Cloud Run Jobs'ı kullanarak LLM'de ince ayar yapma
Bu codelab hakkında
subjectSon güncelleme Mar 21, 2025
1. Giriş
Genel Bakış
Bu örnekte, doğal dilde soru sorulduğunda LLM'nin SQL sorgusuyla yanıt vermesini sağlamak için bir gemma-2b modelinde metinden SQL'ye dönüştürme veri kümesiyle hassas ayar yapacaksınız. Ardından, hassas ayarlanmış modeli alıp vLLM'yi kullanarak Cloud Run'da sunarsınız.
Neler öğreneceksiniz?
- Cloud Run Jobs GPU'yu kullanarak hassas ayarlama yapma
- Modelin daha hızlı yüklenmesi ve sunulması için GPU işi için doğrudan VPC yapılandırması nasıl kullanılır?
2. Başlamadan önce
GPU'lar özelliğini kullanmak için desteklenen bir bölge için kota artışı isteğinde bulunmanız gerekir. Gerekli kota, Cloud Run Admin API kapsamındaki nvidia_l4_gpu_allocation_no_zonal_redundancy'dir. Kota isteğinde bulunmaya yönelik doğrudan bağlantıyı burada bulabilirsiniz.
3. Kurulum ve Gereksinimler
Bu codelab boyunca kullanılacak ortam değişkenlerini ayarlayın.
PROJECT_ID=<YOUR_PROJECT_ID>
REGION=<YOUR_REGION>
HF_TOKEN=<YOUR_HF_TOKEN>
AR_REPO=codelab-finetuning-jobs
IMAGE_NAME=finetune-to-gcs
JOB_NAME=finetuning-to-gcs-job
BUCKET_NAME=$PROJECT_ID-codelab-finetuning-jobs
SECRET_ID=HF_TOKEN
SERVICE_ACCOUNT="finetune-job-sa"
SERVICE_ACCOUNT_ADDRESS=$SERVICE_ACCOUNT@$PROJECT_ID.iam.gserviceaccount.com
Aşağıdaki komutu çalıştırarak hizmet hesabını oluşturun:
gcloud iam service-accounts create $SERVICE_ACCOUNT \
--display-name="Cloud Run job to access HF_TOKEN Secret ID"
HuggingFace erişim jetonunu depolamak için Secret Manager'ı kullanın.
Gizli anahtar oluşturma ve kullanma hakkında daha fazla bilgiyi Gizli Anahtar Yöneticisi dokümanlarından edinebilirsiniz.
gcloud secrets create $SECRET_ID \
--replication-policy="automatic"
printf $HF_TOKEN | gcloud secrets versions add $SECRET_ID --data-file=-
Aşağıdakine benzer bir çıkış görürsünüz:
you'll see output similar to
Created secret [HF_TOKEN].
Created version [1] of the secret [HF_TOKEN].
Varsayılan Compute hizmet hesabınıza Secret Manager Gizli Anahtar Erişimi rolünü verme
gcloud secrets add-iam-policy-binding $SECRET_ID \
--member serviceAccount:$SERVICE_ACCOUNT_ADDRESS \
--role='roles/secretmanager.secretAccessor'
İnce ayarlanmış modelinizi barındıracak bir paket oluşturun
gsutil mb -l $REGION gs://$BUCKET_NAME
Ardından SA'ya pakete erişim izni verin.
gcloud storage buckets add-iam-policy-binding gs://$BUCKET_NAME \
--member=serviceAccount:$SERVICE_ACCOUNT_ADDRESS \
--role=roles/storage.objectAdmin
İş için bir yapı kayıt arşivi deposu oluşturun
gcloud artifacts repositories create $AR_REPO \
--repository-format=docker \
--location=$REGION \
--description="codelab for finetuning using CR jobs" \
--project=$PROJECT_ID
4. Cloud Run iş resmini oluşturma
Sonraki adımda, aşağıdakileri yapan kodu oluşturacaksınız:
- huggingface'tan gemma-2b'yi içe aktarır.
- huggingface'taki veri kümesini kullanarak metinden SQL'e veri kümesiyle gemma-2b üzerinde hassas ayarlama yapar. İş, hassas ayar için tek bir L4 GPU kullanır.
