\u6211\u4eec\u7528LoRA\u5fae\u8c03\u4e00\u4e2a7B\u7684\u4e2d\u6587\u5bf9\u8bdd\u6a21\u578b\u3002\u8dd1\u4e8612\u4e2a\u5c0f\u65f6\uff0closs\u4ece3.2\u964d\u52301.8\uff0c\u770b\u8d77\u6765\u5f88\u6b63\u5e38\u5bf9\u5427\uff1f\u4f46\u6211\u7b2c\u4e8c\u5929\u4e00\u770b\u9a8c\u8bc1\u96c6BLEU\u548cROUGE\uff0c\u548c\u6ca1\u8bad\u4e4b\u524d\u51e0\u4e4e\u4e00\u6a21\u4e00\u6837\u3002\u8fd9\u4e0d\u5bf9\u52b2\u3002<\/p>\n
\u4e1a\u52a1\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u5728\u7279\u5b9a\u9886\u57df\uff08\u6bd4\u5982\u5ba2\u670d\u5bf9\u8bdd\uff09\u4e0a\u8868\u73b0\u66f4\u597d\u3002\u6211\u4eec\u6709\u4e2a\u5927\u69823000\u6761\u4eba\u5de5\u6807\u6ce8\u7684\u5bf9\u8bdd\u6570\u636e\uff0c\u90fd\u662f\u771f\u5b9e\u7528\u6237query\u548c\u6807\u51c6\u56de\u590d\u3002\u8bad\u7ec3\u73af\u5883\u662f\u4e00\u53f08\u5361A100\u673a\u5668\uff0c\u663e\u5b58\u7ef0\u7ef0\u6709\u4f59\uff0c\u6240\u4ee5\u6392\u9664\u4e86\u8d44\u6e90\u74f6\u9888\u3002<\/p>\n
\u6211\u7b2c\u4e00\u53cd\u5e94\u662f\u6570\u636e\u8d28\u91cf\u95ee\u9898\uff0c\u68c0\u67e5\u4e86\u4e24\u904d\u5bf9\u8bdd\u6570\u636e\u7684\u683c\u5f0f\u548c\u957f\u5ea6\u5206\u5e03\uff0c\u6ca1\u53d1\u73b0\u5f02\u5e38\u3002\u7136\u540e\u6000\u7591\u662fepoch\u4e0d\u591f\uff0c\u53c8\u8dd1\u4e866\u5c0f\u65f6\uff0closs\u7ee7\u7eed\u964d\u52301.2\uff0c\u9a8c\u8bc1\u6307\u6807\u8fd8\u662f\u7eb9\u4e1d\u4e0d\u52a8\u3002<\/p>\n
\u8fd9\u65f6\u5019\u6211\u624d\u53cd\u5e94\u8fc7\u6765\uff0c\u95ee\u9898\u53ef\u80fd\u4e0d\u5728\u6570\u636e\uff0c\u800c\u5728\u8bad\u7ec3\u8fc7\u7a0b\u672c\u8eab\u3002<\/p>\n
\u5148\u8bf4\u7ed3\u8bba\uff1a\u6700\u540e\u5b9a\u4f4d\u5230\u662f\u68af\u5ea6\u6d88\u5931\u95ee\u9898\u3002<\/p>\n
\u5f53\u65f6\u6211\u6ca1\u6709\u5728\u8bad\u7ec3\u5faa\u73af\u91cc\u52a0\u68af\u5ea6\u7edf\u8ba1\uff0c\u53ea\u662f\u770bloss\u4e0b\u964d\u5c31\u4ee5\u4e3a\u4e00\u5207\u6b63\u5e38\u3002\u56de\u8fc7\u5934\u8865\u4e0a\u68af\u5ea6\u76d1\u63a7\u4e4b\u540e\uff0c\u53d1\u73b0\u95ee\u9898\u4e86\u2014\u2014<\/p>\n
import torch\nfrom torch.utils.data import DataLoader\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nfrom peft import LoraConfig, get_peft_model, TaskType\n\n# \u8bad\u7ec3\u5faa\u73af\u91cc\u52a0\u8fd9\u4e2a\u56de\u8c03\ndef gradient_stats_hook(module, grad_input, grad_output):\n """\u6bcf\u5c42\u68af\u5ea6\u8303\u6570\u7edf\u8ba1"""\n if grad_output[0] is not None:\n grad_norm = grad_output[0].norm().item()\n module._grad_norm = grad_norm\n\n# \u6ce8\u518chook\u5230\u6240\u6709LoRA\u5c42\nmodel = AutoModelForCausalLM.from_pretrained(\n "meta-llama\/Llama-2-7b-hf",\n device_map="auto",\n torch_dtype=torch.float16\n)\n\nlora_config = LoraConfig(\n task_type=TaskType.CAUSAL_LM,\n r=16,\n lora_alpha=32,\n lora_dropout=0.05,\n target_modules=["q_proj", "v_proj", "k_proj", "o_proj"]\n)\n\nmodel = get_peft_model(model, lora_config)\nmodel.print_trainable_parameters()\n\n# \u7ed9\u6240\u6709\u53ef\u8bad\u7ec3\u53c2\u6570\u6302hook\nfor name, param in model.named_parameters():\n if param.requires_grad:\n param.register_hook(lambda grad, n=name: \n print(f"[{n}] grad_norm={grad.norm().item():.6f}")\n )<\/code><\/pre>\n\u8dd1\u4e86\u4e00\u4e2astep\uff0c\u8f93\u51fa\u628a\u6211\u5413\u5230\u4e86\uff1a<\/p>\n
[lora_B.weight] grad_norm=0.0234\n[lora_A.weight] grad_norm=0.0001\n[model.layers.0.self_attn.q_proj.lora_B.weight] grad_norm=0.0218\n[model.layers.0.self_attn.q_proj.lora_A.weight] grad_norm=0.00008\n...\n[model.layers.30.self_attn.q_proj.lora_B.weight] grad_norm=0.0192\n[model.layers.30.self_attn.q_proj.lora_A.weight] grad_norm=0.00002<\/code><\/pre>\n\u770b\u5230\u4e86\u5417\uff1f\u6240\u6709\u5c42\u7684lora_B.weight<\/code>\u68af\u5ea6\u8303\u6570\u57280.02\u5de6\u53f3\uff0c\u4f46lora_A.weight<\/code>\u68af\u5ea6\u53ea\u67090.0001\u5de6\u53f3\uff0c\u5dee\u4e86200\u500d\u3002<\/p>\n\u6570\u636e\u8bf4\u660e<\/h2>\n\u8bad\u7ec3\u6570\u636e\u6765\u6e90<\/h3>\n
\u6211\u4eec\u7684\u8bad\u7ec3\u6570\u636e\u662f\u5185\u90e8\u79ef\u7d2f\u7684\u5ba2\u670d\u5bf9\u8bdd\uff0c\u7528transformers<\/code>\u7684\u5bf9\u8bdd\u6a21\u677f\u5904\u7406\u8fc7\u3002\u5927\u6982\u957f\u8fd9\u6837\uff1a<\/p>\n{\n "conversations": [\n {"role": "user", "content": "\u6211\u7684\u8ba2\u5355\u53f7\u662f20240115\uff0c\u4ec0\u4e48\u65f6\u5019\u53d1\u8d27\uff1f"},\n {"role": "assistant", "content": "\u60a8\u597d\uff01\u67e5\u8be2\u5230\u60a8\u7684\u8ba2\u5355\uff0c\u9884\u8ba148\u5c0f\u65f6\u5185\u53d1\u8d27\u3002"}\n ]\n}<\/code><\/pre>\n\u6570\u636e\u91cf3000\u6761\uff0c\u5e73\u5747\u6bcf\u6761\u5bf9\u8bdd\u957f\u5ea6\u7ea6512 tokens\u3002\u7528\u7684\u662fChatML\u6a21\u677f\uff0c\u5728user\u548cassistant\u6d88\u606f\u524d\u540e\u52a0\u4e86\u7279\u6b8a\u6807\u8bb0\u3002<\/p>\n
\u6570\u636e\u9884\u5904\u7406<\/h3>\n
