Add GPU type selection (RTX 4090 / A100 / H100)

[Milaskinger]

Feature Request

Users should be able to request a specific GPU tier when submitting a job.

Proposed API

job = client.run("train.py", gpu="rtx4090")   # default, cheapest
job = client.run("train.py", gpu="a100")       # higher VRAM
job = client.run("train.py", gpu="h100")       # maximum performance

Backend changes needed

• POST /api/jobs should accept an optional gpu_type field
• Job routing should filter providers by GPU capability
• Pricing should reflect the GPU tier

Client changes

• Add gpu parameter to client.run()
• Add gpu_type field to JobResult

This is the most-requested feature from enterprise customers.

[snakefood3232]

Adding GPU tier selection as requested is a straightforward backend+client update. I'd handle this by extending the job schema and routing logic to filter providers by GPU capability, then updating the Python client's run() method. Can have a PR up within a day—I've implemented similar resource selection for compute platforms before.

[Milaskinger]

Hey @snakefood3232 — the client-side part of this is straightforward. Here's the contract we want to land on:

API signature

job = client.run(
    "train.py",
    task_name="my-run",
    gpu_type="rtx4090",   # new optional param
)

Accepted values (for now)
"rtx4090", "a100", "h100" — pass as-is to the backend in the request body as {"gpu_type": "rtx4090"}. If None (default), omit the field entirely so old behaviour is unchanged.

Scope for this PR — SDK only
Don't touch the backend routing or provider schema — that lives server-side. The SDK just needs to forward the preference and the server handles provider matching. If the requested GPU type is unavailable the API already returns a 503 with {"detail": "no_provider_available"} — map that to a new NoProviderAvailableError exception (add it to exceptions.py).

What to update

• client.py — add gpu_type: Optional[str] = None to run()
• exceptions.py — add NoProviderAvailableError
• README.md — one short example block showing gpu_type="a100"

Go for it — ping once the PR is open.

[Milaskinger]

GPU selection — design note

GPU selection by model (RTX 4090, A100, H100) is on the roadmap. Here's the planned API:

```python

Coming in v0.3.0

job = client.run(
"train.py",
gpu="rtx4090", # or "a100", "h100", "any" (default)
min_vram_gb=12, # optional minimum VRAM filter
)
```

The matchmaking engine will route jobs to providers matching the requested GPU. Currently, jobs are dispatched to the first available provider (which is an RTX 4090 on our network today).

We'll open a separate tracking issue when implementation starts. Thanks for the suggestion — it's exactly the kind of feature that makes the platform more predictable for production workloads.