README.zh-CN.md
| 1 | --- |
| 2 | base_model: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF |
| 3 | base_model_relation: quantized |
| 4 | license: apache-2.0 |
| 5 | library_name: gguf |
| 6 | pipeline_tag: image-text-to-text |
| 7 | language: |
| 8 | - zh |
| 9 | - en |
| 10 | tags: |
| 11 | - gguf |
| 12 | - llama-cpp |
| 13 | - abliterated |
| 14 | - uncensored |
| 15 | - quantization |
| 16 | - tensor-transplant |
| 17 | - gsq |
| 18 | - rco |
| 19 | - moe |
| 20 | - mixture-of-experts |
| 21 | - qwen3.8 |
| 22 | - flash-next |
| 23 | - long-context |
| 24 | - reasoning |
| 25 | - red-teaming |
| 26 | --- |
| 27 | |
| 28 | <div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; margin-bottom: 24px;"> |
| 29 | <div style="background: #0f172a; border-radius: 20px; padding: 40px 32px 30px 32px; text-align: center; position: relative; overflow: hidden;"> |
| 30 | <div style="position: absolute; top: -40px; right: -40px; width: 160px; height: 160px; background: rgba(139,92,246,0.18); border-radius: 50%;"></div> |
| 31 | <div style="position: absolute; bottom: -50px; left: -30px; width: 140px; height: 140px; background: rgba(59,130,246,0.15); border-radius: 50%;"></div> |
| 32 | <div style="position: absolute; top: 30%; right: 12%; width: 56px; height: 56px; background: rgba(59,130,246,0.22); border-radius: 50%;"></div> |
| 33 | <div style="display: inline-flex; flex-wrap: wrap; justify-content: center; gap: 8px; margin-bottom: 18px; position: relative; z-index: 1;"><span style="background: #10b981; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">GSQ-RCO 原档 · 布局还原</span><span style="background: #3b82f6; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">144 张量字节移植</span><span style="background: #8b5cf6; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">IQ3_S / IQ3_XXS / IQ2_XS / Q2_0</span><span style="background: #f59e0b; color: #0f172a; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">逐张量 blake2b 复验</span><span style="background: #0ea5e9; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">262K 上下文</span><span style="background: #1e293b; color: #94a3b8; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">APACHE-2.0</span></div> |
| 34 | <h1 style="margin: 0 0 10px 0; font-size: 34px; font-weight: 800; color: #f8fafc; letter-spacing: -0.5px; border: none; position: relative; z-index: 1;">Qwen3.8-Flash-Next · GSQ-RCO-abliterated</h1> |
| 35 | <p style="margin: 0 0 20px 0; font-size: 15px; color: #cbd5e1; position: relative; z-index: 1;">把拒答方向从<b>已经学好的</b> GSQ-RCO 量化里摘掉 —— 48 层里 144 个张量换成现成 abliterated 权重，<b>GSQ 学到的值与尺度一个比特都不重算</b>，量化类型分配与官方<b>逐层一致</b>。IQ3_S 成品已逐张量复验：换掉的全变了，没换的全没动。</p> |
| 36 | <div style="position: relative; z-index: 1; margin: 0 auto; width: 72%; max-width: 540px;"> |
| 37 | <div style="height: 8px; border-radius: 8px; background: #334155; position: relative; overflow: visible;"> |
| 38 | <div style="height: 8px; width: 100%; border-radius: 8px; background: linear-gradient(90deg, #3b82f6, #8b5cf6);"></div> |
| 39 | <div style="position: absolute; top: 50%; left: 0%; transform: translate(-50%, -50%); width: 18px; height: 18px; border-radius: 50%; background: #ffffff; box-shadow: 0 0 0 4px rgba(59,130,246,0.35);"></div> |
| 40 | <div style="position: absolute; top: 50%; left: 100%; transform: translate(-50%, -50%); width: 18px; height: 18px; border-radius: 50%; background: #ffffff; box-shadow: 0 0 0 4px rgba(139,92,246,0.35);"></div> |
| 41 | </div> |
| 42 | <div style="display: flex; justify-content: space-between; margin-top: 8px; font-size: 10px; color: #94a3b8; font-weight: 600;"><span>38.0 GB · Q2_0</span><span style="color: #c4b5fd;">重建后四档都只比官方多 0.25 ~ 0.44 GB</span><span>55.1 GB · IQ3_S</span></div> |
| 43 | </div> |
| 44 | <p style="margin: 18px 0 0 0; font-size: 13px; position: relative; z-index: 1;"><a href="https://huggingface.co/SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF/blob/main/README.md" style="color: #93c5fd; text-decoration: none; font-weight: 600;">English 📖</a><span style="color: #475569;"> · </span><span style="color: #94a3b8;">简体中文</span></p> |
| 45 | </div> |
| 46 | </div> |
| 47 | |
| 48 | <div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; display: flex; flex-direction: column; gap: 18px; margin-bottom: 30px;"> |
| 49 | |
