BentoML

BentoML is a modern, open-source framework for building, shipping, and scaling AI applications. It bridges the gap between ML experimentation and production deployment, letting you package any model from any framework into a production-ready API service in minutes. Run BentoML on Clore.ai's GPU cloud for cost-efficient AI application hosting.

---

What is BentoML?

BentoML makes it easy to take a trained model and turn it into a scalable API service:

• Framework-agnostic: PyTorch, TensorFlow, JAX, scikit-learn, HuggingFace, XGBoost, LightGBM, and more

• Bento: A self-contained, reproducible artifact (model + code + dependencies)

• Runner: Scalable model inference unit with automatic batching

• Service: FastAPI-like HTTP/gRPC service definition

• BentoCloud: Optional managed deployment platform

• Docker-first: Every Bento can be containerized with one command

Key features:

• Adaptive micro-batching for throughput optimization

• Built-in input/output validation with Pydantic

• OpenAPI spec auto-generated

• Prometheus metrics built-in

• Streaming response support (LLMs)

---

Prerequisites

Requirement	Minimum	Recommended
GPU VRAM	8 GB	16–24 GB
GPU	Any NVIDIA	RTX 4090 / A100
RAM	8 GB	16 GB
Storage	20 GB	40 GB
Python	3.9+	3.11+

---

Step 1 — Rent a GPU on Clore.ai

1. Log in to clore.ai.

2. Click Marketplace and select a GPU instance with ≥ 16 GB VRAM.

3. Set Docker image: we'll use a custom build (see Step 2).

4. Set open ports: 22 (SSH) and 3000 (BentoML service).

5. Click Rent.

---

Step 2 — Dockerfile

BentoML doesn't have an official GPU Docker image, so we build one:

FROM pytorch/pytorch:2.11.0-cuda12.8-cudnn9-runtime

ENV DEBIAN_FRONTEND=noninteractive

RUN apt-get update && apt-get install -y \
    git wget curl \
    openssh-server \
    libgl1 libglib2.0-0 \
    && rm -rf /var/lib/apt/lists/*

# Configure SSH
RUN mkdir /var/run/sshd && \
    echo 'root:clore123' | chpasswd && \
    sed -i 's/#PermitRootLogin prohibit-password/PermitRootLogin yes/' /etc/ssh/sshd_config

# Install BentoML and common ML libs
RUN pip install --upgrade pip && \
    pip install \
        bentoml \
        transformers \
        accelerate \
        diffusers \
        Pillow \
        numpy \
        scipy \
        tritonclient[all]

WORKDIR /workspace

EXPOSE 22 3000

CMD service ssh start && tail -f /dev/null

Build and Push

Build the image and push it to your own Docker Hub account (replace YOUR_DOCKERHUB_USERNAME with your actual username):

docker build -t YOUR_DOCKERHUB_USERNAME/bentoml-gpu:latest .
docker push YOUR_DOCKERHUB_USERNAME/bentoml-gpu:latest

BentoML does not provide an official GPU Docker image on Docker Hub. The bentoml/bento-server images on Docker Hub are for serving pre-packaged Bentos and do not include CUDA support. Build the image from the Dockerfile above for GPU-enabled deployments on Clore.ai.

---

Step 3 — Connect via SSH

ssh root@<clore-host> -p <assigned-ssh-port>

Verify BentoML:

bentoml --version
# Expected: bentoml, version 1.x.x

---

Step 4 — Your First BentoML Service

Simple Text Classifier

Create a service file:

mkdir -p /workspace/my-service
cat > /workspace/my-service/service.py << 'EOF'
import bentoml
from bentoml.io import JSON, Text
import numpy as np

# Define a Runner (the model unit)
class TextClassifierRunnable(bentoml.Runnable):
    SUPPORTED_RESOURCES = ("gpu", "cpu")
    SUPPORTS_CPU_MULTI_THREADING = True
    
    def __init__(self):
        import torch
        from transformers import pipeline
        
        self.classifier = pipeline(
            "text-classification",
            model="distilbert-base-uncased-finetuned-sst-2-english",
            device=0 if torch.cuda.is_available() else -1,
        )
    
    @bentoml.Runnable.method(batchable=True, batch_dim=0)
    def classify(self, texts: list[str]) -> list[dict]:
        results = self.classifier(texts)
        return results

# Create Runner
classifier_runner = bentoml.Runner(
    TextClassifierRunnable,
    name="text_classifier",
    max_batch_size=32,
    max_latency_ms=100,
)

# Define Service
svc = bentoml.Service(
    name="text_classifier_service",
    runners=[classifier_runner],
)

@svc.api(input=Text(), output=JSON())
async def classify(text: str) -> dict:
    """Classify sentiment of input text."""
    results = await classifier_runner.classify.async_run([text])
    return results[0]
EOF

Start the Service

cd /workspace/my-service

bentoml serve service:svc \
    --host 0.0.0.0 \
    --port 3000 \
    --reload

The --reload flag enables hot-reload during development. Remove it in production for stability.

