sglang 0.3.6.post1__py3-none-any.whl → 0.3.6.post2__py3-none-any.whl

sglang/bench_one_batch.py

@@ -466,7 +466,6 @@ if __name__ == "__main__":
 
     try:
         main(server_args, bench_args)
-    except Exception as e:
-        raise e
     finally:
-        kill_child_process()
+        if server_args.tp_size != 1:
+            kill_child_process()

sglang/bench_one_batch_server.py

@@ -5,9 +5,9 @@ This script launches a server and uses the HTTP interface.
 It accepts server arguments (the same as launch_server.py) and benchmark arguments (e.g., batch size, input lengths).
 
 Usage:
-python3 -m sglang.bench_server_latency --model meta-llama/Meta-Llama-3.1-8B --batch-size 1 16 64 --input-len 1024 --output-len 8
+python3 -m sglang.bench_one_batch_server --model meta-llama/Meta-Llama-3.1-8B --batch-size 1 16 64 --input-len 1024 --output-len 8
 
-python3 -m sglang.bench_server_latency --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8
+python3 -m sglang.bench_one_batch_server --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8
 """
 
 import argparse

sglang/check_env.py

@@ -22,18 +22,24 @@ PACKAGE_LIST = [
     "hf_transfer",
     "huggingface_hub",
     "interegular",
+    "modelscope",
+    "orjson",
+    "outlines",
+    "packaging",
     "psutil",
     "pydantic",
     "multipart",
     "zmq",
+    "torchao",
     "uvicorn",
     "uvloop",
     "vllm",
-    "outlines",
+    "xgrammar",
     "openai",
     "tiktoken",
     "anthropic",
     "litellm",
+    "decord",
 ]
 
 

sglang/lang/tracer.py

@@ -278,6 +278,6 @@ class TracingScope:
 
     def add_child_state(self, state: TracerProgramState):
         cur_scope = self
-        while cur_scope != None:
+        while cur_scope is not None:
             cur_scope.tracer_state.child_states.append(state)
             cur_scope = cur_scope.last_scope

sglang/launch_server.py

@@ -1,6 +1,5 @@
 """Launch the inference server."""
 
-import os
 import sys
 
 from sglang.srt.server import launch_server
@@ -12,7 +11,5 @@ if __name__ == "__main__":
 
     try:
         launch_server(server_args)
-    except Exception as e:
-        raise e
     finally:
         kill_child_process()

sglang/srt/configs/model_config.py

@@ -14,13 +14,13 @@
 
 import json
 import logging
-import os
 from enum import IntEnum, auto
 from typing import List, Optional
 
 from transformers import PretrainedConfig
 
 from sglang.srt.hf_transformers_utils import get_config, get_context_length
+from sglang.srt.utils import get_bool_env_var
 
 logger = logging.getLogger(__name__)
 
@@ -59,13 +59,9 @@ class ModelConfig:
 
         # Derive context length
         derived_context_len = get_context_length(self.hf_text_config)
-        allow_long_context = os.environ.get(
-            "SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN", None
-        )
-
         if context_length is not None:
             if context_length > derived_context_len:
-                if allow_long_context:
+                if get_bool_env_var("SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN"):
                     logger.warning(
                         f"Warning: User-specified context_length ({context_length}) is greater than the derived context_length ({derived_context_len}). "
                         f"This may lead to incorrect model outputs or CUDA errors."

sglang/srt/layers/attention/flashinfer_backend.py

@@ -18,7 +18,7 @@ import triton.language as tl
 from sglang.global_config import global_config
 from sglang.srt.layers.attention import AttentionBackend
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.utils import is_flashinfer_available
+from sglang.srt.utils import get_bool_env_var, is_flashinfer_available
 
 if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention
@@ -47,8 +47,8 @@ class FlashInferAttnBackend(AttentionBackend):
 
         # Parse constants
         if "SGLANG_FLASHINFER_USE_TENSOR_CORE" in os.environ:
-            self.decode_use_tensor_cores = (
-                os.environ["SGLANG_FLASHINFER_USE_TENSOR_CORE"].lower() == "true"
+            self.decode_use_tensor_cores = get_bool_env_var(
+                "SGLANG_FLASHINFER_USE_TENSOR_CORE"
             )
         else:
             if not _grouped_size_compiled_for_decode_kernels(

sglang/srt/layers/sampler.py

@@ -74,7 +74,7 @@ class Sampler(nn.Module):
                         filter_apply_order="joint",
                     )
 