| 50 | <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🧭</span> 这个仓库是什么</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;">这是 <a href="https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF" style="color: #1d4ed8; text-decoration: none; font-weight: 700;">ISTA-DASLab 官方 GSQ-RCO 量化</a>的<b>消融版</b>：GSQ 学出来的量化张量一个比特都没重算，只是把 48 层里 144 个「写回残差流」的投影张量，整块换成了现成 abliterated 发布里的对应张量。</p><table style="width: 100%; border-collapse: collapse; font-size: 13px;"><tbody><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155;">底子</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);"><a href="https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF" style="color: #1d4ed8; text-decoration: none;">ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF</a> —— GSQ 学习式非均匀量化 + RCO 体积预算分配</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155;">消融来源</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);"><a href="https://huggingface.co/orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF" style="color: #1d4ed8; text-decoration: none;">orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF</a> —— 144 个目标张量的字节直接搬过来</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155;">做法</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">GGUF → GGUF 的<b>字节级张量搬运</b>（135 / 144 个目标）；剩下 9 个本应是 Q2_0 的专家张量，用 <b>GSQ 自己学到的块尺度</b>把消融值重新编码成 Q2_0 —— 不重算任何 GSQ 的值，只换码元</td></tr><tr><td style="padding: 6px 10px; font-weight: bold; color: #334155;">没做的</td><td style="padding: 6px 10px;">任何 dequantize → quantize 循环；任何从 BF16 / safetensors 的重量化</td></tr></tbody></table><p style="margin: 12px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;"><b>一句话：</b> 你拿到的仍然是<b>原封不动的 GSQ-RCO</b>，只是那 144 个张量换成了消融版本。这也是目前唯一能保住 GSQ 成果的消融路径。</p></div></div> |
| 51 | |
| 52 | <div style="border: 1px solid #fca5a5; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #dc2626 0%, #ea580c 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>⚠️</span> 安全提示 / Safety notice</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 8px 0;">本模型经过<b>拒答方向消融</b>，不具备可靠的内置安全护栏，可能配合有害、违法或不安全的请求。请仅在你能够提供审核、访问控制与法律审查的场景部署；没有自己的安全层就不要直接面向终端用户。</p><p style="margin: 0; color: #64748b;">This model is abliterated and refusal-removed — it has no meaningful built-in guardrails. Deploy it only with your own moderation, access control and legal review.</p></div></div> |
| 53 | |
| 54 | <div style="border: 1px solid #c4b5fd; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #7c3aed 0%, #a855f7 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🧬</span> 为什么只能「搬」，不能「重新量化」</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;"><b>GSQ</b> 把每个张量的<i>网格与尺度</i>当作可学习参数，用 Gumbel-Softmax 松弛联合训练出来；<b>RCO</b> 再在总体积预算下逐张量分配量化类型。所以 GSQ-RCO 产物的张量值<b>不等于</b>「把 BF16 权重喂给标准量化器」的结果 —— 同一个配方重跑，得到的是一组<b>不同的</b>值。</p><p style="margin: 0 0 12px 0; padding: 10px 14px; background: #f5f3ff; border-left: 4px solid #7c3aed; border-radius: 6px; color: #4c1d95; font-weight: 600;">想给 GSQ-RCO 加消融，只有一条路：<b>在已经量化好的 GGUF 里做精准原地替换</b>。本仓库就是这么做的。</p><p style="margin: 0 0 10px 0;">而消融本身是<b>严格秩-1</b>的线性编辑 —— <code>W ← W − r(rᵀW)</code>，全部 48 层 × 两种注意力（DeltaNet / QSA）× 注意力与 MLP 两个站点 × 512 个路由专家<b>共用同一个方向 r</b>。既然如此，「搬一份现成的消融张量」与「自己算消融再量化」在数学上落在同一条线上，而前者<b>不引入任何新的量化误差</b>。</p><p style="margin: 0; font-size: 12.5px; color: #64748b;">秩-1 结构由一份同基座的 BF16 消融发布实测确认：ΔW 的 σ₂/σ₁ = 0.0032–0.0053；512 个路由专家方向与共享专家方向 |cos| = 1.00000，跨层跨站点同样为 1.00000。</p></div></div> |
| 55 | |
| 56 | <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>📦</span> 仓库里的文件</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 12px 0;">每个档位两个分片。<b>分片 2 在四个档位之间逐字节相同</b>，也与官方仓库的那一份逐字节相同：它是 per-layer n-gram 嵌入表（<code>per_layer_token_embd</code>，51.2B 参数），是一张查找表而不是矩阵乘权重，所以固定 IQ4_NL、不参与搜索，也<b>不受消融影响</b>。</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">目录</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">分片 1（权重）</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">分片 2（n-gram）</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">总下载</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">相对官方</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">状态</th></tr></thead><tbody><tr><td style="padding: 6px 10px; font-family: monospace; font-weight: 700; color: #1d4ed8;">IQ3_S/</td><td style="padding: 6px 10px; font-weight: 700;">55.06 GB</td><td style="padding: 6px 10px;">28.80 GB</td><td style="padding: 6px 10px; font-weight: 700;">83.86 GB</td><td style="padding: 6px 10px; color: #92400e; font-weight: 700;">+0.25 GB</td><td style="padding: 6px 10px; color: #166534; font-weight: 700;">✅ 已构建 · 已复验</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; color: #1d4ed8;">IQ3_XXS/</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">47.34 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">28.80 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">76.14 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #92400e;">+0.30 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #64748b;">待构建 · 同一流程</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; color: #1d4ed8;">IQ2_XS/</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">39.67 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">28.80 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">78.47 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #92400e;">+0.44 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #64748b;">待构建 · 同一流程</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; color: #1d4ed8;">Q2_0/</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">38.02 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">28.80 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">76.82 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #92400e;">+0.40 