Skip to content

Readiness reports a model as ready when another recipe's alias spells its basename #945

Open

Description

@rominf

Readiness (openai_models_endpoint_has_model in crates/rocm-core/src/endpoint_readiness.rs) accepts a bare basename as a match for a catalog model. Two catalog aliases spell the basename of a different recipe, so a server running one recipe reads as ready for the other:

Alias Recipe it belongs to Recipe whose basename it spells
qwen3-32b Qwen/Qwen3-32B-FP8 Qwen/Qwen3-32B
qwen3.6-35b-a3b unsloth/Qwen3.6-35B-A3B-GGUF:Q4_K_M Qwen/Qwen3.6-35B-A3B

For example, a server launched on the FP8 recipe under its qwen3-32b alias reports qwen3-32b in /v1/models, and that reads as ready for Qwen/Qwen3-32B. Readiness cannot tell the two apart from the name alone.

#803 pins the current behaviour: an_alias_spelling_another_models_basename_reads_as_ready_for_it in crates/rocm-core/tests/model_selector_properties.rs derives the collisions from the catalog, asserts this exact list, and asserts each still reads as ready, so a catalog change that adds or removes a collision fails there.

Two ways to fix it:

  1. Readiness stops accepting a bare basename when it is also another catalog recipe's alias.
  2. The catalog drops or renames the two colliding aliases.

Whichever is chosen, the pinned list in that test should be updated with it.

Activity

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions