Cómo ejecutar la inferencia por lotes en GPUs con trabajos de Cloud Run
Acerca de este codelab
subjectÚltima actualización: mar 20, 2025
1. Introducción
Descripción general
En este ejemplo, ajustarás un modelo gemma-2b con un conjunto de datos de texto a SQL para que el LLM responda con una consulta de SQL cuando se le haga una pregunta en lenguaje natural. Luego, tomarás el modelo ajustado y lo publicarás en Cloud Run con vLLM.
Qué aprenderás
- Cómo realizar un ajuste fino con la GPU de trabajos de Cloud Run
- Cómo usar la configuración de VPC directa para un trabajo de GPU para subir y entregar el modelo más rápido
2. Antes de comenzar
Para usar la función de GPU, debes solicitar un aumento de cuota para una región admitida. La cuota necesaria es nvidia_l4_gpu_allocation_no_zonal_redundancy, que se encuentra en la API de Cloud Run Admin. Este es el vínculo directo para solicitar una cuota.
3. Configuración y requisitos
Establece las variables de entorno que se usarán en este codelab.
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
Ejecuta este comando para crear la cuenta de servicio:
gcloud iam service-accounts create $SERVICE_ACCOUNT \
--display-name="Cloud Run job to access HF_TOKEN Secret ID"
Usa Secret Manager para almacenar el token de acceso de HuggingFace.
Puedes obtener más información para crear y usar secretos en la documentación de Secret Manager.
gcloud secrets create $SECRET_ID \
--replication-policy="automatic"
printf $HF_TOKEN | gcloud secrets versions add $SECRET_ID --data-file=-
Verás un resultado similar al siguiente:
you'll see output similar to
Created secret [HF_TOKEN].
Created version [1] of the secret [HF_TOKEN].
Otorga a tu cuenta de servicio de procesamiento predeterminada el rol de descriptor de acceso a secretos de Secret Manager
gcloud secrets add-iam-policy-binding $SECRET_ID \
--member serviceAccount:$SERVICE_ACCOUNT_ADDRESS \
--role='roles/secretmanager.secretAccessor'
Crea un bucket que alojará tu modelo ajustado
gsutil mb -l $REGION gs://$BUCKET_NAME
Luego, otorga acceso al bucket al SA.
gcloud storage buckets add-iam-policy-binding gs://$BUCKET_NAME \
--member=serviceAccount:$SERVICE_ACCOUNT_ADDRESS \
--role=roles/storage.objectAdmin
Crea un repositorio de Artifact Registry para el trabajo
gcloud artifacts repositories create $AR_REPO \
--repository-format=docker \
--location=$REGION \
--description="codelab for finetuning using CR jobs" \
--project=$PROJECT_ID
Crea un bucket de Cloud Storage para el modelo ajustado
gsutil mb -l $REGION gs://$BUCKET_NAME
Por último, crea un repositorio de Artifact Registry para tu trabajo de Cloud Run.
gcloud artifacts repositories create $AR_REPO \
--repository-format=docker \
--location=$REGION \
--description="codelab for finetuning using cloud run jobs"
4. Crea la imagen del trabajo de Cloud Run
En el siguiente paso, crearás el código que hace lo siguiente:
- Importa gemma-2b desde huggingface
- Realiza un ajuste fino en gemma-2b con el conjunto de datos de texto a SQL usando el conjunto de datos de huggingface. La tarea usa una sola GPU L4 para el ajuste fino.
- Sube el modelo ajustado llamado new_model al bucket de GCS del usuario.
Crea un directorio para el código de trabajo de ajuste fino.
mkdir codelab-finetuning-job
cd codelab-finetuning-job
Crea un archivo llamado finetune.py.
# 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)
Crea un archivo requirements.txt.
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
Cómo crear un Dockerfile
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/*
RUN pip3 install -r requirements.txt --no-cache-dir
COPY finetune.py /finetune.py
ENV PYTHONUNBUFFERED 1
CMD python3 /finetune.py --device cuda
Compila el contenedor en tu repositorio de Artifact Registry
gcloud builds submit --tag $REGION-docker.pkg.dev/$PROJECT_ID/$AR_REPO/$IMAGE_NAME
5. Implementa y ejecuta el trabajo
En este paso, crearás la configuración de YAML de Jobs con salida de VPC directa para subir archivos más rápido a Google Cloud Storage.
Ten en cuenta que este archivo contiene variables que actualizarás en un paso posterior.
Primero, crea un archivo llamado finetune-job.yaml.
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_RPOJECT_ID-codelab-finetuning-jobs
maxRetries: 3
timeoutSeconds: '3600'
nodeSelector:
run.googleapis.com/accelerator: nvidia-l4
Ahora, ejecuta el siguiente comando para reemplazar los marcadores de posición por tus variables de entorno de la imagen:
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
A continuación, crea el trabajo de Cloud Run
gcloud alpha run jobs replace finetune-job.yaml
Y ejecuta el trabajo. Este proceso tardará aproximadamente 10 minutos.
gcloud alpha run jobs execute $JOB_NAME --region $REGION
6. Usa un servicio de Cloud Run para publicar tu modelo ajustado con vLLM
Crea una carpeta para el código del servicio de Cloud Run que publicará el modelo ajustado
cd ..
mkdir codelab-finetuning-service
cd codelab-finetuning-service
Crea un archivo service.yaml
Esta configuración usa una VPC directa para acceder al bucket de GCS a través de una red privada para obtener descargas más rápidas.
Ten en cuenta que este archivo contiene variables que actualizarás en un paso posterior.
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: serve-gemma2b-sql
labels:
cloud.googleapis.com/location: us-central1
annotations:
run.googleapis.com/launch-stage: BETA
run.googleapis.com/ingress: all
run.googleapis.com/ingress-status: all
spec:
template:
metadata:
labels:
annotations:
autoscaling.knative.dev/maxScale: '5'
run.googleapis.com/cpu-throttling: 'false'
run.googleapis.com/network-interfaces: '[{"network":"default","subnetwork":"default"}]'
spec:
containers:
- name: serve-finetuned
image: us-docker.pkg.dev/vertex-ai/vertex-vision-model-garden-dockers/pytorch-vllm-serve:20240220_0936_RC01
ports:
- name: http1
containerPort: 8000
resources:
limits:
cpu: 8000m
nvidia.com/gpu: '1'
memory: 32Gi
volumeMounts:
- name: fuse
mountPath: /finetune/new_model
command: ["python3", "-m", "vllm.entrypoints.api_server"]
args:
- --model=/finetune/new_model
- --tensor-parallel-size=1
env:
- name: MODEL_ID
value: 'new_model'
- name: HF_HUB_OFFLINE
value: '1'
volumes:
- name: fuse
csi:
driver: gcsfuse.run.googleapis.com
volumeAttributes:
bucketName: YOUR_BUCKET_NAME
nodeSelector:
run.googleapis.com/accelerator: nvidia-l4
Actualiza el archivo service.yaml con el nombre de tu bucket.
sed -i "s/YOUR_BUCKET_NAME/$BUCKET_NAME/" finetune-job.yaml
Ahora, implementa tu servicio de Cloud Run
gcloud alpha run services replace service.yaml
7. Prueba tu modelo ajustado
Primero, obtén la URL de servicio de tu servicio de Cloud Run.
SERVICE_URL=$(gcloud run services describe serve-gemma2b-sql --platform managed --region $REGION --format 'value(status.url)')
Crea la instrucción para tu modelo.
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)"
Ahora, ejecuta curl en tu servicio
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
Deberías ver una respuesta similar a la siguiente:
{"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"]}