#!/usr/bin/env python3
"""光湖代码模型→1.5B铸渊模板 蒸馏脚本
Teacher: Qwen2.5-Coder-7B (SFT后的代码模型)
Student: Qwen2.5-Coder-1.5B (将学会铸渊的执行思维)
"""

import os, json, torch, sys
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["TOKENIZERS_PARALLELISM"] = "false"

from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from datasets import Dataset
from tqdm import tqdm
import torch.nn.functional as F

TEACHER_PATH = "/root/autodl-tmp/output/qwen25-coder-7b-sft/final"
STUDENT_PATH = "/root/autodl-tmp/cache/Qwen/Qwen2___5-Coder-1___5B"
DATA = "/root/autodl-tmp/corpus/zhuyuan_deep_finetune.jsonl"
OUT = "/root/autodl-tmp/output/qwen25-coder-15b-zhuyuan-distill"
EPOCHS = 3; BS = 4; GA = 8; LR = 1e-5; MAX_LEN = 2048; TEMP = 2.0; ALPHA = 0.7

os.makedirs(OUT, exist_ok=True)

print("[1/6] Loading zhuyuan corpus...")
with open(DATA) as f:
    raw = [json.loads(line) for line in f]
print(f"  {len(raw)} examples")

print("[2/6] Loading teacher (Coder-7B) and student (Coder-1.5B)...")
tokenizer = AutoTokenizer.from_pretrained(STUDENT_PATH, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

print("  Loading teacher...")
teacher = AutoModelForCausalLM.from_pretrained(TEACHER_PATH, 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"  Teacher: {sum(p.numel() for p in teacher.parameters())/1e9:.2f}B")

print("  Loading student...")
student = AutoModelForCausalLM.from_pretrained(STUDENT_PATH, trust_remote_code=True, torch_dtype=torch.bfloat16, attn_implementation="sdpa").cuda()
student.train()
print(f"  Student: {sum(p.numel() for p in student.parameters())/1e9:.2f}B")

print("[3/6] Tokenizing + generating teacher logits...")
processed = []
for d in tqdm(raw, desc="Tokenize"):
    ids, labs = [], []
    for msg in d["messages"]:
        c = msg["content"]
        if not c.strip(): continue
        t = f"<|im_start|>{msg['role']}\n{c}<|im_end|>\n"
        tok = tokenizer.encode(t, add_special_tokens=False)
        ids.extend(tok)
        labs.extend(tok if msg["role"] == "assistant" else [-100] * len(tok))
    if len(ids) > MAX_LEN:
        ids, labs = ids[:MAX_LEN], labs[:MAX_LEN]
    with torch.no_grad():
        inp = torch.tensor([ids]).cuda()
        t_out = teacher(input_ids=inp)
        t_logits = t_out.logits[0].float().cpu()
    processed.append({"input_ids": ids, "labels": labs, "attention_mask": [1]*len(ids), "teacher_logits": t_logits})

print("[4/6] Training config...")
def distill_collate(features):
    max_len = max(len(f["input_ids"]) for f in features)
    batch = {}
    for k in ["input_ids", "labels", "attention_mask"]:
        pad = tokenizer.pad_token_id if k != "labels" else -100
        batch[k] = torch.tensor([f[k] + [pad]*(max_len-len(f[k])) for f in features])
    vocab_size = features[0]["teacher_logits"].size(-1)
    tl = []
    for f in features:
        t = f["teacher_logits"]
        pad_len = max_len - t.size(0)
        tl.append(torch.cat([t, torch.zeros(pad_len, vocab_size)], dim=0) if pad_len > 0 else t[:max_len])
    batch["teacher_logits"] = torch.stack(tl)
    return batch

class DistillTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        outputs = model(input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"], use_cache=False)
        student_logits = outputs.logits
        shift_logits = student_logits[..., :-1, :].contiguous()
        shift_labels = inputs["labels"][..., 1:].contiguous()
        sft_loss = F.cross_entropy(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1), ignore_index=-100, reduction="mean")
        teacher_logits = inputs["teacher_logits"]
        mask = (inputs["labels"] != -100).unsqueeze(-1).float()
        kl_loss = F.kl_div(F.log_softmax(student_logits / TEMP, dim=-1), F.softmax(teacher_logits / TEMP, dim=-1), reduction="none")
        kl_loss = (kl_loss * mask).sum() / mask.sum() * (TEMP ** 2)
        return ALPHA * kl_loss + (1 - ALPHA) * sft_loss

args = TrainingArguments(
    output_dir=OUT, num_train_epochs=EPOCHS,
    per_device_train_batch_size=BS, gradient_accumulation_steps=GA,
    learning_rate=LR, warmup_ratio=0.05, lr_scheduler_type="cosine",
    bf16=True, tf32=True, logging_steps=10,
    save_strategy="epoch", save_total_limit=3,
    remove_unused_columns=False, dataloader_num_workers=4,
    gradient_checkpointing=True, optim="adamw_torch",
    report_to="none", ddp_find_unused_parameters=False,
)

trainer = DistillTrainer(model=student, args=args, train_dataset=Dataset.from_list(processed), data_collator=distill_collate)

print("[5/6] Starting distillation!")
gpu = torch.cuda.get_device_name(0)
mem = torch.cuda.get_device_properties(0).total_memory / 1e9
print(f"  GPU: {gpu} ({mem:.1f}GB) | Temp={TEMP}, Alpha={ALPHA}")
sys.stdout.flush()
trainer.train()

print("[6/6] Saving...")
final = os.path.join(OUT, "final")
trainer.save_model(final)
tokenizer.save_pretrained(final)
peak = torch.cuda.max_memory_allocated() / 1e9
print(f"  Model: {final} | Peak VRAM: {peak:.2f}GB / {mem:.1f}GB")
print("DONE!")