DeepSeek-V4-Pro-DSpark-Drafter-GGUF

DeepSeek V4 Pro DSpark drafters

This repository contains two DSpark auxiliaries for DeepSeek V4 Pro. They are not standalone language models and do not include target-model weights.

Artifacts

Both 81-tensor GGUFs use the standardized llama.cpp general.architecture=dflash schema. Use MXFP4 for higher routed-expert fidelity when the complete launch fits. Use Q2_K when memory is tighter.

Download

hf download dev7a/DeepSeek-V4-Pro-DSpark-Drafter-GGUF \
  DeepSeek-V4-Pro-DSpark-Drafter-MXFP4-Q8_0-dflash.gguf

Provenance and verification

The source is deepseek-ai/DeepSeek-V4-Pro-DSpark revision 7c09739fd136abfb70a49ec334157f65f45b52cd. Only source shards 64 through 66 contain the complete auxiliary. Their sizes and SHA-256 values are pinned in manifest/source.json.

The repository contains the deterministic converter, standardized dflash rewriter, independent numeric verifier, tests, build manifests, payload comparisons, and checksums. Both quantization recipes and both schema rewrites were repeated and produced byte-identical results.

The standardized files carry the complete target tokenizer copied through the official llama.cpp DeepSeek V4 DSpark converter at commit 15586e2d7165570fb3aa7c26e0d442e289ef69de, including tokenizer.ggml.mask_token_id=128799. The Flash-compatible artifact produced by the same builder passed a real llama.cpp draft-dspark decode with 120 draft tokens generated, 38 accepted, and no invalid -1 token.

Reproduce either artifact on Linux AArch64 with Python 3.14:

uv venv --python 3.14.6 .venv
uv pip install --python .venv/bin/python \
  --require-hashes --only-binary=:all: \
  -r requirements-linux-aarch64-py314.lock
.venv/bin/python scripts/download_sources.py --destination sources
.venv/bin/python -m unittest discover -s tests -v

recipe=mxfp4-q8_0
.venv/bin/python scripts/reproduce.py \
  --sources sources --recipe "$recipe" \
  --manifest-dir manifest --repeat-check
.venv/bin/python scripts/dflash.py \
  --sources sources --recipe "$recipe" \
  --legacy-input DeepSeek-V4-Pro-DSpark-Drafter-MXFP4-Q8_0.gguf \
  --target-tokenizer-gguf tokenizer.gguf \
  --manifest-dir manifest --repeat-check

Use recipe=q2_k-q8_0 and the corresponding Q2_K legacy filename to reproduce the compact variant.

Compatibility

Use these files only with a compatible DeepSeek V4 Pro target. The runtime must support the standardized llama.cpp dflash GGUF schema, MXFP4, and Q2_K as required by the selected file. Always validate the complete target and drafter memory plan before acquisition.

This is a community conversion, not an official DeepSeek release. The source-derived weights remain under DeepSeek's MIT license. Conversion code and third-party notices are provided in LICENSE.code and THIRD_PARTY_NOTICES.md.