- new_model adlı hassas ayarlanmış modeli kullanıcının GCS paketine yükler
İnce ayar iş kodunuz için bir dizin oluşturun.
mkdir codelab-finetuning-job
cd codelab-finetuning-job
finetune.py adlı bir dosya oluşturun
# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import torch
from datasets import load_dataset, Dataset
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
BitsAndBytesConfig,
TrainingArguments,
)
from peft import LoraConfig, PeftModel
from trl import SFTTrainer
from pathlib import Path
# GCS bucket to upload the model
bucket_name = os.getenv("BUCKET_NAME", "YOUR_BUCKET_NAME")
# The model that you want to train from the Hugging Face hub
model_name = os.getenv("MODEL_NAME", "google/gemma-2b")
# The instruction dataset to use
dataset_name = "b-mc2/sql-create-context"
# Fine-tuned model name
new_model = os.getenv("NEW_MODEL", "gemma-2b-sql")
################################################################################
# QLoRA parameters
################################################################################
# LoRA attention dimension
lora_r = int(os.getenv("LORA_R", "4"))
# Alpha parameter for LoRA scaling
lora_alpha = int(os.getenv("LORA_ALPHA", "8"))
# Dropout probability for LoRA layers
lora_dropout = 0.1
################################################################################
# bitsandbytes parameters
################################################################################
# Activate 4-bit precision base model loading
use_4bit = True
# Compute dtype for 4-bit base models
bnb_4bit_compute_dtype = "float16"
# Quantization type (fp4 or nf4)
bnb_4bit_quant_type = "nf4"
# Activate nested quantization for 4-bit base models (double quantization)
use_nested_quant = False
################################################################################
# TrainingArguments parameters
################################################################################
# Output directory where the model predictions and checkpoints will be stored
output_dir = "./results"
# Number of training epochs
num_train_epochs = 1
# Enable fp16/bf16 training (set bf16 to True with an A100)
fp16 = True
bf16 = False
# Batch size per GPU for training
per_device_train_batch_size = int(os.getenv("TRAIN_BATCH_SIZE", "1"))
# Batch size per GPU for evaluation
per_device_eval_batch_size = int(os.getenv("EVAL_BATCH_SIZE", "2"))
# Number of update steps to accumulate the gradients for
gradient_accumulation_steps = int(os.getenv("GRADIENT_ACCUMULATION_STEPS", "1"))
# Enable gradient checkpointing
gradient_checkpointing = True
# Maximum gradient normal (gradient clipping)
max_grad_norm = 0.3
# Initial learning rate (AdamW optimizer)
learning_rate = 2e-4
# Weight decay to apply to all layers except bias/LayerNorm weights
weight_decay = 0.001
# Optimizer to use
optim = "paged_adamw_32bit"
# Learning rate schedule
lr_scheduler_type = "cosine"
# Number of training steps (overrides num_train_epochs)
max_steps = -1
# Ratio of steps for a linear warmup (from 0 to learning rate)
warmup_ratio = 0.03
# Group sequences into batches with same length
# Saves memory and speeds up training considerably
group_by_length = True
# Save checkpoint every X updates steps
save_steps = 0
# Log every X updates steps
logging_steps = int(os.getenv("LOGGING_STEPS", "50"))
################################################################################
# SFT parameters
################################################################################
# Maximum sequence length to use
max_seq_length = int(os.getenv("MAX_SEQ_LENGTH", "512"))
# Pack multiple short examples in the same input sequence to increase efficiency
packing = False
# Load the entire model on the GPU 0
device_map = {'':torch.cuda.current_device()}
# Set limit to a positive number
limit = int(os.getenv("DATASET_LIMIT", "5000"))
dataset = load_dataset(dataset_name, split="train")
if limit != -1:
dataset = dataset.shuffle(seed=42).select(range(limit))
def transform(data):
question = data['question']
context = data['context']
answer = data['answer']
template = "Question: {question}\nContext: {context}\nAnswer: {answer}"
return {'text': template.format(question=question, context=context, answer=answer)}
transformed = dataset.map(transform)
# Load tokenizer and model with QLoRA configuration
compute_dtype = getattr(torch, bnb_4bit_compute_dtype)
bnb_config = BitsAndBytesConfig(
load_in_4bit=use_4bit,
bnb_4bit_quant_type=bnb_4bit_quant_type,
bnb_4bit_compute_dtype=compute_dtype,
bnb_4bit_use_double_quant=use_nested_quant,
)
# Check GPU compatibility with bfloat16
if compute_dtype == torch.float16 and use_4bit:
major, _ = torch.cuda.get_device_capability()
if major >= 8:
print("=" * 80)
print("Your GPU supports bfloat16")
print("=" * 80)
