|
| 1 | +# Copyright (c) Microsoft Corporation. All rights reserved. |
| 2 | +# Licensed under the MIT License. |
| 3 | + |
| 4 | +import argparse |
| 5 | +import json |
| 6 | +import random |
| 7 | +import threading |
| 8 | +import time |
| 9 | + |
| 10 | +import onnxruntime_genai as og |
| 11 | +import tqdm |
| 12 | +from datasets import load_dataset |
| 13 | + |
| 14 | + |
| 15 | +def get_random_prompts(num_questions: int, split="validation") -> list[str]: |
| 16 | + dataset = load_dataset("squad_v2", split=split) |
| 17 | + questions = [item["question"] for item in dataset] |
| 18 | + return random.sample(questions, min(num_questions, len(questions))) |
| 19 | + |
| 20 | + |
| 21 | +class ClientRequest: |
| 22 | + def __init__( |
| 23 | + self, prompt: str, model: og.Model, tokenizer: og.Tokenizer, opaque_data: any |
| 24 | + ): |
| 25 | + self.prompt = prompt |
| 26 | + self.params = og.GeneratorParams(model) |
| 27 | + self.params.set_search_options( |
| 28 | + do_sample=False, |
| 29 | + max_length=256, |
| 30 | + ) |
| 31 | + |
| 32 | + messages = [ |
| 33 | + {"role": "system", "content": ""}, |
| 34 | + {"role": "user", "content": f"{prompt}"}, |
| 35 | + ] |
| 36 | + messages = json.dumps(messages) |
| 37 | + |
| 38 | + self.request = og.Request(self.params) |
| 39 | + self.request.add_tokens( |
| 40 | + tokenizer.encode( |
| 41 | + tokenizer.apply_chat_template( |
| 42 | + messages=messages, add_generation_prompt=True |
| 43 | + ) |
| 44 | + ) |
| 45 | + ) |
| 46 | + self.request.set_opaque_data(opaque_data) |
| 47 | + self.streaming_tokenizer = tokenizer.create_stream() |
| 48 | + self.token_stream = "" |
| 49 | + |
| 50 | + |
| 51 | +class RequestPool: |
| 52 | + def __init__( |
| 53 | + self, |
| 54 | + model: og.Model, |
| 55 | + tokenizer: og.Tokenizer, |
| 56 | + engine: og.Engine, |
| 57 | + num_requests: int, |
| 58 | + load_factor: float = 0.2, |
| 59 | + debug: bool = False, |
| 60 | + ): |
| 61 | + self.model = model |
| 62 | + self.tokenizer = tokenizer |
| 63 | + self.engine = engine |
| 64 | + self.num_requests = num_requests |
| 65 | + self.requests: list[ClientRequest] = [] |
| 66 | + self.prompts = get_random_prompts(num_requests) |
| 67 | + self.load_factor = load_factor |
| 68 | + self.lock = threading.Lock() |
| 69 | + self.bar = tqdm.tqdm(total=len(self.prompts)) |
| 70 | + self.debug = debug |
| 71 | + |
| 72 | + # Add load_factor * num_requests requests to the engine |
| 73 | + for prompt in self.prompts[: int(num_requests * load_factor)]: |
| 74 | + request = ClientRequest(prompt, model, tokenizer, self) |
| 75 | + self.requests.append(request) |
| 76 | + self.engine.add_request(request.request) |
| 77 | + |
| 78 | + def fill(self): |
| 79 | + for i, prompt in enumerate( |
| 80 | + self.prompts[int(len(self.prompts) * self.load_factor) :] |
| 81 | + ): |
| 82 | + request = ClientRequest(prompt, self.model, self.tokenizer, self) |
| 83 | + with self.lock: |
| 84 | + self.requests.append(request) |
| 85 | + self.engine.add_request(request.request) |
| 86 | + time.sleep(1) # Simulate some delay in request generation |
| 87 | + |
| 88 | + def drain(self, request: og.Request): |
| 89 | + with self.lock: |
| 90 | + client_request = next( |
| 91 | + (r for r in self.requests if r.request == request), None |
