sgl-project · JustinTong0323 · Jan 26, 2026 · Jan 26, 2026 · Jan 27, 2026 · Jan 27, 2026
@@ -30,8 +30,6 @@ dependencies = [
   "flashinfer_python==0.6.4", # keep it aligned with jit-cache version in Dockerfile
   "flashinfer_cubin==0.6.4",
   "gguf",
-  "hf_transfer",
-  "huggingface_hub",
   "interegular",
   "llguidance>=0.7.11,<0.8.0",
   "modelscope",
@@ -72,7 +70,7 @@ dependencies = [
   "av ; sys_platform == 'linux' and (platform_machine == 'aarch64' or platform_machine == 'arm64' or platform_machine == 'armv7l')",
   "torchvision",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.3.0",
   "uvicorn",
   "uvloop",
   "watchfiles",
@@ -135,14 +133,15 @@ tracing = [
 
 test = [
   "accelerate",
+  "addict",
   "bitsandbytes",
   "expecttest",
   "jsonlines",
   "lm-eval[api]>=0.4.9.2",
   "matplotlib",
   "pandas",
   "parameterized",
-  "peft",
+  "peft>=0.18.0",
   "pytest",
   "sentence_transformers",
   "tabulate",

diff --git a/python/pyproject_cpu.toml b/python/pyproject_cpu.toml
@@ -26,8 +26,6 @@ dependencies = [
   "einops",
   "fastapi",
   "gguf",
-  "hf_transfer",
-  "huggingface_hub",
   "intel-openmp; platform_machine == 'x86_64'",
   "interegular",
   "llguidance>=0.7.11,<0.8.0",
@@ -62,7 +60,7 @@ dependencies = [
   "torchaudio==2.9.0",
   "torchvision==0.24.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.3.0",
   "triton==3.5.0",
   "uvicorn",
   "uvloop",
@@ -85,7 +83,7 @@ test = [
   "jsonlines",
   "matplotlib",
   "pandas",
-  "peft",
+  "peft>=0.18.0",
   "pytest",
   "sentence_transformers",
 ]

diff --git a/python/pyproject_npu.toml b/python/pyproject_npu.toml
@@ -26,8 +26,6 @@ dependencies = [
   "einops",
   "fastapi",
   "gguf",
-  "hf_transfer",
-  "huggingface_hub",
   "interegular",
   "llguidance>=0.7.11,<0.8.0",
   "modelscope",
@@ -57,7 +55,7 @@ dependencies = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.3.0",
   "uvicorn",
   "uvloop",
   "xgrammar==0.1.27",
@@ -98,7 +96,7 @@ test = [
   "jsonlines",
   "matplotlib",
   "pandas",
-  "peft",
+  "peft>=0.18.0",
   "pytest",
   "sentence_transformers",
   "tabulate",

diff --git a/python/pyproject_other.toml b/python/pyproject_other.toml
@@ -28,8 +28,6 @@ runtime_common = [
   "einops",
   "fastapi",
   "gguf",
-  "hf_transfer",
-  "huggingface_hub",
   "interegular",
   "llguidance>=0.7.11,<0.8.0",
   "modelscope",
@@ -59,7 +57,7 @@ runtime_common = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.3.0",
   "uvicorn",
   "uvloop",
   "xgrammar==0.1.27",
@@ -143,7 +141,7 @@ test = [
   "jsonlines",
   "matplotlib",
   "pandas",
-  "peft",
+  "peft>=0.18.0",
   "pytest",
   "sentence_transformers",
   "tabulate",

diff --git a/python/pyproject_xpu.toml b/python/pyproject_xpu.toml
@@ -31,8 +31,6 @@ dependencies = [
   "einops",
   "fastapi",
   "gguf",
-  "hf_transfer",
-  "huggingface_hub",
   "interegular",
   "llguidance>=0.7.11,<0.8.0",
   "modelscope",
@@ -62,7 +60,7 @@ dependencies = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.3.0",
   "uvicorn",
   "uvloop",
   # "xgrammar==0.1.24", , xgrammar depends on CUDA PyTorch and Triton only
@@ -84,7 +82,7 @@ test = [
   "jsonlines",
   "matplotlib",
   "pandas",
-  "peft",
+  "peft>=0.18.0",
   "pytest",
   "sentence_transformers",
   "tabulate",

diff --git a/python/sglang/check_env.py b/python/sglang/check_env.py
@@ -30,7 +30,6 @@ def is_cuda_v2():
     "numpy",
     "aiohttp",
     "fastapi",
-    "hf_transfer",
     "huggingface_hub",
     "interegular",
     "modelscope",

@@ -12,7 +12,6 @@
 from pathlib import Path
 
 import filelock
-import huggingface_hub.constants
 import torch
 from safetensors.torch import safe_open
 from torch.distributed.tensor import DTensor
@@ -37,21 +36,6 @@
 temp_dir = tempfile.gettempdir()
 
 
-def enable_hf_transfer() -> None:
-    """automatically activates hf_transfer"""
-    if "HF_HUB_ENABLE_HF_TRANSFER" not in os.environ:
-        try:
-            # enable hf hub transfer if available
-            import hf_transfer  # type: ignore # noqa
-
-            huggingface_hub.constants.HF_HUB_ENABLE_HF_TRANSFER = True
-        except ImportError:
-            pass
-
-
-enable_hf_transfer()
-
-
 class DisabledTqdm(tqdm):
 
     def __init__(self, *args, **kwargs):

