[DO NOT MERGE] [TIMED LAUNCH Oct 21] Getting Started with NVIDIA Kumo Graph Neural Networks in Snowflake

[sfc-gh-aramasagaram]

Caution

DO NOT MERGE. TIMED LAUNCH. This guide is embargoed until October 21, 2026. Opened only to generate a preview link for review. The author will request merge on launch day.

status: Hidden is intentional: it is a safeguard so the guide cannot go live through an accidental merge. Please review and approve ahead of the date; the author will switch it to Published and replace the partner link placeholders on launch day.

Summary

• New quickstart: train an NVIDIA Kumo graph neural network on relational tables in Snowflake Notebooks (Container Runtime, GPU) and write predictions back to a Snowflake table.
• Covers getting the Kumo wheel, publishing it to a Workspace with Cortex Code, installing the Kumo skill, then SnowflakeConnector, Table, Graph, PredictiveQuery, fit, and predict.
• status: Hidden until launch.

Before merge (launch day)

• Replace KUMO_WHEEL_LINK_PLACEHOLDER with the NGC or Hugging Face link
• Replace KUMO_SKILL_LINK_PLACEHOLDER and BLOG_LINK_PLACEHOLDER
• Set status: Published
• Remove the DO NOT MERGE label/title

Generated with Snowflake CoCo

[github-actions]

🔗 Staging Preview Ready

Quickstart: getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake
Language: en

👉 Preview URL: https://publish-p57963-e462098.adobeaemcloud.com/en/developers/guides/getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake-6768e721bd324e7e63e3edd8a1af7d3541bd9ec5

Note: Images are uploaded asynchronously. If images are missing, wait a few minutes and refresh.

---

Generated by GitHub Actions

[github-actions]

🔗 Staging Preview Ready

Quickstart: getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake
Language: en

👉 Preview URL: https://publish-p57963-e462098.adobeaemcloud.com/en/developers/guides/getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake-99aaf9efa0b6bb789dd43c967ac5dc6c03a3f5db

Note: Images are uploaded asynchronously. If images are missing, wait a few minutes and refresh.

---

Generated by GitHub Actions

[sfc-gh-kenguyen]

Hi @sfc-gh-aramasagaram

Blocking

• Placeholders: KUMO_WHEEL_LINK_PLACEHOLDER (Get Kumo section and Resources), KUMO_SKILL_LINK_PLACEHOLDER, and BLOG_LINK_PLACEHOLDER are still in the body. Readers can't get the wheel or the skill, so the guide can't be followed. Replace them with the real links before merge.
• status: Hidden: Hidden is no longer supported. Set status: Published so the page actually goes live on merge.
• Step 2, cp -r <PATH_TO_SKILL>/kumo-data-science ...: readers have no way to find <PATH_TO_SKILL>. Point this at the skill's download location, the same link as KUMO_SKILL_LINK_PLACEHOLDER.

Suggested

• Overview: the bold labels (**What you will build**, **Prerequisites**) should be ### What You'll Learn, ### What You'll Build, ### What You'll Need, and ### Prerequisites H3s, which the site uses for consistent rendering. Make the first prerequisite a Snowflake account link.
• Cleanup: the guide creates a predictions table and runs on a GPU compute pool but never removes them. Add a section that drops <PREDICTIONS_TABLE_NAME> and suspends the compute pool (ALTER COMPUTE POOL ... SUSPEND) so readers stop paying for GPU credits.
• Conclusion: rename to ## Conclusion And Resources, add a ### What You Learned list, and include a link to the Snowflake Documentation.
• prediction_time=datetime(2025, 6, 30): this is specific to the AdventureWorks data. Say so inline, so readers using their own tables know to change it.

[sfc-gh-aramasagaram]

Thanks for the review! Pushed fixes:

Addressed

• Overview now uses ### What You'll Learn, ### What You'll Build, ### What You'll Need, and ### Prerequisites; first prerequisite links to a Snowflake account signup.
• Step 2 no longer uses <PATH_TO_SKILL>; it points to the skill download link (same link as the skill resource) and copies from the downloaded folder.
• prediction_time: added an inline comment and a note that the date fits the sample data, and readers should pick a date for prediction (usually the day after their latest record).
• Added a Save GPU Cost section with ALTER COMPUTE POOL <GPU_COMPUTE_POOL_NAME> SUSPEND;.
• Renamed to ## Conclusion And Resources, added ### What You Learned, and linked the Snowflake ML documentation.