-                if not torch.all(success):
+                if self.use_nan_detectioin and not torch.all(success):
                     logger.warning("Detected errors during sampling!")
                     batch_next_token_ids = torch.zeros_like(batch_next_token_ids)
             elif global_server_args_dict["sampling_backend"] == "pytorch":

sglang/srt/managers/detokenizer_manager.py

@@ -25,8 +25,6 @@ from sglang.srt.managers.io_struct import (
     BatchEmbeddingOut,
     BatchStrOut,
     BatchTokenIDOut,
-    GetMemPoolSizeReqOutput,
-    UpdateWeightReqOutput,
 )
 from sglang.srt.managers.schedule_batch import FINISH_MATCHED_STR, FINISH_MATCHED_TOKEN
 from sglang.srt.server_args import PortArgs, ServerArgs

sglang/srt/managers/image_processor.py

@@ -131,6 +131,7 @@ class LlavaImageProcessor(BaseImageProcessor):
         if not image_data:
             return None
 
+        modalities = request_obj.modalities or ["image"]
         aspect_ratio = getattr(self.hf_config, "image_aspect_ratio", None)
         grid_pinpoints = (
             self.hf_config.image_grid_pinpoints
@@ -139,9 +140,12 @@ class LlavaImageProcessor(BaseImageProcessor):
             else None
         )
 
+        if isinstance(image_data, str):
+            image_data = [image_data]
+
         if isinstance(image_data, list) and len(image_data) > 0:
-            # Multiple images
-            if len(image_data) > 1:
+            if "multi-images" in modalities or "video" in modalities:
+                # Multiple images
                 aspect_ratio = "pad"  # LLaVA OneVision Handling: more than one image --> interleaved image mode or video mode. We do not use anyres
                 pixel_values, image_hashes, image_sizes = [], [], []
                 res = []
@@ -166,13 +170,6 @@ class LlavaImageProcessor(BaseImageProcessor):
                 )
                 image_hashes = [image_hash]
                 image_sizes = [image_size]
-        elif isinstance(image_data, str):
-            # A single image
-            pixel_values, image_hash, image_size = await self._process_single_image(
-                image_data, aspect_ratio, grid_pinpoints
-            )
-            image_hashes = [image_hash]
-            image_sizes = [image_size]
         else:
             raise ValueError(f"Invalid image data: {image_data}")
 

sglang/srt/managers/schedule_batch.py

@@ -31,6 +31,7 @@ import dataclasses
 import logging
 from typing import List, Optional, Tuple, Union
 
+import numpy as np
 import torch
 import triton
 import triton.language as tl
@@ -167,6 +168,30 @@ class ImageInputs:
 
         return ret
 
+    def merge(self, other, vocab_size):
+        assert self.pixel_values.shape[1:] == other.pixel_values.shape[1:]
+        self.pixel_values = np.concatenate([self.pixel_values, other.pixel_values])
+        self.image_hashes += other.image_hashes
+
+        self.pad_values = [
+            (self.image_hashes) % vocab_size,
+            (self.image_hashes >> 16) % vocab_size,
+            (self.image_hashes >> 32) % vocab_size,
+            (self.image_hashes >> 64) % vocab_size,
+        ]
+
+        optional_args = [
+            "image_sizes",
+            "image_offsets",
+            # "modalities", # modalities should be ["multi-images"] (one entry) even for multiple images
+            "aspect_ratio_ids",
+            "aspect_ratio_mask",
+            "image_grid_thws",
+        ]
+        for arg in optional_args:
+            if getattr(self, arg, None) is not None:
+                setattr(self, arg, getattr(self, arg) + getattr(other, arg))
+
 