| 92 | + ) |
| 93 | + while request.has_unseen_tokens(): |
| 94 | + token = request.get_unseen_token() |
| 95 | + client_request.token_stream += ( |
| 96 | + client_request.streaming_tokenizer.decode(token) |
| 97 | + ) |
| 98 | + |
| 99 | + if request.is_done(): |
| 100 | + assert ( |
| 101 | + client_request is not None |
| 102 | + ), "Client request not found in the pool" |
| 103 | + |
| 104 | + if self.debug: |
| 105 | + print(f"🫵 : {client_request.prompt}") |
| 106 | + print(f"🤖 : {client_request.token_stream}") |
| 107 | + self.engine.remove_request(request) |
| 108 | + self.requests.remove(client_request) |
| 109 | + self.bar.update(1) |
| 110 | + |
| 111 | + |
| 112 | +class Engine: |
| 113 | + def __init__(self, model_path: str, execution_provider: str, debug: bool): |
| 114 | + self.config = og.Config(model_path) |
| 115 | + self.config.clear_providers() |
| 116 | + if execution_provider != "cpu": |
| 117 | + self.config.append_provider(execution_provider) |
| 118 | + self.model = og.Model(self.config) |
| 119 | + self.tokenizer = og.Tokenizer(self.model) |
| 120 | + self.engine = og.Engine(self.model) |
| 121 | + self.debug = debug |
| 122 | + self.tokens_decoded = 0 |
| 123 | + |
| 124 | + def run(self): |
| 125 | + while request := self.engine.step(): |
| 126 | + request_pool = request.get_opaque_data() |
| 127 | + request_pool.drain(request) |
| 128 | + self.tokens_decoded += 1 |
| 129 | + |
| 130 | + |
| 131 | +def run(args: argparse.Namespace): |
| 132 | + engine = Engine(args.model_path, args.execution_provider, args.debug) |
| 133 | + request_pool = RequestPool( |
| 134 | + engine.model, |
| 135 | + engine.tokenizer, |
| 136 | + engine.engine, |
| 137 | + args.num_requests, |
| 138 | + debug=args.debug, |
| 139 | + ) |
| 140 | + |
| 141 | + producer_thread = threading.Thread(target=request_pool.fill) |
| 142 | + producer_thread.start() |
| 143 | + |
| 144 | + start = time.time() |
| 145 | + engine.run() |
| 146 | + end = time.time() |
| 147 | + |
| 148 | + request_pool.bar.close() |
| 149 | + print(f"⌛Tokens per second: {engine.tokens_decoded / (end - start):.2f}") |
| 150 | + |
| 151 | + |
| 152 | +if __name__ == "__main__": |
| 153 | + parser = argparse.ArgumentParser( |
| 154 | + description="End-to-end AI Question/Answer example for gen-ai", |
| 155 | + ) |
| 156 | + parser.add_argument( |
| 157 | + "-m", |
| 158 | + "--model_path", |
| 159 | + type=str, |
| 160 | + required=True, |
| 161 | + help="Onnx model folder path (must contain genai_config.json and model.onnx)", |
| 162 | + ) |
| 163 | + parser.add_argument( |
| 164 | + "-e", |
| 165 | + "--execution_provider", |
| 166 | + type=str, |
| 167 | + required=True, |
| 168 | + choices=["cpu", "cuda", "dml", "webgpu"], |
| 169 | + help="Execution provider to run ONNX model with", |
| 170 | + ) |
| 171 | + parser.add_argument( |
| 172 | + "-d", |
| 173 | + "--debug", |
| 174 | + action="store_true", |
| 175 | + help="Enable debug logging", |
| 176 | + ) |
| 177 | + parser.add_argument( |
| 178 | + "-n", |
| 179 | + "--num_requests", |
| 180 | + type=int, |
| 181 | + default=1, |
| 182 | + help="Number of requests to process in the pool", |
| 183 | + ) |
| 184 | + args = parser.parse_args() |
| 185 | + |
| 186 | + run(args) |
0 commit comments