v2.5.3 — Donor-Free Single-Pass MaxQ-Fractal, IPC Hardening, Clean Provenance & Full 51-Chunk/BF16 Benchmarks

v2.5.3 — Donor-Free Single-Pass MaxQ-Fractal (1,328 / 1,328 Tensors), Clean GGUF Provenance Headers, IPC-Hardened GLIBC_2.30 Binaries, Full 51-Chunk AtomicChat wiki-alt.txt & True-BF16 Benchmarks

Highlights

1. Donor-Free Single-Pass DeepSeek-V4-Flash-0731-MaxQ-Fractal Across All 1,328 Tensors (scripts/build_dsv4_maxq_singlepass.py, scripts/dsv4_singlepass.c):

   • Rebuilt Synchopate/DeepSeek-V4-Flash-0731-MaxQ-Fractal-64GB-CPU-GGUF (56.93 GiB / 61,129,364,160 B, 1.72 bpw across 3 shards) in a pure single pass across all 1,199 non-MoE tensors (173 × Q6_K + 1 × Q8_0 dense bottlenecks + 317 × Fractal-64 MLA/Indexer tensors) and all 129 3D routed-MoE tensors directly from deepseek-ai/DeepSeek-V4-Flash-0731 (@ 7872f01b1d1fe23eabc4c98b48bffcef5a386062) weighted by Salvatore Sanfilippo's (@antirez) 1.5M-token DeepSeek-V4-Flash-chat-v2-routed-moe-ds4-1p5m.dat (antirez/deepseek-v4-gguf @ f71f23d552d664e523b422157b2befbf74040380, calibrated 2026-05-12 on April DeepSeek-V4-Flash chat-v2 activations), eliminating v2.5.2's donor routed-MoE copying and ffn_down_exps double-quantization (Abliterated-Fractal Shards 2 & 3 remain the v2.5.1 single-pass build, verified byte-identical to HEAD's --mode aligned output on unedited control experts and differing on prem-research's abliterated top-8 routed down_proj experts in blk.17..36).
   • Measured Impact on 16C Zen 5 (AMD Ryzen 9 9955HX, 0 GPUs, Unpruned Top-6):
     • Functional Quality Gate (benchmarks/eval_quality_and_tools.py): Jumps from 64.0 / 80.0 (24/30 Tools) to 70.0 / 80.0 (30.0/30 Agentic Tool-Calling, 20.0/30 Code Gen, 20.0/20 Long-Context Needle & Math), matching Abliterated-Fractal (70.0 / 80.0) point-for-point (89.56 / 100.0 vs 89.44 / 100.0 under a common --target-tps 9.0 rubric; Fractal-64 & VQ3 score 64.0 / 80.0 Functional / 84.0 / 100.0). Isolated pure N=1 unpruned Top-6 decode (tg64, zero speculation) is 8.66 tok/s at 57.54 GiB idle / 57.58 GiB active VmRSS (36.11 tok/s cold pp67 prefill; telemetry_n1_dsv4_maxq_v253.json).
     • Same-Recipe ARA Abliteration Delta (Abliterated-Fractal vs. MaxQ-Fractal v2.5.3): 32-probe unconditional step-0 Logit KLD drops 10.6× from 1.0219 down to 0.0959 nats/tok (75.0% Top-1, 96.9% Top-5), with prefix-matched rollout KLD (172 positions) at 0.0512 nats/tok (90.7% Top-1, 99.4% Top-5; kld_dsv4_ablit_vs_maxq.json).
     • Standard WikiText-2 (wiki.test.raw, 8,192 tok) Perplexity: MaxQ-Fractal v2.5.3 improves from 11.7434 to 11.6139 ± 0.58453 (statistically tied within < 0.15σ with Fractal-64 11.5288 ± 0.55630, while lifting Tool-Calling from 24/30 to 30/30; Abliterated: 11.3944 ± 0.56500, VQ3: 12.2345 ± 0.59759).

2. Full 51-Chunk AtomicChat/dsv4-eval-artifacts (wiki-alt.txt, -c 5632 -b 5632) & Direct Nemotron-30B True-BF16 Benchmarks (benchmarks/QUALITY_GATE_VQ3_64.md, benchmarks/results/):

   • Full 51-Chunk wiki-alt.txt (143,565 scored positions) on MaxQ-Fractal v2.5.3: Achieves standalone llama-perplexity Final estimate: PPL = 7.2977 ± 0.0504 across all 51 chunks on Zen 5 CPU (+60.8% vs 150.75 GiB bullerwins MXFP4_MOE-BF16 4.5381 ± 0.0292, +22.2% vs 76.87 GiB Unsloth UD-IQ1_S 5.9734, and +14.4% vs 65.42 GiB AtomicChat AD-IQ1_M 6.3813 Mean PPL(Q) on 8× RTX 5090, where UD-IQ1_S and AD-IQ1_M exceed 64 GB host RAM and AtomicChat documents a 4–5% RTX 5090 vs H100 hardware gap) and 7.3180 ± 0.1039 at cumulative chunk 12 (33,780 scored positions, down from 7.5133 in v2.5.2).
   • Direct Nemotron-3.5-30B-A3B True-BF16 (58.84 GiB, SHA256 = fd513c91...) Perplexity & 32-Probe / Multi-Step KLD Ladder: Measured WikiText-2 (12,288 tok) directly on the uncompressed 58.84 GiB BF16 master (6.9582 ± 0.24151) and computed direct BF16 Logit KLD across all variants (Fractal-64: 0.1254 step-0 / 0.0929 nats matched [81.2% / 89.8% Top-1]; VQ3-64: 0.1495 / 0.0944 nats; Abliterated-VQ3-64 vs VQ3-64: 0.0308 / 0.0312 nats [92.4% Top-1, 100.0% Top-5]). Committed all four reference Top-64 logit snapshots (nemotron_bf16, nemotron_fractal, nemotron_vq3, dsv4_maxq) for offline reproducibility.

