Repository navigation
workflows: pin actions by commit SHA; Dependabot keeps the pins current - #143
Merged
Merged
Conversation
actions/checkout v4.4.0, actions/upload-pages-artifact v3.0.1 and actions/deploy-pages v4.0.5, each by the commit its tag points to today, so a moved or compromised tag cannot change what runs. .github/dependabot.yml proposes new pins weekly as PRs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR Reviewer Guide 🔍Here are some key observations to aid the review process:
|
bong-water-water-bong
pushed a commit
that referenced
this pull request
Oct 25, 2026
… MTP cap flag `match.gguf_keys` is a new, optional recipe match key: the recipe applies only when every listed GGUF metadata key is present in the `-m` file. It is what separates PrismML's Ternary-Bonsai-2-27B from Qwen3.8-27B, which has the same `general.architecture = qwen35`: the ternary files carry `prism.hadamard.version`, Qwen3.8 files do not. `app/gguf_meta.h` gains `gguf_has_keys()`, which walks only the key/value section (same assumptions as `gguf_architecture`), and serve fills `RecipeFacts::gguf_keys` only when a recipe matches on them, so a launch without such a recipe does not read the header twice. The Bonsai recipe comes first in file order because the first recipe to add a flag wins: bonsai-ternary-mtp-n2 hrx + qwen35 + mtp + prism.hadamard.version -> --spec-draft-n-max 2 Ternary-Bonsai-2-27B accepts 0.76 of two-token MTP drafts against 0.67 at three, so two wins prose, short answers and translation (40.4 / 38.3 / 30.4 / 34.7 tok/s) and three to four wins code by 4-8% (43.7 / 36.0 / 24.6 / 30.1 at four). The plain `qwen35-mtp-max-3` recipe then finds the flag already set on those files and adds nothing, so Qwen3.8-27B keeps three. Also fixes `qwen35-mtp-max-3` itself: it added `--mtp-max 3` to **llama-server**, which has no such flag (serve's own `--mtp-max` maps to `--spec-draft-n-max`, app/serve.cpp); it now adds the flag llama-server really reads. DFlash's `--spec-draft-p-min 0` was already correct. Docs: `docs/recipes.md` documents `gguf_keys` and lists the Bonsai recipe; `docs/serve.md` gains a "Ternary-Bonsai-2-27B" subsection in the drafters section (the three files hold identical weights, PQ2_0 is the pick at pp512 553 / tg128 21.4, the Qwen3.8 MTP head attaches, two drafts by default and four for code, DFlash2 works since fork #143 but trails MTP at 0.37 against 0.67 acceptance, and both come with the next private-add-on pin); `docs/hrx.md`'s PrismML bullet records PQ2_0 on the dense prefill kernels and ternary weights on the 2..16-token verify kernels. Verified locally without a GPU: a standalone harness linking the real `app/recipes.cpp` builds synthetic GGUFs with and without `prism.hadamard.version` and checks that the recipe matches only the file that carries it, that an empty list matches anything, that a missing key does not match, and that a non-list `gguf_keys` is refused; `tests/recipes_route.sh` covers the same through `1bit serve` with its fake backend (present, absent and non-list cases added), and its header comment now names the key. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
bong-water-water-bong
added a commit
that referenced
this pull request
Oct 25, 2026
llama.cpp along 1bit/hrx-vulkan-patched 851fa6c8d8f2 -> e7e5cd08f9bb: DFlash2 drafters on the pinned branch (#139) + FILL cache key / TOP_K tests (#140), Qwen3.8-Flash-Next on HRX with the HIP-defer guard (#142) and IQ2_S routing (#144), DFlash2 Hadamard folds for Ternary-Bonsai (#143), HRX test-suite / Flint / jinja changes (#126-#135). hrx-system held at 82804197bcbe. hw-gate green after the KLD base logits were restored on the gate box. The private kernel pin follows in its own PR. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The workflows used
actions/checkout@v4,upload-pages-artifact@v3anddeploy-pages@v4. A tag can be moved, so the code that runs could change underneath us. This pins each one to the commit its tag points to today:actions/checkout11d5960aactions/upload-pages-artifact56afc609actions/deploy-pagesd6db9016.github/dependabot.ymlproposes new pins weekly, as PRs that run the usual checks. Once this merges, the repo setting that requires SHA-pinned actions gets turned on.🤖 Generated with Claude Code