sglang 0.4.3.post4__py3-none-any.whl → 0.4.4.post1__py3-none-any.whl

sglang/bench_serving.py

@@ -1006,7 +1006,7 @@ async def benchmark(
 
     # Flush cache
     if "sglang" in backend:
-        requests.post(base_url + "/flush_cache")
+        requests.post(base_url + "/flush_cache", headers=get_auth_headers())
 
     time.sleep(1.0)
 

sglang/lang/chat_template.py

@@ -230,6 +230,29 @@ register_chat_template(
     )
 )
 
+register_chat_template(
+    ChatTemplate(
+        name="janus-pro",
+        default_system_prompt=None,
+        role_prefix_and_suffix={
+            "system": (
+                "",
+                "",
+            ),
+            "User": (
+                "<｜User｜>",
+                "",
+            ),
+            "assistant": (
+                "<｜Assistant｜>",
+                "<｜end▁of▁sentence｜>",
+            ),
+        },
+        stop_str=("<｜end▁of▁sentence｜>",),
+        image_token="<image_placeholder>\n",
+    )
+)
+
 # The difference between "llama-3-instruct-llava" and "llama-3-instruct" is that llava uses a different image_token.
 register_chat_template(
     ChatTemplate(
@@ -384,6 +407,12 @@ def match_deepseek(model_path: str):
         return get_chat_template("deepseek-v3")
 
 
+@register_chat_template_matching_function
+def match_deepseek_janus_pro(model_path: str):
+    if "janus" in model_path.lower():
+        return get_chat_template("janus-pro")
+
+
 @register_chat_template_matching_function
 def match_dbrx(model_path: str):
     if "dbrx" in model_path.lower() and "instruct" in model_path.lower():

sglang/srt/_custom_ops.py

@@ -6,10 +6,12 @@ from typing import List, Tuple
 import torch
 import torch.library
 
-from sglang.srt.utils import is_hip, is_hpu
+from sglang.srt.utils import get_bool_env_var, is_hip, is_hpu
 
 logger = logging.getLogger(__name__)
-use_vllm_custom_allreduce = os.environ.get("USE_VLLM_CUSTOM_ALLREDUCE", default=True)
+use_vllm_custom_allreduce = get_bool_env_var(
+    "USE_VLLM_CUSTOM_ALLREDUCE", default="true"
+)
 
 if not is_hpu():
     # ROCm does not use vllm custom allreduce
@@ -75,42 +77,42 @@ else:
             rank: int,
             full_nvlink: bool,
         ) -> int:
-            return sgl_kernel.ops.allreduce.init_custom_ar(
+            return sgl_kernel.allreduce.init_custom_ar(
                 meta, rank_data, handles, offsets, rank, full_nvlink
             )
 
         def all_reduce_reg(fa: int, inp: torch.Tensor, out: torch.Tensor) -> None:
-            sgl_kernel.ops.allreduce.all_reduce_reg(fa, inp, out)
+            sgl_kernel.allreduce.all_reduce_reg(fa, inp, out)
 
         def all_reduce_unreg(
             fa: int, inp: torch.Tensor, reg_buffer: torch.Tensor, out: torch.Tensor
         ) -> None:
-            sgl_kernel.ops.allreduce.all_reduce_unreg(fa, inp, reg_buffer, out)
+            sgl_kernel.allreduce.all_reduce_unreg(fa, inp, reg_buffer, out)
 
         def dispose(fa: int) -> None:
-            sgl_kernel.ops.allreduce.dispose(fa)
+            sgl_kernel.allreduce.dispose(fa)
 
         def meta_size() -> int:
-            return sgl_kernel.ops.allreduce.meta_size()
+            return sgl_kernel.allreduce.meta_size()
 
         def register_buffer(
             fa: int, t: torch.Tensor, handles: List[str], offsets: List[int]
         ) -> None:
-            return sgl_kernel.ops.allreduce.register_buffer(fa, t, handles, offsets)
+            return sgl_kernel.allreduce.register_buffer(fa, t, handles, offsets)
 
         def get_graph_buffer_ipc_meta(fa: int) -> Tuple[torch.Tensor, List[int]]:
-            return sgl_kernel.ops.allreduce.get_graph_buffer_ipc_meta(fa)
+            return sgl_kernel.allreduce.get_graph_buffer_ipc_meta(fa)
 
         def register_graph_buffers(
             fa: int, handles: List[str], offsets: List[List[int]]
         ) -> None:
-            sgl_kernel.ops.allreduce.register_graph_buffers(fa, handles, offsets)
+            sgl_kernel.allreduce.register_graph_buffers(fa, handles, offsets)
 
         def allocate_meta_buffer(size: int) -> torch.Tensor:
-            return sgl_kernel.ops.allreduce.allocate_meta_buffer(size)
+            return sgl_kernel.allreduce.allocate_meta_buffer(size)
 
         def get_meta_buffer_ipc_handle(inp: torch.Tensor) -> torch.Tensor:
-            return sgl_kernel.ops.allreduce.get_meta_buffer_ipc_handle(inp)
+            return sgl_kernel.allreduce.get_meta_buffer_ipc_handle(inp)
 
     else:
         # TRTLLM custom allreduce
@@ -123,7 +125,7 @@ else:
             barrier_in: List[int],
             barrier_out: List[int],
         ) -> int:
-            return sgl_kernel.ops.init_custom_reduce(
+            return sgl_kernel.init_custom_reduce(
                 rank_id,
                 world_size,
                 rank_data_base,
@@ -134,15 +136,15 @@ else:
             )
 
         def all_reduce(fa: int, inp: torch.Tensor, out: torch.Tensor) -> None:
-            sgl_kernel.ops.custom_reduce(fa, inp, out)
+            sgl_kernel.custom_reduce(fa, inp, out)
 
         def dispose(fa: int) -> None:
-            sgl_kernel.ops.custom_dispose(fa)
+            sgl_kernel.custom_dispose(fa)
 
         def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
-            return sgl_kernel.ops.get_graph_buffer_ipc_meta(fa)
+            return sgl_kernel.get_graph_buffer_ipc_meta(fa)
 
         def register_graph_buffers(
             fa: int, handles: List[List[int]], offsets: List[List[int]]
         ) -> None:
-            sgl_kernel.ops.register_graph_buffers(fa, handles, offsets)
+            sgl_kernel.register_graph_buffers(fa, handles, offsets)

sglang/srt/configs/__init__.py

@@ -1,6 +1,7 @@
 from sglang.srt.configs.chatglm import ChatGLMConfig
 from sglang.srt.configs.dbrx import DbrxConfig
 from sglang.srt.configs.exaone import ExaoneConfig
+from sglang.srt.configs.janus_pro import MultiModalityConfig
 from sglang.srt.configs.qwen2_5_vl_config import (
     Qwen2_5_VLConfig,
     Qwen2_5_VLVisionConfig,
@@ -12,4 +13,5 @@ __all__ = [
     "DbrxConfig",
     "Qwen2_5_VLConfig",
     "Qwen2_5_VLVisionConfig",
+    "MultiModalityConfig",
 ]