File size: 18,456 Bytes
72c0672 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 | import json
import base64
import argparse
import os
import sys
import gzip
import time
import math
import gc
import torch
import torch.multiprocessing as mp
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
from tqdm import tqdm
from typing import List, Dict, Any, Callable, Tuple, Optional
import Levenshtein
from collections import defaultdict
# ==========================================
# 0. 系统路径与环境修复
# ==========================================
current_dir = os.getcwd()
if current_dir not in sys.path:
sys.path.append(current_dir)
script_dir = os.path.dirname(os.path.abspath(__file__))
if script_dir not in sys.path:
sys.path.append(script_dir)
print(f"🔧 System Path Fixed. CWD: {current_dir}")
# ==========================================
# 1. 依赖检查
# ==========================================
try:
from transformers import AutoTokenizer
except ImportError:
print("❌ Error: 'transformers' not installed.")
sys.exit(1)
try:
from m1_compression import utils
from m1_compression.compressor import (
load_m1_model_and_tokenizer,
ALPHABET_SIZE,
ARITHMETIC_CODER_BASE,
ARITHMETIC_CODER_PRECISION
)
from m1_compression.hybrid_arithmetic_coder import CPUArithmeticEncoder
from m1_compression.batched_arithmetic_coder import _pdf_to_cdf
print("✅ Successfully imported m1_compression modules.")
except ImportError as e:
print(f"❌ FATAL ERROR: {e}")
sys.exit(1)
# ==========================================
# 2. 辅助函数
# ==========================================
def vread(buf: bytes, i: int):
shift = val = 0
while True:
b = buf[i]
i += 1
val |= (b & 0x7F) << shift
if b < 0x80: return val, i
shift += 7
def unpack_windows(input_bytes: bytes, b64_stream: str) -> List[Tuple[bytes, int]]:
try:
if not b64_stream: return []
buf, i, cursor, byte_windows = base64.b64decode(b64_stream), 0, 0, []
while i < len(buf):
gap, i = vread(buf, i)
size, i = vread(buf, i)
start = cursor + gap
if gap > 0: byte_windows.append((input_bytes[cursor:start], 0))
end = start + size
byte_windows.append((input_bytes[start:end], 1))
cursor = end
if cursor < len(input_bytes): byte_windows.append((input_bytes[cursor:], 0))
return byte_windows
except (base64.binascii.Error, IndexError): return []
def list_to_comparable_str(int_list: List[int]) -> str:
return "".join([chr(min(x, 0x10FFFF)) for x in int_list])
def pad_batch(batch: List[bytes]):
batch_tensors = [torch.tensor(list(data), dtype=torch.int64) for data in batch]
lengths = torch.tensor([len(data) for data in batch], dtype=torch.int64)
padded_batch = torch.nn.utils.rnn.pad_sequence(
batch_tensors,
batch_first=True,
padding_value=0
)
return padded_batch, lengths
# ==========================================
# 3. 核心压缩逻辑 (AC)
# ==========================================
def batched_m1_compress_predict_fn(model):
def predict_fn(input_tensor: torch.Tensor, **kwargs) -> torch.Tensor:
if input_tensor.dim() == 1: input_tensor = input_tensor.unsqueeze(0)
with torch.no_grad():
logits = model(input_tensor, **kwargs)
logits = logits[..., :256].float()
probs = torch.softmax(logits, dim=-1)
return probs
return predict_fn
def compress_segments_ac_impl(
sorted_segments: List[bytes],
batched_predict_fn: Callable,
first_byte_prob: torch.Tensor,
device: torch.device
) -> List[List[int]]:
"""
底层批处理函数:接收一大堆 segments,分批送入 GPU 计算,再用 CPU 编码
"""
M = len(sorted_segments)
if M == 0: return []
# 去重
segment_to_indices = defaultdict(list)
for i, seg in enumerate(sorted_segments):
segment_to_indices[seg].append(i)
unique_segments = [seg for seg in segment_to_indices.keys() if len(seg) > 0]
segment_to_compressed = {}
encoder = CPUArithmeticEncoder(base=ARITHMETIC_CODER_BASE, precision=ARITHMETIC_CODER_PRECISION)
# 这里的 Batch Size 是送入 GPU 进行推理的 Batch,取决于显存大小
# 建议设大一点,比如 64 或 128,因为是在显存允许范围内并行
GPU_BATCH_SIZE = 256
for i in range(0, len(unique_segments), GPU_BATCH_SIZE):
batch_segments = unique_segments[i : i + GPU_BATCH_SIZE]
try:
padded_batch, lengths = pad_batch(batch_segments)
padded_batch = padded_batch.to(device)
# lengths 在 CPU 上给 encoder 用
with torch.no_grad():
prompt_probs = batched_predict_fn(padded_batch)
final_probs = torch.cat(
[
first_byte_prob.expand(prompt_probs.shape[0], -1, -1),
prompt_probs[:, :-1, ...]