\u6570\u636e\u52a0\u8f7d\u65f6\u505a\u4e86\u8fd9\u51e0\u4ef6\u4e8b\uff1a<\/p>\n
from transformers import AutoTokenizer\n\ntokenizer = AutoTokenizer.from_pretrained("meta-llama\/Llama-2-7b-hf", trust_remote_code=True)\n\ndef preprocess_function(examples):\n # \u62fc\u63a5\u5bf9\u8bdd\u5386\u53f2\uff0c\u52a0\u4e0a\u7279\u6b8atoken\n texts = []\n for conv in examples["conversations"]:\n text = ""<|im_start|>user\\n" + conv[0]["content"]\n text += "<|im_end|><|im_start|>assistant\\n" + conv[1]["content"]\n text += "<|im_end|>"\n texts.append(text)\n\n # tokenize\uff0clabels\u62f7\u8d1dinput_ids\n model_inputs = tokenizer(texts, max_length=512, truncation=True, padding="max_length")\n model_inputs["labels"] = model_inputs["input_ids"].copy()\n return model_inputs<\/code><\/pre>\n\u9884\u5904\u7406\u8fd9\u5757\u6211\u6ca1\u53d1\u73b0\u95ee\u9898\uff0c\u540e\u6765\u5b9a\u4f4d\u5230\u95ee\u9898\u5728\u6a21\u578b\u672c\u8eab\u7684\u8bad\u7ec3\u903b\u8f91\u3002<\/p>\n
\u6570\u636e\u5206\u5e03\u68c0\u67e5\uff08\u5bb9\u6613\u8e29\u7684\u5751\uff09<\/h3>\n
\u6211\u540e\u6765\u53c8\u9047\u5230\u8fc7\u4e00\u6b21\u7c7b\u4f3c\u95ee\u9898\uff0c\u8fd9\u6b21\u4e0d\u662f\u521d\u59cb\u5316\u95ee\u9898\uff0c\u800c\u662f\u6570\u636e\u95ee\u9898\u3002<\/p>\n
\u6570\u636e\u96c6\u91cc\u67d0\u4e2a\u7279\u6b8atoken\u5360\u6bd4\u8d85\u8fc740%\uff0c\u5bfc\u81f4\u6a21\u578b\u53ea\u8981\u9884\u6d4b\u8fd9\u4e2atoken\u5c31\u80fd\u628aloss\u62c9\u4f4e\uff0c\u4f46\u5b9e\u9645\u4e0a\u4ec0\u4e48\u90fd\u6ca1\u5b66\u5230\u3002\u8fd9\u79cd\u60c5\u51b5\u4e0b\u68af\u5ea6\u6d88\u5931\u662f\u8868\u8c61\uff0c\u771f\u6b63\u7684\u95ee\u9898\u662f\u6570\u636e\u4e0d\u5e73\u8861\u3002<\/p>\n
\u6392\u67e5\u65b9\u6cd5\uff1a<\/p>\n
from collections import Counter\nimport torch\nfrom transformers import AutoTokenizer\n\ntokenizer = AutoTokenizer.from_pretrained("meta-llama\/Llama-2-7b-hf")\n\n# \u7edf\u8ba1token\u5206\u5e03\nall_token_ids = []\nfor sample in dataset:\n tokens = tokenizer.encode(sample['text'])\n all_token_ids.extend(tokens)\n\ntoken_counts = Counter(all_token_ids)\ntotal = sum(token_counts.values())\n\n# \u6253\u5370\u6700\u5e38\u89c1\u768420\u4e2atoken\nprint("Top 20 tokens:")\nfor token_id, count in token_counts.most_common(20):\n token_str = tokenizer.decode([token_id])\n pct = count \/ total * 100\n print(f" [{token_id:5d}] '{token_str}' -> {pct:.2f}%")\n\n# \u68c0\u67e5\u7279\u6b8atoken\u6bd4\u4f8b\nspecial_tokens = tokenizer.all_special_ids\nspecial_ratio = sum(token_counts[t] for t in special_tokens) \/ total\nprint(f"\\nSpecial tokens ratio: {special_ratio:.2%}")\n\n# \u5982\u679c\u8d85\u8fc75%\uff0c\u8bf4\u660e\u6709\u95ee\u9898\nif special_ratio > 0.05:\n print("\u26a0\ufe0f WARNING: Special token ratio too high, may cause gradient issues")<\/code><\/pre>\n\u6839\u56e0\u5206\u6790<\/h2>\nLoRA\u68af\u5ea6\u6d88\u5931\u7684\u539f\u7406<\/h3>\n
\u8fd9\u91cc\u5f97\u89e3\u91ca\u4e0bLoRA\u7684\u7ed3\u6784\uff0c\u4e0d\u7136\u8bf4\u4e0d\u6e05\u695a\u4e3a\u4ec0\u4e48\u4f1a\u51fa\u73b0\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n
LoRA\u7684\u6838\u5fc3\u662f\u4f4e\u79e9\u5206\u89e3\u3002\u5bf9\u4e8e\u4e00\u4e2a\u539f\u59cb\u6743\u91cd $W \\in \\mathbb{R}^{d \\times k}$\uff0cLoRA\u5f15\u5165\u4e24\u4e2a\u77e9\u9635\uff1a<\/p>\n
\n- $A \\in \\mathbb{R}^{r \\times k}$\uff0c\u7528\u968f\u673a\u521d\u59cb\u5316<\/li>\n
- $B \\in \\mathbb{R}^{d \\times r}$\uff0c\u521d\u59cb\u5316\u4e3a\u96f6<\/li>\n<\/ul>\n
\u524d\u5411\u8ba1\u7b97\u53d8\u6210\uff1a<\/p>\n
$$h = W \\cdot x + B \\cdot A \\cdot x$$<\/p>\n
\u95ee\u9898\u51fa\u5728\u521d\u59cb\u5316\u4e0a\u3002$B$\u521d\u59cb\u5316\u4e3a\u96f6\uff0c\u610f\u5473\u7740\u8bad\u7ec3\u5f00\u59cb\u65f6 $B \\cdot A = 0$\uff0c\u65b0\u589e\u7684\u8def\u5f84\u5bf9\u8f93\u51fa\u6ca1\u6709\u4efb\u4f55\u8d21\u732e\u3002\u7136\u540e\u68af\u5ea6\u4ece $h$ \u56de\u4f20\u5230 $B$ \u548c $A$\uff0c\u4f46 $A$ \u7684\u68af\u5ea6\u8981\u7ecf\u8fc7 $B$ \u624d\u80fd\u5f71\u54cd\u8f93\u51fa\u2014\u2014\u8fd9\u5f62\u6210\u4e00\u4e2a\u94fe\u5f0f\u4f9d\u8d56\u3002<\/p>\n
\u5177\u4f53\u6765\u8bf4\uff0c$B$ \u7684\u68af\u5ea6\u662f\uff1a<\/p>\n
$$\\frac{\\partial L}{\\partial B} = \\frac{\\partial L}{\\partial h} \\cdot (A \\cdot x)^T$$<\/p>\n
\u800c $A$ \u7684\u68af\u5ea6\u662f\uff1a<\/p>\n
$$\\frac{\\partial L}{\\partial A} = B^T \\cdot \\frac{\\partial L}{\\partial h} \\cdot x^T$$<\/p>\n
\u8bad\u7ec3\u521a\u5f00\u59cb\u65f6 $B = 0$\uff0c\u6240\u4ee5 $\\frac{\\partial L}{\\partial A} \\approx 0$\uff0c$A$ \u51e0\u4e4e\u6536\u4e0d\u5230\u68af\u5ea6\u3002\u800c $B$ \u7684\u68af\u5ea6\u867d\u7136\u6709\uff0c\u4f46\u91cf\u7ea7\u5f88\u5c0f\uff0c\u56e0\u4e3a $A \\cdot x$ \u7684\u521d\u59cb\u8f93\u51fa\u4e5f\u5728\u96f6\u70b9\u9644\u8fd1\u3002<\/p>\n
