[Core] Support TPU v7x accelerator type for device discovery

[ryanaoleary]

Description

Briefly describe what this PR accomplishes and why it's needed.
This PR adds support for Google Cloud's 7th generation TPU (Ironwood).

The TPU 7x generation introduces a change in the accelerator type naming convention reported by the environment. Unlike previous generations (v6e-16, v5p-8, etc.), 7x instances report types starting with tpu (e.g. tpu7x-16).

This PR accounts for the new format and enables Ray to auto-detect the v7x hardware automatically (users don't have to manually configure env vars). This is critical for libraries like Ray Train and for vLLM support - where the automatic device discovery is utilized during JAX initialization.

Related issues

Fixes #59964

Additional information

For more info about TPU v7x: https://docs.cloud.google.com/tpu/docs/tpu7x.

[gemini-code-assist]

Code Review

This pull request successfully introduces support for Google Cloud's 7th generation TPU (Ironwood), which uses a new naming convention (e.g., tpu7x-16). The changes correctly update the regex for accelerator type validation and the logic for converting the new tpu prefixed types to the internal v prefixed format. The addition of new test cases for tpu7x-16 is also a welcome improvement, ensuring the new functionality works as expected. However, there is a critical oversight regarding the VALID_TPU_TYPES tuple, which needs to be updated to fully support the new v7x generation.

[ryanaoleary]

In this PR we convert the tpu7x accelerator type format to v7x for consistency with all the other types, which is then what's expected by the rest of the Ray functions that call it internally. When creating a SlicePlacementGroup, we allow users to specify either tpu7x, tpu-v7x, or v7x, but internally this string will always get converted to v7x.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

[cursor]

Missing tpu7x conversion in SlicePlacementGroup validation

Medium Severity

The PR description states users can specify tpu7x, tpu-v7x, or v7x as accelerator_version when creating a SlicePlacementGroup. However, _accelerator_version_check directly checks membership in VALID_TPU_TYPES without converting tpu7x to v7x first. While get_tpu_worker_resources (called earlier in __init__) uses get_tpu_version_from_type to convert tpu7x → v7x, the subsequent validation in _accelerator_version_check fails because it doesn't perform the same conversion. Users passing accelerator_version="tpu7x" will get a ValueError despite the documented support.

Additional Locations (1)

• python/ray/util/tpu.py#L39-L42