---

Step 5 — Access the Service

Open the auto-generated Swagger UI:

http://<clore-host>:<public-port-3000>

Or test via curl:

curl -X POST http://<clore-host>:<public-port-3000>/classify \
    -H "Content-Type: text/plain" \
    -d "This GPU cloud service is amazing!"

Expected response:

{"label": "POSITIVE", "score": 0.9986}

---

Step 6 — Image Classification Service

Vision Model Service

# /workspace/vision-service/service.py
import bentoml
from bentoml.io import Image, JSON
from PIL import Image as PILImage
import numpy as np

class ImageClassifierRunnable(bentoml.Runnable):
    SUPPORTED_RESOURCES = ("gpu",)
    SUPPORTS_CPU_MULTI_THREADING = False
    
    def __init__(self):
        import torch
        import torchvision.transforms as transforms
        from torchvision.models import resnet50, ResNet50_Weights
        
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        weights = ResNet50_Weights.DEFAULT
        self.model = resnet50(weights=weights).to(self.device)
        self.model.eval()
        self.preprocess = weights.transforms()
        self.categories = weights.meta["categories"]
    
    @bentoml.Runnable.method(batchable=True, batch_dim=0)
    def predict(self, images: list) -> list[dict]:
        import torch
        
        batch = torch.stack([self.preprocess(img) for img in images]).to(self.device)
        
        with torch.no_grad():
            predictions = self.model(batch).softmax(dim=1)
        
        results = []
        for pred in predictions:
            top5 = pred.topk(5)
            results.append({
                "predictions": [
                    {"label": self.categories[idx], "score": round(score.item(), 4)}
                    for score, idx in zip(top5.values, top5.indices)
                ]
            })
        return results


image_runner = bentoml.Runner(
    ImageClassifierRunnable,
    name="image_classifier",
    max_batch_size=16,
)

svc = bentoml.Service(
    name="image_classifier_service",
    runners=[image_runner],
)

@svc.api(input=Image(), output=JSON())
async def classify(image: PILImage.Image) -> dict:
    """Classify an image with ResNet50."""
    results = await image_runner.predict.async_run([image])
    return results[0]

bentoml serve service:svc --host 0.0.0.0 --port 3000

Test with an image:

curl -X POST http://<clore-host>:<public-port-3000>/classify \
    -H "Content-Type: image/jpeg" \
    --data-binary @/path/to/image.jpg

---

Step 7 — LLM Streaming Service

For language models with streaming responses:

# /workspace/llm-service/service.py
import bentoml
from bentoml.io import JSON, Text
from typing import AsyncGenerator

class LLMRunnable(bentoml.Runnable):
    SUPPORTED_RESOURCES = ("gpu",)
    SUPPORTS_CPU_MULTI_THREADING = False
    
    def __init__(self):
        from transformers import AutoModelForCausalLM, AutoTokenizer
        import torch
        
        model_name = "microsoft/phi-2"
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForCausalLM.from_pretrained(
            model_name,
            torch_dtype=torch.float16,
            device_map="auto"
        )
    
    @bentoml.Runnable.method(batchable=False)
    def generate(self, prompt: str, max_tokens: int = 200) -> str:
        import torch
        
        inputs = self.tokenizer(prompt, return_tensors="pt").to("cuda")
        
        with torch.no_grad():
            outputs = self.model.generate(
                **inputs,
                max_new_tokens=max_tokens,
                do_sample=True,
                temperature=0.7,
                pad_token_id=self.tokenizer.eos_token_id,
            )
        
        return self.tokenizer.decode(outputs[0], skip_special_tokens=True)


llm_runner = bentoml.Runner(LLMRunnable, name="llm")

svc = bentoml.Service("llm_service", runners=[llm_runner])

