API Client

Source vllm-project/vllm.

  1.  1"""Example Python client for `vllm.entrypoints.api_server`
  2.  2NOTE: The API server is used only for demonstration and simple performance
  3.  3benchmarks. It is not intended for production use.
  4.  4For production use, we recommend `vllm serve` and the OpenAI client API.
  5.  5"""
  6.  6
  7.  7import argparse
  8.  8import json
  9.  9from typing import Iterable, List
  10. 10
  11. 11import requests
  12. 12
  13. 13
  14. 14def clear_line(n: int = 1) -> None:
  15. 15    LINE_UP = '\033[1A'
  16. 16    LINE_CLEAR = '\x1b[2K'
  17. 17    for _ in range(n):
  18. 18        print(LINE_UP, end=LINE_CLEAR, flush=True)
  19. 19
  20. 20
  21. 21def post_http_request(prompt: str,
  22. 22                      api_url: str,
  23. 23                      n: int = 1,
  24. 24                      stream: bool = False) -> requests.Response:
  25. 25    headers = {"User-Agent": "Test Client"}
  26. 26    pload = {
  27. 27        "prompt": prompt,
  28. 28        "n": n,
  29. 29        "use_beam_search": True,
  30. 30        "temperature": 0.0,
  31. 31        "max_tokens": 16,
  32. 32        "stream": stream,
  33. 33    }
  34. 34    response = requests.post(api_url, headers=headers, json=pload, stream=True)
  35. 35    return response
  36. 36
  37. 37
  38. 38def get_streaming_response(response: requests.Response) -> Iterable[List[str]]:
  39. 39    for chunk in response.iter_lines(chunk_size=8192,
  40. 40                                     decode_unicode=False,
  41. 41                                     delimiter=b"\0"):
  42. 42        if chunk:
  43. 43            data = json.loads(chunk.decode("utf-8"))
  44. 44            output = data["text"]
  45. 45            yield output
  46. 46
  47. 47
  48. 48def get_response(response: requests.Response) -> List[str]:
  49. 49    data = json.loads(response.content)
  50. 50    output = data["text"]
  51. 51    return output
  52. 52
  53. 53
  54. 54if __name__ == "__main__":
  55. 55    parser = argparse.ArgumentParser()
  56. 56    parser.add_argument("--host", type=str, default="localhost")
  57. 57    parser.add_argument("--port", type=int, default=8000)
  58. 58    parser.add_argument("--n", type=int, default=4)
  59. 59    parser.add_argument("--prompt", type=str, default="San Francisco is a")
  60. 60    parser.add_argument("--stream", action="store_true")
  61. 61    args = parser.parse_args()
  62. 62    prompt = args.prompt
  63. 63    api_url = f"http://{args.host}:{args.port}/generate"
  64. 64    n = args.n
  65. 65    stream = args.stream
  66. 66
  67. 67    print(f"Prompt: {prompt!r}\n", flush=True)
  68. 68    response = post_http_request(prompt, api_url, n, stream)
  69. 69
  70. 70    if stream:
  71. 71        num_printed_lines = 0
  72. 72        for h in get_streaming_response(response):
  73. 73            clear_line(num_printed_lines)
  74. 74            num_printed_lines = 0
  75. 75            for i, line in enumerate(h):
  76. 76                num_printed_lines += 1
  77. 77                print(f"Beam candidate {i}: {line!r}", flush=True)
  78. 78    else:
  79. 79        output = get_response(response)
  80. 80        for i, line in enumerate(output):
  81. 81            print(f"Beam candidate {i}: {line!r}", flush=True)