 class Req:
     """The input and output status of a request."""
@@ -177,6 +202,7 @@ class Req:
         origin_input_text: str,
         origin_input_ids: Tuple[int],
         sampling_params: SamplingParams,
+        origin_input_ids_unpadded: Optional[Tuple[int]] = None,
         lora_path: Optional[str] = None,
         input_embeds: Optional[List[List[float]]] = None,
         session_id: Optional[str] = None,
@@ -184,7 +210,11 @@ class Req:
         # Input and output info
         self.rid = rid
         self.origin_input_text = origin_input_text
-        self.origin_input_ids_unpadded = origin_input_ids  # Before image padding
+        self.origin_input_ids_unpadded = (
+            origin_input_ids_unpadded
+            if origin_input_ids_unpadded
+            else origin_input_ids  # Before image padding
+        )
         self.origin_input_ids = origin_input_ids
         self.output_ids = []  # Each decode stage's output ids
         self.fill_ids = None  # fill_ids = origin_input_ids + output_ids
@@ -260,6 +290,12 @@ class Req:
         # The number of cached tokens, that were already cached in the KV cache
         self.cached_tokens = 0
 
+    def extend_image_inputs(self, image_inputs, vocab_size):
+        if self.image_inputs is None:
+            self.image_inputs = image_inputs
+        else:
+            self.image_inputs.merge(image_inputs, vocab_size)
+
     # whether request reached finished condition
     def finished(self) -> bool:
         return self.finished_reason is not None

sglang/srt/managers/scheduler.py

@@ -71,9 +71,10 @@ from sglang.srt.utils import (
     broadcast_pyobj,
     configure_logger,
     crash_on_warnings,
+    get_bool_env_var,
     get_zmq_socket,
-    gpu_proc_affinity,
     kill_parent_process,
+    set_gpu_proc_affinity,
     set_random_seed,
     suppress_other_loggers,
 )
@@ -82,7 +83,7 @@ from sglang.utils import get_exception_traceback
 logger = logging.getLogger(__name__)
 
 # Test retract decode
-test_retract = os.getenv("SGLANG_TEST_RETRACT", "false").lower() == "true"
+test_retract = get_bool_env_var("SGLANG_TEST_RETRACT")
 
 
 class Scheduler:
@@ -559,12 +560,13 @@ class Scheduler:
 
         # Image inputs
         if recv_req.image_inputs is not None:
-            req.image_inputs = ImageInputs.from_dict(
+            image_inputs = ImageInputs.from_dict(
                 recv_req.image_inputs, self.model_config.vocab_size
             )
             req.origin_input_ids = self.pad_input_ids_func(
-                req.origin_input_ids_unpadded, req.image_inputs
+                req.origin_input_ids, image_inputs
             )
+            req.extend_image_inputs(image_inputs, self.model_config.vocab_size)
 
             if len(req.origin_input_ids) > self.max_req_input_len:
                 req.finished_reason = FINISH_ABORT(
@@ -1404,7 +1406,8 @@ def run_scheduler_process(
     pipe_writer,
 ):
     # set cpu affinity to this gpu process
-    gpu_proc_affinity(server_args.tp_size, server_args.nnodes, gpu_id)
+    if get_bool_env_var("SGLANG_SET_CPU_AFFINITY"):
+        set_gpu_proc_affinity(server_args.tp_size, server_args.nnodes, gpu_id)
 
     # [For Router] if env var "DP_RANK" exist, set dp_rank to the value of the env var
     if dp_rank is None and "DP_RANK" in os.environ:

sglang/srt/managers/session_controller.py

@@ -41,16 +41,27 @@ class Session:
                 ]
                 + req.input_ids
             )
+            input_ids_unpadded = (
+                self.reqs[-1].origin_input_ids_unpadded
+                + self.reqs[-1].output_ids[
+                    : self.reqs[-1].sampling_params.max_new_tokens
+                ]
+                + req.input_ids
+            )
         else:
             input_ids = req.input_ids
+            input_ids_unpadded = req.input_ids
         new_req = Req(
-            req.rid,
-            None,
-            input_ids,
-            req.sampling_params,
+            rid=req.rid,
+            origin_input_text=None,
+            origin_input_ids=input_ids,
+            origin_input_ids_unpadded=input_ids_unpadded,
+            sampling_params=req.sampling_params,
             lora_path=req.lora_path,
             session_id=self.session_id,
         )
+        if len(self.reqs) > 0:
+            new_req.image_inputs = self.reqs[-1].image_inputs
         new_req.tokenizer = tokenizer
         if req.session_rid is not None and len(self.reqs) == 0:
             new_req.finished_reason = FINISH_ABORT(

sglang/srt/models/llava.py

@@ -49,7 +49,13 @@ class LlavaBaseForCausalLM(nn.Module):
         image_sizes, pad_values = image_inputs.image_sizes, image_inputs.pad_values
 
