%env LOCATION <YOUR_LOCATION>
%env SUBSCRIPTION_ID <YOUR_SUBSCRIPTION_ID>
%env RESOURCE_GROUP <YOUR_RESOURCE_GROUP>
%env WORKSPACE_NAME <YOUR_WORKSPACE_NAME>
%env MODEL_ID nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16

import os
from uuid import uuid4

os.environ["ENDPOINT_NAME"] = f"nemotron-omni-{str(uuid4())[:8]}"
os.environ["DEPLOYMENT_NAME"] = f"nemotron-omni-{str(uuid4())[:8]}"

%pip install azure-ai-ml azure-identity --upgrade --quiet

import os
from azure.ai.ml import MLClient
from azure.identity import DefaultAzureCredential

client = MLClient(
    credential=DefaultAzureCredential(),
    subscription_id=os.getenv("SUBSCRIPTION_ID"),
    resource_group_name=os.getenv("RESOURCE_GROUP"),
    workspace_name=os.getenv("WORKSPACE_NAME"),
)

from azure.ai.ml.entities import ManagedOnlineEndpoint

endpoint = ManagedOnlineEndpoint(name=os.getenv("ENDPOINT_NAME"))
client.begin_create_or_update(endpoint).wait()

from azure.ai.ml.entities import ManagedOnlineDeployment, OnlineRequestSettings

deployment = ManagedOnlineDeployment(
    name=os.getenv("DEPLOYMENT_NAME"),
    endpoint_name=os.getenv("ENDPOINT_NAME"),
    model=f"azureml://registries/HuggingFace/models/{os.getenv('MODEL_ID').replace('/', '-').replace('_', '-').lower()}/labels/latest",
    instance_type="Standard_NC24ads_A100_v4",
    instance_count=1,
    request_settings=OnlineRequestSettings(request_timeout_ms=180000),
)
client.online_deployments.begin_create_or_update(deployment).result()

%pip install openai --upgrade --quiet

from urllib.parse import urlsplit
from openai import OpenAI
import os

api_key = client.online_endpoints.get_keys(os.getenv("ENDPOINT_NAME")).primary_key
url_parts = urlsplit(client.online_endpoints.get(os.getenv("ENDPOINT_NAME")).scoring_uri)
api_url = f"{url_parts.scheme}://{url_parts.netloc}/v1"

openai_client = OpenAI(
    base_url=api_url,
    api_key=api_key,
    default_headers={"azureml-model-deployment": os.getenv("DEPLOYMENT_NAME")},
)

image_url = "https://recursos.museoreinasofia.es/styles/large_landscape/public/Obra/DE00050_2.jpg.webp"

response = openai_client.chat.completions.create(
    model=os.getenv("MODEL_ID"),
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": 
                        {
                            "url": image_url
                        }
                },
                {
                    "type": "text",
                    "text": "You are a high school history teacher. Given this image, generate 5 high-school level short questions and answers covering authorship, artistic movement, historical context and symbolism."
                },
            ],
        }
    ],
)
print(response.choices[0].message.content)

# Response:
# 
# **1. Who created this painting?**  
# *Pablo Picasso* painted it in 1937.
# 
# **2. What artistic movement is it associated with?**  
# It is a prime example of *Cubism* (specifically the analytical, monochrome phase).
# 
# **3. When and why was it made?**  
# Created in 1937, it was a response to the bombing of the town of Guernica during the Spanish Civil War.
# 
# **4. What might the bull and the horse represent?**  
# The bull is often seen as a symbol of Spain (or brutality), while the horse represents the suffering and the innocent victims of the conflict.
# 
# **5. How does the black‑and‑white color scheme affect the work’s impact?**  
# The stark, limited palette heightens the sense of tragedy and urgency, focusing the viewer on form and emotion rather than color.

