|
| 1 | +import contextlib |
| 2 | +import csv |
| 3 | +import json |
| 4 | +import os |
| 5 | +import random |
| 6 | +import re |
| 7 | +import time |
| 8 | +from dataclasses import asdict |
| 9 | + |
| 10 | +from tqdm import tqdm |
| 11 | +from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Vector |
| 12 | + |
| 13 | +random.seed(0) |
| 14 | + |
| 15 | +import sys |
| 16 | + |
| 17 | +from transformers import AutoTokenizer |
| 18 | +from vllm import LLM, SamplingParams |
| 19 | +from vllm.config import KVTransferConfig |
| 20 | +from vllm.engine.arg_utils import EngineArgs |
| 21 | +from vllm.inputs import TokensPrompt |
| 22 | + |
| 23 | +from ucm.logger import init_logger |
| 24 | + |
| 25 | +logger = init_logger(__name__) |
| 26 | + |
| 27 | +model = "" |
| 28 | +data_dir = "" |
| 29 | +path_to_dataset = "" |
| 30 | +tokenizer = None |
| 31 | +# 28705 is the token id for <space> char in llama model |
| 32 | +# 151643 is the pad token id in qwen model |
| 33 | +chunk_end_token_id = -1 |
| 34 | +chunk_pad_token_id = -1 |
| 35 | +block_size = 64 |
| 36 | + |
| 37 | + |
| 38 | +def setup_environment_variables(): |
| 39 | + os.environ["VLLM_USE_V1"] = "1" |
| 40 | + os.environ["PYTHONHASHSEED"] = "123456" |
| 41 | + |
| 42 | + global model, data_dir, path_to_dataset, tokenizer, chunk_end_token_id, chunk_pad_token_id |
| 43 | + model = os.getenv("MODEL_PATH", "/home/models/mistralai/Mistral-7B-Instruct-v0.2") |
| 44 | + if not os.path.isdir(model): |
| 45 | + model = input( |
| 46 | + "Enter path to model, e.g./home/models/mistralai/Mistral-7B-Instruct-v0.2: " |
| 47 | + ) |
| 48 | + if not os.path.isdir(model): |
| 49 | + print("Exiting. Incorrect model_path") |
| 50 | + sys.exit(1) |
| 51 | + |
| 52 | + data_dir = os.getenv("DATA_DIR", "/home/data/kv_cache") |
| 53 | + if not os.path.isdir(data_dir): |
| 54 | + data_dir = input( |
| 55 | + "Enter the directory for UCMStore to save kv cache, e.g. /home/data/kv_cache: " |
| 56 | + ) |
| 57 | + create = input(f"Directory {data_dir} dose not exist. Create it? (Y/n): ") |
| 58 | + if create.lower() == "y": |
| 59 | + os.makedirs(data_dir, exist_ok=True) |
| 60 | + else: |
| 61 | + print("Exiting. Directory not created.") |
| 62 | + sys.exit(1) |
| 63 | + |
| 64 | + # now support wikimqa |
| 65 | + path_to_dataset = os.getenv( |
| 66 | + "BLEND_DATASET_PATH", "/home/data/Longbench/data/2wikimqa.jsonl" |
| 67 | + ) |
| 68 | + if not os.path.isfile(path_to_dataset): |
| 69 | + path_to_dataset = input( |
| 70 | + "Enter path of one of 2wikimqa dataset in longbench, e.g. /home/data/Longbench/data/2wikimqa.jsonl: " |
| 71 | + ) |
| 72 | + if not os.path.isfile(path_to_dataset): |
| 73 | + print("Exiting. Incorrect dataset path") |
| 74 | + sys.exit(1) |
| 75 | + |
| 76 | + tokenizer = AutoTokenizer.from_pretrained(model, use_chat_template=True) |
| 77 | + # as for Qwen model, use pad_token_id for padding block |
| 78 | + # as for Llama model, current use unk_token for padding block |
| 79 | + chunk_pad_token_id = tokenizer.encode("▁", add_special_tokens=False)[0] |
| 80 | + chunk_end_token_id = chunk_pad_token_id |
| 81 | + |
| 82 | + if tokenizer.pad_token_id is not None: |
| 83 | + chunk_pad_token_id = tokenizer.pad_token_id |
| 84 | + chunk_end_token_id = tokenizer.pad_token_id |
| 85 | + |
| 86 | + |
| 87 | +@contextlib.contextmanager |
| 88 | +def build_llm_with_uc(module_path: str, name: str, model: str): |
| 89 | + ktc = KVTransferConfig( |
| 90 | + kv_connector=name, |
| 91 | + kv_connector_module_path=module_path, |
| 92 | + kv_role="kv_both", |
| 93 | + kv_connector_extra_config={ |
| 94 | + "ucm_connectors": [ |
| 95 | + { |
