Model Deployment

Complete guide for exporting, optimizing, and deploying fine-tuned LLMs to production environments.

Overview

After fine-tuning your model with Unsloth, deploy it efficiently:

• GGUF export - For llama.cpp, Ollama, local inference
• vLLM deployment - For high-throughput production serving
• HuggingFace Hub - For sharing and version control
• Quantization - Reduce size while maintaining quality
• Platform selection - Choose the right infrastructure
• Monitoring - Track performance and costs

Quick Start

Export to GGUF (Ollama/llama.cpp)

from unsloth import FastLanguageModel

# Load your fine-tuned model
model, tokenizer = FastLanguageModel.from_pretrained(
    "./fine_tuned_model",
    max_seq_length=2048
)

# Export to GGUF format
model.save_pretrained_gguf(
    "./gguf_output",
    tokenizer,
    quantization_method="q4_k_m"  # 4-bit quantization
)

# Use with Ollama
# ollama create my-model -f ./gguf_output/Modelfile
# ollama run my-model

Deploy with vLLM

from unsloth import FastLanguageModel

# Save for vLLM
model.save_pretrained("./vllm_model")
tokenizer.save_pretrained("./vllm_model")

# Start vLLM server
# python -m vllm.entrypoints.openai.api_server \
#   --model ./vllm_model \
#   --tensor-parallel-size 1 \
#   --dtype bfloat16

Push to HuggingFace Hub

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "./fine_tuned_model",
    max_seq_length=2048
)

# Push to Hub
model.push_to_hub(
    "your-username/model-name",
    token="hf_...",
    private=False
)
tokenizer.push_to_hub(
    "your-username/model-name",
    token="hf_..."
)

Export Formats

1. GGUF (llama.cpp / Ollama)

Best for: Local deployment, edge devices, CPU inference

# Export with different quantization levels
quantization_methods = {
    "q4_k_m": "4-bit, medium quality (recommended)",
    "q5_k_m": "5-bit, higher quality",
    "q8_0": "8-bit, near-original quality",
    "f16": "16-bit float, full quality",
    "f32": "32-bit float, highest quality"
}

# Export
model.save_pretrained_gguf(
    "./gguf_output",
    tokenizer,
    quantization_method="q4_k_m"
)

# Creates:
# - model-q4_k_m.gguf (quantized model)
# - Modelfile (for Ollama)

Use with Ollama:

# Create Ollama model
cd gguf_output
ollama create my-medical-model -f Modelfile

# Run
ollama run my-medical-model "What are the symptoms of pneumonia?"

# API server
ollama serve
# curl http://localhost:11434/api/generate -d '{"model": "my-medical-model", "prompt": "..."}'

Use with llama.cpp:

# Build llama.cpp
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make

# Run inference
./main -m ../gguf_output/model-q4_k_m.gguf -p "Your prompt here"

# Server mode
./server -m ../gguf_output/model-q4_k_m.gguf --host 0.0.0.0 --port 8080

2. vLLM (Production Serving)

Best for: High-throughput production, API serving, multi-user

# Prepare model for vLLM
model.save_pretrained("./vllm_model")
tokenizer.save_pretrained("./vllm_model")

# Optional: Merge LoRA weights into base model
model = FastLanguageModel.merge_and_unload(model)
model.save_pretrained("./vllm_model_merged")

Deploy vLLM Server:

# Single GPU
python -m vllm.entrypoints.openai.api_server \
  --model ./vllm_model \
  --dtype bfloat16 \
  --max-model-len 4096

# Multi-GPU (tensor parallelism)
python -m vllm.entrypoints.openai.api_server \
  --model ./vllm_model \
  --tensor-parallel-size 4 \
  --dtype bfloat16

# With quantization (AWQ)
python -m vllm.entrypoints.openai.api_server \
  --model ./vllm_model \
  --quantization awq \
  --dtype half

Use vLLM API:

import openai

# Configure client
openai.api_key = "EMPTY"
openai.api_base = "http://localhost:8000/v1"

# Generate
response = openai.Completion.create(
    model="./vllm_model",
    prompt="Your prompt here",
    max_tokens=512,
    temperature=0.7
)

print(response.choices[0].text)

3. HuggingFace Hub

Best for: Sharing, version control, collaboration

from unsloth import FastLanguageModel
from huggingface_hub import HfApi

# Load and push
model, tokenizer = FastLanguageModel.from_pretrained("./fine_tuned_model")

# Push to Hub
model.push_to_hub(
    "username/model-name",
    token="hf_...",
    private=True,  # or False for public
    commit_message="Initial upload of medical model"
)
tokenizer.push_to_hub("username/model-name", token="hf_...")