],
dim=1
)
final_probs = utils.batched_normalize_pdf_for_arithmetic_coding(final_probs)
cdfs_gpu = _pdf_to_cdf(final_probs)
cdf_low = cdfs_gpu.gather(2, padded_batch.unsqueeze(-1)).squeeze(-1)
cdf_high = cdfs_gpu.gather(2, (padded_batch + 1).unsqueeze(-1)).squeeze(-1)
cdf_ends = torch.stack([cdf_low, cdf_high], dim=-1)
chunked_compressed_bytes, _, _ = encoder.incremental_batched_encode(
cdf_ends.cpu(),
ALPHABET_SIZE,
lengths,
bit_threshold=16,
force_padding_to_threshold=False,
return_num_padded_bits=True
)
for seg, code in zip(batch_segments, chunked_compressed_bytes):
segment_to_compressed[seg] = list(code)
except Exception as e:
# print(f"Batch Error: {e}")
for seg in batch_segments:
segment_to_compressed[seg] = list(seg) # 降级
all_results = [None] * M
for seg, indices in segment_to_indices.items():
res = segment_to_compressed.get(seg, list(seg))
for idx in indices:
all_results[idx] = res
return all_results
class M1ACManager:
def __init__(self, model_path, first_prob_path, device_id):
self.device = torch.device(f"cuda:{device_id}")
print(f"[GPU {device_id}] Loading M1 Model...")
self.model, _, _ = load_m1_model_and_tokenizer(model_path)
self.model.to(self.device)
self.model.eval()
self.predict_fn = batched_m1_compress_predict_fn(self.model)
if first_prob_path and os.path.exists(first_prob_path):
with open(first_prob_path, 'r') as f:
prob_data = json.load(f)
self.first_byte_prob = torch.tensor(prob_data, dtype=torch.float32, device=self.device)
if self.first_byte_prob.dim() == 1:
self.first_byte_prob = self.first_byte_prob.unsqueeze(0).unsqueeze(0)
else:
self.first_byte_prob = torch.ones((1, 1, ALPHABET_SIZE), dtype=torch.float32, device=self.device) / ALPHABET_SIZE
def compress_batch(self, inputs: List[Tuple[str, Optional[str]]]) -> List[List[int]]:
"""
新的批量压缩接口。
inputs: List of (text, windows_b64)
Returns: List of compressed int lists
"""
all_segments_flat = []
# 记录每个 sample 对应的 segments 在 flat 列表中的起止位置
# map: sample_idx -> (start_idx, end_idx)
sample_segment_map = []
current_idx = 0
# 1. 准备所有 Segments
for text, windows_b64 in inputs:
raw_bytes = text.encode('utf-8')
sample_segs = []
if windows_b64:
# Case 1: 原始数据 (Metadata Split)
for seg, ind in unpack_windows(raw_bytes, windows_b64):
if len(seg) > 0: sample_segs.append(seg)
else:
# Case 2: 扰动数据 (Fixed Chunking)
CHUNK = 512
for i in range(0, len(raw_bytes), CHUNK):
sample_segs.append(raw_bytes[i : i + CHUNK])
count = len(sample_segs)
sample_segment_map.append((current_idx, current_idx + count))
all_segments_flat.extend(sample_segs)
current_idx += count
if not all_segments_flat:
return [[] for _ in inputs]
# 2. 批量调用 GPU 压缩
# compress_segments_ac_impl 内部会处理 GPU mini-batch,所以这里可以传入大列表
compressed_chunks_flat = compress_segments_ac_impl(
all_segments_flat, self.predict_fn, self.first_byte_prob, self.device
)
# 3. 结果重组
results = []
for start, end in sample_segment_map:
# 将属于该 sample 的所有 chunk 拼起来
sample_chunks = compressed_chunks_flat[start:end]
full_stream = [x for chunk in sample_chunks for x in chunk]
results.append(full_stream)
return results