@svc.api(input=JSON(), output=Text())
async def generate(body: dict) -> str:
    prompt = body.get("prompt", "")
    max_tokens = body.get("max_tokens", 200)
    return await llm_runner.generate.async_run(prompt, max_tokens)

---

Step 8 — Save and Build a Bento

A Bento is a packaged, reproducible artifact:

# /workspace/build_bento.py
import bentoml

# Save model to BentoML model store
import torch
from torchvision.models import resnet50, ResNet50_Weights

model = resnet50(weights=ResNet50_Weights.DEFAULT)
model.eval()

saved_model = bentoml.pytorch.save_model(
    name="resnet50",
    model=model,
    labels={"framework": "pytorch", "task": "image-classification"},
    metadata={"accuracy": 0.80, "dataset": "ImageNet"}
)
print(f"Model saved: {saved_model.tag}")

python /workspace/build_bento.py

# List saved models
bentoml models list

# Build a Bento (requires bentofile.yaml)
bentoml build

bentofile.yaml

service: "service:svc"
labels:
  owner: "ml-team"
  stage: "production"
include:
  - "*.py"
python:
  packages:
    - torch
    - torchvision
    - transformers
    - Pillow
    - numpy
docker:
  python_version: "3.11"
  cuda_version: "12.1"
  system_packages:
    - libgl1

bentoml build

# List built bentos
bentoml list

# Containerize
bentoml containerize image_classifier_service:latest \
    --image-tag YOUR_DOCKERHUB_USERNAME/my-bento:latest

---

Monitoring and Metrics

BentoML exposes Prometheus metrics at /metrics:

curl http://<clore-host>:<public-port-3000>/metrics

Key metrics:

# Request rate
bentoml_service_request_total{endpoint="classify", http_status_code="200"}
# Latency
bentoml_service_request_duration_seconds{endpoint="classify"}
# Runner throughput  
bentoml_runner_request_total{runner_name="image_classifier"}

---

Adaptive Batching Configuration

# Fine-tune batching behavior
image_runner = bentoml.Runner(
    ImageClassifierRunnable,
    name="image_classifier",
    max_batch_size=64,          # Max requests per batch
    max_latency_ms=50,          # Max wait before dispatching
)

---

Troubleshooting

Service Won't Start

ERROR - Failed to initialize runner

Solutions:

• Check CUDA availability: python -c "import torch; print(torch.cuda.is_available())"

• Verify GPU VRAM: nvidia-smi

• Check model download completed (look for download progress in logs)

Port 3000 Not Accessible

# Ensure service binds to 0.0.0.0 (not localhost)
bentoml serve service:svc --host 0.0.0.0 --port 3000

High Latency on First Request

This is normal — the first request triggers model loading (warm-up). All subsequent requests will be fast. Add a warm-up endpoint call after start:

# Warm up after starting
sleep 10 && curl -s -o /dev/null http://localhost:3000/healthz

Import Errors

ModuleNotFoundError: No module named 'transformers'

Solution:

pip install transformers accelerate

---

Clore.ai GPU Recommendations

BentoML is a serving framework — GPU requirements depend entirely on the model you deploy. Here's what to expect for common workloads:

GPU	VRAM	Clore.ai Price	LLM (7B Q4) Throughput	Diffusion (SDXL)	Vision (ResNet50)
RTX 3090	24 GB	$0.07–0.21/hr	~80 tok/s	~4 img/min	~400 req/s
RTX 4090	24 GB	$0.14–0.42/hr	~140 tok/s	~8 img/min	~700 req/s
A100 40GB	40 GB	bare metal	~110 tok/s	~6 img/min	~1200 req/s
A100 80GB	80 GB	bare metal	~130 tok/s	~7 img/min	~1400 req/s

Use case guidance:

• LLM API serving (7B–13B): RTX 3090 ($0.07–0.21/hr) — optimal price-performance

• Image generation APIs: RTX 3090 or RTX 4090 depending on throughput needs

• Large models (34B–70B Q4): A100 40GB (bare metal) — fits comfortably

• Production multi-model serving: A100 80GB for memory headroom

BentoML's adaptive micro-batching is particularly effective on A100s — the hardware scheduler handles batching efficiently, extracting more throughput per dollar than naive single-request serving. For high-traffic APIs, A100 40GB often delivers better ROI than two RTX 4090s.

---

Useful Resources

• BentoML Official Documentation

• BentoML GitHub

• BentoML Examples

• BentoML Discord Community

• BentoML Gallery

• Quickstart: Serving LLMs