# MagicQuant adapted-downward manifest

This directory documents the Qwen3.8-27B Quark/AWQ native-MXFP4 release built from [`amd/Qwen3.8-27B-Quark-AWQ-MXFP4`](https://huggingface.co/amd/Qwen3.8-27B-Quark-AWQ-MXFP4).

Most recipe identities were selected from [`magiccodingman/Qwen3.8-27B-MagicQuant-GGUF`](https://huggingface.co/magiccodingman/Qwen3.8-27B-MagicQuant-GGUF), but this is not a pure clone. The source recipe types were adapted to an already-quantized MXFP4 parent using a strict downward-only rule: a recipe may reduce a tensor, but it may never increase a smaller source tensor merely to match the original recipe.

This release did not rerun the full MagicQuant evolution/search. It is a controlled, fixed-configuration downward ladder with fresh native-MXFP4 benchmarks.

## Public release files

- `magicquant.final-survivors.json` — ten published hybrids plus the documented, rejected `UD-IQ2_XXS` experiment, including fresh KLD/PPL, exact sizes, savings, hashes, imatrix identities, verification counts, and publication status. The standard filename is retained for compatibility with other MagicQuant releases; the rows are complete tested results, not a newly dominance-pruned survivor search.
- `magicquant.clone-configs.json` — exact effective type maps for all 866 tensors in every output. It records both the original recipe metrics and the fresh adapted-clone metrics.
- `magicquant.hybrid-map.json` — summarized tensor-group/type distributions, MTP block treatment, source recipe provenance, and strict-downward adaptation details.
- `magicquant.clone-benchmarks.json` — full fresh benchmark summaries against the native MXFP4 reference logits.
- `magicquant.imatrix.json` — generic and Unsloth imatrix provenance, hashes, sizes, coverage, and per-artifact use.
- `magicquant.projectors.json` — BF16 and Q8_0 vision projector construction, tensor types, hashes, and CPU-only load validation.
- `magicquant.provenance.json` — base model, AMD checkpoint, recipe repository, native repack, llama.cpp commit, and policy lineage.
- `magicquant.hashes.json` — exact release artifact byte sizes and SHA-256 values.
- `magicquant.release-boundary.json` — explicit public include/private exclude patterns for a safe Hugging Face upload.
- `index.json` — compact local experiment index used to assemble the release.
- `ladder-status.json` — sanitized final controller state showing that all planned ladder jobs completed.

`UD-IQ2_XXS` completed and passed file-integrity verification, but failed the release quality floor: KLD was `1.172122` and PPL was `17.87418` versus native PPL `5.801511`. It was also strictly dominated by the smaller `MQ-IQ2_XXS_1` (8.22 GB, KLD `0.321797`). Its GGUF is intentionally absent from the release; the experiment records and historical hash remain for transparency.

## Detailed records

`experiments/<name>/` contains the durable record for each GGUF:

- `config.json` — source, recipe, imatrix, runtime, and benchmark configuration;
- `commands.json` — reproducible command shapes with public placeholders instead of machine paths;
- `effective-tensor-types.txt` — direct llama.cpp tensor override file;
- `tensor-decisions.json` — all 866 source/requested/effective decisions and reasons;
- `type-summary.json` — source and effective type counts;
- `results.json` — size, KLD/PPL distribution, benchmark scope, and artifact hashes;
- `verification.json` — changed and byte-exact preserved payload counts.

`source-recipes/` contains the selected recipe snapshots from the source MagicQuant repository. These snapshots preserve source-run names, original metrics, and original 866-tensor type maps.

## Publication safety

Public JSON records use repository-relative names, `hf://` provenance identifiers, or explicit placeholders such as `<REPO_ROOT>` and `<SCRATCH_HOME>`. Local benchmark logs, scratch paths, the virtual environment, and the llama.cpp worktree are intentionally excluded from publication by the repository ignore files.

## Benchmark scope

All language KLD results use identical native MXFP4 reference logits, a 32,768-token normalized `Salesforce/wikitext` corpus, context 2048, four CPU threads, and zero GPU layers. The language benchmark does not evaluate the vision projector or MTP prediction path; those scope limits are present in every experiment result.