Deploying DeepSeek R1 on Databricks: A Step-by-Step Guide

Deploying the DeepSeek R1 Model on Databricks: A Step-by-Step Guide

Databricks, a popular data engineering platform, is increasingly used for AI and machine learning tasks. This tutorial guides you through deploying the distributed DeepSeek R1 model on Databricks, a powerful large language model often preferred for on-premise deployment. This avoids sending data to external servers. For a deeper dive into DeepSeek R1's features and comparisons, see the DeepSeek-R1: Features, Comparison, Distilled Models & More blog.

This guide covers account setup, model registration using the UI, and access via the playground and local CURL commands. New to Databricks? The Introduction to Databricks course provides a comprehensive overview of the Databricks Lakehouse platform and its data management capabilities. For a deeper understanding of data management within Databricks, consider the Data Management in Databricks course.

Registering the DeepSeek R1 Model

1. Launch a Notebook: After creating your Databricks workspace, click " New" and select a notebook.

2. Install Packages: Install necessary Python libraries:

<code class="language-python">%%capture
!pip install torch transformers mlflow accelerate torchvision
%restart_python</code>

3. Load Model and Tokenizer: Load the DeepSeek R1 model and tokenizer from Hugging Face:

<code class="language-python">import pandas as pd
import mlflow
import mlflow.transformers
import torch
from mlflow.models.signature import infer_signature
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig, pipeline

model_name = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
config = AutoConfig.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, config=config, torch_dtype=torch.float16)</code>

4. Test the Model: Test with a sample prompt and generate a signature for model registration:

<code class="language-python">text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
example_prompt = "How does a computer work?"
example_inputs = pd.DataFrame({"inputs": [example_prompt]})
example_outputs = text_generator(example_prompt, max_length=200)
signature = infer_signature(example_inputs, example_outputs)
print(example_outputs)</code>

Expected output (may vary slightly):

<code>[{'generated_text': "How does a computer work? What is the computer? What is the computer used for? What is the computer used for in real life?\n\nI need to answer this question, but I need to do it step by step. I need to start with the very basic level and build up from there. I need to make sure I understand each concept before moving on. I need to use a lot of examples to explain each idea. I need to write my thoughts as if I'm explaining them to someone else, but I need to make sure I understand how to structure the answer properly.\n\nOkay, let's start with the basic level. What is a computer? It's an electronic device, right? And it has a central processing unit (CPU) that does the processing. But I think the central processing unit is more efficient, so maybe it's the CPU. Then, it has memory and storage. I remember that memory is like RAM and storage is like ROM. But wait, I think"}]</code>

5. Conda Environment: Define a conda environment:

<code class="language-python">conda_env = {
    "name": "mlflow-env",
    "channels": ["defaults", "conda-forge"],
    "dependencies": [
        "python=3.11",
        "pip",
        {"pip": ["mlflow", "transformers", "accelerate", "torch", "torchvision"]}
    ]
}</code>

6. Register the Model: Register the model using mlflow.transformers.log_model:

<code class="language-python">with mlflow.start_run() as run:
    mlflow.transformers.log_model(
        transformers_model=text_generator,
        artifact_path="deepseek_model",
        signature=signature,
        input_example=example_inputs,
        registered_model_name="deepseek_r1_llama_8b",
        conda_env=conda_env
    )</code>

Deploying DeepSeek R1

1. Navigate to Models: In the Databricks dashboard, go to the "Models" tab.

2. Serve the Model: Select your model and click "Serve this model."

3. Configure Endpoint: Name your endpoint, choose compute options, set concurrency, and click "Create."

For fine-tuning on a custom dataset, refer to the Fine-Tuning DeepSeek R1 tutorial.

Accessing the Deployed Model

1. Databricks Playground: Test directly in the Databricks playground.

2. CURL Command: Generate a Databricks API key (Settings > Developer), set it as an environment variable ($DATABRICKS_TOKEN), and use CURL:

<code class="language-python">%%capture
!pip install torch transformers mlflow accelerate torchvision
%restart_python</code>

For information on DeepSeek R1 vs. V3, see the DeepSeek R1 vs V3 blog. New to LLMs? The Introduction to LLMs in Python course is a great starting point. Remember that while CPU deployment is possible, it might be slower.

The above is the detailed content of Deploying DeepSeek R1 on Databricks: A Step-by-Step Guide. For more information, please follow other related articles on the PHP Chinese website!