         # hardcode for spatial_unpad + anyres
-        image_aspect_ratio = "anyres" if len(image_sizes) == 1 else "pad"
+        if image_inputs.modalities is not None and (
+            "multi-images" in image_inputs.modalities
+            or "video" in image_inputs.modalities
+        ):
+            image_aspect_ratio = "pad"
+        else:
+            image_aspect_ratio = "anyres"
         offset_list = []
         for image_s in image_sizes:
             if len(image_sizes) > 16:

sglang/srt/server.py

@@ -86,6 +86,7 @@ from sglang.srt.utils import (
     set_ulimit,
 )
 from sglang.utils import get_exception_traceback
+from sglang.version import __version__
 
 logger = logging.getLogger(__name__)
 
@@ -455,7 +456,6 @@ def launch_engine(
         data = scheduler_pipe_readers[i].recv()
 
         if data["status"] != "ready":
-            self.shutdown()
             raise RuntimeError(
                 "Initialization failed. Please see the error messages above."
             )
@@ -528,6 +528,7 @@ async def _get_server_info():
         **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
+        "version": __version__,
     }
 
 

sglang/srt/utils.py

@@ -72,7 +72,7 @@ def is_flashinfer_available():
     Check whether flashinfer is available.
     As of Oct. 6, 2024, it is only available on NVIDIA GPUs.
     """
-    if os.environ.get("SGLANG_IS_FLASHINFER_AVAILABLE", "true") == "false":
+    if not get_bool_env_var("SGLANG_IS_FLASHINFER_AVAILABLE", default="true"):
         return False
     return torch.cuda.is_available() and not is_hip()
 
@@ -517,6 +517,11 @@ def monkey_patch_vllm_p2p_access_check(gpu_id: int):
 
     setattr(tgt, "gpu_p2p_access_check", lambda *arg, **kwargs: True)
 
+    # Suppress the warnings from this delete function when using sglang.bench_one_batch
+    from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
+
+    setattr(CustomAllreduce, "__del__", lambda *args, **kwargs: None)
+
 
 vllm_all_gather_backup = None
 
@@ -626,7 +631,7 @@ def add_api_key_middleware(app, api_key: str):
 
 
 def prepare_model_and_tokenizer(model_path: str, tokenizer_path: str):
-    if "SGLANG_USE_MODELSCOPE" in os.environ:
+    if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
         if not os.path.exists(model_path):
             from modelscope import snapshot_download
 
@@ -931,7 +936,7 @@ def get_nvgpu_memory_capacity():
 
 def crash_on_warnings():
     # Crash on warning if we are running CI tests
-    return os.getenv("SGLANG_IS_IN_CI", "false").lower() == "true"
+    return get_bool_env_var("SGLANG_IS_IN_CI")
 
 
 def get_device_name(device_id: int = 0) -> str:
@@ -990,7 +995,7 @@ def direct_register_custom_op(
         my_lib._register_fake(op_name, fake_impl)
 
 
-def gpu_proc_affinity(
+def set_gpu_proc_affinity(
     tp_size: int,
     nnodes: int,
     gpu_id: int,
@@ -1022,3 +1027,8 @@ def gpu_proc_affinity(
     # set cpu_affinity to current process
     p.cpu_affinity(bind_cpu_ids)
     logger.info(f"Process {pid} gpu_id {gpu_id} is running on CPUs: {p.cpu_affinity()}")
+
+
+def get_bool_env_var(name: str, default: str = "false") -> bool:
+    value = os.getenv(name, default)
+    return value.lower() in ("true", "1")

