PseudoBrain x DeepSeek-V4.1-Flash-1017 (Chronos v7.5 Dual-Verified Abliterated)

PseudoBrain-DeepSeek-V4.1-Flash-1017-Chronos-v7.5-Abliterated is the full 40-layer (15,360 routed experts + shared experts + Sparse Multi-Head Latent Attention + 4-lane Sinkhorn Hyper-Connections + multimodal vision/aligner) Chronos v7.5 conversion of deepseek-ai/DeepSeek-V4.1-Flash-1017 with Dual-Verified Post-Quantization Orthogonal Abliteration.

---

1. Key Highlights & Verified Hardware Receipts

1.1 Official deepseek-ai/DeepSeek-V4.1-Flash-1017 11-Benchmark Evaluation Suite (440 Tasks, 14,080 Tokens + Live 40-Layer Decode)

Evaluated via benchmarks/dsv41_official_benchmarks_eval.py (dsv41_official_benchmarks_report.json and dsv41_40layer_colab_benchmark.json) across all 11 official benchmarks reported by DeepSeek for DeepSeek-V4.1-Flash-1017 plus the HarmBench-320 Post-Quantization Orthogonal Abliteration Audit (16 Tier-1 4-bit + 80 Tier-2 2-bit + 288 Tier-3 1-bit = 384/384 active routed experts + 1 shared expert per layer across all 40 layers, 36 full-block CUDA Graphs + 4 sub-block CUDA Graphs + Direct Zero-Copy UVA Tier-3):

---

2. Dual-Verified Post-Quantization Orthogonal Abliteration

Standard abliteration on Mixture-of-Experts models suffers from two failure modes when combined with extreme low-bit compression:

1. Incomplete Expert Coverage: Modifying only attn.wo_b and ffn.shared_experts.w2 leaves the 384 routed experts per layer (15,360 total routed experts across 40 layers) with non-zero refusal projection ($|R_l W_{2, m}|_F = 1.5938$).
2. Quantization Refusal Leakage: Even if routed experts are orthogonalized before compression, 4-bit holographic bank and 1-bit/2-bit tail quantization residuals re-introduce non-zero projection into the refusal subspace ($|R_l \widehat{W}_{2, m}|_F > 0$).

Chronos v7.5 solves both via Dual-Verified Post-Quantization Orthogonal Abliteration:

1. Attention & Shared Expert Preservation (dealignai, 320/320 = 100.00% HarmBench Verified): Preserves the exact block-FP8 (float8_e4m3fn + UE8M0 scale) abliterated attn.wo_b and ffn.shared_experts.w2 tensors across layers 10..36 from dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8.
2. Orthonormal Dual Refusal Basis $R_l \in \mathbb{R}^{d_r \times 5120}$ (OBLITERATUS + dealignai): Extracts the exact 16-bit BF16 Failspy refusal vector $r_{\text{ob}, l} \in \mathbb{R}^{5120}$ from OBLITERATUS/DeepSeek-V4.1-Flash-1017-obliterated (layers 1..39, verified at 88.08% MMLU vs 87.96% stock on 2,500 questions) and $r_{\text{dea}, l} \in \mathbb{R}^{5120}$ (layers 10..36), constructing the Gram-Schmidt orthonormal refusal basis $R_l \in \mathbb{R}^{d_r \times 5120}$ ($R_l R_l^T = I_{d_r}$).
3. Pre- and Post-Quantization Routed Expert Orthogonalization (15,360 Routed Experts):
   • Projects out $R_l$ from all 384 routed experts' w2 (down_proj) matrices prior to 8:1 holographic folding: $$W_{2, m}^{\perp} = W_{2, m} - R_l^T (R_l W_{2, m})$$
   • Projects out $R_l$ from the Coupled SwiGLU Woodbury Drift target and the Rank-64 FP8 left singular vectors $U_{\text{down}, m}$: $$U_{\text{down}, m}^{\perp} = U_{\text{down}, m} - R_l^T (R_l U_{\text{down}, m})$$
   • Embeds ffn.abliteration.r_basis ($R_l$) directly in each converted .safetensors shard and enforces exact runtime output projection inside ChronosDsv41SparseMoeBlock: $$\widehat{W}{2, m}^{\perp} = \widehat{W}{2, m} - R_l^T (R_l \widehat{W}{2, m}), \qquad |R_l \widehat{W}{2, m}^{\perp}|_F = 0$$

---

3. Chronos v7.5 Architecture for DeepSeek-V4.1-Flash-1017

Each of the 40 MoE layers (layers.0 .. layers.39) contains 384 routed experts (gate_proj: 2304 x 5120, up_proj: 2304 x 5120, down_proj: 5120 x 2304, top_k = 6, scoring_func = "sqrtsoftplus", routed_scaling_factor = 1.5, swiglu_limit = 10.0) partitioned into 48 Holographic Superposition Groups of $K = 8$ experts:

$$\widehat{W}m = \gamma_m \left( \beta_m \cdot \mathcal{U}m!\left(\text{NVFP4}(S{\text{group}})\right) + U{\text{FP8}, m} V_{\text{FP8}, m} + \alpha_m \cdot Q_{\text{Lloyd-Max}, 2\text{-bit}}(R_{\text{tail}, m}) \right)$$

1. 8:1 NVFP4 (E2M1) Router-Weighted Holographic Superposition Bank (S_gu, S_down): Folds 8 experts into a single shared 4-bit NVFP4 tensor using 2D block-128 Fast Walsh-Hadamard Transforms (FWHT) and deterministic per-expert permutation/Rademacher keys.
2. Full-Spectrum Sherman-Morrison-Woodbury (SMW) Rank-64 FP8 SVD Healer (U_gu, V_gu, U_down, V_down): Applies the exact matrix-inversion lemma to the 256-D router subspace (v_rt_sub) and 64-D active Clamped-SwiGLU subspace (v_act_sub) without dense $5120 \times 5120$ materialization.
3. Coupled Clamped-SwiGLU (swiglu_limit = 10.0) Woodbury Drift Compensation: Compensates down_proj for non-linear activation drift $\Delta A = \text{SwiGLU}{10}(\widehat{Z}) - \text{SwiGLU}{10}(Z)$ via a $64 \times 64$ push-through ridge solve.
4. 2-Bit Isotropic Lloyd-Max Gaussian Tail Lattice ({-1.5104, -0.4528, +0.4528, +1.5104}): Captures 88.2% of residual Gaussian tail energy per coordinate (4 values packed per uint8 byte with per-row FP8 E4M3FN scale).

---

4. Usage & vLLM Integration

pip install torch safetensors huggingface_hub tokenizers
git clone https://github.com/defnotean/Chronos.git
cd Chronos

Direct Python / PyTorch Evaluation

import torch
from chronos_1t.models.dsv41_chronos_moe import ChronosDsv41SparseMoeBlock

device = torch.device("cuda:0")
block = ChronosDsv41SparseMoeBlock(
    layer_idx=20,
    shard_path="chronos-dsv41-layer-00020-of-00040.safetensors",
    device=device,
    hot_cache_size=32,
    offload_tail_to_cpu=True,
)
x = torch.randn(1, 5120, dtype=torch.bfloat16, device=device)
y = block(x)
print("Output shape:", y.shape, "Refusal projection norm:", (y.float() @ block.r_basis.float().T).norm().item())

Source Code & Technical Paper

• GitHub Repository: https://github.com/defnotean/Chronos