Run LLMxCPG-Q by vllm

queries/README.md

@@ -33,6 +33,29 @@ python generate_and_run_queries.py -d /path/to/your/dataset.json -o /path/to/you
 
 The script offers several optional arguments to customize its behavior.  For example, you can adjust the number of worker threads (-n), specify a different Docker Compose file for Joern (-c), or change the port used for the LLM (--llm-port).  To see a full list of available arguments and their descriptions, run the script with the -h or --help flag:
 
+### Run LLMxCPG-Q by vLLM
+
+You can launch a local OpenAI-compatible endpoint with the helper script:
+
+```bash
+python run_vllm_server.py --model-id QCRI/LLMxCPG-Q --port 9001
+```
+
+Then run query generation with the local endpoint:
+
+```bash
+python generate_and_run_queries.py \
+	-d /path/to/your/dataset.json \
+	-o /path/to/your/output_dir \
+	--llm-model-type vLLM \
+	--llm-model-name LLMxCPG-Q \
+	--llm-port 9001
+```
+
+Notes:
+- Need [vllm](https://github.com/vllm-project/vllm) : uv pip install vllm --torch-backend=auto
+- Endpoint URL is `http://127.0.0.1:9001/v1/chat/completions` by default.
+
 ### Generate queries for your custom vulnerability dataset
 
 To be able to use our script `generate_and_run_queries.py`, please follow these steps:

queries/run_vllm_server.py

@@ -0,0 +1,149 @@
+import argparse
+import shlex
+import subprocess
+import sys
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Launch a local OpenAI-compatible vLLM server for LLMxCPG-Q."
+    )
+
+    mode = parser.add_mutually_exclusive_group(required=True)
+    mode.add_argument(
+        "--model-id",
+        type=str,
+        help="HF/local model path when using a merged/full model.",
+    )
+    mode.add_argument(
+        "--lora-adapter",
+        type=str,
+        help="HF/local LoRA adapter path for LLMxCPG-Q.",
+    )
+    parser.add_argument(
+        "--base-model",
+        type=str,
+        default="Qwen/Qwen2.5-Coder-32B-Instruct",
+        help="Base model used with --lora-adapter.",
+    )
+    parser.add_argument(
+        "--served-model-name",
+        type=str,
+        default="LLMxCPG-Q",
+        help="Model name exposed by the OpenAI API server.",
+    )
+    parser.add_argument("--host", type=str, default="127.0.0.1", help="Server host.")
+    parser.add_argument("--port", type=int, default=9001, help="Server port.")
+    parser.add_argument("--dtype", type=str, default="auto", help="vLLM dtype.")
+    parser.add_argument(
+        "--max-model-len",
+        type=int,
+        default=32768,
+        help="Maximum context length for vLLM.",
+    )
+    parser.add_argument(
+        "--tensor-parallel-size",
+        type=int,
+        default=1,
+        help="Tensor parallel size for multi-GPU runs.",
+    )
+    parser.add_argument(
+        "--gpu-memory-utilization",
+        type=float,
+        default=0.9,
+        help="Target GPU memory utilization for vLLM.",
+    )
+    parser.add_argument(
+        "--max-lora-rank",
+        type=int,
+        default=64,
+        help="Maximum LoRA rank when using --lora-adapter.",
+    )
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Pass --trust-remote-code to vLLM.",
+    )
+    parser.add_argument(
+        "--extra-vllm-args",
+        type=str,
+        default="",
+        help="Extra raw arguments appended to the vLLM command.",
+    )
+
+    args = parser.parse_args()
+
+    return args
+
+
+def build_command(args: argparse.Namespace) -> list[str]:
+    cmd = [
+        sys.executable,
+        "-m",
+        "vllm.entrypoints.openai.api_server",
+        "--host",
+        args.host,
+        "--port",
+        str(args.port),
+        "--served-model-name",
+        args.served_model_name,
+        "--dtype",
+        args.dtype,
+        "--max-model-len",
+        str(args.max_model_len),
+        "--tensor-parallel-size",
+        str(args.tensor_parallel_size),
+        "--gpu-memory-utilization",
+        str(args.gpu_memory_utilization),
+    ]
+
+    if args.trust_remote_code:
+        cmd.append("--trust-remote-code")
+
+    if args.lora_adapter:
+        cmd.extend(
+            [
+                "--model",
+                args.base_model,
+                "--enable-lora",
+                "--max-lora-rank",
+                str(args.max_lora_rank),
+                "--lora-modules",
+                f"{args.served_model_name}={args.lora_adapter}",
+            ]
+        )
+    else:
+        cmd.extend(["--model", args.model_id])
+
+    if args.extra_vllm_args.strip():
+        cmd.extend(shlex.split(args.extra_vllm_args))
+
+    return cmd
+
+
+def main() -> None:
+    args = parse_args()
+    cmd = build_command(args)
+
+    print("Starting local LLMxCPG-Q endpoint...")
+    print("Endpoint:", f"http://{args.host}:{args.port}/v1/chat/completions")
+    print("Model name for --llm-endpoint:", args.served_model_name)
+    print()
+    print("Run generate_and_run_queries.py with:")
+    print(
+        "python generate_and_run_queries.py "
+        "-d /path/to/dataset.json "
+        "-o /path/to/output_dir "
+        "--llm-model-type vLLM "
+        f"--llm-endpoint {args.served_model_name} "
+        f"--llm-port {args.port}"
+    )
+    print()
+    print("vLLM command:")
+    print(" ".join(shlex.quote(part) for part in cmd))
+
+    subprocess.run(cmd, check=True)
+
+
+if __name__ == "__main__":
+    main()