diff --git a/docs/models/samples/dummy_llama_2.py b/docs/models/samples/dummy_llama_2.py
new file mode 100644
index 000000000..7304ff41c
--- /dev/null
+++ b/docs/models/samples/dummy_llama_2.py
@@ -0,0 +1,98 @@
+# -*- coding: utf-8 -*-
+"""
+This example demonstrates how to use mgleize/fairseq2-dummy-Llama-3.2-1B model with vLLM.
+This model weight is randomly initialized and is only for testing purposes.
+PS:
+    this model is not supported yet by modelscope, so if you have applied modelscope in your codebase, like:
+    ```bash
+    export VLLM_USE_MODELSCOPE=True
+    ```
+    you should set it to false, make the vllm download model weight from huggingface directly.
+
+
+Requirements:
+- vLLM: v0.11.0 or higher
+- vLLM-metax: v0.11.0 or higher
+- MACA SDK: 3.2.x.x or higher
+"""
+
+import argparse
+import os
+import sys
+from typing import List
+import torchvision
+torchvision.disable_beta_transforms_warning()  # silence annoying warning
+
+from vllm import LLM, SamplingParams
+
+DEFAULT_PROMPTS: List[str] = [
+    "Describe vLLM's core advantages in one paragraph.",
+    "Give me 3 engineering recommendations to improve large language model throughput, and briefly explain why.",
+    "Explain what tensor parallelism is and when to use it.",
+    "Translate this sentence into Chinese: We are using vLLM to perform offline inference validation testing.",
+    "Write a Python snippet (no more than 5 lines) showing how to use vLLM to call a local model to generate text.",
+]
+
+def load_prompts(args) -> List[str]:
+    prompts: List[str] = []
+    if args.prompt:
+        prompts.extend([p.strip() for p in args.prompt if p.strip()])
+    if args.prompt_file:
+        path = args.prompt_file
+        if os.path.isfile(path):
+            with open(path, "r", encoding="utf-8") as f:
+                for line in f:
+                    p = line.strip()
+                    if p:
+                        prompts.append(p)
+        else:
+            print(f"[WARN] Prompt file does not exist: {path}")
+    if not prompts:
+        prompts = DEFAULT_PROMPTS.copy()
+    return prompts
+
+def main():
+    # Add some args to make program configurable from command line
+    parser = argparse.ArgumentParser(description="Minimal offline inference with vLLM (batch prompts, simplified)")
+    parser.add_argument("--model", default="mgleize/fairseq2-dummy-Llama-3.2-1B",
+                        help="Model name or local directory. If not provided, a small demo model is used. You can replace with e.g. meta-llama/Llama-3.2-1B-Instruct")
+    parser.add_argument("-p", "--prompt", action="append",
+                        help="Repeatable; provide multiple prompts by specifying multiple times")
+    parser.add_argument("--prompt-file",
+                        help="Load prompts from a file, one per line")
+    parser.add_argument("--max-tokens", type=int, default=128,
+                        help="Maximum number of generated tokens")
+    parser.add_argument("--temperature", type=float, default=0.2,
+                        help="Sampling temperature")
+    args = parser.parse_args()
+
+    # Inference configures
+    sampling = SamplingParams(
+        temperature=args.temperature,
+        max_tokens=args.max_tokens,
+    )
+    prompts = load_prompts(args)
+    print(f"[INFO] Loading model: {args.model}")
+    llm = LLM(
+        model=args.model,          # online or local
+        dtype="auto",
+        trust_remote_code=True,    # allow custom model code from online repo
+        tensor_parallel_size=1,    # adjustable by number of GPUs
+        gpu_memory_utilization=0.90,
+    )
+
+    print(f"[INFO] Batch generation, total {len(prompts)} prompts")
+    outputs = llm.generate(prompts, sampling)
+
+    print("\n===== OUTPUT =====")
+    for i, (inp, out) in enumerate(zip(prompts, outputs), 1):
+        text = out.outputs[0].text.strip() if out.outputs else ""
+        print(f"\n--- #{i} ---")
+        print(f"Prompt:\n<< {inp}")
+        print(f"Output:\n>> {text}")
+
+    success = any(o.outputs and (o.outputs[0].text.strip() != "") for o in outputs)
+    sys.exit(0 if success else 2)
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file
diff --git a/docs/models/supported_models.md b/docs/models/supported_models.md
index f2974db78..06cbebde5 100644
--- a/docs/models/supported_models.md
+++ b/docs/models/supported_models.md
@@ -45,6 +45,7 @@ Here the plugin would list all the **tested** model on Maca.
 | `Qwen3ForCausalLM` | Qwen3 | `Qwen/Qwen3-8B`, etc. | ✅︎ | ✅︎ |
 | `Qwen3MoeForCausalLM` | Qwen3MoE | `Qwen/Qwen3-30B-A3B`, etc. | ✅︎ | ✅︎ |
 | `Qwen3NextForCausalLM` | Qwen3NextMoE | `Qwen/Qwen3-Next-80B-A3B-Instruct`, etc. | ✅︎ | ✅︎ |
+| `Fairseq2LlamaForCausalLM` | Llama(fairseq2 format) | `mgleize/fairseq2-dummy-Llama-3.2-1B`, etc. | ✅ | ✅ |
 
 ## List of Multimodal Language Models