# Add model card
api = HfApi()
api.upload_file(
    path_or_fileobj="README.md",
    path_in_repo="README.md",
    repo_id="username/model-name",
    token="hf_..."
)

Download from Hub:

from unsloth import FastLanguageModel

# Anyone can now load your model
model, tokenizer = FastLanguageModel.from_pretrained(
    "username/model-name",
    max_seq_length=2048
)

4. Docker Deployment

Best for: Reproducible deployments, cloud platforms

Dockerfile for vLLM:

FROM vllm/vllm-openai:latest

# Copy model
COPY ./vllm_model /app/model

# Expose port
EXPOSE 8000

# Run server
CMD ["python", "-m", "vllm.entrypoints.openai.api_server", \
     "--model", "/app/model", \
     "--host", "0.0.0.0", \
     "--port", "8000"]

Build and run:

# Build
docker build -t my-model-server .

# Run
docker run -d \
  --gpus all \
  -p 8000:8000 \
  -v $(pwd)/vllm_model:/app/model \
  my-model-server

# Test
curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "/app/model", "prompt": "Hello", "max_tokens": 50}'

Quantization Strategies

Quantization Methods Comparison

GGUF Quantization

# Export multiple quantization levels
quantization_levels = ["q4_k_m", "q5_k_m", "q8_0"]

for quant in quantization_levels:
    model.save_pretrained_gguf(
        f"./gguf_output_{quant}",
        tokenizer,
        quantization_method=quant
    )
    print(f"Exported {quant}")

# Compare file sizes
# q4_k_m: ~4GB (7B model)
# q5_k_m: ~5GB
# q8_0: ~8GB

GPTQ Quantization

Best for: GPU inference with high throughput

from transformers import GPTQConfig

# Configure GPTQ
gptq_config = GPTQConfig(
    bits=4,
    dataset="c4",  # Calibration dataset
    tokenizer=tokenizer,
    group_size=128
)

# Quantize
quantized_model = model.quantize(gptq_config)

# Save
quantized_model.save_pretrained("./gptq_model")
tokenizer.save_pretrained("./gptq_model")

AWQ Quantization

Best for: vLLM deployment

from awq import AutoAWQForCausalLM

# Load model
model = AutoAWQForCausalLM.from_pretrained("./fine_tuned_model")

# Quantize
model.quantize(
    tokenizer,
    quant_config={
        "zero_point": True,
        "q_group_size": 128,
        "w_bit": 4
    }
)

# Save
model.save_quantized("./awq_model")
tokenizer.save_pretrained("./awq_model")

# Use with vLLM
# python -m vllm.entrypoints.openai.api_server \
#   --model ./awq_model --quantization awq

Deployment Platforms

Local Deployment

Pros: Full control, no API costs, data privacy Cons: Limited scale, hardware costs

# Ollama (easiest)
ollama create my-model -f Modelfile
ollama run my-model

# llama.cpp (most flexible)
./server -m model.gguf --host 0.0.0.0 --port 8080

# vLLM (best performance)
python -m vllm.entrypoints.openai.api_server --model ./model

Hardware Requirements:

Cloud Platforms

Modal

Best for: Serverless, pay-per-use

import modal

stub = modal.Stub("my-model")

@stub.function(
    image=modal.Image.debian_slim().pip_install("vllm"),
    gpu="A100",
    timeout=600
)
def generate(prompt: str) -> str:
    from vllm import LLM

    llm = LLM(model="./model")
    output = llm.generate(prompt)
    return output[0].outputs[0].text

# Deploy
# modal deploy app.py

Pricing: ~$1-3/hour A100, pay only for usage

RunPod

Best for: Persistent endpoints, GPU pods

# Deploy via RunPod UI or API
curl -X POST https://api.runpod.io/v2/endpoints \
  -H "Authorization: Bearer $RUNPOD_API_KEY" \
  -d '{
    "name": "my-model",
    "gpu_type": "RTX_4090",
    "docker_image": "vllm/vllm-openai:latest",
    "env": {
      "MODEL_NAME": "./model"
    }
  }'

