[Packaging] Remove hf dependency and add py.typed

pyproject.toml

@@ -17,7 +17,6 @@ requires-python = ">=3.8, <4"
 dependencies = [
   "pydantic",
   "torch>=1.10.0",
-  "transformers>=4.38.0",
   "triton; platform_system == 'Linux' and platform_machine == 'x86_64'",
   "mlx-lm; platform_system == 'Darwin' and platform_machine == 'arm64'",
   "ninja",
@@ -34,11 +33,11 @@ test = [
   "huggingface-hub[cli]",
   "protobuf",
   "pytest",
-  "sentencepiece",
-  "tiktoken",
   # transformers==4.50.0 has error on MacOS.
   # https://github.com/huggingface/transformers/issues/36906
-  "transformers<4.50.0; platform_system == 'Darwin'",
+  "transformers!=4.50.0",
+  "sentencepiece",
+  "tiktoken",
 ]
 
 [build-system]
@@ -78,6 +77,7 @@ sdist.include = [
   "/cpp/**/*.h",
   "/include/**/*",
   "/python/xgrammar/**/*.py",
+  "/python/xgrammar/py.typed",
 
   # Third party files
   "/3rdparty/**/*",

python/xgrammar/tokenizer_info.py

@@ -2,18 +2,40 @@
 
 import json
 from enum import Enum
-from typing import Any, Dict, List, Optional, Union
-
-try:
-    import sentencepiece
-except ImportError:
-    sentencepiece = None
-try:
-    import tiktoken
-except ImportError:
-    tiktoken = None
-
-from transformers import PreTrainedTokenizerBase, PreTrainedTokenizerFast
+from types import ModuleType
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
+
+if TYPE_CHECKING:
+    try:
+        import sentencepiece
+    except ImportError:
+        pass
+
+    try:
+        import tiktoken
+    except ImportError:
+        pass
+
+    try:
+        from transformers import PreTrainedTokenizerBase, PreTrainedTokenizerFast
+    except ImportError:
+        pass
+else:
+    try:
+        import sentencepiece
+    except ImportError:
+        sentencepiece = None
+
+    try:
+        import tiktoken
+    except ImportError:
+        tiktoken = None
+
+    try:
+        from transformers import PreTrainedTokenizerBase, PreTrainedTokenizerFast
+    except ImportError:
+        PreTrainedTokenizerBase = Any
+        PreTrainedTokenizerFast = Any
 
 from .base import XGRObject, _core
 from .support import logging
@@ -162,7 +184,7 @@
 
         Parameters
         ----------
-        tokenizer : PreTrainedTokenizerBase
+        tokenizer : transformers.PreTrainedTokenizerBase
             The huggingface tokenizer.
 
         vocab_size : Optional[int], default: None
@@ -258,6 +280,7 @@
                         "stop_token_ids is neither provided by user nor found from the tokenizer. "
                         "It will be automatically detected."
                     )
+
             return TokenizerInfo(
                 encoded_vocab,
                 VocabType.RAW,