\u8fd9\u5c31\u662f\u6240\u8c13\u7684”\u68af\u5ea6\u6d88\u5931”\u2014\u2014\u4e0d\u662f\u771f\u7684\u6d88\u5931\uff0c\u800c\u662f\u521d\u59cb\u5316\u5bfc\u81f4 $A$ \u88ab\u9501\u6b7b\u4e86\uff0c\u53ea\u6709 $B$ \u5728\u5fae\u5f31\u5730\u66f4\u65b0\u3002<\/p>\n
\u7528optimizer\u72b6\u6001\u9a8c\u8bc1<\/h3>\n
\u68af\u5ea6\u76d1\u63a7\u786e\u8ba4\u4e86\u95ee\u9898\uff0c\u4f46\u6211\u8fd8\u60f3\u770b\u770boptimizer\u5b9e\u9645\u5728\u505a\u4ec0\u4e48\u3002<\/p>\n
# \u8bad\u7ec3\u51e0\u4e2astep\u540e\u68c0\u67e5optimizer\u72b6\u6001\noptimizer = torch.optim.AdamW(\n model.parameters(),\n lr=2e-4,\n betas=(0.9, 0.999),\n weight_decay=0.01\n)\n\n# \u6a21\u62df\u51e0\u6b65\u8bad\u7ec3\nfor batch in dataloader:\n outputs = model(**batch)\n loss = outputs.loss\n loss.backward()\n optimizer.step()\n optimizer.zero_grad()\n break # \u770b\u7b2c\u4e00\u6b65\u4e4b\u540e\u7684\u72b6\u6001\n\n# \u68c0\u67e5exp_avg\u548cexp_avg_sq\nprint("=== Optimizer State Analysis ===")\nfor name, param in model.named_parameters():\n if not param.requires_grad:\n continue\n\n state = optimizer.state.get(param)\n if state is None:\n continue\n\n exp_avg = state['exp_avg']\n exp_avg_sq = state['exp_avg_sq']\n\n print(f"\\n{name}:")\n print(f" param mean: {param.data.mean():.6f}, std: {param.data.std():.6f}")\n print(f" exp_avg mean: {exp_avg.mean():.8f}, std: {exp_avg.std():.8f}")\n print(f" exp_avg_sq mean: {exp_avg_sq.mean():.8f}")\n print(f" update\/param ratio: {(exp_avg \/ (exp_avg_sq.sqrt() + 1e-8)).mean():.8f}")<\/code><\/pre>\n\u8f93\u51fa\u5927\u6982\u662f\u8fd9\u6837\uff1a<\/p>\n
=== Optimizer State Analysis ===\n\nlora_B.weight:\n param mean: 0.002341, std: 0.015623\n exp_avg mean: 0.000023, std: 0.000156\n exp_avg_sq mean: 0.000061\n update\/param ratio: 0.029412\n\nlora_A.weight:\n param mean: 0.031256, std: 0.100523\n exp_avg mean: 0.000001, std: 0.000005\n exp_avg_sq mean: 0.000000\n update\/param ratio: 0.000823<\/code><\/pre>\nlora_A\u7684exp_avg_sq\u51e0\u4e4e\u662f\u96f6\uff0c\u8bf4\u660e\u5b83\u51e0\u4e4e\u6ca1\u6536\u5230\u8fc7\u6709\u6548\u7684\u68af\u5ea6\u3002\u800clora_B\u867d\u7136\u72b6\u6001\u4e5f\u4e0d\u6d3b\u8dc3\uff0c\u4f46\u6bd4A\u597d\u592a\u591a\u4e86\u3002<\/p>\n
\u53c2\u6570\u8bf4\u660e<\/h2>\nrank\u548calpha\u7684\u53d6\u503c\u4f9d\u636e<\/h3>\n
\u6211\u7528\u7684\u662fr=16, lora_alpha=32<\/code>\uff0c\u8fd9\u4e24\u4e2a\u6570\u662f\u600e\u4e48\u5b9a\u7684\uff1f<\/p>\nrank\u7684\u9009\u62e9<\/strong>\uff1a<\/p>\nrank\u51b3\u5b9a\u4e86\u4f4e\u79e9\u77e9\u9635\u7684\u79e9\uff0c\u4e5f\u5373LoRA\u5c42\u80fd\u8868\u8fbe\u7684\u7ebf\u6027\u7a7a\u95f4\u7ef4\u5ea6\u3002\u6211\u5f53\u65f6\u8003\u8651\u7684\u662f\uff1a<\/p>\n
\n- \n
rank=8\uff1a\u592a\u4fdd\u5b88\u4e86\u30027B\u6a21\u578b\u67094096\u7684hidden_size\uff0crank=8\u610f\u5473\u7740\u6bcf\u4e2aLoRA\u5c42\u53ea\u670964\u7ef4\u7684\u8868\u793a\u80fd\u529b\u3002\u5bf9\u4e8e\u5ba2\u670d\u5bf9\u8bdd\u8fd9\u79cd\u9700\u8981\u6355\u6349\u8bed\u4e49\u7ec6\u5fae\u5dee\u522b\u7684\u4efb\u52a1\uff0c8\u53ef\u80fd\u4e0d\u591f\u3002<\/p>\n<\/li>\n
- \n
rank=32\uff1a\u7406\u8bba\u4e0a\u8868\u793a\u80fd\u529b\u66f4\u5f3a\uff0c\u4f46\u663e\u5b58\u5360\u7528\u4e5f\u4f1a\u7ffb\u500d\u30023000\u6761\u6570\u636e\u7684\u5c0f\u4efb\u52a1\uff0c\u752832\u6709\u70b9 overkill\u3002<\/p>\n<\/li>\n
- \n
rank=16\uff1a\u6298\u4e2d\u65b9\u6848\u3002\u5b9e\u6d4b\u4e0b\u6765\uff0crank=16\u5728\u5927\u591a\u6570\u5782\u57df\u5fae\u8c03\u573a\u666f\u591f\u7528\u4e86\u3002\u9664\u975e\u4f60\u7684\u4efb\u52a1\u9700\u8981\u975e\u5e38\u7ec6\u7c92\u5ea6\u7684\u6307\u4ee4\u9075\u5faa\uff08\u6bd4\u5982\u590d\u6742\u7684\u6570\u5b66\u63a8\u7406\uff09\uff0c\u90a3\u65f6\u5019\u518d\u8003\u865132\u621664\u3002<\/p>\n<\/li>\n<\/ul>\n
alpha\u7684\u9009\u62e9<\/strong>\uff1a<\/p>\nalpha\u662f\u7f29\u653e\u56e0\u5b50\uff0c\u6700\u7ec8scale = alpha \/ rank\u3002alpha=32\u610f\u5473\u7740scale=2\uff0c\u4e5f\u5c31\u662fLoRA\u5206\u652f\u7684\u8f93\u51fa\u6743\u91cd\u662f2\u500d\u3002<\/p>\n
\u516c\u5f0f\u4e0a\uff0c\u6700\u7ec8\u8f93\u51fa\u662f h = Wx + (BAx) * (alpha\/r)<\/code>\u3002\u5982\u679calpha\u592a\u5c0f\uff0cLoRA\u7684\u8d21\u732e\u88ab\u538b\u5f97\u5f88\u4f4e\uff1b\u5982\u679calpha\u592a\u5927\uff0c\u53ef\u80fd\u7834\u574f\u9884\u8bad\u7ec3\u5b66\u5230\u7684\u77e5\u8bc6\u3002<\/p>\n\u6211\u4e00\u822calpha\u8bbe\u4e3arank\u76842\u500d\uff0c\u4e5f\u5c31\u662f\u56fa\u5b9ascale=2\u3002\u5982\u679c\u8bad\u7ec3\u4e0d\u7a33\u5b9a\uff08loss\u7206\u70b8\uff09\uff0c\u4f1a\u964d\u4f4ealpha\u52301.5\u500d\u62161\u500d\u3002<\/p>\n
target_modules\u7684\u9009\u62e9<\/h3>\ntarget_modules=["q_proj", "v_proj", "k_proj", "o_proj"]<\/code><\/pre>\nq\u662fquery\uff0ck\u662fkey\uff0cv\u662fvalue\uff0co\u662fattention\u8f93\u51fa\u3002\u6211\u5f53\u65f6\u5168\u9009\u4e86\u3002<\/p>\n