# ==========================================
# 4. Worker Process (重构为 Batch 处理)
# ==========================================
def process_file_worker(rank, gpu_id, file_path, output_dir, model_path, prob_path, max_lines):
try:
torch.cuda.set_device(gpu_id)
# Tokenizer Init
try:
tokenizer = AutoTokenizer.from_pretrained("infly/OpenCoder-1.5B-Base", trust_remote_code=True)
if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token
except:
tokenizer = AutoTokenizer.from_pretrained("gpt2")
# AC Init
try:
ac_manager = M1ACManager(model_path, prob_path, gpu_id)
except Exception as e:
print(f"❌ [GPU {gpu_id}] Init Failed: {e}")
return
results = {"Gzip": [], "Tokenizer": [], "AC_M1": []}
filename = os.path.basename(file_path)
print(f"[GPU {gpu_id}] Processing {filename}...")
# Buffer for batch processing
# 这里的 Batch 是指 "一次性读取多少行文本然后送去压缩"
# 建议设大一点,比如 50 或 100,取决于每行文本的长度
WORKER_BATCH_SIZE = 200
batch_texts = [] # [text1, text2, ...]
batch_pert_texts = [] # [text1_p, text2_p, ...]
batch_metas = [] # [meta1, meta2, ...]
processed_count = 0
def flush_batch():
nonlocal batch_texts, batch_pert_texts, batch_metas
if not batch_texts: return
# 1. Gzip (CPU fast enough, loop is fine)
for t, tp in zip(batch_texts, batch_pert_texts):
gz1 = list(gzip.compress(t.encode('utf-8')))
gz2 = list(gzip.compress(tp.encode('utf-8')))
if gz1:
d = Levenshtein.distance(list_to_comparable_str(gz1), list_to_comparable_str(gz2))
results["Gzip"].append(d / len(gz1))
# 2. Tokenizer (CPU/GPU)
# Tokenizer 通常很快,或者可以用 tokenizer.batch_encode_plus
for t, tp in zip(batch_texts, batch_pert_texts):
tok1 = tokenizer.encode(t, add_special_tokens=False)
tok2 = tokenizer.encode(tp, add_special_tokens=False)
if tok1:
d = Levenshtein.distance(list_to_comparable_str(tok1), list_to_comparable_str(tok2))
results["Tokenizer"].append(d / len(tok1))
# 3. AC (GPU BATCHING IS HERE)
# 准备输入数据
orig_inputs = list(zip(batch_texts, batch_metas)) # (text, meta)
pert_inputs = list(zip(batch_pert_texts, [None]*len(batch_pert_texts))) # (text, None)
try:
# 批量压缩
ac1_list = ac_manager.compress_batch(orig_inputs)
ac2_list = ac_manager.compress_batch(pert_inputs)
for ac1, ac2 in zip(ac1_list, ac2_list):
if ac1 and len(ac1) > 0:
d = Levenshtein.distance(list_to_comparable_str(ac1), list_to_comparable_str(ac2))
results["AC_M1"].append(d / len(ac1))
except Exception as e:
print(f"[GPU {gpu_id}] AC Batch Error: {e}")
# Clear buffer
batch_texts, batch_pert_texts, batch_metas = [], [], []
with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
for i, line in enumerate(f):
if max_lines > 0 and i >= max_lines: break
try:
data = json.loads(line)
text = data.get('text', '')
windows_b64 = data.get('windows_starts_lens_b64')
if not text or len(text) < 100: continue
cut_idx = max(1, int(len(text) * 0.1))
text_pert = text[cut_idx:]
# Add to buffer
batch_texts.append(text)
batch_pert_texts.append(text_pert)
batch_metas.append(windows_b64)
processed_count += 1
# Flush if full
if len(batch_texts) >= WORKER_BATCH_SIZE:
flush_batch()
if processed_count % 500 == 0:
print(f"[GPU {gpu_id}] Processed {processed_count} lines...")
