keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors
Drop-in abliterated weights for official DeepSeek-V4-Flash-0731 (GA) that still run MiaAI-Lab’s production 2× DGX Spark recipe with DSpark speculative decoding intact.
The ablit is L10–35 only. DSpark target layers 40–42 (and L36–39) are restored hash-identical to official stock safetensors. MTP / draft tensors were never edited (edit_mtp: false).
This replaces the earlier L10–42 drop on this repo (…-Abliterated-32-32). Same gated terms. Old URL redirects here.
| HF | https://huggingface.co/drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors |
| Base | deepseek-ai/DeepSeek-V4-Flash-0731 @ 9e165c30 |
| Ablit | L10–35 · λ=3.5 · k=1 · 26× attn.wo_b · mean Δrel ≈ 0.057 |
| Anchors | L36–42 shards = official SHA-256 · DSpark targets [40, 41, 42] · MTP stock |
| Scales | float8_e8m0fnu preserved |
| Runtime | Anemll ghcr.io/anemll/dspark-vllm-gx10:0.1.1 + MiaAI 2× recipe · TP=2 · MTP=5 · nvfp4_ds_mla · util ≤ 0.85 |
Full credit: Anemll/dspark-vllm-gx10 · MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark · DeepSeek-AI
See RESPONSIBLE_USE.md and the gate form above.
Why “anchored tensors”
0731 ships DSpark inside the checkpoint (dspark_target_layer_ids: [40, 41, 42]). Projecting wo_b through L42 walks through the drafter. The target stops refusing; the stock drafter keeps proposing refusal-shaped tokens; acceptance dies.
This drop restores those safetensor shards to official stock:
| Layer | Shard | dest SHA-256 prefix |
|---|---|---|
| 36 | model-00038-of-00048.safetensors | 137fa617a74ba8e7 |
| 37 | model-00039-of-00048.safetensors | a29af1aa519d7ce7 |
| 38 | model-00040-of-00048.safetensors | 8bc93d8a7d1987dc |
| 39 | model-00041-of-00048.safetensors | fd312e7fdd6cb579 |
| 40 | model-00042-of-00048.safetensors | 4d19bf368083c9a1 |
| 41 | model-00043-of-00048.safetensors | b7103842ceb70848 |
| 42 | model-00044-of-00048.safetensors | 422d3889fa20c238 |
Invariant (DSPARK_ANCHOR_MANIFEST.json): restored target shards hash-identical to official stock; live source and all MTP/draft tensors unchanged.
Live vs official 0731 (same Mia production stack, 2026-08-23)
2× DGX Spark · Anemll 0.1.1@sha256:a83948492cf1… · vLLM 0.25.2.dev0+g752a3a504 · 1M ctx · seqs=6 · util 0.835 · thinking off.
Real tasks (C1 decode / C6 agg):
| Task | Official 0731 | This checkpoint |
|---|---|---|
| structural (JSON fill-512) | 64.1 / 173 | 66.0 / 171 |
| list | 84.7 / 213 | 81.1 / 315 |
| essay | 31.9 / 90 | 34.1 / 94 |
| reading | 35.2 / 99 | 34.6 / 93 |
MiaAI matrix (bench-miaai.py, 128 tok, thinking false):
| Cell | Official | This checkpoint |
|---|---|---|
| c=1 p=256 | 67.6 | 69.2 |
| c=6 p=256 best agg | 167.5 | 173.9 |
| c=1 p=128K | 76.1 · TTFT 85s | 67.4 · TTFT 83s |
| KV tokens | 2,490,368 (2.38× @ 1M) | 2,471,790 (2.36×) |
Smoke 6/6 both legs. Long c=6 is Mia issue #27 (serialized long prefill), not an ablit regression.
Download
# after access is approved
hf download drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors \
--local-dir ~/models/DeepSeek-V4-Flash-0731-ablit-anchored
The previous id drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-32-32 redirects here.
Serve (Mia + Anemll, 2× DGX Spark)
docker pull ghcr.io/anemll/dspark-vllm-gx10:0.1.1
# In MiaAI DeepSeek-v4-Flash-DSpark-2x-DGX-Spark .env.dspark:
# ABLITERATED=1
# DSPARK_MODEL_ABLITERATED=/path/to/DeepSeek-V4-Flash-0731-ablit-anchored
# DSPARK_REVISION_ABLITERATED=
# DSPARK_ENCODING_FILE=.../encoding/encoding_dsv4.py
# MTP_NUM_TOKENS=5 GPU_MEMORY_UTILIZATION_TEXT=0.835 MAX_MODEL_LEN=1048576
Files
| Path | Purpose |
|---|---|
model-*-of-00048.safetensors + index | Full FP8 checkpoint (DSpark MTP stock + restored L36–42) |
DSPARK_ANCHOR_MANIFEST.json | SHA-256 invariant for restored shards |
encoding/encoding_dsv4.py | 0731 tokenizer encoding for Anemll/Mia |
ABLIT_META.json | Edit stats / recipe fingerprint |
ablit/ | Refusal-direction artifacts |
Credits
- DeepSeek-AI — DeepSeek-V4-Flash-0731
- Anemll — https://github.com/Anemll/dspark-vllm-gx10
- MiaAI-Lab — https://github.com/MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark
- Keys (drowzeys) — abliteration + safetensor anchors