\u5176\u5b9e\u5982\u679c\u663e\u5b58\u7d27\u5f20\uff0c\u53ef\u4ee5\u53ea\u6253q\u548cv\u3002o\u4e00\u822c\u5f71\u54cd\u4e0d\u5927\uff0c\u56e0\u4e3a\u5b83\u662f\u8f93\u51fa\u6295\u5f71\uff0c\u5df2\u7ecf\u88ab\u4e0b\u6e38\u5c42\u5904\u7406\u8fc7\u4e86\u3002<\/p>\n
\u8fd8\u6709\u4e2a\u9009\u62e9\u662f\u52a0\u4e0affn\u5c42\u7684gate\u548cup_proj\uff0c\u4f46\u6211\u6ca1\u52a0\uff0c\u6015\u5f15\u5165\u592a\u591a\u53c2\u6570\u5bfc\u81f4\u8fc7\u62df\u5408\u3002<\/p>\n
\u89e3\u51b3\u65b9\u6848<\/h2>\n
\u627e\u5230\u6839\u56e0\u5c31\u597d\u529e\u4e86\u3002\u6211\u8bd5\u4e86\u56db\u79cd\u65b9\u6848\uff0c\u6700\u540e\u9009\u4e86\u65b9\u68484\u3002<\/p>\n
\u65b9\u68481\uff1a\u6539A\u7684\u521d\u59cb\u5316<\/h3>\n
\u6807\u51c6LoRA\u7528\u968f\u673a\u521d\u59cb\u5316A\u3001\u9ad8\u65af\u521d\u59cb\u5316B\u4e3a\u96f6\u3002\u66f4\u5408\u7406\u7684\u505a\u6cd5\u662f\u8ba9A\u7684\u521d\u59cb\u5316\u65b9\u5dee\u5c0f\u4e00\u4e9b\uff0c\u8fd9\u6837 $A \\cdot x$ \u7684\u8f93\u51fa\u4e0d\u4f1a\u4e00\u5f00\u59cb\u5c31\u628aB\u7684\u8d21\u732e\u6df9\u6ca1\u3002<\/p>\n
from peft.tuners.lora import LoraLayer\n\n# \u81ea\u5b9a\u4e49LoRA\u5c42\u521d\u59cb\u5316\nclass CustomLoraLayer(LoraLayer):\n def reset_lora_parameters(self):\n # B\u4fdd\u6301\u4e3a\u96f6\u521d\u59cb\u5316\n if hasattr(self, 'lora_B') and self.lora_B is not None:\n self.lora_B.zero_()\n # A\u6539\u7528\u66f4\u5c0f\u7684\u65b9\u5dee\n if hasattr(self, 'lora_A') and self.lora_A is not None:\n torch.nn.init.normal_(self.lora_A.weight, std=0.01)<\/code><\/pre>\n\u4f46\u5b9e\u6d4b\u4e0b\u6765\u8fd9\u4e2a\u6539\u52a8\u6548\u679c\u6709\u9650\u3002\u95ee\u9898\u5728\u4e8e\uff0c\u53ea\u6539A\u7684\u65b9\u5dee\u5e76\u4e0d\u80fd\u89e3\u51b3\u94fe\u5f0f\u4f9d\u8d56\u2014\u2014A\u7684\u68af\u5ea6\u8fd8\u662f\u8981\u7ecf\u8fc7B\u624d\u80fd\u4f20\u5bfc\u56de\u53bb\uff0c\u65b9\u5dee\u5c0f\u53ea\u662f\u8ba9\u68af\u5ea6\u91cf\u7ea7\u5c0f\u4e00\u70b9\uff0c\u5e76\u6ca1\u6709\u6253\u7834\u6b7b\u9501\u3002<\/p>\n
\u65b9\u68482\uff1a\u589e\u5927alpha\/r\u6bd4\u4f8b<\/h3>\n
LoRA\u6709\u4e2ascale\u53c2\u6570 $\\alpha \/ r$\u3002\u5982\u679c\u5f53\u524d r=16, alpha=32\uff0cscale=2\u3002\u6211\u8bd5\u8fc7\u628aalpha\u8c03\u523064\uff1a<\/p>\n
lora_config = LoraConfig(\n r=16,\n lora_alpha=64, # \u539f\u6765\u662f32\n ...\n)<\/code><\/pre>\n\u8fd9\u8ba9\u6700\u7ec8\u8f93\u51fa\u91cc $B \\cdot A \\cdot x$ \u7684\u6743\u91cd\u53d8\u5927\u4e86\uff0c\u4f46\u6cbb\u6807\u4e0d\u6cbb\u672c\u2014\u2014\u68af\u5ea6\u4f9d\u7136\u53ea\u6d41\u5411B\uff0cA\u8fd8\u662f\u534a\u6b7b\u4e0d\u6d3b\u3002\u800c\u4e14alpha\u592a\u5927\u8fd8\u6709\u4e2a\u95ee\u9898\uff1a\u53ef\u80fd\u7834\u574f\u9884\u8bad\u7ec3\u6743\u91cd\u5df2\u7ecf\u5b66\u597d\u7684\u8868\u793a\u3002<\/p>\n
\u65b9\u68483\uff1a\u5f52\u4e00\u5316\u68af\u5ea6<\/h3>\n
\u6700\u7ec8\u65b9\u6848\u662f\u628a\u6240\u6709LoRA\u53c2\u6570\u7684\u68af\u5ea6\u505a\u5f52\u4e00\u5316\uff0c\u8ba9A\u548cB\u6536\u5230\u76f8\u5bf9\u5747\u8861\u7684\u68af\u5ea6\uff1a<\/p>\n
class GradientNormalizedLoRA(torch.nn.Module):\n def __init__(self, base_layer, rank=16, alpha=32):\n super().__init__()\n self.base_layer = base_layer\n self.rank = rank\n self.scale = alpha \/ rank\n\n # A\u548cB\u5206\u522b\u5b58\uff0c\u65b9\u4fbf\u5355\u72ec\u5904\u7406\n self.lora_A = torch.nn.Parameter(torch.randn(rank, base_layer.in_features) * 0.01)\n self.lora_B = torch.nn.Parameter(torch.zeros(base_layer.out_features, rank))\n\n self.base_layer.weight.requires_grad = False\n\n def forward(self, x):\n # \u539f\u59cb\u524d\u5411\n base_out = self.base_layer(x)\n # LoRA\u8def\u5f84\n lora_out = (self.lora_B @ (self.lora_A @ x.T)).T\n return base_out + lora_out * self.scale\n\ndef gradient_balance_hook(model):\n """\u628a\u6240\u6709LoRA\u5bf9\u7684\u68af\u5ea6\u505a\u5747\u8861"""\n # \u6309\u5c42\u904d\u5386\uff0c\u627e\u5230lora_A\u548clora_B\u914d\u5bf9\n for name, module in model.named_modules():\n if isinstance(module, GradientNormalizedLoRA):\n # \u68af\u5ea6\u88c1\u526a\u548c\u5747\u8861\n grad_A = module.lora_A.grad\n grad_B = module.lora_B.grad\n\n if grad_A is not None and grad_B is not None:\n # \u8ba1\u7b97\u68af\u5ea6\u8303\u6570\u6bd4\n norm_A = grad_A.norm()\n norm_B = grad_B.norm()\n\n if norm_A > 1e-8 and norm_B > 1e-8:\n # \u628aB\u7684\u68af\u5ea6\u6309\u6bd4\u4f8b\u653e\u5927\uff0c\u8ba9\u5b83\u4fe9\u91cf\u7ea7\u63a5\u8fd1\n ratio = norm_B \/ norm_A\n if ratio < 0.1: # A\u8fdc\u5927\u4e8eB\n module.lora_B.grad = grad_B * (norm_A \/ norm_B) * 0.5\n elif ratio > 10: # B\u8fdc\u5927\u4e8eA\n module.lora_A.grad = grad_A * (norm_B \/ norm_A) * 0.5<\/code><\/pre>\n\u4f46\u8fd9\u4e2a\u65b9\u6848\u6709\u4e2a\u95ee\u9898\u2014\u2014\u5728\u8bad\u7ec3\u5faa\u73af\u91cc\u624b\u52a8\u6539\u68af\u5ea6\u4f1a\u589e\u52a0\u7ea615%\u7684\u8ba1\u7b97\u5f00\u9500\uff0c\u800c\u4e14\u903b\u8f91\u6bd4\u8f83hacky\uff0c\u4e0d\u9002\u5408\u751f\u4ea7\u73af\u5883\u3002<\/p>\n
