# Swift1.5 Q5 conversion

Source: `ukisai/Swift1.5-Qwen3.8-Flash-Next@0bd4fe22431372cdad1979267d3ab45aa7e6150a`.
Recipe: published Qwen `MQ-Q5-SSD-PLE-BF16`, source-map SHA-256
`bedb8b03cd6a06072c1e0b31eda42fc8be4642d22a5889ebd5da10d902fd3de9`.

Keep `source/`, `work/` and this package under `publish/` in one parent directory.
Install `requirements.txt` into `work/venv`. The release used CPU-only
`torch==2.11.0+cpu`; install torch from `https://download.pytorch.org/whl/cpu`.

```bash
hf download ukisai/Swift1.5-Qwen3.8-Flash-Next \
  --revision 0bd4fe22431372cdad1979267d3ab45aa7e6150a \
  --local-dir "$root/source" --exclude '*.mp4' --exclude '*.png' --exclude 'benchmarks/*'
```

`manifests/swift/recipe-base-source-map.json` ports the reference tensor map
without changing roles, shapes, slices or quantization targets. The planner
reads the actual Swift integer PLE controls and verifies exact Q5/PLE-plan
equivalence to the reference. The converter refuses to overwrite output GGUFs.

Before copying PLE, run:

```bash
python publish/reproduction/scripts/verify_ple_identity.py \
  --model-dir "$root/source" \
  --reference publish/reproduction/manifests/reference/bf16-ple-manifest.json \
  --reference-files publish/reproduction/manifests/reference/base-source-info.json \
  --candidate-files publish/reproduction/manifests/swift/upstream-source-info.json \
  --out publish/reproduction/manifests/swift/verify-ple-identity.json
```

The 31 complete PLE-only source shards are byte-identical to pinned Qwen BF16
shards. The four PLE tensors in modified mixed shards are hashed directly.
After all 128 parts pass, copy the existing Qwen FP8 sidecar:

```bash
cp -r ~/workspace/ds4-exaone/models/Qwen3.8-Flash-Next-Mixed-Quant-SSD-PLE-GGUF/PLE-FP8 \
  "$root/publish/PLE-FP8"
bash "$root/publish/reproduction/run.sh" "$root" build
```

`run.sh` verifies all selected source files and all 1,658 source tensors,
builds the main map, checks exact recipe/PLE layout equivalence, verifies PLE
identity and copied hashes, converts 1,628 main tensors, audits GGUF structure,
independently hashes weights and writes the artifact manifest/card.

Main GGUF metadata keeps the existing BF16 packaging contract. Runtime FP8
selection requires `DS4_QWEN_PLE_DIR`; no BF16 PLE files are bundled here.
There is no imatrix dependency. Source BF16 data remain mmap-backed.

`publish` mode uploads to the fixed Baekpica repository after local gates,
checks public visibility, all seven large-file LFS hashes/sizes, and fully
reads/hashes every smaller staged file without authentication. It then
publishes the verification receipt and verifies the final revision.

Run focused packaging tests with:

```bash
cd publish/reproduction/scripts
python -m unittest test_ple_identity test_release_manifest test_publication
```

Conversion checks do not qualify Swift GPU inference, serving, performance
or task quality. Original Qwen FP8 extraction receipts are retained separately
from the fresh package checks.