feat(turn-detector): drop transformers dependency

livekit-plugins/livekit-plugins-turn-detector/livekit/plugins/turn_detector/base.py

@@ -2,7 +2,6 @@
 
 import asyncio
 import json
-import logging
 import math
 import re
 import time
@@ -25,6 +24,9 @@
 MAX_HISTORY_TOKENS = 128
 MAX_HISTORY_TURNS = 6
 
+# files we need from the HF repo besides the ONNX model
+_TOKENIZER_FILES = ("tokenizer.json", "tokenizer_config.json")
+
 
 def _download_from_hf_hub(repo_id: str, filename: str, **kwargs: Any) -> str:
     from huggingface_hub import hf_hub_download
@@ -83,34 +85,22 @@ def _format_chat_ctx(self, chat_ctx: list[dict[str, Any]]) -> str:
                 new_chat_ctx.append(msg)
                 last_msg = msg
 
-        convo_text = self._tokenizer.apply_chat_template(
-            new_chat_ctx, add_generation_prompt=False, add_special_tokens=False, tokenize=False
+        convo_text: str = self._chat_template.render(
+            messages=new_chat_ctx, add_generation_prompt=False
         )
 
         # remove the EOU token from current utterance
         ix = convo_text.rfind("<|im_end|>")
-        text = convo_text[:ix]
-        return text  # type: ignore
+        return convo_text[:ix]
 
     def initialize(self) -> None:
-        logger = logging.getLogger("transformers")
+        import numpy as np
+        import onnxruntime as ort  # type: ignore
+        from huggingface_hub import errors
+        from jinja2 import Environment
+        from tokenizers import Tokenizer  # type: ignore[import-untyped]
 
-        class _SuppressSpecific(logging.Filter):
-            def filter(self, record: logging.LogRecord) -> bool:
-                msg = record.getMessage()
-                return not msg.startswith(
-                    "None of PyTorch, TensorFlow >= 2.0, or Flax have been found."
-                )
-
-        filt = _SuppressSpecific()
-        # filter this log since it conflicts with the console CLI (since it directly prints to stdout)
-        logger.addFilter(filt)
-        try:
-            import onnxruntime as ort  # type: ignore
-            from huggingface_hub import errors
-            from transformers import AutoTokenizer
-        finally:
-            logger.removeFilter(filt)
+        self._np = np
 
         revision = self.__class__.model_revision()
         try:
@@ -121,6 +111,13 @@ def filter(self, record: logging.LogRecord) -> bool:
                 revision=revision,
                 local_files_only=True,
             )
+            tokenizer_json_path = _download_from_hf_hub(
+                HG_MODEL, "tokenizer.json", revision=revision, local_files_only=True
+            )
+            tokenizer_config_path = _download_from_hf_hub(
+                HG_MODEL, "tokenizer_config.json", revision=revision, local_files_only=True
+            )
+
             sess_options = ort.SessionOptions()
             sess_options.intra_op_num_threads = max(
                 1, min(math.ceil(hw.get_cpu_monitor().cpu_count()) // 2, 4)
@@ -130,12 +127,20 @@ def filter(self, record: logging.LogRecord) -> bool:
             self._session = ort.InferenceSession(
                 local_path_onnx, providers=["CPUExecutionProvider"], sess_options=sess_options
             )
-            self._tokenizer = AutoTokenizer.from_pretrained(  # type: ignore[no-untyped-call]
-                HG_MODEL,
-                revision=revision,
-                local_files_only=True,
-                truncation_side="left",
-            )
+
+            self._tokenizer = Tokenizer.from_file(tokenizer_json_path)
+            # match the previous transformers behavior: left-truncate to MAX_HISTORY_TOKENS
+            self._tokenizer.enable_truncation(max_length=MAX_HISTORY_TOKENS, direction="left")
+
+            with open(tokenizer_config_path) as f:
+                tokenizer_config = json.load(f)
+            chat_template = tokenizer_config.get("chat_template")
+            if not chat_template:
+                raise RuntimeError(
+                    f"tokenizer_config.json for {HG_MODEL}@{revision} has no chat_template"
+                )
+            # the EOU model templates are simple message loops; no custom helpers needed
+            self._chat_template = Environment(autoescape=False).from_string(chat_template)
 
         except (errors.LocalEntryNotFoundError, OSError):
             logger.error(
@@ -157,15 +162,9 @@ def run(self, data: bytes) -> bytes | None:
 
         start_time = time.perf_counter()
         text = self._format_chat_ctx(chat_ctx)
-        inputs = self._tokenizer(
-            text,
-            add_special_tokens=False,
-            return_tensors="np",
-            max_length=MAX_HISTORY_TOKENS,
-            truncation=True,
-        )
-        # run inference
-        outputs = self._session.run(None, {"input_ids": inputs["input_ids"].astype("int64")})
+        encoding = self._tokenizer.encode(text, add_special_tokens=False)
+        input_ids = self._np.asarray(encoding.ids, dtype=self._np.int64).reshape(1, -1)
+        outputs = self._session.run(None, {"input_ids": input_ids})
         eou_probability = outputs[0].flatten()[-1]
         end_time = time.perf_counter()
 
@@ -178,14 +177,11 @@ def run(self, data: bytes) -> bytes | None:
 
     @classmethod
     def _download_files(cls) -> None:
-        from transformers import AutoTokenizer
-
-        # ensure the tokenizer is downloaded
-        AutoTokenizer.from_pretrained(HG_MODEL, revision=cls.model_revision())  # type: ignore[no-untyped-call]
-        _download_from_hf_hub(
-            HG_MODEL, ONNX_FILENAME, subfolder="onnx", revision=cls.model_revision()
-        )
-        _download_from_hf_hub(HG_MODEL, "languages.json", revision=cls.model_revision())
+        revision = cls.model_revision()
+        for filename in _TOKENIZER_FILES:
+            _download_from_hf_hub(HG_MODEL, filename, revision=revision)
+        _download_from_hf_hub(HG_MODEL, ONNX_FILENAME, subfolder="onnx", revision=revision)
+        _download_from_hf_hub(HG_MODEL, "languages.json", revision=revision)
 
 
 class EOUPlugin(Plugin):

livekit-plugins/livekit-plugins-turn-detector/pyproject.toml

@@ -23,7 +23,8 @@ classifiers = [
 ]
 dependencies = [
     "livekit-agents>=1.5.8",
-    "transformers>=4.47.1,!=4.57.2,!=4.57.3",  # 4.57.2-4.57.3 have a bug with local_files_only=True (huggingface/transformers#42369)
+    "tokenizers>=0.20,<1",
+    "huggingface-hub>=0.25",
     "numpy>=1.26",
     "onnxruntime>=1.18",
     "jinja2",