Pricing: ~$0.30-0.50/hour RTX 4090, ~$1.50/hour A100

Vast.ai

Best for: Lowest cost, spot instances

# Search for instances
vastai search offers 'gpu_name=RTX_4090 num_gpus=1'

# Rent instance
vastai create instance <instance_id> \
  --image vllm/vllm-openai:latest \
  --env MODEL_NAME=./model

Pricing: ~$0.15-0.30/hour RTX 4090, ~$0.80/hour A100

AWS/GCP/Azure

Best for: Enterprise, compliance, scale

AWS SageMaker:

from sagemaker.huggingface import HuggingFaceModel

# Create model
huggingface_model = HuggingFaceModel(
    model_data="s3://bucket/model.tar.gz",
    role=role,
    transformers_version="4.37",
    pytorch_version="2.1",
    py_version="py310"
)

# Deploy
predictor = huggingface_model.deploy(
    initial_instance_count=1,
    instance_type="ml.g5.xlarge"
)

# Generate
result = predictor.predict({
    "inputs": "Your prompt here"
})

Pricing: ~$1-5/hour depending on instance type

Platform Comparison

Optimization Strategies

1. Merge LoRA Adapters

Before deployment, merge LoRA weights:

from unsloth import FastLanguageModel

# Load with LoRA
model, tokenizer = FastLanguageModel.from_pretrained(
    "./fine_tuned_model",
    max_seq_length=2048
)

# Merge LoRA into base weights
model = FastLanguageModel.merge_and_unload(model)

# Save merged model (no LoRA overhead)
model.save_pretrained("./merged_model")
tokenizer.save_pretrained("./merged_model")

Benefits:

• Faster inference (no adapter computation)
• Simpler deployment (single model file)
• Broader compatibility

2. Enable Flash Attention

# During model loading
model, tokenizer = FastLanguageModel.from_pretrained(
    "model-name",
    max_seq_length=2048,
    use_flash_attention_2=True  # 2-3x faster attention
)

# For vLLM deployment
# vLLM automatically uses flash attention if available

3. Batch Processing

For high throughput:

from vllm import LLM, SamplingParams

llm = LLM(model="./model")

# Batch prompts
prompts = [
    "Prompt 1",
    "Prompt 2",
    # ... up to 100s of prompts
]

# Generate in batch (much faster than sequential)
outputs = llm.generate(prompts, SamplingParams(temperature=0.7))

for output in outputs:
    print(output.outputs[0].text)

4. Continuous Batching

vLLM automatically does continuous batching:

# Just configure for optimal throughput
python -m vllm.entrypoints.openai.api_server \
  --model ./model \
  --max-num-batched-tokens 8192 \
  --max-num-seqs 256

Load Testing & Benchmarking

Benchmark Inference Speed

import time
from vllm import LLM

llm = LLM(model="./model")

# Test prompts
prompts = ["Test prompt"] * 100

# Benchmark
start = time.time()
outputs = llm.generate(prompts)
end = time.time()

total_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)
tokens_per_sec = total_tokens / (end - start)

print(f"Throughput: {tokens_per_sec:.2f} tokens/sec")

Load Testing with Locust

from locust import HttpUser, task, between

class ModelUser(HttpUser):
    wait_time = between(1, 3)

    @task
    def generate(self):
        self.client.post("/v1/completions", json={
            "model": "./model",
            "prompt": "What is the capital of France?",
            "max_tokens": 50
        })

# Run: locust -f loadtest.py --host http://localhost:8000

Performance Targets

Monitoring & Observability

Basic Monitoring

import prometheus_client
from prometheus_client import Counter, Histogram

# Metrics
REQUEST_COUNT = Counter('model_requests_total', 'Total requests')
REQUEST_DURATION = Histogram('model_request_duration_seconds', 'Request duration')
TOKENS_GENERATED = Counter('model_tokens_generated_total', 'Total tokens')

# Instrument your endpoint
@REQUEST_DURATION.time()
def generate(prompt: str):
    REQUEST_COUNT.inc()
    output = model.generate(prompt)
    TOKENS_GENERATED.inc(len(output.token_ids))
    return output

# Expose metrics
prometheus_client.start_http_server(9090)

vLLM Metrics

vLLM exposes metrics automatically:

curl http://localhost:8000/metrics

# Key metrics:
# - vllm:num_requests_running
# - vllm:num_requests_waiting
# - vllm:gpu_cache_usage_perc
# - vllm:time_to_first_token_seconds
# - vllm:time_per_output_token_seconds