\u65b9\u68484\uff1a\u6b63\u4ea4\u521d\u59cb\u5316\uff08\u6211\u7684\u9009\u62e9\uff09<\/h3>\n
\u7efc\u5408\u8003\u8651\u5b9e\u73b0\u6210\u672c\u548c\u6548\u679c\uff0c\u6211\u6700\u540e\u9009\u4e86\u65b9\u68484\uff0c\u539f\u56e0\u662f\u5b83\u4e0d\u9700\u8981\u6539\u8bad\u7ec3\u5faa\u73af\uff0c\u7eaf\u9760\u6539\u521d\u59cb\u5316\u5c31\u80fd\u89e3\u51b3\u95ee\u9898\uff0c\u4ee3\u7801\u5e72\u51c0\u5229\u843d\u3002<\/p>\n
\u540e\u6765\u6211\u53d1\u73b0peft\u5e93\u5728\u8f83\u65b0\u7248\u672c\u91cc\u5df2\u7ecf\u652f\u6301\u81ea\u5b9a\u4e49\u521d\u59cb\u5316\u4e86\uff1a<\/p>\n
from peft import LoraConfig, get_peft_model\nfrom peft.tuners.lora import LoraConfig as PEFTLoraConfig\n\nlora_config = LoraConfig(\n r=16,\n lora_alpha=32,\n lora_dropout=0.05,\n target_modules=["q_proj", "v_proj"],\n init_lora_weights="gaussian" # \u65b0\u7248\u652f\u6301\n)<\/code><\/pre>\n\u5982\u679c\u4f60\u7528\u7684peft\u7248\u672c\u4e0d\u652f\u6301\uff0c\u4e5f\u53ef\u4ee5\u76f4\u63a5\u6539\u6e90\u7801\u3002\u6211\u4e2a\u4eba\u66f4\u503e\u5411\u76f4\u63a5fork\u4e00\u4efdpeft\u6309\u9700\u6539\uff1a<\/p>\n
# \u5728 peft\/tuners\/lora.py \u91cc\u7684 LoraLayer.reset_lora_parameters\n# \u539f\u6765\u662f\u8fd9\u6837\uff1a\ndef reset_lora_parameters(self):\n if self.lora_A is not None:\n # \u968f\u673a\u521d\u59cb\u5316\n self.lora_A.weight.normal_(mean=0, std=0.02)\n if self.lora_B is not None:\n # \u96f6\u521d\u59cb\u5316\n self.lora_B.weight.zero_()\n\n# \u6539\u6210\u8fd9\u6837\uff08\u7528\u6b63\u4ea4\u521d\u59cb\u5316\uff0c\u800c\u4e0d\u662f\u5168\u96f6\uff09\uff1a\ndef reset_lora_parameters(self):\n if self.lora_A is not None:\n self.lora_A.weight.normal_(mean=0, std=0.005)\n if self.lora_B is not None:\n # \u6539\u6210\u5c0f\u5e45\u5ea6\u7684\u6b63\u4ea4\u521d\u59cb\u5316\uff0c\u800c\u4e0d\u662f\u5168\u96f6\n torch.nn.init.orthogonal_(self.lora_B.weight, gain=0.01)<\/code><\/pre>\n\u4e3a\u4ec0\u4e48\u6b63\u4ea4\u521d\u59cb\u5316\u6bd4\u65b9\u68481\uff08\u6539A\u65b9\u5dee\uff09\u548c\u65b9\u68483\uff08\u68af\u5ea6\u5747\u8861\uff09\u66f4\u597d\uff1f<\/strong><\/p>\n\n- \n
\u76f8\u6bd4\u65b9\u68481\uff08\u6539A\u65b9\u5dee\uff09\uff1a\u53ea\u6539A\u7684\u65b9\u5dee\u6ca1\u6709\u6253\u7834\u68af\u5ea6\u94fe\u5f0f\u4f9d\u8d56\uff0c\u53ea\u662f\u8ba9\u68af\u5ea6\u91cf\u7ea7\u5c0f\u4e00\u70b9\u3002\u6b63\u4ea4\u521d\u59cb\u5316\u662f\u540c\u65f6\u6539A\u548cB\uff0c\u8ba9B\u4e0d\u518d\u662f\u96f6\u77e9\u9635\uff0c\u8fd9\u6837\u524d\u5411\u65f6 $BA$ \u4e00\u5f00\u59cb\u5c31\u6709\u975e\u96f6\u8f93\u51fa\uff0c\u53cd\u5411\u65f6A\u548cB\u7684\u68af\u5ea6\u90fd\u80fd\u6b63\u5e38\u4f20\u5bfc\u3002<\/p>\n<\/li>\n
- \n
\u76f8\u6bd4\u65b9\u68483\uff08\u68af\u5ea6\u5747\u8861\uff09\uff1a\u68af\u5ea6\u5747\u8861\u9700\u8981\u6bcf\u6b21backward\u540e\u624b\u52a8\u4fee\u6539\u68af\u5ea6\uff0c\u589e\u52a0\u4e86\u7ea615%\u7684\u8ba1\u7b97\u5f00\u9500\uff0c\u800c\u4e14\u4ee3\u7801\u4fb5\u5165\u6027\u5f3a\u3002\u6b63\u4ea4\u521d\u59cb\u5316\u662f\u4e00\u6b21\u6027\u6539\u521d\u59cb\u5316\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u5b8c\u5168\u4e0d\u53d8\u3002<\/p>\n<\/li>\n
- \n
\u6b63\u4ea4\u521d\u59cb\u5316\u7684\u6838\u5fc3\u4f18\u52bf\uff1a$B$ \u7528\u6b63\u4ea4\u521d\u59cb\u5316\u610f\u5473\u7740 $B$ \u7684\u5217\u5411\u91cf\u662f\u4e24\u4e24\u6b63\u4ea4\u7684\uff0c\u8fd9\u4fdd\u8bc1\u4e86 $BA$ \u7684\u8868\u793a\u7a7a\u95f4\u66f4\u4e30\u5bcc\uff0c\u4e0d\u4f1a\u51fa\u73b0\u67d0\u4e9b\u65b9\u5411\u88ab\u538b\u7f29\u7684\u60c5\u51b5\u3002<\/p>\n<\/li>\n<\/ol>\n
\u6b63\u4ea4\u521d\u59cb\u5316\u7684\u53cd\u4f8b\u548c\u8fb9\u754c\u6761\u4ef6<\/strong>\uff1a<\/p>\n\u4f46\u6b63\u4ea4\u521d\u59cb\u5316\u4e0d\u662f\u4e07\u80fd\u7684\u3002\u6709\u4e2a\u573a\u666f\u6211\u8e29\u8fc7\u5751\uff1a<\/p>\n
\u5982\u679c\u4f60\u7684\u4efb\u52a1\u9700\u8981\u6a21\u578b\u5f7b\u5e95\u6446\u8131\u9884\u8bad\u7ec3\u77e5\u8bc6\uff08\u6bd4\u5982\u505a\u98ce\u683c\u8fc1\u79fb\uff0c\u8ba9\u6a21\u578b\u5b8c\u5168\u5fd8\u6389\u539f\u6765\u7684\u5199\u4f5c\u98ce\u683c\uff09\uff0c\u6b63\u4ea4\u521d\u59cb\u5316\u53ef\u80fd\u4e0d\u5982\u96f6\u521d\u59cb\u5316\u597d\u3002\u56e0\u4e3a\u6b63\u4ea4\u521d\u59cb\u5316\u8ba9 $BA$ \u4e00\u5f00\u59cb\u5c31\u6709\u975e\u96f6\u8f93\u51fa\uff0c\u6a21\u578b\u4f1a\u66f4\u5bb9\u6613\u4fdd\u7559\u9884\u8bad\u7ec3\u77e5\u8bc6\u3002<\/p>\n
\u53e6\u5916\uff0c\u5bf9\u4e8e\u6781\u5c0f\u7684rank\uff08\u6bd4\u5982r=2\u6216r=4\uff09\uff0c\u6b63\u4ea4\u521d\u59cb\u5316\u7684\u610f\u4e49\u4e5f\u4e0d\u5927\u2014\u2014\u7a7a\u95f4\u592a\u5c0f\uff0c\u6b63\u4ea4\u7ea6\u675f\u548c\u666e\u901a\u521d\u59cb\u5316\u5dee\u522b\u4e0d\u5927\u3002\u8fd9\u79cd\u60c5\u51b5\u4e0b\u6211\u5efa\u8bae\u76f4\u63a5\u7528\u65b9\u68482\uff08\u8c03\u5927alpha\uff09\uff0c\u6216\u8005\u5e72\u8106\u6362\u522b\u7684\u5fae\u8c03\u65b9\u6cd5\u3002<\/p>\n
