Repository navigation
pr-set: pin server penalize_prompt (#283), drop ggml-org#30670 merged upstream - #284
Conversation
penalize_prompt = false keeps prompt tokens out of the repeat, presence, frequency and DRY history in llama-server, which the FLUX.3 prompt upsampler needs to match its reference sampling. feature_matrix.py gets a penalize-prompt probe: a synthetic llama GGUF from test-llama-archs -o, a context whose greedy next token is already in it, and a presence penalty that must move that token by default and must not with the field false. It is deferred when the build has no llama-server (the CPU gate).
Upstream squash-merged it as 01fc3aa, so the pin no longer merges onto any tag from b11712 on (REFUSED in common/parsers and src/llama-arch*, the 2026-10-17 nightly failure). pin_contract already reports 513/513 of its lines in the base. The beam feature check stays, owned by the base.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5c5e228999
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".
| if not server.exists(): | ||
| raise NeedsServer("needs llama-server, which this build did not produce") |
There was a problem hiding this comment.
When the documented --gpu verification is run against a build that accidentally omitted llama-server, this raises NeedsServer, which subclasses NeedsGPU and is therefore recorded as a deferral; the command exits successfully with report["ok"] == true and can claim all features were demonstrated without testing the newly pinned server feature. Only the known CPU gate should defer this condition; an explicit GPU/full-build run should treat the missing target as unproven.
Useful? React with 👍 / 👎.
…erred Only the CPU gate builds without llama-server; a full build that lacks it has not demonstrated the penalize_prompt pin.
|
@codex review |
|
@codex security review |
|
Codex Review: Didn't find any major issues. Can't wait for the next one! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
🛡️ Codex Security ReviewSecurity review completed. No security issues were found in this pull request. Reviewed commit: Only the user who started this review can view the report in Codex. ℹ️ About Codex security reviews in GitHubThis is an experimental Codex feature. Security reviews are triggered when:
Once complete, Codex will leave suggestions, or a comment if no findings are found. |
pr-set: pin server
penalize_prompt(#283), drop ggml-org#30670The unslothai/llama.cpp prebuilt is upstream ggml-org at an aged
b####tag plus the pins inscripts/unsloth/pr-set.json, merged in order (git diff3, thenadditive_merge.pyfor pure add/add conflicts only). No fork carry: a pin is deleted once upstream has the work.Changes
4f26c617488627c927c96b6f789a064e5d06668a: the per-requestpenalize_promptfield for llama-server. Withfalse, prompt tokens stay out of the repeat, presence, frequency and DRY history, which the FLUX.3 prompt upsampler needs in order to match its reference sampling. Defaulttrueis unchanged behaviour.feature_matrix.pygets apenalize-promptprobe, andfeature-checks.jsonapenalize-promptfeature owned by server: add penalize_prompt to keep prompt tokens out of the penalty history #283. The probe works offline:test-llama-archs -a llama -owrites a synthetic GGUF, llama-server decodes a token-id context whose greedy next token is already in it, and a presence penalty of 100 must move that token by default and must leave it alone withpenalize_prompt: false. A build with no llama-server (the CPU gate) defers it, the same way GPU-only checks are deferred.test_feature_matrix.pycovers that deferral.01fc3aae8, so fromb11712on the pin is REFUSED incommon/parsers/parsers.h,common/parsers/sources.cmake,src/llama-arch.cpp,src/llama-arch.handsrc/llama-model.cpp. That is the cause of the 2026-10-17 nightly failure (run 39066439326).pin_contractalready reported 513/513 of its lines present in the base. Thebeamfeature check stays, now owned by the base.Evidence per tag
llama_pinset.py sim --tags window(b11712 aged to b11716 newest), 18 pins:Before dropping ggml-org#30670, every tag in the window failed on that pin alone. #283 merged cleanly everywhere.
llama_pinset.py verify --tag aged --cuda(b11712) and--tag newest --cuda(b11716), B200:verifyreadsfeature-checks.jsonfrom master, so it ran the 11 existing features. I ran this branch'sfeature_matrix.py --gpuwith this branch's manifest on the same merged CUDA trees: 12/12 features ok, includingpenalize-prompt. On the CPU gate trees, which have no llama-server,penalize-promptis deferred.The probe on its own: it passes on the pin build (prompt token kept with
penalize_prompt: false, moved by the default prompt-wide penalty) and fails on plainb11696("penalize_prompt = false still penalized the prompt").Not covered here
auditalso flags CUDA: faster MoE prompt processing (compact MMQ tiling, device expert sort, GET_ROWS) #264 and llama-mmap: populate the non-lazy ranges when MAP_POPULATE is skipped for lazy tensors #266: their PR heads have moved past the pinned commits. That is already the case on master and is not touched here.