sglang 0.4.3__py3-none-any.whl → 0.4.3.post1__py3-none-any.whl

sglang/lang/backend/openai.py

@@ -161,6 +161,10 @@ class OpenAI(BaseBackend):
                 prompt = s.text_
 
             kwargs = sampling_params.to_openai_kwargs()
+            if self.model_name.startswith("o1") or self.model_name.startswith("o3"):
+                kwargs.pop("max_tokens", None)
+            else:
+                kwargs.pop("max_completion_tokens", None)
             comp = openai_completion(
                 client=self.client,
                 token_usage=self.token_usage,
@@ -175,6 +179,7 @@ class OpenAI(BaseBackend):
             ), "constrained type not supported on chat model"
             kwargs = sampling_params.to_openai_kwargs()
             kwargs.pop("stop")
+
             comp = openai_completion(
                 client=self.client,
                 token_usage=self.token_usage,

sglang/lang/chat_template.py

@@ -353,7 +353,6 @@ register_chat_template(
     )
 )
 
-
 register_chat_template(
     ChatTemplate(
         name="deepseek-v3",
@@ -428,12 +427,15 @@ def match_chat_ml(model_path: str):
     if "tinyllama" in model_path:
         return get_chat_template("chatml")
     # Now the suffix for qwen2 chat model is "instruct"
-    if (
-        "qwen" in model_path
-        and ("chat" in model_path or "instruct" in model_path)
-        and ("llava" not in model_path)
-    ):
-        return get_chat_template("qwen")
+    if "qwen" in model_path and "vl" in model_path:
+        return get_chat_template("qwen2-vl")
+    if "qwen" in model_path:
+        if "vl" in model_path:
+            return get_chat_template("qwen2-vl")
+        if ("chat" in model_path or "instruct" in model_path) and (
+            "llava" not in model_path
+        ):
+            return get_chat_template("qwen")
     if (
         "llava-v1.6-34b" in model_path
         or "llava-v1.6-yi-34b" in model_path
@@ -443,6 +445,12 @@ def match_chat_ml(model_path: str):
         return get_chat_template("chatml-llava")
 
 
+@register_chat_template_matching_function
+def match_chat_minicpm(model_path: str):
+    if "minicpm" in model_path:
+        return get_chat_template("minicpmv")
+
+
 @register_chat_template_matching_function
 def match_chat_yi(model_path: str):
     model_path = model_path.lower()
@@ -459,6 +467,13 @@ def match_gemma_it(model_path: str):
         return get_chat_template("gemma-it")
 
 
+@register_chat_template_matching_function
+def match_openbmb_minicpm(model_path: str):
+    model_path = model_path.lower()
+    if "minicpm" in model_path:
+        return get_chat_template("minicpmv")
+
+
 @register_chat_template_matching_function
 def match_c4ai_command_r(model_path: str):
     model_path = model_path.lower()

sglang/lang/ir.py

@@ -63,6 +63,7 @@ class SglSamplingParams:
             warnings.warn("Regular expression is not supported in the OpenAI backend.")
         return {
             "max_tokens": self.max_new_tokens,
+            "max_completion_tokens": self.max_new_tokens,
             "stop": self.stop or None,
             "temperature": self.temperature,
             "top_p": self.top_p,

sglang/srt/configs/__init__.py

@@ -1,12 +1,15 @@
 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.qwen2vl import Qwen2VLConfig, Qwen2VLVisionConfig
+from sglang.srt.configs.qwen2_5_vl_config import (
+    Qwen2_5_VLConfig,
+    Qwen2_5_VLVisionConfig,
+)
 
 __all__ = [
     "ExaoneConfig",
-    "Qwen2VLConfig",
-    "Qwen2VLVisionConfig",
     "ChatGLMConfig",
     "DbrxConfig",
+    "Qwen2_5_VLConfig",
+    "Qwen2_5_VLVisionConfig",
 ]

sglang/srt/configs/model_config.py

@@ -98,6 +98,7 @@ class ModelConfig:
         if (
             "DeepseekV2ForCausalLM" in self.hf_config.architectures
             or "DeepseekV3ForCausalLM" in self.hf_config.architectures
+            or "DeepseekV3ForCausalLMNextN" in self.hf_config.architectures
         ):
             self.head_dim = 256
             self.attention_arch = AttentionArch.MLA
@@ -402,6 +403,7 @@ def is_multimodal_model(model_architectures: List[str]):
         or "LlavaVidForCausalLM" in model_architectures
         or "MllamaForConditionalGeneration" in model_architectures
         or "Qwen2VLForConditionalGeneration" in model_architectures
+        or "Qwen2_5_VLForConditionalGeneration" in model_architectures
         or "MiniCPMV" in model_architectures
     ):
         return True