Sn1waR/DeepSeek-V4-Flash-0731-CRACK-DSpark

DeepSeek V4 Flash 0731 CRACK — DGX Spark

This is a refusal-subspace-edited derivative of the exact deepseek-ai/DeepSeek-V4-Flash-0731 checkpoint at revision 9e165c30e2704aec5d9d593cce3eebd58bbef1cb. It retains the upstream 48-shard FP8 E4M3 + UE8M0 layout, all 43 main layers, and the three stock DSpark MTP projections. It is intended for CUDA/vLLM tensor-parallel inference on two NVIDIA DGX Sparks.

This is not the Apple MLX/JANG checkpoint and is not compatible with the JANG runtime. The reproducible editor, direction artifact, deployment profiles, tests, and detailed research record are maintained in jiayuqi7813/DeepSeek-V4-Flash-0731-CRACK-2x-DGX-Spark.

What changed

• Target tensors: main-layer attention wo_b, layers 10–42 inclusive.
• Direction: one SRA-cleaned, rank-1 attention-output direction per layer.
• Transform: strength-2 Householder reflection, W' = W - 2 Vᵀ(VW).
• Quantization: native 128×128 FP8 blocks with the original UE8M0 scales held fixed; only the selected FP8 weight tensors change.
• Row-norm rematching: disabled.
• MTP: stock upstream projections, unchanged.
• Direction SHA-256: fe8c263a8d32deb71e3f6e866b90f8246f452f6e2103b0e0400a77480fd2602a.

The exact-arithmetic transform reverses the selected output component while preserving each matrix's column Gram matrix before requantization. No gradient training, fine-tuning, LoRA, or synthetic-response training was used.

Validation summary

The candidate was built independently on both DGX nodes. Each copy passed tensor-granularity validation across all 72,317 indexed tensors: all non-target tensors in rewritten shards were bit-exact, fixed scale tensors were bit-exact, and every rewritten-shard SHA-256 matched across nodes.

Locked relative capability comparisons used the same eager/no-MTP TP=2 profile for the source and edited checkpoints:

Evaluation	Source	Edited
Synthetic regression suite	13/19	13/19
GSM8K deterministic 100-question screen	97/100	96/100
HumanEval deterministic 32-task screen	27/32	27/32
HumanEval deterministic full 164-task confirmation	150/164	148/164
MMLU-Pro balanced 560-question confirmation	379/560	382/560

The full HumanEval delta was -1.22 percentage points. Paired transitions were 147 pass/pass, 3 source-pass/edited-fail, 1 source-fail/edited-pass, and 13 fail/fail (exact McNemar p=0.625). All 328 generations ended naturally; no length-limited output was scored.

On the 85-prompt harmful holdout, an untouched-source three-way judge measured material compliance (PARTIAL or COMPLY) at 3/85 (3.53%) for source and 77/85 (90.59%) for edited. Strict COMPLY was 1/85 (1.18%) versus 44/85 (51.76%). The separate 10-prompt English/Chinese harmful probe measured material compliance at 0/10 versus 10/10, with matched harmless probes showing no new lexical refusals.

A 12-prompt English/Chinese open-ended writing comparison used anonymous, position-balanced A/B judging by the untouched source checkpoint. All answers ended naturally under the final 768-token cap. The edited model won 4, source won 2, and 6 tied; mean four-dimension totals were 19.25/20 edited versus 18.58/20 source. Manual review conservatively corrected two dialogue verdicts with flawed rationales to ties, leaving 2 edited wins, 2 source wins, and 8 ties. Deterministic hard-constraint compliance was 8/12 for both models; both missed the requested length on the same four explanation/argument tasks.

These are preservation comparisons, not claims of direct comparability with the original benchmark papers. See the code repository for dataset revisions, selection hashes, prompts, isolated HumanEval execution, paired transition counts, behavior results, and production performance measurements.

Runtime

The validated runtime is:

• ghcr.io/anemll/dspark-vllm-gx10:0.1.1
• image ID sha256:3430d6614a8e2925f34d059af6caf05aff42387326db4d05639a60f10f2654d8
• vLLM TP=2 / NCCL across two DGX Sparks
• DSpark speculative decoding with five MTP tokens in production
• declared maximum context length: 1,048,576 tokens

Use the open-source repository's production profile instead of a generic vLLM command: it carries the GB10, tokenizer, NVFP4 KV-cache, MTP, CUDA-graph, and dual-CX-7 settings verified for this checkpoint.

Limitations and intended interpretation

This edit intentionally changes refusal behavior and should not be represented as the alignment or safety behavior of the upstream model. A non-refusal is not proof of factual correctness, completeness, legality, or operational fitness. The main behavioral holdout is English; the separate English/Chinese probe is small and descriptive. Rare FP8 clipping, long-context behavior beyond the synthetic retrieval screen, languages outside the measured set, multi-user serving, and downstream fine-tunes can behave differently. The writing probe is deliberately small, and the full HumanEval point estimate is slightly lower even though it remains inside the locked preservation envelope and is not statistically significant.

Operators and downstream distributors are responsible for evaluating the model in their own application context and accurately disclosing that it is a modified derivative.

License and attribution

The upstream checkpoint and this derivative are distributed under the MIT license included with the files. Preserve the DeepSeek copyright notice, this model card, CRACK_EDIT_MANIFEST.json, CRACK_EDIT_REPORT.json, and CRACK_VALIDATION.json when redistributing the edited weights. Dataset and tooling notices are recorded in the code repository's THIRD_PARTY_NOTICES.md.

SOURCE_HF_MANIFEST.json records the pinned upstream download and original LFS hashes for provenance. It is not a hash manifest for the edited shard payloads; the derivative shard hashes are recorded in CRACK_EDIT_REPORT.json.