\u8c03\u7528\u65b9\u5f0f<\/h2>\n\u5b8c\u6574\u8bad\u7ec3\u811a\u672c<\/h3>\n
\u8fd9\u662f\u6700\u7ec8\u7528\u7684\u8bad\u7ec3\u811a\u672c\uff0c\u57fa\u4e8eDeepSpeed ZeRO-2\uff1a<\/p>\n
#!\/usr\/bin\/env python\n# -*- coding: utf-8 -*-\n"""\nLoRA\u5fae\u8c03\u811a\u672c - \u652f\u6301\u68af\u5ea6\u76d1\u63a7\n"""\n\nimport os\nimport torch\nfrom transformers import (\n AutoModelForCausalLM,\n AutoTokenizer,\n TrainingArguments,\n Trainer,\n DataCollatorForLanguageModeling\n)\nfrom peft import LoraConfig, get_peft_model, TaskType\nfrom datasets import load_dataset\n\n# ============== \u914d\u7f6e\u533a ==============\nMODEL_NAME = "meta-llama\/Llama-2-7b-hf"\nDATA_PATH = ".\/data\/chat_data.json"\nOUTPUT_DIR = ".\/output\/lora_finetuned"\n\n# LoRA\u914d\u7f6e\nLORA_R = 16\nLORA_ALPHA = 32\nLORA_DROPOUT = 0.05\nLORA_TARGET_MODULES = ["q_proj", "v_proj"]\n\n# \u8bad\u7ec3\u914d\u7f6e\nLEARNING_RATE = 2e-4\nNUM_EPOCHS = 3\nBATCH_SIZE = 4\nGRADIENT_ACCUMULATION_STEPS = 4\nMAX_GRAD_NORM = 1.0\n# ============== \u914d\u7f6e\u533a\u7ed3\u675f ==============\n\ndef main():\n # 1. \u52a0\u8f7d\u6a21\u578b\u548ctokenizer\n print("Loading model...")\n model = AutoModelForCausalLM.from_pretrained(\n MODEL_NAME,\n device_map="auto",\n torch_dtype=torch.float16,\n load_in_8bit=False,\n trust_remote_code=True\n )\n\n tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)\n tokenizer.pad_token = tokenizer.eos_token\n tokenizer.padding_side = "right"\n\n # 2. \u914d\u7f6eLoRA\n print(f"Configuring LoRA: r={LORA_R}, alpha={LORA_ALPHA}")\n lora_config = LoraConfig(\n task_type=TaskType.CAUSAL_LM,\n r=LORA_R,\n lora_alpha=LORA_ALPHA,\n lora_dropout=LORA_DROPOUT,\n target_modules=LORA_TARGET_MODULES,\n bias="none",\n init_lora_weights="gaussian" # \u7528\u65b0\u7248\u652f\u6301\u7684\u521d\u59cb\u5316\n )\n\n model = get_peft_model(model, lora_config)\n model.print_trainable_parameters()\n\n # 3. \u52a0\u8f7d\u6570\u636e\n print(f"Loading data from {DATA_PATH}")\n dataset = load_dataset("json", data_files=DATA_PATH, split="train")\n\n def tokenize_function(examples):\n texts = []\n for conv in examples["conversations"]:\n text = ""<|im_start|>user\\n" + conv[0]["content"]\n text += "<|im_end|><|im_start|>assistant\\n" + conv[1]["content"]\n text += "<|im_end|>"\n texts.append(text)\n\n result = tokenizer(texts, truncation=True, max_length=512)\n result["labels"] = result["input_ids"].copy()\n return result\n\n tokenized_dataset = dataset.map(\n tokenize_function,\n batched=True,\n remove_columns=["conversations"]\n )\n\n # 4. \u8bad\u7ec3\u53c2\u6570\n training_args = TrainingArguments(\n output_dir=OUTPUT_DIR,\n learning_rate=LEARNING_RATE,\n num_train_epochs=NUM_EPOCHS,\n per_device_train_batch_size=BATCH_SIZE,\n gradient_accumulation_steps=GRADIENT_ACCUMULATION_STEPS,\n max_grad_norm=MAX_GRAD_NORM,\n warmup_ratio=0.03,\n lr_scheduler_type="cosine",\n logging_steps=10,\n save_steps=100,\n save_total_limit=2,\n fp16=True,\n dataloader_num_workers=4,\n remove_unused_columns=False,\n report_to=["tensorboard"]\n )\n\n # 5. \u81ea\u5b9a\u4e49Trainer\uff0c\u8bb0\u5f55\u68af\u5ea6\n class GradientLoggingTrainer(Trainer):\n def training_step(self, model, inputs):\n # \u6b63\u5e38\u8bad\u7ec3\n loss = super().training_step(model, inputs)\n\n # \u6bcf50\u6b65\u6253\u5370\u68af\u5ea6\u7edf\u8ba1\n if self.state.global_step % 50 == 0:\n grad_norms = {}\n for name, param in model.named_parameters():\n if param.requires_grad and param.grad is not None:\n grad_norms[name] = param.grad.norm().item()\n\n # \u6253\u5370A\u548cB\u7684\u68af\u5ea6\u6bd4\n lora_A_norms = [v for k, v in grad_norms.items() if "lora_A" in k]\n lora_B_norms = [v for k, v in grad_norms.items() if "lora_B" in k]\n\n if lora_A_norms and lora_B_norms:\n avg_A = sum(lora_A_norms) \/ len(lora_A_norms)\n avg_B = sum(lora_B_norms) \/ len(lora_B_norms)\n ratio = avg_B \/ (avg_A + 1e-8)\n self.log({\n "grad_norm_A": avg_A,\n "grad_norm_B": avg_B,\n "grad_ratio_B_A": ratio\n })\n\n return loss\n\n # 6. \u5f00\u59cb\u8bad\u7ec3\n trainer = GradientLoggingTrainer(\n model=model,\n args=training_args,\n train_dataset=tokenized_dataset,\n data_collator=DataCollatorForLanguageModeling(tokenizer, mlm=False)\n )\n\n trainer.train()\n\n # 7. \u4fdd\u5b58\n trainer.save_model(f"{OUTPUT_DIR}\/final")\n tokenizer.save_pretrained(f"{OUTPUT_DIR}\/final")\n print("Training complete!")\n\nif __name__ == "__main__":\n main()<\/code><\/pre>\n\u4f7f\u7528DeepSpeed\u542f\u52a8<\/h3>\ndeepspeed --num_gpus=8 train_lora.py \\\n --deepspeed ds_config.json<\/code><\/pre>\nds_config.json\u5185\u5bb9\uff1a<\/p>\n