| 96 | + "ucm_connector_name": "UcmNfsStore", |
| 97 | + "ucm_connector_config": { |
| 98 | + "storage_backends": data_dir, |
| 99 | + "kv_block_size": 33554432, |
| 100 | + }, |
| 101 | + } |
| 102 | + ], |
| 103 | + "load_only_first_rank": False, |
| 104 | + "ucm_sparse_config": { |
| 105 | + "Blend": { |
| 106 | + "chunk_end_token_id": chunk_end_token_id, |
| 107 | + "compute_meta": { |
| 108 | + "model.layers.1.self_attn.attn": { |
| 109 | + "ratio": 0.2, |
| 110 | + }, |
| 111 | + }, |
| 112 | + } |
| 113 | + }, |
| 114 | + "use_layerwise": True, |
| 115 | + }, |
| 116 | + ) |
| 117 | + |
| 118 | + llm_args = EngineArgs( |
| 119 | + model=model, |
| 120 | + enforce_eager=True, |
| 121 | + kv_transfer_config=ktc, |
| 122 | + max_model_len=16384 * 2, |
| 123 | + max_num_batched_tokens=16384 * 2, |
| 124 | + gpu_memory_utilization=0.8, |
| 125 | + block_size=block_size, |
| 126 | + enable_prefix_caching=False, |
| 127 | + distributed_executor_backend="mp", |
| 128 | + tensor_parallel_size=1, |
| 129 | + trust_remote_code=True, |
| 130 | + ) |
| 131 | + |
| 132 | + llm = LLM(**asdict(llm_args)) |
| 133 | + try: |
| 134 | + yield llm |
| 135 | + finally: |
| 136 | + logger.info("LLM engine is exiting.") |
| 137 | + |
| 138 | + |
| 139 | +def get_output( |
| 140 | + llm: LLM, |
| 141 | + prompt, |
| 142 | + sampling_params: SamplingParams, |
| 143 | +): |
| 144 | + start = time.time() |
| 145 | + outputs = llm.generate(prompt, sampling_params) |
| 146 | + print("-" * 50) |
| 147 | + generated_text = None |
| 148 | + for output in outputs: |
| 149 | + generated_text = output.outputs[0].text |
| 150 | + e2e_time = time.time() - start |
| 151 | + print("-" * 50) |
| 152 | + return e2e_time, generated_text |
| 153 | + |
| 154 | + |
| 155 | +def pad_rag_chunks(token_ids, block_size, pad_id, end_id): |
| 156 | + """ |
| 157 | + pad token_ids with pad_id and end up with end_id |
| 158 | + """ |
| 159 | + # assert pad_id != end_id |
| 160 | + remainder = len(token_ids) % block_size |
| 161 | + |
| 162 | + if remainder == 0 and token_ids[-1] in [pad_id, end_id]: |
| 163 | + # no need to pad |
| 164 | + token_ids[-1] = end_id |
| 165 | + return token_ids |
| 166 | + |
| 167 | + pad_len = block_size - remainder - 1 |
| 168 | + padded = token_ids + [pad_id] * pad_len + [end_id] |
| 169 | + return padded |
| 170 | + |
| 171 | + |
| 172 | +systemPrompt = "Answer the question based on the given passages. Only give me the answer and do not output any other words.\n\nThe following are given passages.\n" |
| 173 | + |
| 174 | + |
| 175 | +def main(): |
| 176 | + module_path = "ucm.integration.vllm.blend_connector" |
| 177 | + name = "UCMBlendConnector" |
| 178 | + |
| 179 | + setup_environment_variables() |
| 180 | + |
| 181 | + with build_llm_with_uc(module_path, name, model) as llm: |
| 182 | + prefill_sampling_params = SamplingParams( |
| 183 | + temperature=0.0, top_p=0.95, max_tokens=1 |
| 184 | + ) |
| 185 | + sampling_params = SamplingParams(temperature=0, top_p=0.95, max_tokens=128) |
| 186 | + # choose one data row in LongBenchV1 (wikimqa) |
| 187 | + assert os.path.isfile( |
| 188 | + path_to_dataset |
| 189 | + ), f"Incorrect dataset path. Please specify the dataset path by `export DATASET_PATH=/path/to/longbench/multifieldqa_zh.jsonl`" |
| 190 | + with open(path_to_dataset, "r") as f: |
| 191 | + lines = f.readlines() |
| 192 | + dataset_row = json.loads(lines[0]) |
| 193 | + |
| 194 | + passages = re.findall( |
| 195 | + r"Passage\s+(\d+):(.*?)(?=Passage\s+\d+:|$)", dataset_row["context"], re.S |
| 196 | + ) |
| 197 | + chunks = [f"Passage {i}:{passages[i][1]}" for i in range(len(passages))] |