Intentionally not changed

• Placeholders (wheel, skill, blog links): we are waiting on the partner (NVIDIA) to provide the final URLs. They will be replaced before merge.
• status: Hidden: this is a timed launch (Oct 21). We want this reviewed and approved ahead of the date, but it must not go live early. The status is a deliberate safeguard against an accidental merge and will be switched to Published on launch day.
• Predictions table cleanup: the predictions table is the artifact that powers downstream applications, so the guide keeps it on purpose. The cleanup step suspends the GPU compute pool to stop credit usage instead.

[github-actions]

🔗 Staging Preview Ready

Quickstart: getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake
Language: en

👉 Preview URL: https://publish-p57963-e462098.adobeaemcloud.com/en/developers/guides/getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake-46cb85235ba17af93509ea60ecebf926be628dae

Note: Images are uploaded asynchronously. If images are missing, wait a few minutes and refresh.

---

Generated by GitHub Actions

[sfc-gh-amadan]

Blocking

• Step 1 (lines 55–65): the prompt has CoCo rewrite a vendor wheel's METADATA pins and rebuild RECORD hashes, so every reader ends up with a different, LLM-patched dependency set that can't be reproduced or supported. Ask Kumo/NVIDIA to ship a Container Runtime–compatible wheel plus a pinned requirements.txt, and reduce this step to uploading those two files.

Suggested

• Prerequisites: "load them into Snowflake before you start" carries the most setup work in the guide, with no instructions. Link a loader script or add the SQL to create and load the four tables.
• Overview has both ### What You'll Need and ### Prerequisites. Merge them into What You'll Need.
• Before merge on Oct 21: replace KUMO_WHEEL_LINK_PLACEHOLDER (lines 43, 231), KUMO_SKILL_LINK_PLACEHOLDER (70, 233), and BLOG_LINK_PLACEHOLDER (234), and set status: Published.

[sfc-gh-aramasagaram]

Thanks again. Pushed updates:

Addressed

• Sample data: added assets/load_sample_data.sql and a new Step 1 that downloads the four Microsoft sample files and runs the script with snow sql. Tested end to end: row counts (19,820 / 504 / 31,465 / 121,317), date range, and totals match the tables the guide was validated on.
• Overview: merged Prerequisites into ### What You'll Need.
• Skill link: KUMO_SKILL_LINK_PLACEHOLDER is replaced with the NVIDIA Kumo agent skills repo (https://github.com/NVIDIA/kumo-engine/tree/main/agent-skills).

Step 1 wheel prompt (keeping for this release)
This is the v1 wheel. NVIDIA built a single wheel that works across platforms, and that is what they are taking GA at GTC Berlin on Oct 21, so there is no Container Runtime specific wheel for this launch. The CoCo step is how readers get that wheel working on Container Runtime today. Agreed that a Container Runtime optimized wheel with a pinned requirements.txt is the better long term answer; I'll propose it to NVIDIA for a later version, and this step will shrink to an upload once it exists.

Still pending before merge on Oct 21

• KUMO_WHEEL_LINK_PLACEHOLDER and BLOG_LINK_PLACEHOLDER (waiting on the partner for URLs)
• status: Published

[github-actions]

Helpful Information

Non-Image Files in Assets Folder

Non-image files found in /assets folders will **NOT** be uploaded to snowflake.com. If you are referencing them in your guide, you should link to them directly (e.g. [using permanent GitHub links](htt…

• site/sfguides/src/getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake/assets/load_sample_data.sql

These are informational messages and will not block your PR from being merged.

[github-actions]

🔗 Staging Preview Ready

Quickstart: getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake
Language: en

👉 Preview URL: https://publish-p57963-e462098.adobeaemcloud.com/en/developers/guides/getting-started-with-nvidia-kumo-graph-neural-networks-in-snowflake-a537771f4ffd9cb27502d21b9f60f7d0198930b4

Note: Images are uploaded asynchronously. If images are missing, wait a few minutes and refresh.

---

Generated by GitHub Actions