Pin unslothai/llama.cpp#210 (server-side preemption and exact concurrency) at afd3248b82

[danielhanchen]

Adds one pin: #210 at 15bcc6c, the tree of #197 (server-side preemption: parking a slot instead of ending every conversation when the KV cache fills, per-request park and resume notices, asynchronous parks, the opt-in exact concurrency mode) composed with the pinned commit of ggml-org#25731 so it merges after it on b10871. #197 carries #184, #190 and #192 in its history, so those are not pinned; #194 is superseded by #197 (its code fixes were ported) and conflicts with it, so it is not pinned either.

Every preemption process in the pinned tree is off by default: --preempt-ram defaults to 0 (a server without the flag behaves exactly as upstream, unified cache included) and exact concurrency is an environment opt-in. Studio passes --preempt-ram explicitly when it wants the server to park.

Also records #210 under unchecked in feature-checks.json: a server park has no arch, backend-op or mtmd probe surface; the feature is covered by test-exact-geometry and the server preemption tests in #197, which need a CUDA host for the exact path.

Reproduced the preflight's resolve loop (fresh clone of upstream at the base tag, every pin fetched and merged in listed order with diff3, additive_merge.py on conflict, then merge_checks.py and pin_contract.py) on b10871, the base the preflight resolves at the time of writing:

ok ggml-org/llama.cpp#24423
ok ggml-org/llama.cpp#25731 (additive resolve)
ok unslothai/llama.cpp#70 (additive resolve)
ok unslothai/llama.cpp#61
ok unslothai/llama.cpp#95
ok ggml-org/llama.cpp#27754
ok unslothai/llama.cpp#137
ok unslothai/llama.cpp#158
ok unslothai/llama.cpp#157
ok unslothai/llama.cpp#149
ok unslothai/llama.cpp#144
ok unslothai/llama.cpp#152
ok unslothai/llama.cpp#176
ok unslothai/llama.cpp#210 (additive resolve)
merge_checks: clean
pin_contract: all 14 pins are intact in the merged tree

The new pin's two residual hunks are pure add/add against earlier pins (the CUDA proc-address table entry next to ggml-org#24423's, the exact-concurrency refusals in llama-memory-hybrid.cpp next to the kpool-dirty blocks), which additive_merge.py resolves.

Why the pin is composed with inkling rather than a single commit on the tag: inkling's pinned commit edits the mul_mat dispatch #197 restructured (its f32_pedantic guards) and adds declarations and a test struct at the same insertion points, a modify/modify overlap additive_merge.py cannot resolve. The pin commit therefore has the inkling pin as its second parent, with the guard folded into ggml_cuda_mul_mat_path, so the composed merge has one base. The cost: #210 has to be recomposed whenever ggml-org#25731 is repinned or dropped, which the repin bot does not do on its own.

[chatgpt-codex-connector]

You have reached your Codex usage limits for security reviews. Please try again later.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-09T13:24:17.549287Z	de95de4	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: de95de4612

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Register the new pin in feature-checks

When unsloth-pr-set-lint.yml runs for this pr-set.json change, it derives unslothai#210 from the new URL, but that identifier appears in neither features nor unchecked in scripts/unsloth/feature-checks.json; the "Every pin is either checked or knowingly unchecked" step therefore exits with status 1. Add a runtime feature check for the preemption/concurrency behavior, or explicitly record why this pin is unchecked, so the pin validation can pass.

Useful? React with 👍 / 👎.

[danielhanchen]

Fixed in e2e0390: #210 is recorded under unchecked in feature-checks.json with the reason (a server park has no arch, backend-op or mtmd probe surface; it is covered by test-exact-geometry and the server preemption tests in #197, which need a CUDA host). The lint pr-set.json run on that push passes.