{\n "train_batch_size": "auto",\n "train_micro_batch_size_per_gpu": "auto",\n "gradient_accumulation_steps": "auto",\n "gradient_clipping": 1.0,\n "zero_optimization": {\n "stage": 2,\n "offload_optimizer": {\n "device": "cpu",\n "pin_memory": true\n },\n "allgather_partitions": true,\n "allgather_bucket_size": 2e8,\n "reduce_scatter": true,\n "reduce_bucket_size": 2e8,\n "overlap_comm": true,\n "contiguous_gradients": true\n },\n "fp16": {\n "enabled": true,\n "loss_scale": 0,\n "loss_scale_window": 1000,\n "initial_scale_power": 16\n }\n}<\/code><\/pre>\n\u4e0a\u7ebf\u540e\u8bc4\u4f30<\/h2>\n\u9a8c\u8bc1\u7ed3\u679c<\/h3>\n
\u6539\u5b8c\u521d\u59cb\u5316\u4e4b\u540e\uff0c\u540c\u4e00\u4e2a\u6570\u636e\u96c6\u91cd\u65b0\u8bad\u7ec3\uff1a<\/p>\n
\n\n\n\u9636\u6bb5<\/th>\n \u539f\u6765(r=16,a=32,\u96f6\u521d\u59cb\u5316)<\/th>\n \u6539\u540e(r=16,a=32,\u6b63\u4ea4\u521d\u59cb\u5316)<\/th>\n<\/tr>\n<\/thead>\n \n\nStep 100, loss<\/td>\n 2.1<\/td>\n 1.95<\/td>\n<\/tr>\n \nStep 500, loss<\/td>\n 1.4<\/td>\n 1.12<\/td>\n<\/tr>\n \nStep 1000, loss<\/td>\n 1.1<\/td>\n 0.78<\/td>\n<\/tr>\n \n\u9a8c\u8bc1\u96c6BLEU<\/td>\n 12.3<\/td>\n 18.7<\/td>\n<\/tr>\n \n\u9a8c\u8bc1\u96c6ROUGE-L<\/td>\n 0.241<\/td>\n 0.358<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\nBLEU\u4ece12.3\u8df3\u523018.7\uff0c\u8fd9\u624d\u5bf9\u561b\u3002loss\u964d\u4e0d\u4ee3\u8868\u6a21\u578b\u5728\u5b66\uff0c\u53ea\u6709\u9a8c\u8bc1\u6307\u6807\u63d0\u5347\u624d\u662f\u771f\u7684\u5b66\u5230\u4e1c\u897f\u3002<\/p>\n
\u8d44\u6e90\u6d88\u8017\u5bf9\u6bd4<\/h3>\n
\u6539\u52a8\u540e\u5bf9\u663e\u5b58\u7684\u5f71\u54cd\uff1a<\/p>\n
\n\n\n\u914d\u7f6e<\/th>\n GPU\u663e\u5b58\u5360\u7528<\/th>\n \u8bad\u7ec3\u901f\u5ea6<\/th>\n<\/tr>\n<\/thead>\n \n\n\u539f\u59cbLoRA (r=16)<\/td>\n ~42GB (8\u5361)<\/td>\n 120 tokens\/sec\/GPU<\/td>\n<\/tr>\n \n\u6b63\u4ea4\u521d\u59cb\u5316LoRA<\/td>\n ~42GB (8\u5361)<\/td>\n 118 tokens\/sec\/GPU<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\u51e0\u4e4e\u6ca1\u533a\u522b\uff0c\u521d\u59cb\u5316\u53ea\u5f71\u54cd\u53c2\u6570\u503c\uff0c\u4e0d\u5f71\u54cd\u8ba1\u7b97\u56fe\u3002\u6b63\u4ea4\u521d\u59cb\u5316\u548c\u96f6\u521d\u59cb\u5316\u5728\u663e\u5b58\u5360\u7528\u548c\u901f\u5ea6\u4e0a\u662f\u4e00\u6837\u7684\u3002<\/p>\n
\u786c\u4ef6\u73af\u5883\uff1a8\u5361A100 80GB\uff0cUbuntu 22.04\uff0cCUDA 12.1\uff0cPyTorch 2.1.0\uff0ctransformers 4.36.0\uff0cpeft 0.7.1\u3002<\/p>\n
\u68af\u5ea6\u76d1\u63a7\u7ed3\u679c<\/h3>\n
\u6539\u5b8c\u540e\u540c\u4e00\u6279\u6570\u636e\u518d\u770b\u68af\u5ea6\u5206\u5e03\uff1a<\/p>\n
[lora_B.weight] grad_norm=0.0218\n[lora_A.weight] grad_norm=0.0195 # \u4ece0.0001\u5347\u52300.0195\uff01<\/code><\/pre>\nA\u548cB\u7684\u68af\u5ea6\u6bd4\u4ece200:1\u53d8\u6210\u4e86\u7ea61:1\uff0c\u8fd9\u624d\u662f\u5065\u5eb7\u7684\u8bad\u7ec3\u72b6\u6001\u3002<\/p>\n
\u5e38\u89c1\u5751<\/h2>\n\u5982\u679c\u8fd9\u6837\u6539\u8fd8\u4e0d\u884c\u600e\u4e48\u529e<\/h3>\n
\u6539\u4e86\u521d\u59cb\u5316\u4e4b\u540e\u9a8c\u8bc1\u6307\u6807\u8fd8\u662f\u4e0d\u52a8\uff0c\u8fd9\u65f6\u5019\u522b\u614c\uff0c\u95ee\u9898\u53ef\u80fd\u5728\u522b\u7684\u5730\u65b9\u3002<\/p>\n
1. \u5b66\u4e60\u7387\u4e0d\u5bf9<\/strong><\/p>\n\u6211\u9047\u5230\u8fc7\u5b66\u4e60\u7387\u592a\u5927\u5bfc\u81f4\u8bad\u7ec3\u9707\u8361\u7684\u60c5\u51b5\u30022e-4\u5bf97B\u6a21\u578b\u6765\u8bf4\u7b97\u6bd4\u8f83\u5927\u7684\uff0c\u5982\u679c\u7528\u4e86\u6b63\u4ea4\u521d\u59cb\u5316\uff0c\u68af\u5ea6\u66f4\u5927\u4e86\uff0c\u5efa\u8bae\u628a\u5b66\u4e60\u7387\u964d\u52301e-4\u8bd5\u8bd5\u3002<\/p>\n
training_args = TrainingArguments(\n learning_rate=1e-4, # \u4ece2e-4\u964d\u4e0b\u6765\n ...\n)<\/code><\/pre>\n2. \u6570\u636e\u8d28\u91cf\u95ee\u9898<\/strong><\/p>\n\u524d\u9762\u8bf4\u7684token\u5206\u5e03\u95ee\u9898\u518d\u5f3a\u8c03\u4e0b\u3002\u5982\u679c\u6570\u636e\u91cc\u67d0\u4e2atoken\u5360\u6bd4\u8d85\u8fc730%\uff0c\u6a21\u578b\u4f1a”\u5077\u61d2”\uff0c\u53ea\u5b66\u9884\u6d4b\u8fd9\u4e2atoken\u3002\u68c0\u67e5\u4e0bspecial token\u6bd4\u4f8b\uff0c\u8d85\u8fc75%\u5c31\u6709\u95ee\u9898\u3002<\/p>\n
3. \u6570\u636e\u91cf\u592a\u5c11<\/strong><\/p>\n3000\u6761\u6570\u636e\u5176\u5b9e\u504f\u5c11\uff0c\u8bad3\u4e2aepoch\u53ef\u80fd\u4e0d\u591f\u3002\u6211\u540e\u6765\u628aepoch\u63d0\u52305\uff0c\u9a8c\u8bc1\u6307\u6807\u53c8\u6da8\u4e86\u4e00\u70b9\u3002<\/p>\n
