# Upload manifest

This local release candidate is intended to contain the following published payload.

## Upload

- `model.safetensors` — the final packed W4A16 model payload.
- `config.json`, `generation_config.json` — architecture/generation metadata including the compressed-tensors quantization configuration.
- `tokenizer.json`, `tokenizer_config.json`, `processor_config.json`, `chat_template.jinja` — inherited tokenizer, processor, and template files required for normal Gemma 4/vLLM use.
- `README.md` — model card.
- `LICENSE` — the inherited Apache-2.0 license text.
- `SHA256SUMS` — checksums of all uploaded payload files.
- `benchmarks/rtx3090-performance.json` — public-safe performance data.
- `benchmarks/functional-canary.json` and `benchmarks/rtx3090-128k-validation.json` — public-safe functional and near-limit validation records.
- `benchmarks/perplexity.json` — public-safe matched perplexity result.
- `provenance/calibration-and-source.json` and `provenance/artifact-audit.json` — compact reproducibility records.
- `eval/evaluate_gemma4_ppl_vllm.py` — perplexity implementation.
- `examples/` — small public client/serve examples if added.

## Keep private

- All calibration images, image manifests containing captions/prompts, and text calibration windows.
- The 49 GB BF16 tuned/QAT parent checkpoint and its 11 source shards.
- Raw conversion logs, full vLLM logs, offload directories, environment caches, and compiler caches.
- Original absolute-path provenance files, host/user/kernel details, and any credentials/tokens.
- Trace artifacts, failed attempts, and internal planning notes.

The staged `model.safetensors` must be the byte-identical final artifact; no model-weight file is modified by release packaging.