docs(operate): relayminer exposing deepseek service + datanode deployment on Kubernetes (using ollama)

[eddyzags]

Summary

This PR adds an example to deploy a Relayminer exposing a deepseek service as well as how to deploy a deepseek datanode on Kubernetes (using Ollama deepseek-r1:1.5b model)

Primary Changes:

• poktroll documentation.

Type of change

Select one or more from the following:

• New feature, functionality or library
• Bug fix
• Code health or cleanup
• Documentation
• Other (specify)

Sanity Checklist

• I have updated the GitHub Issue Metadata: assignees, reviewers, labels, project, iteration and milestone
• For docs: make docusaurus_start
• For small changes: make go_develop_and_test and make test_e2e
• For major changes: devnet-test-e2e label to run E2E tests in CI
• For migration changes: make test_e2e_oneshot
• 'TODO's, configurations and other docs

[RawthiL]

Hi @eddyzags , there are several issues with the approach presented in this document:

1. Ollama is not a production ready product, so presenting it to be staked in a network where you expect multiple concurrent calls does not seem like a good idea. There are better alternatives to do this, like vLLM or TGI.

2. You are showing a service that goes straight to /api/generate which results in the service only being able to respond in Ollama's API format (in a very specific endpoint), something that is not being accepted as the industry standard. We proposed to use OpenAI API standard which is much more broadly accepted. Also, staking a fixed endpoint cripples the users' ability to select a desired type of completion, like /v1/completions or /v1/chat/completions.

3. You are staking a deepSeek-r1:1.5b model into a service that states in it's description that it is Deepseek V3 0324 (a ~600B parameter model). While there is no way to enforce models on services (as we lengthy discussed in the forum), presenting a guide that offends the service description does not seem correct. There is another service that can be used to stake language models models in mainnet which is text-generation, in this one the only assumption is that you provide an OpenAI API compatible endpoint.

4. The endpoint you are staking responds to "model": "deepseek-r1:1.5b", there is no way for a app or gateway to know the value of this parameter. We have created a solution for this, a sidecar for ML services that override the deployed model name so you can accept an standard name, like pocket_network.

5. Regarding streaming, there is a very old implementation that you can use to allow streaming, and even if you deploy the basic relay miner, setting "stream": true won't break the relayminer, yo will just receive the whole stream in a single payload. So, this clarification is confusing and will get old fast.

6. A minor detail, 1000069upokt is not enough to stake a Supplier in mainnet.

There was a long conversation around the topic of deploying language models in the network, while it was done in the Morse era, none of the conclusions changed with Shannon. There we explain why some things needs to be standardized (API format) and why using service names with vendor names makes no sense.

[Olshansk]

@eddyzags What's the next step here?