Skip to content

Daytona: derive GPU types from the pricing feed - #286

Open
mislavivanda wants to merge 1 commit into
dstackai:mainfrom
mislavivanda:daytona-feed-gpu-types
Open

mislavivanda wants to merge 1 commit into
dstackai:mainfrom
mislavivanda:daytona-feed-gpu-types

Conversation

@mislavivanda

Copy link
Copy Markdown
Contributor

What

Makes the Daytona provider derive its GPU types from Daytona's public pricing feed (https://billing.app.daytona.io/gpu-pricing) instead of the hardcoded GPU_MAP.

Every GPU type in that feed is orderable, so the feed alone can decide what is offered. This adds the now-available B200 and L40S (both creatable since Daytona 0.225.0 / 0.231.0), and future Daytona GPUs that gpuhunt already knows appear without a code change.

How

  • Daytona type identifiers are resolved against gpuhunt's known accelerators (find_accelerators) by dropping separators: RTX-PRO-6000 → RTXPRO6000, RTX-5090 → RTX5090. Vendor and GPU memory come from the matched AcceleratorInfo.
  • provider_data["gpu_type"] keeps the original Daytona identifier whenever it differs from the gpuhunt name, so dstack's Daytona backend keeps passing the right gpuType to the create API.
  • Representative memory shape is now derived from GPU memory: 50 GB per GPU for GPUs with ≤ 32 GB, 100 GB otherwise. For every current type this produces the same offers as before.
  • GPU types gpuhunt does not know are still skipped with a warning.

Testing

  • Unit tests updated (B200/L40S now offered, capacity for types missing from the feed is ignored) plus new tests for feed-only GPU types and the memory shape: 76 passed.
  • Integrity tests no longer import GPU_MAP. They validate offers against the live feed and add a check that every feed GPU type resolves to a known accelerator, so a brand-new chip surfaces as a CI failure rather than being silently dropped: 11 passed against the live feed.
  • python -m gpuhunt daytona: 9 GPU types × 8 counts × on-demand/spot + 2 CPU offers = 146 offers.

🤖 Generated with Claude Code

Every GPU type in Daytona's public pricing feed is orderable, so the feed
alone now decides which GPUs are offered instead of a hardcoded GPU_MAP.
Daytona type identifiers are resolved against gpuhunt's known
accelerators by dropping separators (RTX-PRO-6000 -> RTXPRO6000), which
also provides vendor and GPU memory. This adds the now-available B200
and L40S, and future Daytona GPUs that gpuhunt already knows appear
without code changes.

- Memory shape: 50 GB per GPU for GPUs with <= 32 GB GPU memory,
  100 GB otherwise (unchanged results for all current types).
- Unknown accelerators are still skipped with a warning.
- Integrity tests check offers against the live feed and fail if the
  feed contains a GPU type gpuhunt does not know.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Mislav Ivanda <mislavivanda454@gmail.com>
@peterschmidt85

Copy link
Copy Markdown
Contributor

@mislavivanda, the PR looks good. A real E2E test uncovered a separate issue affecting the existing dstack integration, which I've reported in dstackai/dstack#4494.

When starting a Daytona run, the sandbox starts, but SSH setup fails with:

expiresInMinutes must be between 1 and 1440

dstack requests a ten-year expiry, which previously worked. The current 24-hour limit prevents runs from completing provisioning, independently of this PR.

Could Daytona restore support for long SSH-token expiries? Requesting 24 hours would fix startup, but reliable access for longer runs would then require credential renewal in dstack.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants