From 8d16fdb71c155e81d87a557e74fb2eb9c8776d30 Mon Sep 17 00:00:00 2001
From: gufengc <gufeng@gradient.network>
Date: Tue, 18 Nov 2025 22:08:43 +0800
Subject: [PATCH] fix(tokenizer): fix tokenizer bug

---
 src/parallax/server/executor.py       | 1 +
 src/parallax/sglang/model_runner.py   | 1 +
 src/parallax/utils/tokenizer_utils.py | 2 +-
 3 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/parallax/server/executor.py b/src/parallax/server/executor.py
index d2cda590..08dda6f6 100644
--- a/src/parallax/server/executor.py
+++ b/src/parallax/server/executor.py
@@ -858,6 +858,7 @@ def _handle_raw_request(self, raw_request: Dict):
             prompt = self.tokenizer.apply_chat_template(
                 messages,
                 raw_request.get("tools") or None,
+                tokenize=True,
                 add_generation_prompt=True,
                 **chat_template_kwargs,
             )
diff --git a/src/parallax/sglang/model_runner.py b/src/parallax/sglang/model_runner.py
index d1624a13..818e251d 100755
--- a/src/parallax/sglang/model_runner.py
+++ b/src/parallax/sglang/model_runner.py
@@ -217,6 +217,7 @@ def form_sgl_server_args(
         mem_fraction_static=0.85,
         moe_runner_backend=moe_runner_backend,
         tp_size=tp_size,
+        trust_remote_code=True,
     )
     return sgl_server_args
 
diff --git a/src/parallax/utils/tokenizer_utils.py b/src/parallax/utils/tokenizer_utils.py
index 63bee300..e2ea3216 100755
--- a/src/parallax/utils/tokenizer_utils.py
+++ b/src/parallax/utils/tokenizer_utils.py
@@ -75,7 +75,7 @@ def load_detokenizer(model_path, tokenizer):
     Note, to use a fast streaming tokenizer, pass a local file path rather than
     a Hugging Face repo ID.
     """
-    detokenizer_class = NaiveStreamingDetokenizer
+    detokenizer_class = ParallaxNaiveStreamingDetokenizer
     tokenmap = None
 
     tokenizer_file = model_path / "tokenizer.json"