alessandrobologna/DeepSeek-V4-Pro-DSpark-Drafter-GGUF

DeepSeek V4 Pro DSpark Drafter GGUF

This repository contains only the DSpark auxiliary drafter for DeepSeek V4 Pro. It is not a standalone language model and does not contain the 893 GB target checkpoint.

The GGUF is built from the official deepseek-ai/DeepSeek-V4-Pro-DSpark checkpoint using the standalone Python scripts included here. The source manifest, deterministic tensor recipe, structural parser, numeric verifier, dependency pin, and build evidence are published alongside the artifact.

Artifact

File	Bytes	SHA-256
DeepSeek-V4-Pro-DSpark-Drafter-Q2_K-Q8_0.gguf	26,616,105,472	1ddd0ad10583658809ec6b8c69b02214c83e1d2e0cd89e0c197d86eda10cede7

The artifact uses GGUF v3 and the deepseek_v4_pro_dspark_draft architecture. Its 81 tensors use this policy:

Output type	Tensors	Transformation
F32	45	BF16 expanded exactly to F32, or source F32 bytes preserved
F16	2	BF16 Markov heads rounded to F16
Q8_0	25	FP8 E4M3 weights with E8M0 block scales dequantized, then quantized to Q8_0
Q2_K	9	Routed-expert MXFP4 decoded and requantized with the unweighted DS4 reference Q2_K algorithm

The F16, Q8_0, and Q2_K transformations are lossy and are checked against explicit numeric error bounds. This repository does not claim that the complete conversion is lossless.

Pinned source

• Repository: deepseek-ai/DeepSeek-V4-Pro-DSpark
• Revision: 7c09739fd136abfb70a49ec334157f65f45b52cd
• Input: only safetensors shards 64, 65, and 66 of 66
• DSpark source tensors: 7,009
• Total source size: 41,980,133,916 bytes

The three tail shards contain the complete DSpark module and no target-model tensors. The other 63 shards are not downloaded. Exact input sizes and SHA-256 values are in manifest/source.json; the downloader and converter reject any mismatch.

Reproduce

The reference build uses Linux aarch64, Python 3.14.6, uv 0.11.29, and NumPy 2.5.1 from the hash-pinned binary wheel in requirements-linux-aarch64-py314.lock. Source download requires the hf CLI; the reference build used version 1.24.0. The upstream repository is public, so a Hugging Face token is not required.

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
.venv/bin/python scripts/reproduce.py \
  --sources sources \
  --recipe q2_k-q8_0 \
  --manifest-dir manifest \
  --repeat-check
sha256sum -c SHA256SUMS

Each downloaded byte is checked against manifest/source.json before use. --repeat-check performs a second clean conversion and requires the complete GGUF SHA-256 to match byte for byte.

The converter memory-maps the source shards, processes FP8 tensors in 128-row bands, and uses at most sixteen ordered worker threads for Q2_K. Each worker keeps at most one decoded routed expert resident, and results are written in expert order so scheduling cannot affect the GGUF bytes. The build does not require PyTorch, CUDA, safetensors, a GGUF package, or a GPU.

Verification

The included verifier checks:

• all pinned source sizes and hashes;
• all 7,009 source tensor names, dtypes, shapes, and byte ranges;
• GGUF version, metadata, tensor order, dimensions, types, offsets, alignment, non-overlap, zero padding, and final length;
• exact F32 and F16 output bytes;
• independent Q2_K block decoding and block-normalized reconstruction error against a fresh MXFP4 source decode for all nine routed-expert tensors; and
• exact Q8_0 scale/code encoding plus block-relative error against a fresh FP8 source decode.

Machine-readable evidence is written to:

• manifest/build-q2_k-q8_0.json
• manifest/validation-q2_k-q8_0.json
• SHA256SUMS

The Q2_K verifier enforces a maximum block-relative reconstruction error of 0.55; the measured maximum was 0.5107014775276184. Q8_0 enforces 0.01; its measured maximum was 0.00396728515625. The Pro Q2_K bound covers the measured maximum that exceeded the Flash recipe's 0.50 bound; it is not a claim about end-to-end model quality. The focused unit test also requires byte-for-byte agreement with a nontrivial 84-byte block emitted by the attributed C reference quantizer.

Compatibility and limitations

• This is an auxiliary drafter, not a complete checkpoint.
• A consumer must explicitly support the deepseek_v4_pro_dspark_draft architecture and dspark.* tensor namespace.
• Generic GGUF support alone is insufficient. No compatibility with llama.cpp, Ollama, LM Studio, vLLM, or other runtimes is implied.
• The drafter must be paired with a compatible DeepSeek V4 Pro target model.
• The initial runtime validation target is antirez's DeepSeek-V4-Pro-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-Instruct-imatrix.gguf at revision a88c423b511666d7ff7a4dcaee651669312bea97, SHA-256 a0314d9c0e16122cd60071079124a2d17185d317c55a8f95ecb3ed3506278a96.
• This repository makes no universal quality, acceptance-rate, or speed claim across target quantizations or runtimes.
• The artifact is a community conversion and is not an official DeepSeek release or endorsement.

DSpark configuration

The pinned source defines:

• block size: 5
• target layers: 58, 59, and 60
• Markov rank: 512
• hidden size: 7,168
• routed experts per layer: 384
• vocabulary size: 129,280

These values are embedded in the GGUF metadata and validated during the build.

License and references

The source-derived weights remain under DeepSeek's MIT terms in LICENSE. The conversion and verification scripts are under LICENSE.code. The pinned checkpoint's upstream license and the DS4/GGML attribution in THIRD_PARTY_NOTICES.md provide the applicable notices.

• Official DeepSeek V4 Pro DSpark checkpoint
• DeepSeek V4 technical report
• DeepSpec