Making sgl-kernel work by chaging torch back to 2.9.1

python/pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
   "blobfile==3.0.0",
   "build",
   "compressed-tensors",
-  "cuda-python>=13.1,<13.3",
+  "cuda-python>=13.0",
   "decord2",
   "datasets",
   "einops",
@@ -64,11 +64,11 @@ dependencies = [
   "tiktoken",
   "timm==1.0.16",
   "torch_memory_saver==0.0.9",
-  "torch>=2.10,<2.12",
+  "torch==2.9.1",
   "torchao==0.9.0",
-  "torchaudio>=2.10,<2.12",
+  "torchaudio==2.9.1",
   "torchcodec==0.8.0 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')", # torchcodec does not exist in those systems. If not provided, transformer will use torchvision instead by default.
-  "torchvision>=0.25,<0.27",
+  "torchvision",
   "tqdm",
   "transformers==4.57.1",
   "uvicorn",

python/sglang/srt/configs/__init__.py

@@ -2,7 +2,7 @@
 from sglang.srt.configs.bailing_hybrid import BailingHybridConfig
 from sglang.srt.configs.chatglm import ChatGLMConfig
 from sglang.srt.configs.dbrx import DbrxConfig
-from sglang.srt.configs.deepseekvl2 import DeepseekVL2Config
+from sglang.srt.configs.deepseekvl2 import DeepseekVL2Config, GlmMoeDsaConfig
 from sglang.srt.configs.dots_ocr import DotsOCRConfig
 from sglang.srt.configs.dots_vlm import DotsVLMConfig
 from sglang.srt.configs.exaone import ExaoneConfig
@@ -38,6 +38,7 @@
     "ChatGLMConfig",
     "DbrxConfig",
     "DeepseekVL2Config",
+    "GlmMoeDsaConfig",
     "LongcatFlashConfig",
     "MultiModalityConfig",
     "KimiVLConfig",

python/sglang/srt/configs/deepseekvl2.py

@@ -647,6 +647,10 @@ def __init__(
         )
 
 
+class GlmMoeDsaConfig(DeepseekV2Config):
+    model_type = "glm_moe_dsa"
+
+
 class DeepseekVL2Config(PretrainedConfig):
     model_type = "deepseek_vl_v2"
     vision_config: DeepseekVL2VisionEncoderConfig

python/sglang/srt/layers/moe/kt_ep_wrapper.py

@@ -1646,7 +1646,14 @@ def create_kt_config_from_server_args(
     if is_kt_ep_wrapper_disabled():
         return None
 
-    if server_args.kt_weight_path is None:
+    kt_method = (server_args.kt_method or "").upper()
+    weight_path = (
+        server_args.pagedmoe_storage_root
+        if kt_method == "PAGEDMOE" and server_args.pagedmoe_storage_root is not None
+        else server_args.kt_weight_path
+    )
+
+    if weight_path is None:
         return None
 
     # Get GPU experts masks (initializes if needed)
@@ -1669,7 +1676,7 @@ def create_kt_config_from_server_args(
         cpuinfer_threads=server_args.kt_cpuinfer,
         threadpool_count=server_args.kt_threadpool_count,
         numa_nodes=server_args.kt_numa_nodes,
-        weight_path=server_args.kt_weight_path,
+        weight_path=weight_path,
         chunked_prefill_size=server_args.chunked_prefill_size,
         method=server_args.kt_method,
         max_deferred_experts_per_token=server_args.kt_max_deferred_experts_per_token,

python/sglang/srt/server_args.py

@@ -551,6 +551,7 @@ class ServerArgs:
 
     # Ktransformers/AMX expert parallelism
     kt_weight_path: Optional[str] = None
+    pagedmoe_storage_root: Optional[str] = None
     kt_method: Optional[str] = None
     kt_cpuinfer: Optional[int] = None
     kt_threadpool_count: Optional[int] = None
@@ -4436,6 +4437,11 @@ def add_cli_args(parser: argparse.ArgumentParser):
             type=str,
             help="[ktransformers parameter] The path of the quantized expert weights for amx kernel. A local folder.",
         )
+        parser.add_argument(
+            "--pagedmoe-storage-root",
+            type=str,
+            help="[pagedmoe parameter] The root directory containing pagedmoe manifest.json and layer storage.",
+        )
         parser.add_argument(
             "--kt-method",
             type=str,

python/sglang/srt/utils/hf_transformers_utils.py

@@ -53,6 +53,7 @@
     DotsVLMConfig,
     ExaoneConfig,
     FalconH1Config,
+    GlmMoeDsaConfig,
     GraniteMoeHybridConfig,
     JetNemotronConfig,
     JetVLMConfig,
@@ -86,6 +87,7 @@
     ChatGLMConfig,
     DbrxConfig,
     ExaoneConfig,
+    GlmMoeDsaConfig,
     DeepseekVL2Config,
     MultiModalityConfig,
     KimiVLConfig,
@@ -541,9 +543,32 @@ def get_tokenizer(
         )
         raise RuntimeError(err_msg) from e
     except ValueError as e:
+        tokenizer_json = Path(tokenizer_name) / "tokenizer.json"
+        tokenizer_config = Path(tokenizer_name) / "tokenizer_config.json"
+        if "TokenizersBackend" in str(e) and tokenizer_json.exists():
+            tokenizer_kwargs: Dict[str, Any] = {
+                "tokenizer_file": str(tokenizer_json),
+                "clean_up_tokenization_spaces": False,
+            }
+            if tokenizer_config.exists():
+                with tokenizer_config.open() as f:
+                    tokenizer_config_data = json.load(f)
+                for key in (
+                    "bos_token",
+                    "eos_token",
+                    "pad_token",
+                    "unk_token",
+                    "model_max_length",
+                ):
+                    if key in tokenizer_config_data:
+                        tokenizer_kwargs[key] = tokenizer_config_data[key]
+            tokenizer = PreTrainedTokenizerFast(**tokenizer_kwargs)
+            logging.getLogger(tokenizer.__class__.__module__).addFilter(
+                TokenizerWarningsFilter()
+            )
         # If the error pertains to the tokenizer class not existing or not
         # currently being imported, suggest using the --trust-remote-code flag.
-        if not trust_remote_code and (
+        elif not trust_remote_code and (
             "does not exist or is not currently imported." in str(e)
             or "requires you to execute the tokenizer file" in str(e)
         ):