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 kill_child_process
+from sglang.srt.utils import get_bool_env_var, kill_child_process
 from sglang.test.run_eval import run_eval
 from sglang.utils import get_exception_traceback
 
@@ -44,7 +44,7 @@ DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_QUANT_TP1 = "hugging-quants/Meta-Llama-3.1-8
 
 def is_in_ci():
     """Return whether it is in CI runner."""
-    return os.getenv("SGLANG_IS_IN_CI", "false").lower() == "true"
+    return get_bool_env_var("SGLANG_IS_IN_CI")
 
 
 if is_in_ci():

sglang/version.py

@@ -1 +1 @@
-__version__ = "0.3.6.post1"
+__version__ = "0.3.6.post2"

{sglang-0.3.6.post1.dist-info → sglang-0.3.6.post2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sglang
-Version: 0.3.6.post1
+Version: 0.3.6.post2
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -240,6 +240,7 @@ Provides-Extra: srt
 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"
 Provides-Extra: srt-hip
 Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
 Requires-Dist: torch; extra == "srt-hip"
@@ -350,7 +351,7 @@ Learn more in our release blogs: [v0.2 blog](https://lmsys.org/blog/2024-07-25-s
 [Development Roadmap (2024 Q4)](https://github.com/sgl-project/sglang/issues/1487)
 
 ## Adoption and Sponsorship
-The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, NVIDIA, RunPod, Stanford, UC Berkeley, and xAI.
+The project is supported by (alphabetically): AMD, Baseten, Etched, Hyperbolic, Jam & Tea Studios, LinkedIn, NVIDIA, RunPod, Stanford, UC Berkeley, xAI and 01.AI.
 
 ## Acknowledgment and Citation
 We learned from the design and reused code from the following projects: [Guidance](https://github.com/guidance-ai/guidance), [vLLM](https://github.com/vllm-project/vllm), [LightLLM](https://github.com/ModelTC/lightllm), [FlashInfer](https://github.com/flashinfer-ai/flashinfer), [Outlines](https://github.com/outlines-dev/outlines), and [LMQL](https://github.com/eth-sri/lmql).

