Fixes #10631

Problem

Studio already saves each model's run settings (context length, KV cache, speculative decoding, GPU placement), and the OpenAI-compatible auto-switch applies them when an API request loads that model (#7473, #11904, #10216). Sampling is the gap:

• Chat only remembers the last values used with each model. There is no way to say "this is how I want this model to sample" outside the chat sidebar.
• An API request that switches to a model and leaves temperature, top_p and so on unset always gets the model's recommended values. Nothing the user set in Studio reaches it.

What this adds

A Sampling Defaults row on the model's run settings page, which is where the Hub's Run button and the picker's gear already go. It opens a small editor for Temperature, Top P, Top K, Min P, Repetition Penalty and Presence Penalty.

Saving writes the values to two places:

• API requests. The values go to a per-model, per-account server store (model_sampling_defaults). resolve_effective_sampling reads it, so /v1/chat/completions, /v1/completions, /v1/responses and /v1/messages fill unset fields from it, including the request that switches the model in. The order is: operator UNSLOTH_SAMPLING_* pin, then the request's own value, then the saved default, then the model's recommendation, then the schema default. The /sampling-defaults route behind the API page's example request uses the same function.
• Chat. The same values go into the per-model memory Chat already restores on every load, through the setRememberedParamsForModel path the connected-model settings dialog uses. They are written under both ids a model can have in Chat (repo id and cache snapshot path). A field that was saved and is then cleared goes back to the model's recommended value. A field that was never saved keeps Chat's last-used value, as today, and the editor says so.

How defaults are keyed:

• One row per model, not per quant.
• A cache snapshot path is stored under its repo id, so a new revision still finds the row.
• Repo ids and Windows, UNC and WSL paths match regardless of case, using the launch overrides' _fold_case_insensitive_path rule. POSIX paths match exactly.
• The server resolves reads itself, so the browser does not mirror these rules.

Not changed

• Nothing changes for anyone who never saves a default. The store is empty, the lookup returns early and resolution is the same as before.
• Run settings and load overrides are untouched. The sampling store is separate from the launch override map on purpose: a row there counts as a load override, can shadow a broader repo row and re-ticks "Remember for this model".
• Explicit request values and operator pins still win.
• Diffusion, audio and NPU models do not show the row.
• /v1/completions on a vLLM or SGLang engine still forwards the body without any per-model fill, as before.

End to end

Real Studio installs at the merge base and at this branch, CPU only, unsloth/gemma-3-270m-it-GGUF Q8_0, auto-switch on. Values are what llama-server reports in /slots after the request:

Cost

resolve_effective_sampling runs once per request, never per token. The store read reuses the 2 s cached setting the auto-switch already uses, and an index is rebuilt only when that cache refreshes. Per-call timing on this host (median of 5 x 20k calls, base and head alternating, two rounds):

Tests

studio/backend/tests/test_model_sampling_defaults.py (16 tests) covers:

• precedence against explicit values, operator pins and the Qwen3 thinking table;
• the /v1/completions passthrough repeat_penalty alias;
• snapshot path and quant-qualified ids;
• POSIX and Windows path matching, and ambiguous case folds;
• out-of-range and NaN values dropped, idempotent saves, clearing;
• an unreadable store never failing a request;
• /sampling-defaults reporting the saved value.

Reverting only the resolution hook makes 5 of them fail. Backend tests touching the settings routes fail the same 27 pre-existing cases on the merge base and on this branch, and nothing new. Frontend: tsc clean, vite build OK, npm test 13242 passed.

Screenshots

Run settings panel for the loaded model, merge base vs this PR:

The editor with three values set:

After Save, Chat's own Sampling section for the loaded model picks up the saved values (right), against the merge base (left):