# Load base model
# model = AutoModelForCausalLM.from_pretrained("google/gemma-7b")
model = AutoModelForCausalLM.from_pretrained(
model_name,
quantization_config=bnb_config,
device_map=device_map,
torch_dtype=torch.float16,
)
model.config.use_cache = False
model.config.pretraining_tp = 1
# Load LLaMA tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right" # Fix weird overflow issue with fp16 training
# Load LoRA configuration
peft_config = LoraConfig(
lora_alpha=lora_alpha,
lora_dropout=lora_dropout,
r=lora_r,
bias="none",
task_type="CAUSAL_LM",
target_modules=["q_proj", "v_proj"]
)
# Set training parameters
training_arguments = TrainingArguments(
output_dir=output_dir,
num_train_epochs=num_train_epochs,
per_device_train_batch_size=per_device_train_batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
optim=optim,
save_steps=save_steps,
logging_steps=logging_steps,
learning_rate=learning_rate,
weight_decay=weight_decay,
fp16=fp16,
bf16=bf16,
max_grad_norm=max_grad_norm,
max_steps=max_steps,
warmup_ratio=warmup_ratio,
group_by_length=group_by_length,
lr_scheduler_type=lr_scheduler_type,
)
trainer = SFTTrainer(
model=model,
train_dataset=transformed,
peft_config=peft_config,
dataset_text_field="text",
max_seq_length=max_seq_length,
tokenizer=tokenizer,
args=training_arguments,
packing=packing,
)
trainer.train()
trainer.model.save_pretrained(new_model)
# Reload model in FP16 and merge it with LoRA weights
base_model = AutoModelForCausalLM.from_pretrained(
model_name,
low_cpu_mem_usage=True,
return_dict=True,
torch_dtype=torch.float16,
device_map=device_map,
)
model = PeftModel.from_pretrained(base_model, new_model)
model = model.merge_and_unload()
# Push to HF
# model.push_to_hub(new_model, check_pr=True)
# tokenizer.push_to_hub(new_model, check_pr=True)
# push to GCS
file_path_to_save_the_model = '/finetune/new_model'
model.save_pretrained(file_path_to_save_the_model)
tokenizer.save_pretrained(file_path_to_save_the_model)
requirements.txt dosyası oluşturun.
accelerate==0.30.1
bitsandbytes==0.43.1
datasets==2.19.1
transformers==4.41.0
peft==0.11.1
trl==0.8.6
torch==2.3.0
Dockerfile oluşturun
FROM nvidia/cuda:12.6.2-runtime-ubuntu22.04
RUN apt-get update && \
apt-get -y --no-install-recommends install python3-dev gcc python3-pip git && \
rm -rf /var/lib/apt/lists/*
COPY requirements.txt /requirements.txt
RUN pip3 install -r requirements.txt --no-cache-dir
COPY finetune.py /finetune.py
ENV PYTHONUNBUFFERED 1
CMD python3 /finetune.py --device cuda
Artifact Registry deponuzda kapsayıcıyı derleyin
gcloud builds submit --tag $REGION-docker.pkg.dev/$PROJECT_ID/$AR_REPO/$IMAGE_NAME
5. İşi dağıtma ve yürütme
Bu adımda, Google Cloud Storage'a daha hızlı yüklemeler yapmak için doğrudan VPC çıkışı içeren Jobs YAML yapılandırmasını oluşturacaksınız.
Bu dosyanın, sonraki bir adımda güncelleyeceğiniz değişkenler içerdiğini unutmayın.
Öncelikle finetune-job.yaml adlı bir dosya oluşturun.
apiVersion: run.googleapis.com/v1
kind: Job
metadata:
name: finetuning-to-gcs-job
labels:
cloud.googleapis.com/location: us-central1
annotations:
run.googleapis.com/launch-stage: ALPHA
spec:
template:
metadata:
annotations:
run.googleapis.com/execution-environment: gen2
run.googleapis.com/network-interfaces: '[{"network":"default","subnetwork":"default"}]'
spec:
parallelism: 1
taskCount: 1
template:
spec:
serviceAccountName: YOUR_SERVICE_ACCOUNT_NAME@YOUR_PROJECT_ID.iam.gserviceaccount.com
containers:
- name: finetune-to-gcs
image: YOUR_REGION-docker.pkg.dev/YOUR_PROJECT_ID/YOUR_AR_REPO/YOUR_IMAGE_NAME
env:
- name: MODEL_NAME
value: "google/gemma-2b"
- name: NEW_MODEL
value: "gemma-2b-sql-finetuned"
- name: LORA_R
value: "8"
- name: LORA_ALPHA
value: "16"
- name: TRAIN_BATCH_SIZE
value: "1"
- name: EVAL_BATCH_SIZE
value: "2"
- name: GRADIENT_ACCUMULATION_STEPS
value: "2"
- name: DATASET_LIMIT
value: "1000"
- name: MAX_SEQ_LENGTH
value: "512"
- name: LOGGING_STEPS
value: "5"
- name: HF_TOKEN
valueFrom:
secretKeyRef:
key: 'latest'
name: HF_TOKEN
resources:
limits:
cpu: 8000m
nvidia.com/gpu: '1'
memory: 32Gi
volumeMounts:
- mountPath: /finetune/new_model
name: finetuned_model
volumes:
- name: finetuned_model
csi:
driver: gcsfuse.run.googleapis.com
readOnly: false
volumeAttributes:
bucketName: YOUR_PROJECT_ID-codelab-finetuning-jobs
maxRetries: 3
timeoutSeconds: '3600'
nodeSelector:
run.googleapis.com/accelerator: nvidia-l4
Ardından, aşağıdaki komutu çalıştırarak yer tutucuları görüntünün ortam değişkenlerinizle değiştirin:
sed -i "s/YOUR_SERVICE_ACCOUNT_NAME/$SERVICE_ACCOUNT/; s/YOUR_PROJECT_ID/$PROJECT_ID/; s/YOUR_PROJECT_ID/$PROJECT_ID/; s/YOUR_REGION/$REGION/; s/YOUR_AR_REPO/$AR_REPO/; s/YOUR_IMAGE_NAME/$IMAGE_NAME/; s/YOUR_PROJECT_ID/$PROJECT_ID/" finetune-job.yaml
Ardından Cloud Run işini oluşturun
gcloud alpha run jobs replace finetune-job.yaml
Ve işi yürütün. Bu işlem yaklaşık 10 dakika sürer.