| 198 | + question = f"\n\nAnswer the question based on the given passages. Answer the question within 5 words. Do NOT repeat the question or output any other words. Question: {dataset_row["input"]}\nAnswer:" |
| 199 | + origin_sys_prompt_ids = tokenizer.encode(systemPrompt) |
| 200 | + padded_sys_prompt_ids = pad_rag_chunks( |
| 201 | + origin_sys_prompt_ids, block_size, chunk_pad_token_id, chunk_end_token_id |
| 202 | + ) |
| 203 | + # 1. sys prompt warm up |
| 204 | + print(f"---------------1. sys prompt: warm up---------------") |
| 205 | + get_output( |
| 206 | + llm, |
| 207 | + TokensPrompt(prompt_token_ids=padded_sys_prompt_ids), |
| 208 | + prefill_sampling_params, |
| 209 | + ) |
| 210 | + time.sleep(0.5) |
| 211 | + |
| 212 | + padded_contexts_ids = [] |
| 213 | + padded_prompt_ids = padded_sys_prompt_ids |
| 214 | + origin_prompt_ids = origin_sys_prompt_ids |
| 215 | + for text_chunk in chunks: |
| 216 | + un_pad_ids = tokenizer.encode(text_chunk, add_special_tokens=False) |
| 217 | + padded_ids = pad_rag_chunks( |
| 218 | + un_pad_ids, block_size, chunk_pad_token_id, chunk_end_token_id |
| 219 | + ) |
| 220 | + padded_prompt_ids = padded_prompt_ids + padded_ids |
| 221 | + origin_prompt_ids = origin_prompt_ids + un_pad_ids |
| 222 | + padded_contexts_ids.append(padded_ids) |
| 223 | + |
| 224 | + question_ids = tokenizer.encode(question, add_special_tokens=False) |
| 225 | + padded_prompt_ids = padded_prompt_ids + question_ids |
| 226 | + origin_prompt_ids = origin_prompt_ids + question_ids |
| 227 | + |
| 228 | + print(f"--------------- baseline with no cache blend ---------------") |
| 229 | + baseline_time, baseline_gen_text = get_output( |
| 230 | + llm, TokensPrompt(prompt_token_ids=origin_prompt_ids), sampling_params |
| 231 | + ) |
| 232 | + time.sleep(0.5) |
| 233 | + |
| 234 | + print(f"--------------- cache rag chunks ---------------") |
| 235 | + llm.generate( |
| 236 | + [TokensPrompt(prompt_token_ids=ids) for ids in padded_contexts_ids], |
| 237 | + sampling_params, |
| 238 | + ) |
| 239 | + time.sleep(0.5) |
| 240 | + |
| 241 | + print(f"--------------- warm up blend code ---------------") |
| 242 | + warm_up_blend_prompt_ids = padded_sys_prompt_ids |
| 243 | + for ids in reversed(padded_contexts_ids): |
| 244 | + warm_up_blend_prompt_ids = warm_up_blend_prompt_ids + ids |
| 245 | + warm_up_blend_prompt_ids = warm_up_blend_prompt_ids + question_ids |
| 246 | + llm.generate( |
| 247 | + TokensPrompt(prompt_token_ids=warm_up_blend_prompt_ids), sampling_params |
| 248 | + ) |
| 249 | + time.sleep(0.5) |
| 250 | + |
| 251 | + print(f"--------------- cache blend ---------------") |
| 252 | + blend_time, blend_gen_text = get_output( |
| 253 | + llm, TokensPrompt(prompt_token_ids=padded_prompt_ids), sampling_params |
| 254 | + ) |
| 255 | + time.sleep(0.5) |
| 256 | + |
| 257 | + print(f"--------------- prefix cache ---------------") |
| 258 | + pc_time, pc_gen_text = get_output( |
| 259 | + llm, TokensPrompt(prompt_token_ids=origin_prompt_ids), sampling_params |
| 260 | + ) |
| 261 | + |
| 262 | + print(f"Baseline generated text: {baseline_gen_text!r}") |
| 263 | + print(f"Baseline generated cost time: {baseline_time:.2f} seconds") |
| 264 | + print(f"Blend generated text: {blend_gen_text!r}") |
| 265 | + print(f"Blend generated cost time: {blend_time:.2f} seconds") |
| 266 | + print(f"Prefix Cache generated text: {pc_gen_text!r}") |
| 267 | + print(f"Prefix Cache generated cost time: {pc_time:.2f} seconds") |
| 268 | + print(f"Question:{dataset_row['input']}") |
| 269 | + print(f"Golden answer:{dataset_row["answers"]}") |
| 270 | + |
| 271 | + |
| 272 | +if __name__ == "__main__": |
| 273 | + main() |
0 commit comments