SEOKDONG/Qwen3.6-27B-QLoRA-Adaper

### model
model_name_or_path: Qwen/Qwen3.6-27B
trust_remote_code: true

### method
stage: sft
finetuning_type: lora
lora_target: all          # 모든 linear에 LoRA — 표현력 / 안정성 균형
lora_rank: 32
lora_alpha: 64            # alpha = 2 × rank (스케일 1/2)
lora_dropout: 0.05        # 약한 정규화

### quantization (QLoRA)
quantization_bit: 4
quantization_method: bnb
double_quantization: true
quantization_type: nf4

### dataset
template: qwen3_6
cutoff_len: 2048
packing: false            # 멀티턴 손실 왜곡 방지

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 16   # effective batch = 1 × 16 × 2 GPU = 32
learning_rate: 5.0e-5
num_train_epochs: 2.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true                       # fp16 금지 (학습 진동 방지)
max_grad_norm: 1.0               # 그래디언트 클리핑
weight_decay: 0.01
optim: adamw_torch
adam_beta2: 0.95                 # 손실 스파이크 완화
gradient_checkpointing: true     # VRAM 절감

### eval
eval_strategy: steps
eval_steps: 500

### acceleration
flash_attn: sdpa
enable_liger_kernel: true        # VRAM 절감 + throughput 향상

huggingface-cli download <YOUR_HF_ID>/<THIS_REPO_NAME> \
  --local-dir ./adapters/qwen3-ft

from huggingface_hub import snapshot_download

adapter_path = snapshot_download(
    repo_id="<YOUR_HF_ID>/<THIS_REPO_NAME>",
    local_dir="./adapters/qwen3-ft",
)
print(adapter_path)

#!/bin/bash
# Qwen3.6-27B-FP8 + Korean Enterprise QLoRA Adapter

ADAPTER_PATH="./adapters/qwen3-ft"
MODEL_NAME="Qwen/Qwen3.6-27B-FP8"      # FP8 베이스: ~27GB, BF16 대비 절반
LORA_MODULE_NAME="qwen3-ft"            # API 호출 시 model= 에 사용할 이름
PORT=8000
GPU_COUNT=2                            # RTX 5090 × 2 (단일 GPU도 가능)

source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate vllm

vllm serve "${MODEL_NAME}" \
    --enable-lora \
    --lora-modules "${LORA_MODULE_NAME}=${ADAPTER_PATH}" \
    --max-lora-rank 32 \
    --dtype auto \
    --kv-cache-dtype auto \
    --tensor-parallel-size ${GPU_COUNT} \
    --gpu-memory-utilization 0.90 \
    --max-model-len 8192 \
    --max-num-seqs 32 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_xml \
    --host 0.0.0.0 \
    --port ${PORT} \
    --trust-remote-code

# pip install openai
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",  # vLLM은 키 검증을 하지 않음
)

SYSTEM_PROMPT = (
    "당신은 한국 기업 환경에 정통한 업무 어시스턴트입니다. "
    "한국식 보고체계, 직급 관행, 정중한 비즈니스 톤을 유지하세요."
)

resp = client.chat.completions.create(
    model="qwen3-ft",   # ← 어댑터 적용. 베이스만 쓰려면 "Qwen/Qwen3.6-27B-FP8"
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content":
            "신규 SaaS 도입 품의서를 작성해 주세요. "
            "도입 배경, 기대효과, 예산(연 1.2억), 리스크, 결재선 포함."},
    ],
    temperature=0.3,
    top_p=0.9,
    max_tokens=1500,
)

print(resp.choices[0].message.content)

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

prompt = "팀장에게 일정 지연을 보고하는 정중한 메일을 작성해줘. 사유는 외주사 산출물 지연."

for model_name in ["Qwen/Qwen3.6-27B-FP8", "qwen3-ft"]:
    out = client.chat.completions.create(
        model=model_name,
        messages=[{"role": "user", "content": prompt}],
        temperature=0.2, max_tokens=600,
    )
    print(f"\n===== {model_name} =====")
    print(out.choices[0].message.content)

# pip install transformers peft accelerate bitsandbytes
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

BASE = "Qwen/Qwen3.6-27B"
ADAPTER = "<YOUR_HF_ID>/<THIS_REPO_NAME>"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tok = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    BASE, quantization_config=bnb, device_map="auto", trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, ADAPTER)   # 어댑터 결합 (머지 X)
model.eval()

msgs = [
    {"role": "system", "content": "당신은 한국 기업 환경에 정통한 어시스턴트입니다."},
    {"role": "user", "content": "주간 업무보고 양식을 한국 대기업 스타일로 만들어줘."},
]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=800, do_sample=True, temperature=0.3, top_p=0.9)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

@misc{qwen3_27b_korean_enterprise_qlora,
  title  = {Qwen3.6-27B Korean Enterprise QLoRA Adapter},
  author = {<YOUR NAME>},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/<YOUR_HF_ID>/<THIS_REPO_NAME>}},
  note   = {QLoRA NF4 4-bit fine-tuning on ~360K Korean social/enterprise instances, served via vLLM Multi-LoRA}
}