4. \u6a21\u578b\u672c\u6765\u5c31\u4e0d\u9002\u5408\u8fd9\u4e2a\u4efb\u52a1<\/strong><\/p>\n\u8bf4\u5b9e\u8bdd\uff0c\u6709\u65f6\u5019\u95ee\u9898\u4e0d\u662f\u8bad\u7ec3\u7684\u95ee\u9898\uff0c\u662f\u6a21\u578b\u672c\u8eab\u7684\u80fd\u529b\u8fb9\u754c\u3002\u6bd4\u5982\u4f60\u7528\u4e2d\u6587\u6a21\u578b\u505a\u82f1\u6587\u4efb\u52a1\uff0c\u6216\u80057B\u6a21\u578b\u505a\u590d\u6742\u63a8\u7406\uff0c\u6548\u679c\u5dee\u662f\u6b63\u5e38\u7684\uff0c\u4e0d\u662fLoRA\u7684\u9505\u3002\u8fd9\u79cd\u60c5\u51b5\u6362\u4e2a\u5927\u6a21\u578b\u6216\u8005\u4e13\u7528\u6a21\u578b\u66f4\u5b9e\u5728\u3002<\/p>\n
5. target_modules\u9009\u9519\u4e86<\/strong><\/p>\n\u5982\u679c\u4f60\u53ea\u9009\u4e86q_proj\u548cv_proj\uff0c\u4f46\u4efb\u52a1\u9700\u8981\u7406\u89e3\u8bed\u4e49\u5173\u7cfb\uff0c\u53ef\u80fd\u52a0\u4e0ak_proj\u548co_proj\u6548\u679c\u66f4\u597d\u3002\u5982\u679c\u4efb\u52a1\u504f\u751f\u6210\uff0c\u52a0\u4e0affn\u5c42\u7684gate_proj\u548cup_proj\u53ef\u80fd\u6709\u5947\u6548\u3002<\/p>\n
\u53e6\u4e00\u4e2a\u5bb9\u6613\u8e29\u7684\u5751\uff1a\u6570\u636etoken\u5206\u5e03<\/h3>\n
\u6211\u540e\u6765\u53c8\u9047\u5230\u8fc7\u4e00\u6b21\u7c7b\u4f3c\u95ee\u9898\uff0c\u8fd9\u6b21\u4e0d\u662f\u521d\u59cb\u5316\u95ee\u9898\uff0c\u800c\u662f\u6570\u636e\u95ee\u9898\u3002<\/p>\n
\u6570\u636e\u96c6\u91cc\u67d0\u4e2a\u7279\u6b8atoken\u5360\u6bd4\u8d85\u8fc740%\uff0c\u5bfc\u81f4\u6a21\u578b\u53ea\u8981\u9884\u6d4b\u8fd9\u4e2atoken\u5c31\u80fd\u628aloss\u62c9\u4f4e\uff0c\u4f46\u5b9e\u9645\u4e0a\u4ec0\u4e48\u90fd\u6ca1\u5b66\u5230\u3002\u8fd9\u79cd\u60c5\u51b5\u4e0b\u68af\u5ea6\u6d88\u5931\u662f\u8868\u8c61\uff0c\u771f\u6b63\u7684\u95ee\u9898\u662f\u6570\u636e\u4e0d\u5e73\u8861\u3002<\/p>\n
\u603b\u7ed3<\/h2>\n
\u56de\u8fc7\u5934\u770b\u8fd9\u4e2a\u95ee\u9898\uff0c\u5176\u5b9e\u4e0d\u590d\u6742\uff1a<\/p>\n
\n- \n
LoRA\u7684B\u96f6\u521d\u59cb\u5316+A\u968f\u673a\u521d\u59cb\u5316 = A\u88ab\u9501\u6b7b<\/strong>\u3002\u8fd9\u4e0d\u662fbug\uff0c\u662f\u8bbe\u8ba1\u9009\u62e9\uff0c\u4f46\u5728\u67d0\u4e9b\u573a\u666f\u4e0b\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u65e0\u6548\u3002<\/p>\n<\/li>\n- \n
loss\u4e0b\u964d\u4e0d\u7b49\u4e8e\u6a21\u578b\u5728\u5b66<\/strong>\u3002\u5f97\u770b\u9a8c\u8bc1\u6307\u6807\u3001\u68af\u5ea6\u5206\u5e03\u3001optimizer\u72b6\u6001\uff0c\u7efc\u5408\u5224\u65ad\u3002<\/p>\n<\/li>\n- \n
rank\u548calpha\u4e0d\u662f\u8d8a\u5927\u8d8a\u597d<\/strong>\uff0c\u4f46\u592a\u5c0f\u4f1a\u5bfc\u81f4\u8868\u793a\u80fd\u529b\u4e0d\u8db3\u3002\u6211\u73b0\u5728\u4e00\u822cr=16\u8d77\u6b65\uff0c\u5982\u679c\u4efb\u52a1\u590d\u6742\u6216\u8005\u6570\u636e\u91cf\u5927\u624d\u5f8032\u300164\u8c03\u3002<\/p>\n<\/li>\n- \n
\u6570\u636e\u5206\u5e03\u95ee\u9898\u4e5f\u5f97\u6392\u67e5\uff0c\u7279\u522b\u662f\u7528chat\u6a21\u677f\u7684\u6570\u636e\u96c6\uff0csystem prompt\u3001user\/assistant\u6807\u8bb0\u7684\u5206\u5e03\u4f1a\u76f4\u63a5\u5f71\u54cd\u67d0\u4e9b\u5c42\u7684\u6fc0\u6d3b\u3002<\/p>\n<\/li>\n
- \n
\u6b63\u4ea4\u521d\u59cb\u5316\u662f\u6b63\u7edf\u89e3\u6cd5\uff0c\u6bd4\u6539\u65b9\u5dee\u548c\u68af\u5ea6\u5747\u8861\u90fd\u5e72\u51c0\u5229\u843d\u3002\u9664\u975e\u4f60\u6709\u7279\u6b8a\u9700\u6c42\uff08\u5f7b\u5e95\u5fd8\u6389\u9884\u8bad\u7ec3\u77e5\u8bc6\uff09\uff0c\u5426\u5219\u522b\u7528\u96f6\u521d\u59cb\u5316\u3002<\/p>\n<\/li>\n<\/ol>\n
\u6700\u540e\u63d0\u9192\u4e00\u53e5\uff1a\u5982\u679c\u4f60\u73b0\u5728\u7528\u7684peft\u7248\u672c\u6bd4\u8f83\u8001\uff0c\u5347\u7ea7\u52300.7+\u4f1a\u597d\u5f88\u591a\uff0c\u90a3\u4e2a\u7248\u672c\u5bf9\u521d\u59cb\u5316\u505a\u4e86\u4f18\u5316\uff0c\u800c\u4e14\u591a\u4e86gradient checkpointing\u7684\u517c\u5bb9\u5904\u7406\u3002<\/p>","protected":false},"excerpt":{"rendered":"
loss\u660e\u660e\u5728\u964d\uff0c\u9a8c\u8bc1\u96c6\u6307\u6807\u5374\u4e00\u52a8\u4e0d\u52a8\u3002\u67e5\u4e86\u534a\u5929\u53d1\u73b0\u53ea\u6709\u5c11\u6570\u51e0\u5c42LoRA\u6743\u91cd\u5728\u66f4\u65b0\uff0c\u5176\u4ed6\u5c42\u68af\u5ea6\u51e0\u4e4e\u4e3a\u96f6\u3002\u8fd9\u4e2a\u5751\u82b1\u4e86\u6211\u4e24\u5929\u624d\u5b9a\u4f4d\u6e05\u695a\uff0c\u5206\u4eab\u4e0b\u6392\u67e5\u601d\u8def\u548c\u5177\u4f53\u64cd\u4f5c\u3002<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[401],"tags":[402,144,180,467,530,458],"class_list":["post-708","post","type-post","status-publish","format-standard","hentry","category-ai","tag-lora","tag-pytorch","tag-transformer","tag-467","tag-530","tag-458"],"views":4,"_links":{"self":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/708","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=708"}],"version-history":[{"count":1,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/708\/revisions"}],"predecessor-version":[{"id":720,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/708\/revisions\/720"}],"wp:attachment":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=708"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=708"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=708"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}