Fine-tuning 實戰：LoRA 與 Adapter
加上 ASR 語音辨識基礎與 Whisper 微調

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import get_peft_model, LoraConfig, TaskType

# 1. 載入基礎模型（以 LLaMA 為例）
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    torch_dtype=torch.float16,
    device_map="auto",
)

# 2. 設定 LoRA 配置
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,                         # rank
    lora_alpha=32,                # scaling factor = alpha / r
    lora_dropout=0.05,
    target_modules=["q_proj", "v_proj"],  # 只加在 Q 和 V 投影矩陣上
    bias="none",
)

# 3. 套用 LoRA（凍結原始權重，只有 LoRA 矩陣可訓練）
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 4,194,304 || all params: 6,742,609,920 || trainable%: 0.0622%

# 4. 正常訓練...
# trainer = Trainer(model=model, ...)

# 5. 推理前合併（消除推理開銷）
model = model.merge_and_unload()  # W' = W₀ + BA，回到普通模型

from transformers import BitsAndBytesConfig
from peft import prepare_model_for_kbit_training

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,   # 雙重量化
    bnb_4bit_quant_type="nf4",        # NF4 格式
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-70b-hf",
    quantization_config=bnb_config,
    device_map="auto",
)

# 為 k-bit 訓練做準備（凍結非 LoRA 層、設定 gradient checkpointing）
model = prepare_model_for_kbit_training(model)
# 接著套用 LoRA config，同前...

from transformers import (
    WhisperForConditionalGeneration, WhisperProcessor,
    Seq2SeqTrainer, Seq2SeqTrainingArguments
)
from peft import get_peft_model, LoraConfig, TaskType
import torch

# ── 1. 載入模型與處理器 ──────────────────────────────────
model = WhisperForConditionalGeneration.from_pretrained(
    "openai/whisper-small",          # 可換成 medium / large
    torch_dtype=torch.float16,
)
processor = WhisperProcessor.from_pretrained("openai/whisper-small")

# ── 2. 套用 LoRA（只加在 Decoder 的 Q/V 投影上）──────────
lora_config = LoraConfig(
    r=8,
    lora_alpha=16,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type=TaskType.SEQ_2_SEQ_LM,
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 884,736 / 241,734,912 (0.37%)

# ── 3. 準備數據集（假設已有 audio + transcript 對）────────
def preprocess(batch):
    audio = batch["audio"]["array"]
    # 轉換音頻為 Mel 頻譜輸入特徵
    inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
    batch["input_features"] = inputs.input_features[0]
    # Tokenize 文字標籤
    batch["labels"] = processor.tokenizer(batch["sentence"]).input_ids
    return batch

# ── 4. 訓練 ─────────────────────────────────────────────
training_args = Seq2SeqTrainingArguments(
    output_dir="./whisper-lora-finetuned",
    per_device_train_batch_size=8,
    num_train_epochs=3,
    learning_rate=1e-3,          # LoRA 通常用比全量微調更高的 lr
    fp16=True,
    predict_with_generate=True,
)
trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
)
trainer.train()

# ── 5. 合併 LoRA 權重並儲存 ─────────────────────────────
model = model.merge_and_unload()
model.save_pretrained("./whisper-lora-merged")