@@ -227,8 +227,8 @@ def __init__(
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = getattr(config, "rope_theta", 10000)
-        rope_scaling = getattr(config, "rope_scaling", None)
+        rope_theta = config.rope_parameters["rope_theta"]
+        rope_scaling = config.rope_parameters
         if rope_scaling is not None and getattr(
             config, "original_max_position_embeddings", None
         ):

@@ -798,6 +798,11 @@ def get_image_features(
         """
         pixel_values = pixel_values.type(self.visual.dtype)
         image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+        if not isinstance(image_embeds, torch.Tensor):
+            # In transformers v5, the visual encoder returns BaseModelOutputWithPooling.
+            # pooler_output contains the spatially merged embeddings (what we need),
+            # while last_hidden_state contains the raw unmerged output.
+            image_embeds = image_embeds.pooler_output
         split_sizes = (
             image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2
         ).tolist()

@@ -204,8 +204,8 @@ def __init__(
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = getattr(config, "rope_theta", 1000000.0)
-        rope_scaling = getattr(config, "rope_scaling", None)
+        rope_theta = config.rope_parameters["rope_theta"]
+        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(config, "max_position_embeddings", 40960)
         attention_bias = getattr(config, "attention_bias", False)
 

@@ -7,6 +7,8 @@
 This module contains implementations of image encoding stages for diffusion pipelines.
 """
 
+import inspect
+
 import PIL
 import torch
 from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution
@@ -111,9 +113,15 @@ def forward(
             server_args.pipeline_config.prepare_image_processor_kwargs(batch)
         )
 
+        params = inspect.signature(self.image_processor.__call__).parameters
+        image_processor_kwargs = {
+            k: v for k, v in image_processor_kwargs.items() if k in params
+        }
+
         image_inputs = self.image_processor(
             images=image, return_tensors="pt", **image_processor_kwargs
         ).to(cuda_device)
+
         if self.image_encoder:
             # if an image encoder is provided
             with set_forward_context(current_timestep=0, attn_metadata=None):

diff --git a/python/sglang/srt/configs/internvl.py b/python/sglang/srt/configs/internvl.py
@@ -593,7 +593,6 @@ def convert_tokens_to_string(self, tokens):
                 current_sub_tokens.append(token)
                 prev_is_special = False
         out_string += self.sp_model.decode(current_sub_tokens)
-        out_string = self.clean_up_tokenization(out_string)
         out_string = self._maybe_add_prefix_space(tokens=tokens, decoded=out_string)
         return out_string[1:]
 

diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py
@@ -51,10 +51,15 @@ class ModelImpl(str, Enum):
     MINDSPORE = "mindspore"
 
 
-def is_deepseek_nsa(config: PretrainedConfig) -> bool:
+def is_deepseek_nsa(config) -> bool:
+    architectures = (
+        config.get("architectures")
+        if isinstance(config, dict)
+        else getattr(config, "architectures", None)
+    )
     return (
-        config.architectures is not None
-        and config.architectures[0]
+        architectures is not None
+        and architectures[0]
         in [
             "DeepseekV3ForCausalLM",
             "DeepseekV32ForCausalLM",
@@ -63,7 +68,12 @@ def is_deepseek_nsa(config: PretrainedConfig) -> bool:
             "PixtralForConditionalGeneration",
             "GlmMoeDsaForCausalLM",
         ]
-        and getattr(config, "index_topk", None) is not None
+        and (
+            config.get("index_topk")
+            if isinstance(config, dict)
+            else getattr(config, "index_topk", None)
+        )
+        is not None
     )
 
 
@@ -458,7 +468,13 @@ def _derive_model_shapes(self):
                 )
                 if rope_type != "default":
                     mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
-                    scaling_factor = rope_scaling["factor"]
+                    if "factor" not in rope_scaling:
+                        logger.warning(
+                            "rope_scaling (type=%s) missing 'factor', defaulting to 1.0. "
+                            "Check model accuracy.",
+                            rope_type,
+                        )
+                    scaling_factor = rope_scaling.get("factor", 1.0)
                     mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
                     self.scaling = self.scaling * mscale * mscale
         elif "MiniCPM3ForCausalLM" in self.hf_config.architectures:
@@ -504,7 +520,12 @@ def _derive_model_shapes(self):
                 mscale_all_dim = self.hf_config.rope_scaling.get(
                     "mscale_all_dim", False
                 )
-                scaling_factor = self.hf_config.rope_scaling["factor"]
+                if "factor" not in self.hf_config.rope_scaling:
+                    logger.warning(
+                        "BailingMoe rope_scaling missing 'factor', defaulting to 1.0. "
+                        "Check model accuracy.",
+                    )
+                scaling_factor = self.hf_config.rope_scaling.get("factor", 1.0)
                 mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
                 self.scaling = self.scaling * mscale * mscale
         elif "SarvamMLAForCausalLM" in self.hf_config.architectures:
@@ -521,7 +542,12 @@ def _derive_model_shapes(self):
                 mscale_all_dim = self.hf_config.rope_scaling.get(
                     "mscale_all_dim", False
                 )
-                scaling_factor = self.hf_config.rope_scaling["factor"]
+                if "factor" not in self.hf_config.rope_scaling:
+                    logger.warning(
+                        "SarvamMLA rope_scaling missing 'factor', defaulting to 1.0. "
+                        "Check model accuracy.",
+                    )
+                scaling_factor = self.hf_config.rope_scaling.get("factor", 1.0)
                 mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
                 self.scaling = self.scaling * mscale * mscale
         else: