|
| 1 | +import os |
| 2 | +from itertools import count |
| 3 | +from pathlib import Path |
| 4 | +from threading import Thread |
| 5 | + |
| 6 | +import numpy as np |
| 7 | +from client import TritonClient, UserData |
| 8 | +from transformers import AutoTokenizer |
| 9 | +from utils import download_engine, prepare_grpc_tensor, server_loaded |
| 10 | + |
| 11 | +TRITON_MODEL_REPOSITORY_PATH = Path("/packages/inflight_batcher_llm/") |
| 12 | + |
| 13 | + |
| 14 | +class Model: |
| 15 | + def __init__(self, **kwargs): |
| 16 | + self._data_dir = kwargs["data_dir"] |
| 17 | + self._config = kwargs["config"] |
| 18 | + self._secrets = kwargs["secrets"] |
| 19 | + self._request_id_counter = count(start=1) |
| 20 | + self.triton_client = None |
| 21 | + self.tokenizer = None |
| 22 | + self.uses_openai_api = ( |
| 23 | + "openai-compatible" in self._config["model_metadata"]["tags"] |
| 24 | + ) |
| 25 | + |
| 26 | + def load(self): |
| 27 | + tensor_parallel_count = self._config["model_metadata"].get( |
| 28 | + "tensor_parallelism", 1 |
| 29 | + ) |
| 30 | + pipeline_parallel_count = self._config["model_metadata"].get( |
| 31 | + "pipeline_parallelism", 1 |
| 32 | + ) |
| 33 | + if "hf_access_token" in self._secrets._base_secrets.keys(): |
| 34 | + hf_access_token = self._secrets["hf_access_token"] |
| 35 | + else: |
| 36 | + hf_access_token = None |
| 37 | + is_external_engine_repo = "engine_repository" in self._config["model_metadata"] |
| 38 | + |
| 39 | + # Instantiate TritonClient |
| 40 | + self.triton_client = TritonClient( |
| 41 | + data_dir=self._data_dir, |
| 42 | + model_repository_dir=TRITON_MODEL_REPOSITORY_PATH, |
| 43 | + parallel_count=tensor_parallel_count * pipeline_parallel_count, |
| 44 | + ) |
| 45 | + |
| 46 | + # Download model from Hugging Face Hub if specified |
| 47 | + if is_external_engine_repo: |
| 48 | + if not server_loaded(): |
| 49 | + download_engine( |
| 50 | + engine_repository=self._config["model_metadata"][ |
| 51 | + "engine_repository" |
| 52 | + ], |
| 53 | + fp=self._data_dir, |
| 54 | + auth_token=hf_access_token, |
| 55 | + ) |
| 56 | + |
| 57 | + # Load Triton Server and model |
| 58 | + tokenizer_repository = self._config["model_metadata"]["tokenizer_repository"] |
| 59 | + env = {"triton_tokenizer_repository": tokenizer_repository} |
| 60 | + if hf_access_token is not None: |
| 61 | + env["HUGGING_FACE_HUB_TOKEN"] = hf_access_token |
| 62 | + |
| 63 | + self.triton_client.load_server_and_model(env=env) |
| 64 | + |
| 65 | + # setup eos token |
| 66 | + self.tokenizer = AutoTokenizer.from_pretrained( |
| 67 | + tokenizer_repository, token=hf_access_token |
| 68 | + ) |
| 69 | + self.eos_token_id = self.tokenizer.eos_token_id |
| 70 | + |
| 71 | + def predict(self, model_input): |
| 72 | + user_data = UserData() |
| 73 | + model_name = "ensemble" |
| 74 | + stream_uuid = str(os.getpid()) + str(next(self._request_id_counter)) |
| 75 | + |
| 76 | + if self.uses_openai_api: |
| 77 | + prompt = self.tokenizer.apply_chat_template( |
| 78 | + model_input.get("messages"), |
| 79 | + tokenize=False, |
| 80 | + ) |
| 81 | + else: |
| 82 | + prompt = model_input.get("prompt") |
| 83 | + |
| 84 | + max_tokens = model_input.get("max_tokens", 50) |
| 85 | + beam_width = model_input.get("beam_width", 1) |
| 86 | + bad_words_list = model_input.get("bad_words_list", [""]) |
| 87 | + stop_words_list = model_input.get("stop_words_list", [""]) |
| 88 | + repetition_penalty = model_input.get("repetition_penalty", 1.0) |
| 89 | + ignore_eos = model_input.get("ignore_eos", False) |
| 90 | + stream = model_input.get("stream", True) |
| 91 | + |
| 92 | + input0 = [[prompt]] |
| 93 | + input0_data = np.array(input0).astype(object) |
| 94 | + output0_len = np.ones_like(input0).astype(np.uint32) * max_tokens |
| 95 | + bad_words_list = np.array([bad_words_list], dtype=object) |
| 96 | + stop_words_list = np.array([stop_words_list], dtype=object) |
| 97 | + stream_data = np.array([[stream]], dtype=bool) |
| 98 | + beam_width_data = np.array([[beam_width]], dtype=np.uint32) |
| 99 | + repetition_penalty_data = np.array([[repetition_penalty]], dtype=np.float32) |
| 100 | + |
| 101 | + inputs = [ |
| 102 | + prepare_grpc_tensor("text_input", input0_data), |
| 103 | + prepare_grpc_tensor("max_tokens", output0_len), |
| 104 | + prepare_grpc_tensor("bad_words", bad_words_list), |
| 105 | + prepare_grpc_tensor("stop_words", stop_words_list), |
| 106 | + prepare_grpc_tensor("stream", stream_data), |
| 107 | + prepare_grpc_tensor("beam_width", beam_width_data), |
| 108 | + prepare_grpc_tensor("repetition_penalty", repetition_penalty_data), |
| 109 | + ] |
| 110 | + |
| 111 | + if not ignore_eos: |
| 112 | + end_id_data = np.array([[self.eos_token_id]], dtype=np.uint32) |
| 113 | + inputs.append(prepare_grpc_tensor("end_id", end_id_data)) |
| 114 | + else: |
| 115 | + # do nothing, trt-llm by default doesn't stop on `eos` |
| 116 | + pass |
| 117 | + |
| 118 | + # Start GRPC stream in a separate thread |
| 119 | + stream_thread = Thread( |
| 120 | + target=self.triton_client.start_grpc_stream, |
| 121 | + args=(user_data, model_name, inputs, stream_uuid), |
| 122 | + ) |
| 123 | + stream_thread.start() |
| 124 | + |
| 125 | + def generate(): |
| 126 | + # Yield results from the queue |
| 127 | + for i in TritonClient.stream_predict(user_data): |
| 128 | + yield i |
| 129 | + |
| 130 | + # Clean up GRPC stream and thread |
| 131 | + self.triton_client.stop_grpc_stream(stream_uuid, stream_thread) |
| 132 | + |
| 133 | + if stream: |
| 134 | + return generate() |
| 135 | + else: |
| 136 | + if self.uses_openai_api: |
| 137 | + return "".join(generate()) |
| 138 | + else: |
| 139 | + return {"text": "".join(generate())} |
0 commit comments