Use paginated Vast.ai v1 instances endpoint

[7Z0nE]

Summary

• Move get_instances() and get_instance(id) in the Vast.ai backend off GET /api/v0/instances/ and onto GET /api/v1/instances/. Vast.ai's April 2026 product update announces that the v0 list endpoint will start returning paginated results (max 25 per response), so callers must handle the keyset cursor. The v1 endpoint already enforces that contract via after_token / next_token.
• get_instances() walks pages until next_token is empty so the existing 3 s cache fills with the full list.
• get_instance(id) issues a targeted v1 query with select_filters={"id": {"eq": id}} and limit=1, so a single-row lookup stays one HTTP request instead of paginating through every owned instance.
• Other Vast.ai endpoints used by the backend — PUT /asks/{id}/ (create), DELETE /instances/{id}/ (destroy), POST /bundles/ (search offers), PUT /instances/request_logs/{id}/ — have no v1 replacement in the published OpenAPI yet, so they remain on v0.

Test plan

• uv run pytest src/tests/_internal/core/backends/vastai/ -v — 15 passed (4 new test_api_client.py cases cover pagination, single-instance filter, 404 handling, and that destroy still hits v0).
• uv run pytest src/tests/_internal/core/backends/ -q — 244 passed (no regressions in sibling backends).
• uv run pre-commit run --files <changed> — ruff/format clean.
• Smoke test against a real Vast.ai account (would appreciate a maintainer doing this — I do not have credentials).

🤖 Generated with Claude Code

[r4victor]

@7Z0nE, thanks for the PR. The change looks solid and can be merged. Although I'm not sure if we need get_instances() and thus _list_instances_v1() and the cache at all since it's used only for get_instance(). Apparently, it was needed when there was no show-instance API. I'm happy with either of these:

• Drop get_instances() and _list_instances_v1() and the cache in favor of show-instance.
• Implement get_instance() via _list_instances_v1() as you did but drop get_instances().
• Keep the PR as is in case get_instances() is useful in future.

WDYT?

[r4victor]

Let's drop this new test file completely – the tests don't seem very useful.

[7Z0nE]

will do

[7Z0nE]

I'm not sure if we need get_instances() and thus _list_instances_v1() and the cache at all since it's used only for get_instance(). Apparently, it was needed when there was no show-instance API

The show-instance API (GET /api/v0/instances/{id}) still has a "v0" in the path. Requests to this v0 endpoint will not work with API keys generted from the new v1 endpoints. Thus we cannot use show-instance and need to keep the workaround with the v1/instances endpoint (https://docs.vast.ai/api-reference/instances/show-instances).

I will test this again to clarify. If it works, then I would do a simplification.

Drop get_instances() and _list_instances_v1() and the cache in favor of show-instance.
I assumed that get_instances is called by users of the VastAiAPIClient and therefore needs to be supported.
A possibility I see would be to migrate all usages to list_instances with a pagination interface.

will post more updates when I have time to go deeper.