diff --git a/.gitignore b/.gitignore
index ff773c6684..ae2966deb2 100644
--- a/.gitignore
+++ b/.gitignore
@@ -179,3 +179,5 @@ cython_debug/
 
 # downloaded model .bin files
 docker/open_llama/*.bin
+
+.tests/
\ No newline at end of file
diff --git a/.gitmodules b/.gitmodules
index f56cca32df..c48980502d 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,3 +1,3 @@
 [submodule "vendor/llama.cpp"]
 	path = vendor/llama.cpp
-	url = https://github.com/ggml-org/llama.cpp.git
+	url = git@github.com:Archie-9527/llama.cpp.git
diff --git a/tests/test_llama_base_chat.py b/tests/test_llama_base_chat.py
new file mode 100644
index 0000000000..4d117a9e18
--- /dev/null
+++ b/tests/test_llama_base_chat.py
@@ -0,0 +1,34 @@
+from llama_cpp import Llama
+
+llm = Llama(
+    model_path="/Users/heart/Code/CPP-Project/llama-agent/models/Qwen3.5-4B-UD-Q8_K_XL.gguf",
+    n_gpu_layers=-1,
+    n_ctx=4096,
+    n_threads=8,
+    use_mmap=True,
+    chat_format="chatml",
+    verbose=False
+)
+
+print("模型加载完成，开始生成...\n")
+
+prompt = "请用一句话解释 Transformer 的注意力机制。"
+
+# 1) 显式看 tokenize
+toks = llm.tokenize(prompt.encode("utf-8"), add_bos=True, special=False)
+print("TOKENS_IN:", toks[:20], "len=", len(toks))
+
+# 2) 生成（流式）看 token 输出过程
+out = []
+for chunk in llm.create_completion(
+    prompt=prompt,
+    max_tokens=64,
+    temperature=0.7,
+    top_p=0.9,
+    stream=True,
+):
+    text = chunk["choices"][0]["text"]
+    out.append(text)
+    print(text, end="", flush=True)
+
+print("\n\nFINAL:", "".join(out))
\ No newline at end of file