feat(workspaces): add opinionated llm-inference workspace

[mcd01]

Re-integrate LLM Inference Workspace (llm-d based, experimental)

Summary

This PR re-integrates the LLM inference workspace, now built on top of the llm-d reference stack.

The feature should be considered experimental.
Compared to other “exalsius-native” workspaces, the setup is not yet as streamlined, but it already enables Exalsius members to deploy and serve LLMs directly inside a workspace.

The goal is to provide an early, practical way to run model inference workloads while we iterate on UX and integration.

---

What’s included

Workspace template integration

• Add new integrated template: LLM_D
• Add default editing comments for the template
• Register template alongside existing ones (Jupyter, Marimo, Dev Pod, Dist Training)

New configurator

• Introduce LLMInferenceConfigurator
• Automatically configures:
  • Hugging Face token secret
  • Model URI (hf://<repo>/<model>)
  • Model labels
  • InferencePool label matching
  • Tensor parallelism (= num_gpus)
  • Accelerator type (AMD / NVIDIA)

CLI command

New command:

exls workspaces deploy llm-inference

Options:

• --huggingface-token
• --model-name
• --num-gpus
• --wait-for-ready

Includes:

• Model name validation (<repo>/<model>)
• GPU count validation
• Resource selection based on cluster
• Standard confirmation + deployment flow

---

Example usage

exls workspaces deploy llm-inference \
  --model-name Qwen/Qwen3-1.7B \
  --hf-token $HF_TOKEN \
  --num-gpus 2 \
  --wait-for-ready

---

Behavior

The configurator automatically translates inputs into template variables:

Input	Effect
model name	Sets model artifact URI + labels
HF token	Creates auth secret reference
num GPUs	Sets tensor parallelism
GPU vendor	Sets accelerator type (nvidia/amd)

This removes most manual editing from the deployment process.

---

Helm chart

This workspace integrates the existing chart:

https://github.com/exalsius/exalsius-workspace-hub/tree/main/workspace-templates/llm-inference/llm-d-model

---

Notes

• Experimental feature
• Not yet as polished as other workspace types
• Intended for early adopters and internal testing
• Backwards compatible (pure addition)

[srnbckr]

I've rebased the branch and directly fixed some minor issues. Please have a look at my commits and see if that's how you would expect it.

We could discuss to add a short interactive flow to e.g. ask the user for the huggingface token and model name instead of requiring it (similar to how it's done for the marimo or jupyter workspaces) but otherwise I would approve this PR.