84FvV4sae3Rrmdp-Bm-cczynSWSR16Ig-QiSw,25319
 sglang/srt/models/llavavid.py,sha256=DeWqGSmXgIYGuLyy2ZrxjM9WqbRjueP4chNmXt7Bnus,12221
 sglang/srt/models/minicpm.py,sha256=KbiTf-kaDAJxSo9Z4IGMTrs9WrYYji1KXO1kA2iy-as,13816
 sglang/srt/models/minicpm3.py,sha256=C43mTr2Qjccj4sXuTDgzbfZhvCNbsEHNggMRXQ7SrWs,25108
@@ -121,18 +118,19 @@ sglang/srt/models/mixtral.py,sha256=E3d8I7V3Dp1nCEHRbhh-PKBG8UaVK5XOHwl9QyIjcX0,
 sglang/srt/models/mixtral_quant.py,sha256=o-oTG8BGtWuNu-o6muHSarMNBQwrjQowyBFOQhuclZ8,14065
 sglang/srt/models/mllama.py,sha256=pET1x8wY04yoS8HMCncKx0tFPqGp78K8rlA7Eq7XioE,37889
 sglang/srt/models/olmo.py,sha256=DEUPNDM0z83N-Qdhkj2WJMtbiz5JNbSBMIjUaYZN9RM,12068
+sglang/srt/models/olmo2.py,sha256=NriLbVKNGSR9bs0V8feeEorkRSr9BjlYbv50AReo2s4,13469
 sglang/srt/models/olmoe.py,sha256=jVKrjqQQrWLdlkGSGUaMPdT9PHzNH4X-RVwON29eaGw,15412
 sglang/srt/models/phi3_small.py,sha256=fxqGU0xphJzTeuBW38SRRYpRb2rcsg53JxuObK0pZig,15141
 sglang/srt/models/qwen.py,sha256=P9zcFnz_Tsz73tVtLRwZ8uWzCtMxWOrzlv2o9Ys_Gck,9947
 sglang/srt/models/qwen2.py,sha256=ApFFASNwvrkDXi-KkCNA7fTk4uLMuJWoMg15zCaAKdA,12514
 sglang/srt/models/qwen2_moe.py,sha256=1oxDsKDq3jlHKx9jMi1SfHOqCRVyN5n76uw3M-CUODE,17048
-sglang/srt/models/qwen2_vl.py,sha256=G3FNa_N2-CzB56LVrukwBtJazxMrDC_GPNjK6Wqxc4s,26415
+sglang/srt/models/qwen2_vl.py,sha256=wFKBq52nZ5Q1sloDNh9YcYIoJ-4QpGVA15StxRMBuYE,26785
 sglang/srt/models/stablelm.py,sha256=jpmsyWMJo_9JapOESnuV7ObNCh78BRznXY0iFvvIbZE,11354
 sglang/srt/models/torch_native_llama.py,sha256=vNQxsnbVAY1bdyMCCWDZAtWdbaFIiJXhmVxHjk5BB9Y,19400
 sglang/srt/models/xverse.py,sha256=LGe0ma0wOir3x-OLBT_cRocw8JEo9d3AYNxgA2OcLrk,13659
 sglang/srt/models/xverse_moe.py,sha256=YqbzkSsnTFt-8-aI8YobF9qJA70qrBjbS1Kjn1KNqVY,15766
 sglang/srt/models/yivl.py,sha256=yj4aWsOBVGQBLurSrLmYXVC7zGIPH7EYHHtAaAZ7Liw,4859
-sglang/srt/openai_api/adapter.py,sha256=MhOcWZjcLv4_OuvLvDMcAu6K_u2joJvhaZxaKm0hi3M,53634
+sglang/srt/openai_api/adapter.py,sha256=Rit_XJ4h-O1-_fwivIkcGHt1hLLz3Y3XdTtEtDTnBcU,53634
 sglang/srt/openai_api/protocol.py,sha256=vBgrbTqtECsZ5dG0rgP1FHsTBt4eR9zbDX3FBIN-rz4,10172
 sglang/srt/sampling/sampling_batch_info.py,sha256=YC-KPyDWyLGNPL4YVcst4xwP8Wlz2zcCNJHB_5zljXQ,8470
 sglang/srt/sampling/sampling_params.py,sha256=n7RbBg_bS5fYhsiWa8uJYnfoXy_i5DvtTBOkuFnHDNU,5286
@@ -155,10 +153,10 @@ sglang/test/simple_eval_mmlu.py,sha256=FkwamjGMjueTixymkedF-YiPloSLiy4ftILFUrKZ9
 sglang/test/test_activation.py,sha256=jkdNRzJnbd5OgZliQaIXpxovlcky17UrweomcOcMxoE,1442
 sglang/test/test_layernorm.py,sha256=IacByD5d-stXjzBz8Ypamc7povlcedpKPbb_4JLgo3c,3720
 sglang/test/test_programs.py,sha256=1Z0umrsUu9pagzyGH5SrXl_qhKSyTfUv_kWC2mcn0qo,18208
-sglang/test/test_utils.py,sha256=NBEGQC_wtMqODQQZWrxdwmsoLFSZfDlQzIbsQ1kE_Yc,23468
+sglang/test/test_utils.py,sha256=mPRTn1ORMiJODa_wWpH8QQG-IuZuQYZp9nGjyIcIHHU,23645
 sglang/test/srt/sampling/penaltylib/utils.py,sha256=CjxHgywh0hx_87iynzQt_ztHu6zBVuE-YrZ-XPmW6U4,12906
-sglang-0.3.6.post2.dist-info/LICENSE,sha256=FJXh51fvTQklojUFY89XVLsjxRcBqOxPs8XNy-2uZ0c,11346
-sglang-0.3.6.post2.dist-info/METADATA,sha256=3ekB4UX6bNwXzqlRChfxG0R8sme-x0FQAImcw0gpfM8,22122
-sglang-0.3.6.post2.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-sglang-0.3.6.post2.dist-info/top_level.txt,sha256=yxhh3pYQkcnA7v3Bg889C2jZhvtJdEincysO7PEB09M,7
-sglang-0.3.6.post2.dist-info/RECORD,,
+sglang-0.3.6.post3.dist-info/LICENSE,sha256=FJXh51fvTQklojUFY89XVLsjxRcBqOxPs8XNy-2uZ0c,11346
+sglang-0.3.6.post3.dist-info/METADATA,sha256=2EeNTAznmmP399LIX3w0evy2Q2x6IqOSMoRKXJZLrSM,22171
+sglang-0.3.6.post3.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
+sglang-0.3.6.post3.dist-info/top_level.txt,sha256=yxhh3pYQkcnA7v3Bg889C2jZhvtJdEincysO7PEB09M,7
+sglang-0.3.6.post3.dist-info/RECORD,,

sglang/srt/layers/fused_moe_grok/__init__.py

@@ -1 +0,0 @@
-from sglang.srt.layers.fused_moe_grok.layer import FusedMoE, FusedMoEMethodBase