Repository navigation
Model catalog - #1
Merged
Merged
Conversation
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- Measured with gguf-parser v0.26.0, which reads GGUF headers remotely - Parameters: ctx-size 32768, flash-attention on, q8_0 K/V cache, full offload - Records estimate.items[0].vrams[0].nonuma, the discrete-GPU figure - Quants restricted to the UD-Q3_K_XL..Q6_K band Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- One section per model, tuning only; no filesystem paths, so the files work unmodified on any machine via --models-dir - Sampling values taken from each model author's published guidance and cited in the file header - Model ids are the HuggingFace repo name minus -GGUF, lowercased Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Picks the highest-quality quant that fits each tier minus 1 GiB of headroom, then concatenates the matching model sections under a shared [*] block. Generation is necessary because the llama.cpp preset format has no include directive, so multi-file configs cannot be composed at runtime. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Checks each rule the repo relies on: lowercase model ids, file name matching the section name, no filesystem paths, required ctx-size, allowed quants, and keys that llama.cpp actually accepts. Also rejects the three keys llama.cpp silently ignores in presets (host, port, api-key) and values containing ';' or '#', which truncate the rest of the line. Every failure reports the file, the line, the problem and the fix. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Covers the required --models-dir layout, the quant chosen per tier, the build step, and the preset keys that do not work as expected. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Written as short imperative rules so smaller local models can follow it. Records llama.cpp router behaviour that is not documented upstream and was established by reading the source and testing against a running server. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds --notes, which prints a Markdown summary of every profile and the quant chosen per model. Lets the release workflow describe a build without duplicating the tier-selection logic. Generated dist/ output is unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Rebuilds dist/ and fails if it does not match configs/, then runs the validator. An unrecognised key makes llama-server exit 1 at startup, so the key check catches that failure without needing the binary in CI. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Pushes to main that touch configs/ or src/ validate the catalog and publish dist/*.ini plus SHA256SUMS under a CalVer tag, so users can download the current profiles without cloning. Documentation-only commits do not release. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
dist/ is a build artifact and releases already publish it, so tracking it only doubled every diff, invited merge conflicts, and created the staleness problem CI then had to police. - git-ignore dist/ and drop it from the index - remove the dist/ freshness check from both workflows - CI uploads the generated presets as an artifact so reviewers can still see what a change to configs/ produces Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Release notes now combine the generated catalog (usage, profiles, quant per model) with a Conventional Commits changelog produced by git-cliff. The tag does not exist when the changelog is generated, so --unreleased --tag renders the pending commits under the new version. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- Emit only the grouped commit list; the release title already carries the version, so the template no longer duplicates it as a heading - Point issue links at this repository instead of the upstream placeholder - Group build alongside ci, so build commits stop falling into "Other" - Drop the test, perf and body-matched security groups; this repo ships configuration, has no test suite and nothing to profile - Remove the dead "^ci" branch from the skip rule, which never matched because the ci group is evaluated first Co-Authored-By: Claude Opus 5 (1M context) <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
1 participant
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.
No description provided.