Cost Tracking

class CostTracker:
    def __init__(self, cost_per_hour: float):
        self.cost_per_hour = cost_per_hour
        self.start_time = time.time()
        self.total_tokens = 0

    def track_generation(self, num_tokens: int):
        self.total_tokens += num_tokens

    def get_stats(self):
        hours = (time.time() - self.start_time) / 3600
        total_cost = hours * self.cost_per_hour
        cost_per_1k_tokens = (total_cost / self.total_tokens) * 1000

        return {
            'total_cost': total_cost,
            'total_tokens': self.total_tokens,
            'cost_per_1k_tokens': cost_per_1k_tokens,
            'tokens_per_dollar': self.total_tokens / total_cost
        }

# Usage
tracker = CostTracker(cost_per_hour=1.50)  # A100 pricing
tracker.track_generation(512)
print(tracker.get_stats())

Common Deployment Patterns

Pattern 1: Quick Local Demo

# Export to GGUF
python export_gguf.py

# Run with Ollama
ollama create my-demo -f Modelfile
ollama run my-demo

# Share demo
# Users just need: ollama pull username/my-demo

Pattern 2: Production API

# Merge LoRA weights
python merge_lora.py

# Quantize with AWQ
python quantize_awq.py

# Deploy with vLLM
docker run -d --gpus all -p 8000:8000 \
  -v $(pwd)/model:/model \
  vllm/vllm-openai:latest \
  --model /model --quantization awq

# Load balancer + monitoring
# nginx -> vLLM instances -> Prometheus/Grafana

Pattern 3: Multi-Model Serving

from vllm import LLM

# Load multiple models
models = {
    'medical': LLM(model="./medical_model"),
    'legal': LLM(model="./legal_model"),
    'general': LLM(model="./general_model")
}

# Route based on input
def route_and_generate(text: str, domain: str):
    model = models.get(domain, models['general'])
    return model.generate(text)

Pattern 4: Hybrid Deployment

# Small model locally, large model in cloud
class HybridInference:
    def __init__(self):
        self.local = LLM(model="./small_model")  # 3B
        self.cloud_endpoint = "https://api.cloud.com/large-model"

    def generate(self, prompt: str, complexity: str = 'auto'):
        # Simple queries -> local
        # Complex queries -> cloud
        if complexity == 'auto':
            complexity = self.estimate_complexity(prompt)

        if complexity == 'simple':
            return self.local.generate(prompt)
        else:
            return requests.post(self.cloud_endpoint, json={'prompt': prompt})

Troubleshooting

Issue: Out of Memory (OOM)

Solutions:

# 1. Use smaller quantization
model.save_pretrained_gguf("./output", tokenizer, quantization_method="q4_0")

# 2. Reduce max sequence length
python -m vllm.entrypoints.openai.api_server \
  --model ./model \
  --max-model-len 2048  # Instead of 4096

# 3. Enable CPU offloading
model, tokenizer = FastLanguageModel.from_pretrained(
    "./model",
    device_map="auto",  # Automatic CPU/GPU split
    offload_folder="./offload"
)

# 4. Use tensor parallelism (multi-GPU)
python -m vllm.entrypoints.openai.api_server \
  --model ./model \
  --tensor-parallel-size 2  # Split across 2 GPUs

Issue: Slow Inference

Solutions:

# 1. Enable flash attention
model, tokenizer = FastLanguageModel.from_pretrained(
    "./model",
    use_flash_attention_2=True
)

# 2. Use GPTQ/AWQ quantization (faster than GGUF on GPU)
# See quantization section above

# 3. Batch requests
# See batch processing section

# 4. Use vLLM instead of HuggingFace transformers
# vLLM is 10-20x faster for serving

Issue: Model Quality Degradation

Solutions:

# 1. Use higher quantization
# q4_k_m -> q5_k_m -> q8_0

# 2. Don't quantize twice
# If model is already quantized (e.g., bnb-4bit), export to f16 or f32

# 3. Test quantization quality
def test_quantization(original_model, quantized_model, test_prompts):
    results = []
    for prompt in test_prompts:
        orig_out = original_model.generate(prompt)
        quant_out = quantized_model.generate(prompt)
        similarity = calculate_similarity(orig_out, quant_out)
        results.append(similarity)
    return np.mean(results)

# Target: >90% similarity for production use

Issue: High Latency

Solutions:

# 1. Use smaller model
# 7B instead of 13B often has similar quality with 2x lower latency

# 2. Reduce max_tokens
# Lower max_tokens = faster generation

# 3. Use local deployment
# Eliminates network latency

# 4. Optimize GPU settings
python -m vllm.entrypoints.openai.api_server \
  --model ./model \
  --gpu-memory-utilization 0.95 \
  --max-num-batched-tokens 8192