gcloud alpha run jobs execute $JOB_NAME --region $REGION
6. İnce ayarlanmış modelinizi vLLM ile yayınlamak için Cloud Run hizmeti kullanma
İnce ayarlanmış modeli sunacak Cloud Run hizmet kodu için bir klasör oluşturun
cd ..
mkdir codelab-finetuning-service
cd codelab-finetuning-service
service.yaml dosyası oluşturma
Bu yapılandırma, daha hızlı indirmeler için özel ağ üzerinden GCS paketine erişmek üzere doğrudan VPC kullanır.
Bu dosyanın, sonraki bir adımda güncelleyeceğiniz değişkenler içerdiğini unutmayın.
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: serve-gemma2b-sql
labels:
cloud.googleapis.com/location: us-central1
annotations:
run.googleapis.com/launch-stage: BETA
run.googleapis.com/ingress: all
run.googleapis.com/ingress-status: all
spec:
template:
metadata:
labels:
annotations:
autoscaling.knative.dev/maxScale: '5'
run.googleapis.com/cpu-throttling: 'false'
run.googleapis.com/network-interfaces: '[{"network":"default","subnetwork":"default"}]'
spec:
containers:
- name: serve-finetuned
image: us-docker.pkg.dev/vertex-ai/vertex-vision-model-garden-dockers/pytorch-vllm-serve:20240220_0936_RC01
ports:
- name: http1
containerPort: 8000
resources:
limits:
cpu: 8000m
nvidia.com/gpu: '1'
memory: 32Gi
volumeMounts:
- name: fuse
mountPath: /finetune/new_model
command: ["python3", "-m", "vllm.entrypoints.api_server"]
args:
- --model=/finetune/new_model
- --tensor-parallel-size=1
env:
- name: MODEL_ID
value: 'new_model'
- name: HF_HUB_OFFLINE
value: '1'
volumes:
- name: fuse
csi:
driver: gcsfuse.run.googleapis.com
volumeAttributes:
bucketName: YOUR_BUCKET_NAME
nodeSelector:
run.googleapis.com/accelerator: nvidia-l4
service.yaml dosyasını paketinizin adıyla güncelleyin.
sed -i "s/YOUR_BUCKET_NAME/$BUCKET_NAME/" service.yaml
Şimdi Cloud Run hizmetinizi dağıtın
gcloud alpha run services replace service.yaml
7. İnce ayarlanmış modelinizi test etme
Öncelikle Cloud Run hizmetinizin hizmet URL'sini alın.
SERVICE_URL=$(gcloud run services describe serve-gemma2b-sql --platform managed --region $REGION --format 'value(status.url)')
Modeliniz için isteminizi oluşturun.
USER_PROMPT="Question: What are the first name and last name of all candidates? Context: CREATE TABLE candidates (candidate_id VARCHAR); CREATE TABLE people (first_name VARCHAR, last_name VARCHAR, person_id VARCHAR)"
Şimdi hizmetinizi ör.
curl -X POST $SERVICE_URL/generate \
-H "Content-Type: application/json" \
-H "Authorization: bearer $(gcloud auth print-identity-token)" \
-d @- <<EOF
{
"prompt": "${USER_PROMPT}",
"temperature": 0.1,
"top_p": 1.0,
"max_tokens": 56
}
EOF
Aşağıdakine benzer bir yanıt görürsünüz:
{"predictions":["Prompt:\nQuestion: What are the first name and last name of all candidates? Context: CREATE TABLE candidates (candidate_id VARCHAR); CREATE TABLE people (first_name VARCHAR, last_name VARCHAR, person_id VARCHAR)\nOutput:\n CREATE TABLE people_to_candidates (candidate_id VARCHAR, person_id VARCHAR) CREATE TABLE people_to_people (person_id VARCHAR, person_id VARCHAR) CREATE TABLE people_to_people_to_candidates (person_id VARCHAR, candidate_id"]}