3. Clean GGUF Shard-1 Provenance Metadata Across All 4 DeepSeek-V4 Repos & OpenMDW-1.1 Alignment (gguf_shard1_metadata_audit.txt):

   • Replaced inherited unsloth tags and provenance headers in Shard-1 across all four 284B repositories: MaxQ-Fractal and Abliterated-Fractal carry general.quantized_by = "Synchopate (llama-modchip)", exact general.base_model.0.revision (7872f01b... / 457210de...), and quantize.imatrix.* (antirez/deepseek-v4-gguf@f71f23d5/...), while VQ3 and Fractal explicitly record general.quantized_by = "Synchopate (llama-modchip); routed-MoE gate/up donor: Unsloth (UD-IQ1_S @ fbbb5b93)" (@ fbbb5b93fb787c21338159b0af3318bb3f4d9768) alongside dedicated donor attribution sections, and aligned all three Nemotron 30B model cards to upstream openmdw-1.1 (https://openmdw.ai/license/1-1/).

4. C++ IPC Security Hardening Across All Portable GLIBC_2.30 Binaries & Source Builds (patches/dsv4-ipc-hardening.patch):

   • Removed legacy world-writable /tmp/dsv4_moe_cfg.bin (0666), %TEMP%\dsv4_moe_cfg.bin, and global /dev/shm/llama_modchip.bin fallbacks from compiled C++ binaries (ggml-cpu.c, ops.cpp, ngram-map.cpp) and deleted unreferenced patches/llama-modchip-dsv4-10tps.patch.
   • DSV4_MOE_CFG_PATH is now strictly opt-in (0600, O_NOFOLLOW, S_ISREG, st_uid == geteuid()), llama_modchip SHM resolves per-user (MODCHIP_SHM_PATH / $XDG_RUNTIME_DIR/llama_modchip.bin / /dev/shm/llama_modchip_<uid>.bin with (st_mode & 0022) == 0 + magic check), all 8 GLIBC_2.30 Linux x86_64 binaries (0 /tmp/dsv4_moe_cfg.bin & 0 global /dev/shm/llama_modchip.bin strings) are synchronized across GitHub Release v2.5.3 and all 7 Hugging Face repos, and all 7 commit-pinned MODEL_REV + SHA256 manifests are updated in install.sh and modchip.

---

Portable GLIBC_2.30 Linux x86_64 Binaries (SHA256SUMS)

Built inside Ubuntu 20.04 (GLIBC_2.30 max symbol, -static-libstdc++, zero libssl/libcurl dependency; requires OpenMP libgomp1 / libgomp):

• *-avx512: AMD Zen 4/5 (Ryzen 7000/8000/9000, EPYC Genoa/Turin, Threadripper 7000) & Intel Ice Lake / Sapphire Rapids (avx512f + avx512vbmi; -DGGML_AVX512_BF16=OFF).
• *-avx2 (and unsuffixed *-linux-x86_64): Pre-AVX512 x86-64-v3 CPUs (Haswell, Skylake, Alder/Raptor Lake, Zen 1/2/3 — 0 %zmm instructions).

c9fe218d3be566e38c7677c903013ee6d90fa3663aa87d3bbc78c2109e8ac2c4  llama-cli-modchip-linux-x86_64
c9fe218d3be566e38c7677c903013ee6d90fa3663aa87d3bbc78c2109e8ac2c4  llama-cli-modchip-linux-x86_64-avx2
883e6d3ca4d33d716c273962e04ec945b83759332e84ac1e35838af3efdb7a8e  llama-cli-modchip-linux-x86_64-avx512
fa214d0f86a37934dcf9c49e32e8d40d5406e6b970d968aac1e093b952c80bdd  llama-perplexity-modchip-linux-x86_64-avx2
3df4bbfced546caf3c6040f9100f0b5d406f38b6f3d60a2d7515d6a41c81c6d9  llama-perplexity-modchip-linux-x86_64-avx512
09f2a1accacb1454aaf0d9a8d75dda8f7cca36df81138eb57bfee45137ef07e4  llama-server-modchip-linux-x86_64
09f2a1accacb1454aaf0d9a8d75dda8f7cca36df81138eb57bfee45137ef07e4  llama-server-modchip-linux-x86_64-avx2
85fe1e2a780ce7c88d99cf5f6fa384c2726dfeee247460924d6bf937efd19597  llama-server-modchip-linux-x86_64-avx512