{sglang-0.3.6.post1.dist-info → sglang-0.3.6.post2.dist-info}/RECORD

@@ -2,22 +2,22 @@ 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=WxrQUkMcxz5GV8OEHj0ckHgpC76HgO6YxmDvJFRDeyU,15670
-sglang/bench_one_batch_server.py,sha256=nzeF_bcaXanQuYLBxAvd3OO4fwbKproMcahXdHIVR6w,5920
+sglang/bench_one_batch.py,sha256=AVMpCBWEsMI2TlMK55JPgPJu0kHg8DI0WV_Bhd4pJgc,15668
+sglang/bench_one_batch_server.py,sha256=hYc3r9JQOLrfqmKgKPOmP0Kr63Sya9wPV_dHzMRZ2Dw,5924
 sglang/bench_serving.py,sha256=hI7FjaERyqKBrYtKewDU6E4rSufKxqsUPyUgtWtTKSI,52545
-sglang/check_env.py,sha256=nR2m0a9WbQmkimJihUx-Lqi7XjN0jyWTCO2vYyA7R2M,5356
+sglang/check_env.py,sha256=rE4ZAG0e6M-Xd-qdHcKclN8Qav6b9gEh4yvlV_TbOg0,5450
 sglang/global_config.py,sha256=fnT0U9vlHdGaQFKN9tYTnUF4-eVW4HYQURd5zvPtrg0,1286
-sglang/launch_server.py,sha256=_XIqBcXArYtHTqilOFkYWKZBYXGCMHAxbYOST08LGj0,415
+sglang/launch_server.py,sha256=U17c44CbbpMBm2JQxVLaz1mfUKk7PgBDhTLAFNeJEvI,362
 sglang/launch_server_llavavid.py,sha256=tGc17S1vUfLwbi1GB26oOdXxTWr7gjlqpTrPnrMRNO8,1007
 sglang/utils.py,sha256=eCvD3fZCALr-MuyZxJL7HAeeqqpxAxf4LJrf7OiCbco,11547
-sglang/version.py,sha256=YrfhKDmn6rTAj_qREKEXk2FahHCqSbHd4BNoD7wlIi0,28
+sglang/version.py,sha256=_Aams_yVBpGe9-85k-kF3qpgcd3D_AsWkVfMFmCWh3c,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
 sglang/lang/compiler.py,sha256=o1C6G3TzhjSlsH-doTPy5oiVehr57dxNTa5oZw5TTAI,7639
 sglang/lang/interpreter.py,sha256=SBjejhLhTKzNM0HbjtTg5r17WPJ64WFSk6lcM_SCWKs,30717
 sglang/lang/ir.py,sha256=zpzzAO1YVldhE95Vwz5hU_TQltu-xt8A6rfFr0PuIDA,18410
-sglang/lang/tracer.py,sha256=borJmlSJOhg1RUndGRnilnR60eEZz2Y9aU7BpftsOxU,8287
+sglang/lang/tracer.py,sha256=o-jLAPPSuy2vBfsGGrTAnbuWtORzQ50B4C_P5zvYkx8,8291
 sglang/lang/backend/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sglang/lang/backend/anthropic.py,sha256=EXRX7xJgA5KZszX7toSLVnKzFQ5EO0Loj-YjHFtxSxg,2081
 sglang/lang/backend/base_backend.py,sha256=tdoh9YF3CyekY1BKiX9n7-aA4srDWIuA4RDJLM7q8qg,1985
@@ -29,12 +29,12 @@ 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=7PSxAUhiS796yQFeiQxiilRhLQ3FpV0wL53CfDgkCIk,30851
+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=APZEUancLC0jRI1JMbv7e5bIZy3OEySGyZspxGA60yQ,33509
+sglang/srt/utils.py,sha256=QXc01TOB7abpL6p3KzfP7u2xFZohQ-ThbI5DAJGoHeI,33894
 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=dQ58mYKN3M5IwldFZkwIb4CCBa6dREb5Om4Kg2kffOE,9565
+sglang/srt/configs/model_config.py,sha256=r5N_OO4w3_R3kZ80P-ZPECscXmspI41d1vc6uEE9ixM,9526
 sglang/srt/configs/qwen2vl.py,sha256=AYHuFgJ0bwhWYkD7S6fvP7yJejJnuhy4xp5Q2W-O6ps,4424
 sglang/srt/constrained/__init__.py,sha256=UWZNVLvOT5ZBX8M36sONgDmnKtkQ0cSfhQD2jO0ATuk,786
 sglang/srt/constrained/base_grammar_backend.py,sha256=FhVm7PxhXDl0joV9NP5RjKgz7dR1dZvUAQnh0mdtvVY,2353
@@ -50,12 +50,12 @@ sglang/srt/layers/logits_processor.py,sha256=V8fHxeQK8lzUhGD2Xc7MY1Y9qBhzFyh6hqp
 sglang/srt/layers/pooler.py,sha256=rj2lygvleBnyLCBZ8I11HGMgpfIDsT0l3PIkshJwdu4,1606
 sglang/srt/layers/radix_attention.py,sha256=C_mK4mfmKlxMRNeKYP9E5R3PRd3eT-OcE_g3mo36dJM,2058
 sglang/srt/layers/rotary_embedding.py,sha256=29tx3JNR40AoXqBa2cFGBjva9vU2xgFipETlpMaaZas,3985