video_url = "https://ep02.epimg.net/elcomidista/videos/2016/05/23/receta/c8281c0547528149d0d53e90bb8c39b0.mp4"

response = openai_client.chat.completions.create(
    model=os.getenv("MODEL_ID"),
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "video_url",
                    "video_url": 
                        {
                            "url": video_url
                        }
                },
                {
                    "type": "text", 
                    "text": "Write a cuisine-book style recipe in English based on the video. Start with a short introduction of the recipe by translating the speaker's words. After that, add a section listing every ingredient appearing in the video. Finally, explain the preparation process in bullet points, summarizing the instructions in the video."
                },
            ],
        }
    ],
    max_tokens=20480,
    temperature=0.7,
    extra_body={
        "mm_processor_kwargs": {"use_audio_in_video": True},
    },
)

print(response.choices[0].message.content)

# **Introduction**
# Today, we present a variant of tiramisu where the traditional coffee is replaced with black beer. Following a recipe by Barmieller from Barcelona, this version creates a fresher, lighter dessert that isn't overly heavy.
#
# **Ingredients**
# *   3 eggs
# *   6 tablespoons sugar
# *   200g mascarpone cheese
# *   150ml black beer
# *   150ml milk
# *   150g sponge cake (bizcochos)
# *   150g dark chocolate
#
# **Preparation**
# *   Separate the egg whites from the yolks. Whisk the egg whites with a pinch of salt until they form stiff peaks, then set them aside in the refrigerator.
# *   In a separate bowl, whisk the egg yolks with the sugar until the mixture becomes creamy.
# *   Add the mascarpone cheese to the yolk mixture and blend until smooth.
# *   Gently fold the whipped egg whites into the mascarpone mixture to lighten it.
# *   Spread a thin layer of the cream mixture at the bottom of a rectangular glass dish.
# *   Dip the sponge cake pieces into the black beer (and milk mixture) and arrange them in a layer at the bottom of the dish.
# *   Cover the cake layer with another layer of the cream mixture.
# *   Grate the dark chocolate over the top.
# *   Repeat the layering process (cake, cream, chocolate) until the dish is full.
# *   Cover the dish with plastic wrap and refrigerate for a minimum of 2 hours.
# *   Serve the chilled dessert and enjoy.

audio_url= "https://podcast.do-not-panic.com/episodes/2026-04-19-bahdanau-attention-for-neural-machine-tr-dbd9b6.mp3"
image1_url = "https://machinelearningmastery.com/wp-content/uploads/2021/09/bahdanau_1.png"
image2_url = "https://miro.medium.com/v2/resize:fit:1400/1*6ggue1DJVQD8LidvKuPitQ.png"

response = openai_client.chat.completions.create(
    model=os.getenv("MODEL_ID"),
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "audio_url",
                    "audio_url": 
                        {
                            "url": audio_url
                        }
                },
                {
                    "type": "image_url",
                    "image_url": 
                        {
                            "url": image1_url
                        }
                },
                 {
                    "type": "image_url",
                    "image_url": 
                        {
                            "url": image2_url
                        }
                },
                {
                    "type": "text",
                    "text": "Given this audio and these two images, choose which image fits the audio better and explain why."
                },
            ],
        }
    ],
    extra_body={
        "chat_template_kwargs": 
            {
                "enable_thinking": True
            }
    },
)

print(response.choices[0].message.content)

# Based on the audio content, **Image 1** fits better.
#
# **Explanation:**
# The audio discusses the "Attention Paper" (Bahdanau et al., 2015), which introduced the attention mechanism to Neural Machine Translation (NMT). The key concept explained is replacing a **single fixed-length vector** (the "bottleneck") with a **sequence of source representations** (annotations) that the decoder can attend to dynamically.
#
# *   **Image 1** depicts an encoder-decoder architecture where the encoder processes the input sequence ($X_1, \dots, X_T$) into a series of hidden states ($h_1, h_2, \dots, h_T$). It shows attention weights ($a_{t,1}, a_{t,2}, \dots$) connecting these states to the decoder. This visualizes the "row of labeled bins" or sequence of annotations that the audio describes as the solution to the fixed-length vector problem.
# *   **Image 2** illustrates "Scaled Dot-Product Attention" and "Multi-Head Attention," which are specific components of the **Transformer** architecture (Vaswani et al., 2017). The audio explicitly clarifies that the paper being discussed is *not* the full Transformer model ("So attention here is not yet the full Transformer thing..."), making Image 2 less relevant to the specific paper being analyzed.

client.online_endpoints.begin_delete(name=os.getenv("ENDPOINT_NAME")).result()