except Exception:
continue
# Flush remaining
flush_batch()
output_file = os.path.join(output_dir, f"partial_result_{rank}_{filename}.json")
with open(output_file, 'w') as f:
json.dump(results, f)
print(f"✅ [GPU {gpu_id}] Done {filename}. Total: {processed_count}")
except Exception as e:
print(f"❌ [GPU {gpu_id}] Worker failed: {e}")
import traceback
traceback.print_exc()
# ==========================================
# 5. Main
# ==========================================
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--input_dir", type=str, required=True)
parser.add_argument("--m1_model", type=str, required=True)
parser.add_argument("--first_prob_path", type=str, required=True)
parser.add_argument("-o", "--output_dir", type=str, default="analysis_output_parallel")
parser.add_argument("--max_lines", type=int, default=10000)
args = parser.parse_args()
os.makedirs(args.output_dir, exist_ok=True)
files = [os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if f.endswith('.jsonl') and "writer" not in f]
files.sort()
num_gpus = torch.cuda.device_count()
if num_gpus == 0: return
if len(files) > num_gpus:
files = files[:num_gpus]
actual_procs = len(files)
print(f"🚀 Launching {actual_procs} processes (Batch Mode)...")
mp.set_start_method('spawn', force=True)
processes = []
for rank in range(actual_procs):
p = mp.Process(
target=process_file_worker,
args=(rank, rank % num_gpus, files[rank], args.output_dir, args.m1_model, args.first_prob_path, args.max_lines)
)
p.start()
processes.append(p)
for p in processes:
p.join()
print("✅ All workers finished. Merging results...")
final_results = {"Gzip": [], "Tokenizer": [], "AC_M1": []}
for filename in os.listdir(args.output_dir):
if filename.startswith("partial_result_") and filename.endswith(".json"):
try:
with open(os.path.join(args.output_dir, filename), 'r') as f:
data = json.load(f)
for k in final_results:
if k in data: final_results[k].extend(data[k])
except: pass
for k, v in final_results.items():
print(f" -> {k}: {len(v)} samples collected.")
plot_records = []
for algo, vals in final_results.items():
cleaned = [v for v in vals if v < 2.0]
for v in cleaned:
plot_records.append({"Algorithm": algo, "Normalized Edit Distance": v})
if not plot_records:
print("❌ No data collected.")
return
print("📊 Generating plot...")
try:
df = pd.DataFrame(plot_records)
plt.figure(figsize=(12, 7))
sns.set_style("whitegrid")
sns.kdeplot(data=df, x="Normalized Edit Distance", hue="Algorithm", fill=True, common_norm=False, palette="tab10", alpha=0.5)
plt.title("Compression Stability Analysis")
plt.xlabel("Normalized Levenshtein Distance")
plt.xlim(0, 1.2)
plt.savefig(os.path.join(args.output_dir, "stability_parallel_batch.png"), dpi=300)
except Exception as e:
print(f"⚠️ Plotting failed: {e}")
stats = {k: {"mean": float(np.mean(v)), "count": len(v)} for k, v in final_results.items() if v}
with open(os.path.join(args.output_dir, "final_stats.json"), 'w') as f:
json.dump(stats, f, indent=2)
print(f"🎉 Done!")
if __name__ == "__main__":
main()
"""
# 有 8 个json 文件 先测试一个文件
python compare_three_compression_lv.py \
--input_dir /mnt/hdfs/user/linzheng/data/ocpython_subsampled_50G_entropy90_splits_chunk512_ow20_iterative-true_forcepadding-true_merged_ac \
--m1_model /mnt/bn/tiktok-mm-5/aiic/users/linzheng/artifacts/m1_checkpoints/m1_40M_lr1e-3_steps200k_bs8_seqlen2048_python/checkpoints/0000200000 \
--first_prob_path /mnt/bn/tiktok-mm-5/aiic/users/linzheng/artifacts/ac_unigram_probs/python500k_unigram_prob.json
这里可能出现缺少某些模块
pip install xformers==0.0.23.post1 -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
""" |