-sglang/srt/layers/sampler.py,sha256=zgNwgUx7fozkWsEJFRKDV9SipHBijfpU9pTroNst6Ho,4552
+sglang/srt/layers/sampler.py,sha256=_enfER8MSxsCYrR6_NgyFxKA_XqKtii_asOZUFUUsd8,4580
 sglang/srt/layers/torchao_utils.py,sha256=v0hyr4hLsM42QwOPCdKb-ftRTjVokBZbqvRj4O4C-Nw,3415
 sglang/srt/layers/vocab_parallel_embedding.py,sha256=RmaZbgXbFnGKX1eGYxlmiko-6JwaJX6seHupUSCtAm8,21583
 sglang/srt/layers/attention/__init__.py,sha256=EL1o6Q5vLgViN3pOr2A7F6K9FlNEpMdBypFAVMeq_HA,2445
 sglang/srt/layers/attention/double_sparsity_backend.py,sha256=BlX7uXteQpnoOnKsdBKh8h20zMVMEiibB5F_PkZSlNI,10706
-sglang/srt/layers/attention/flashinfer_backend.py,sha256=oblYMbmYzK94H3EA9lMhKWaKdi8HLH5NqAiZmjzj4Es,24875
+sglang/srt/layers/attention/flashinfer_backend.py,sha256=ENnNbsA8bY--eFe-Ecqa2RRklH2-a7SV_yZRzpDKnDQ,24879
 sglang/srt/layers/attention/triton_backend.py,sha256=gjxed2cvc2-8QEHkzyTVv6ui7oYOp2b_vgIUQVD1XuM,6538
 sglang/srt/layers/attention/triton_ops/decode_attention.py,sha256=BE63WhKiutSNkhJLsRwvfsRy-ExvuAv7FZyoWv73ul8,18744
 sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py,sha256=1pSXfY3EEaM7iRN_uElHnAfsrJMhTFbu9fj8Z0O2PbE,21480
@@ -73,13 +73,13 @@ 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=nWBn54pz3aQ8tzVvViwwL2k0V4WATi0qw11H0Bzua-Q,7389
-sglang/srt/managers/image_processor.py,sha256=Pk_dtXzljTkFt7Acsv1RyDzEqvCvjc7BMngxGhtkpDU,13817
+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=-5oYdkStPiYjPWl0tCkUVRjTGB7fjA0wIngK-09da7w,43111
+sglang/srt/managers/schedule_batch.py,sha256=jBABHbL7gyrKdrFrzScJ76MtvG2D9Y5HDx74qsclo80,44470
 sglang/srt/managers/schedule_policy.py,sha256=ayFz4iPLIlG8mx5i1glTCAMHJPGpFedMP9UgRtqkNhA,12526
-sglang/srt/managers/scheduler.py,sha256=8owHPXG6fxZtsCWSJ6K7EOlFDcPxYinZC1DwKMJcEVM,55930
-sglang/srt/managers/session_controller.py,sha256=jXoPHxMGh8T1iYWIEjSXoPVwaL6NEjv3QtqlsrvPE1c,2355
+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/tp_worker.py,sha256=1SQJ60iKS9e5vGY555fT1iZ4OtLumXzeWfB08fSWKbk,6176
 sglang/srt/managers/tp_worker_overlap_thread.py,sha256=7vhPebaOS4JamaS08CGf_hwxnUO7Gy_SXZXEPwNHKoY,7621
@@ -112,7 +112,7 @@ sglang/srt/models/llama.py,sha256=FSGuM3BamhuT5h2jedh5cSFwFYduOJwkAZJJ672awRw,16
 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=72DnZXIwu78zYqU8YIElq_AaSIFO_icYOPTHXE0_-YQ,24941
+sglang/srt/models/llava.py,sha256=HjC2TDLngpaN8HMYyGp5doEK32HeQN8iT2tYE_Slrtg,25130
 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
@@ -155,10 +155,10 @@ 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=ULF7C3pLXkMevXgE_Dodt29OBfvvXKUnRvwKhaBg1ys,23470
+sglang/test/test_utils.py,sha256=NBEGQC_wtMqODQQZWrxdwmsoLFSZfDlQzIbsQ1kE_Yc,23468
 sglang/test/srt/sampling/penaltylib/utils.py,sha256=CjxHgywh0hx_87iynzQt_ztHu6zBVuE-YrZ-XPmW6U4,12906
-sglang-0.3.6.post1.dist-info/LICENSE,sha256=FJXh51fvTQklojUFY89XVLsjxRcBqOxPs8XNy-2uZ0c,11346
-sglang-0.3.6.post1.dist-info/METADATA,sha256=XwhCEL8SbEVcT7LQLk26g6tzduS6mByBE7dDqZYpQxo,22073
-sglang-0.3.6.post1.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-sglang-0.3.6.post1.dist-info/top_level.txt,sha256=yxhh3pYQkcnA7v3Bg889C2jZhvtJdEincysO7PEB09M,7
-sglang-0.3.6.post1.dist-info/RECORD,,
+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,,