Add Daytona catalog provider

[mislavivanda]

What

Adds Daytona (https://daytona.io) as an offline catalog provider — a GPU cloud offering isolated sandboxes with on-demand and spot capacity, billed per second.

Data source

Public, unauthenticated rate card: https://billing.app.daytona.io/gpu-pricing (JSON, served from the same definitions customers are billed against, cached 5 min). Same no-creds offline pattern as CloudRift.

Mapping notes

• Daytona bills GPU sandboxes a la carte (per-GPU rate + separately metered vCPU/RAM/disk), so there are no fixed instance types. The catalog publishes representative configurations: per GPU, 8 vCPU with 100 GB RAM on datacenter-class cards (50 GB on consumer cards) and 256 GB disk, scaled linearly across 1/2/4/8 GPU counts.
• Spot offers price the GPU and the resources at Daytona's spot rates, which are discounted independently of on-demand.
• 8 GPU types, all mapping to existing canonical names: H100, H200, B200, B300 (memory 270 GB per KNOWN_NVIDIA_GPUS), RTXPRO6000, RTX5090, RTX4090, MI355X (AMD). Unknown future types are skipped with a warning.
• GPU capacity is pooled/globally scheduled, so offers are published under a single us location.

Testing

• Unit tests (src/tests/providers/test_daytona.py): payload mapping, exact price pins for both capacity types, unknown-GPU skip, HTTP/shape error handling.
• Live collection verified end to end: python -m gpuhunt daytona → 64 offers, integrity tests pass against the produced CSV.

[peterschmidt85]

I pushed cc9749c to add live GPU availability and regional CPU offers.

Live GPU availability requires DAYTONA_API_KEY. Without it, the provider warns and returns public price estimates.

@mislavivanda, please review the changes and confirm they look good to you.

[mislavivanda]

Hi @peterschmidt85, thanks for the review and follow up commit which moved us to online provider.

Confirmed - the changes look great, thank you. I verified against our production API: org resolution via /api-keys/current works, capacity gating matches live fleet state (sold-out types drop out and return), CPU offers price out exactly per region, and excluding B200 is correct - it is priced ahead of launch and not yet orderable. The online-provider model fits Daytona well given per-second billing and fast capacity churn. If the CI secret ever needs a dedicated key or higher limits, ping us at support@daytona.io.

[peterschmidt85]

Thank you, @mislavivanda ! I’ll merge this shortly.
Would you like us to add Daytona support to dstack as well, or would you prefer to submit a PR yourself?

[mislavivanda]

Thank you, @mislavivanda ! I’ll merge this shortly. Would you like us to add Daytona support to dstack as well, or would you prefer to submit a PR yourself?

Yes, of course, that would be great.

I'm good with both ways, whichever your prefer. If you lift a PR just ping me please so I can review the changes before we merge it.

[peterschmidt85]

If you lift a PR just ping me please so I can review the changes before we merge it.

@mislavivanda Here we go: dstackai/dstack#4321