#!/usr/bin/env python3 """光湖 代码模型(7B)->1.5B铸渊蒸馏 v1 teacher: Qwen2.5-Coder-7B (COS下载) student: Qwen2.5-1.5B-Instruct data: sft.jsonl """ import os, json, torch, sys os.environ['CUDA_VISIBLE_DEVICES'] = '0' os.environ['TOKENIZERS_PARALLELISM'] = 'false' os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True' from transformers import AutoModelForCausalLM, AutoTokenizer from torch.utils.data import Dataset, DataLoader from tqdm import tqdm import torch.nn.functional as F from qcloud_cos import CosConfig, CosS3Client # 代码模型teacher路径（从COS下载） TCH = '/root/autodl-tmp/output/coder-7b-sft-cache/final' STU = '/root/autodl-tmp/models/Qwen/Qwen2___5-1___5B-Instruct' DATA = '/root/autodl-tmp/data/sft.jsonl' OUT = '/root/autodl-tmp/output/qwen25-15b-coder-distill' E, B, GA, LR, ML = 3, 4, 8, 1e-5, 2048 TEMP, ALPHA = 2.0, 0.7 BKT, RG = 'sy-finetune-corpus-1317346199', 'ap-guangzhou' CK = os.environ.get('ZY_OSS_KEY', 'AKIDkQuBQhoiS2OYXWebXLwMbdT7cvAScbbU') CS = os.environ.get('ZY_OSS_SECRET', 'nPoZKArgUJBA4nJenjSxJSQBj5FCj3A4') os.makedirs(OUT, exist_ok=True) os.makedirs(TCH, exist_ok=True) # 1. Tokenize print('[1/5] Tokenize...') with open(DATA) as f: raw = [json.loads(l) for l in f] tok = AutoTokenizer.from_pretrained(STU, trust_remote_code=True) tok.pad_token = tok.eos_token data = [] for d in tqdm(raw): msgs = [m for m in d['messages'] if m['role'] != 'system'] ii, ll = [], [] for m in msgs: c = m['content'] if not c.strip(): continue txt = '<|im_start|>' + m['role'] + '\n' + c + '<|im_end|>\n' tk = tok.encode(txt, add_special_tokens=False) ii.extend(tk) ll.extend(tk if m['role'] == 'assistant' else [-100]*len(tk)) if len(ii) > ML: ii, ll = ii[:ML], ll[:ML] data.append({'input_ids': ii, 'labels': ll}) print(f' {len(data)} examples') class DS(Dataset): def __init__(self, d): self.d = d def __len__(self): return len(self.d) def __getitem__(self, i): return self.d[i] def collate(batch): ml = max(len(x['input_ids']) for x in batch) pad_id = tok.pad_token_id ii = torch.stack([torch.tensor(x['input_ids'] + [pad_id]*(ml-len(x['input_ids']))) for x in batch]) ll = torch.stack([torch.tensor(x['labels'] + [-100]*(ml-len(x['labels']))) for x in batch]) am = (ii != pad_id).long() return {'input_ids': ii.cuda(), 'labels': ll.cuda(), 'attention_mask': am.cuda()} loader = DataLoader(DS(data), B, shuffle=True, collate_fn=collate, num_workers=0) # 2. Load teacher from COS print('[2/5] Load models...') torch.cuda.empty_cache() if not os.path.isdir(STU) or not os.path.isfile(STU + '/config.json'): from modelscope import snapshot_download snapshot_download('Qwen/Qwen2.5-1.5B-Instruct', cache_dir='/root/autodl-tmp/models') if not os.path.isfile(TCH + '/model.safetensors'): print(' DL teacher (code model) from COS...') cfg = CosConfig(Region=RG, SecretId=CK, SecretKey=CS) cl = CosS3Client(cfg) for obj in cl.list_objects(Bucket=BKT, Prefix='models/qwen25-coder-7b-sft/final/').get('Contents', []): fn = obj['Key'].split('/')[-1] if fn == 'DEPLOY_NOTES.md': continue loc = os.path.join(TCH, fn) if not os.path.isfile(loc): print(f' Downloading {fn}...', flush=True) cl.download_file(Bucket=BKT, Key=obj['Key'], DestFilePath=loc) print(' Teacher (code model)...', flush=True) teacher = AutoModelForCausalLM.from_pretrained( TCH, trust_remote_code=True, torch_dtype=torch.bfloat16, attn_implementation='sdpa').cuda() teacher.eval() for p in teacher.parameters(): p.requires_grad_(False) print(f' T: {sum(p.numel() for p in teacher.parameters())/1e9:.2f}B', flush=True) print(' Student...', flush=True) student = AutoModelForCausalLM.from_pretrained( STU, trust_remote_code=True, torch_dtype=torch.bfloat16, attn_implementation='sdpa').cuda() student.train() print(f' S: {sum(p.numel() for p in student.parameters())/1e9:.2f}B', flush=True) print(f' VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB', flush=True) # 3. Train (knowledge distillation) print('[3/5] Train (KD)...', flush=True) opt = torch.optim.AdamW(student.parameters(), LR, weight_decay=0.01) steps_per_epoch = (len(data) // B) total_steps = steps_per_epoch * E global_step = 0 for ep in range(E): print(f'\n===== Epoch {ep+1}/{E} =====', flush=True) loader.dataset.d = data # reshuffle won't hurt for batch in tqdm(loader, total=steps_per_epoch): with torch.no_grad(): t_logits = teacher(**batch).logits # Fix vocab mismatch: teacher(152064) vs student(151936) t_logits = t_logits[:, :, :151936] s_logits = student(**batch).logits # SFT loss sft_loss = F.cross_entropy( s_logits.view(-1, s_logits.size(-1)), batch['labels'].view(-1), ignore_index=-100, reduction='mean') # KL divergence (distillation loss) — use cached t_logits T = TEMP t_prob = F.log_softmax(s_logits[:, :-1] / T, dim=-1) s_prob = F.softmax(t_logits[:, :-1] / T, dim=-1) # Only on non-pad tokens mask = (batch['input_ids'][:, 1:] != tok.pad_token_id).unsqueeze(-1) kl_loss = (F.kl_div(t_prob, s_prob, reduction='none') * mask).sum() / mask.sum() * (T ** 2) loss = sft_loss + ALPHA * kl_loss loss.backward() if (global_step + 1) % GA == 0: torch.nn.utils.clip_grad_norm_(student.parameters(), 1.0) opt.step() opt.zero_grad() if global_step % 50 == 0: print(f' step={global_step} loss={loss.item():.4f} sft={sft_loss.item():.4f} kl={kl_loss.item():.4f}', flush=True) # Periodic cache clear if global_step % 500 == 0 and global_step > 0: torch.cuda.empty_cache() global_step += 1 # Save epoch checkpoint ckpt = os.path.join(OUT, f'ep{ep+1}') student.save_pretrained(ckpt) tok.save_pretrained(ckpt) torch.cuda.empty_cache() print(f' Checkpoint: {ckpt}', flush=True) # 4. Save final print('[4/5] Save final...', flush=True) fnl = os.path.join(OUT, 'final') student.save_pretrained(fnl) tok.save_pretrained(fnl) print(f' Final: {fnl}', flush=True) # 5. Upload to COS print('[5/5] Upload to COS...', flush=True) cfg = CosConfig(Region=RG, SecretId=CK, SecretKey=CS) cl = CosS3Client(cfg) for f in sorted(os.listdir(fnl)): fp = os.path.join(fnl, f) if os.path.isfile(fp): mb = os.path.getsize(fp) / 1024 / 1024 print(f' {f} ({mb:.0f}MB)', flush=True) cl.upload_file(Bucket=BKT, Key=f'models/qwen25-15b-coder-distill/{f}', LocalFilePath=fp) print('ALL DONE', flush=True)