Best Practices

1. Test Before Production

# Always test quantized models
test_prompts = load_test_prompts()

original = LLM(model="./fine_tuned_model")
quantized = LLM(model="./quantized_model")

for prompt in test_prompts:
    orig_out = original.generate(prompt)
    quant_out = quantized.generate(prompt)

    # Compare quality
    print(f"Original: {orig_out}")
    print(f"Quantized: {quant_out}")
    print(f"Similarity: {calculate_similarity(orig_out, quant_out)}")

2. Version Your Models

models/
├── medical-v1.0.0/
│   ├── full/           # Full precision
│   ├── q4_k_m/        # 4-bit GGUF
│   ├── awq/           # AWQ quantized
│   └── README.md      # Model card
├── medical-v1.1.0/
└── production -> medical-v1.0.0/  # Symlink to deployed version

3. Monitor Everything

• Latency (P50, P95, P99)
• Throughput (tokens/sec)
• Error rate
• GPU utilization
• Cost per request
• Quality metrics (if available)

4. Start Small, Scale Up

1. Local testing (Ollama/llama.cpp)
2. Cloud trial (Modal/RunPod single instance)
3. Production (vLLM with load balancer)
4. Scale (Multi-GPU, multi-region)

5. Document Everything

Create a deployment README:

# Model Deployment Guide

## Model Details

- Base: Llama-3.2-7B
- Fine-tuned on: Medical Q&A dataset
- Quantization: Q4_K_M
- Size: 4.2GB

## Deployment

ollama create medical-model -f Modelfile
ollama run medical-model

## Performance

- Latency: ~200ms (TTFT)
- Throughput: 50 tok/s
- Hardware: RTX 4070, 12GB VRAM

## Example Usage

...

Cost Optimization

Estimate Deployment Costs

def estimate_monthly_cost(
    requests_per_day: int,
    avg_tokens_per_request: int,
    platform: str
):
    """
    Estimate monthly deployment costs
    """
    # Platform costs (per hour)
    costs = {
        'local_rtx4090': 0.20,  # Electricity + amortized hardware
        'vast_rtx4090': 0.25,
        'runpod_rtx4090': 0.40,
        'runpod_a100': 1.50,
        'modal_a100': 2.00,
        'aws_g5_xlarge': 1.20
    }

    hourly_cost = costs.get(platform, 1.0)

    # Estimate throughput
    tokens_per_sec = 50  # Conservative estimate
    seconds_per_request = avg_tokens_per_request / tokens_per_sec

    # Calculate usage
    daily_seconds = requests_per_day * seconds_per_request
    daily_hours = daily_seconds / 3600

    # For serverless, only count actual usage
    # For dedicated, count 24/7
    if platform.startswith('modal'):
        monthly_cost = daily_hours * 30 * hourly_cost
    else:
        monthly_cost = 24 * 30 * hourly_cost  # Always-on

    return {
        'monthly_cost': monthly_cost,
        'cost_per_request': monthly_cost / (requests_per_day * 30),
        'daily_hours': daily_hours
    }

# Example
cost = estimate_monthly_cost(
    requests_per_day=10000,
    avg_tokens_per_request=256,
    platform='runpod_rtx4090'
)
print(f"Monthly cost: ${cost['monthly_cost']:.2f}")
print(f"Per request: ${cost['cost_per_request']:.4f}")

Cost Optimization Strategies

1. Use spot instances (Vast.ai) - 50-70% cheaper
2. Scale down during off-peak - 30-50% savings
3. Batch requests - Better GPU utilization
4. Use smaller models - 7B vs 13B often similar quality
5. Aggressive quantization - Q4 often sufficient
6. Multi-tenancy - Share GPU across models

Additional Resources

• Training: See unsloth-finetuning skill for model training
• Tokenizers: See superbpe and unsloth-tokenizer skills
• Optimization: See training-optimization skill for training details
• Datasets: See dataset-engineering skill for data preparation

Summary

Model deployment workflow:

1. ✓ Fine-tune with Unsloth
2. ✓ Merge LoRA adapters
3. ✓ Choose export format (GGUF/vLLM/HF)
4. ✓ Quantize appropriately (Q4_K_M recommended)
5. ✓ Select deployment platform
6. ✓ Deploy and monitor
7. ✓ Optimize costs and performance

Start with local Ollama deployment for testing, then scale to cloud for production.