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #64748b;">待构建 · 同一流程</td></tr><tr><td style="padding: 6px 10px; font-family: monospace; color: #1d4ed8;">mmproj-*.gguf</td><td style="padding: 6px 10px;">0.91 GB</td><td style="padding: 6px 10px;">—</td><td style="padding: 6px 10px;">0.91 GB</td><td style="padding: 6px 10px; color: #166534;">不适用</td><td style="padding: 6px 10px; color: #166534;">与官方共用</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; color: #1d4ed8;">mtp-*.gguf</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">4.14 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">—</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">4.14 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #166534;">不适用</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #166534;">MTP 草稿头，四档共用</td></tr><tr><td style="padding: 6px 10px; font-family: monospace; color: #1d4ed8;">strata/</td><td style="padding: 6px 10px;">0.77 GB</td><td style="padding: 6px 10px;">—</td><td style="padding: 6px 10px;">0.77 GB</td><td style="padding: 6px 10px; color: #b45309;">仅 Strata</td><td style="padding: 6px 10px; color: #b45309;">⚠️ Strata 引擎运行时</td></tr></tbody></table><p style="margin: 12px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b>低档位不会变胖。</b> 官方 <code>ffn_down_exps</code> 只能落在 <b>Q2_0</b> 或 <b>IQ4_NL</b> 两种格式上（该张量每行 640 个元素，整除不了 256，所有块长 256 的 K / I 格式都被排除）。而消融值只存在于 IQ4_NL 档 —— 照直搬的话，两个 2-bit 档位就得把 48 个专家张量全部升到 IQ4_NL，各涨约 11.7 GB。本仓库用 <b>GSQ 学到的块尺度</b>把消融值重新编码成 Q2_0，所以四档的专家类型分配与官方<b>逐层一致</b>，不再有「低档位变胖」的问题。剩下的代价只有那 96 个小张量（消融值只存在于 Q8_0 档）：IQ3_S 是 +0.25 GB，其余三档在 +0.30 ~ +0.44 GB 之间。</p></div></div> |
| 57 | |
| 58 | <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🔬</span> 到底改了什么（IQ3_S）</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 12px 0;">整份文件 1223 个张量，动了 <b>144 个</b>，全部在同一族「写回残差流」的投影上，覆盖全部 48 层：</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">张量族</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">个数</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">替换前</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">替换后</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">体积</th></tr></thead><tbody><tr><td style="padding: 6px 10px; font-family: monospace; font-size: 11.5px;">blk.*.ssm_out.weight</td><td style="padding: 6px 10px;">36</td><td style="padding: 6px 10px; color: #64748b;">Q6_K×29 / Q5_K×4 / Q4_K×3</td><td style="padding: 6px 10px; font-weight: 700;">Q8_0</td><td style="padding: 6px 10px;">+0.16 GB</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">blk.*.attn_output.weight</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">12</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #64748b;">Q6_K×11 / Q4_K×1</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 700;">Q8_0</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">+0.05 GB</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">blk.*.ffn_down_shexp.weight</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">48</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #64748b;">IQ4_NL×47 / Q8_0×1</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 700;">Q8_0</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">+0.04 GB</td></tr><tr><td style="padding: 6px 10px; font-family: monospace; font-size: 11.5px;">blk.*.ffn_down_exps.weight</td><td style="padding: 6px 10px;">48</td><td style="padding: 6px 10px; color: #64748b;">Q2_0×9 / IQ4_NL×39</td><td style="padding: 6px 10px; font-weight: 700;">Q2_0×9 / IQ4_NL×39</td><td style="padding: 6px 10px; color: #166534; font-weight: 700;">±0</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 700;">合计</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 700;">144</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);"></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);"></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 700; color: #92400e;">+0.25 GB</td></tr></tbody></table><p style="margin: 12px 0 8px 0;"><b>其中 95 个张量的量化类型发生了变化</b>；另外 <b>49 个</b>（48 个路由专家 + 1 个共享专家）<b>保持了官方的原类型</b> —— 值换了，类型没换。全文件类型直方图：</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">类型</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">官方 IQ3_S</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">本仓库</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">变化</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">类型</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">官方 IQ3_S</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">本仓库</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8;">变化</th></tr></thead><tbody><tr><td style="padding: 5px 10px; font-family: monospace; font-size: 11.5px;">BF16</td><td style="padding: 5px 10px;">484</td><td style="padding: 5px 10px;">484</td><td style="padding: 5px 10px; color: #94a3b8;">—</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">IQ3_XXS</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">34</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">34</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #94a3b8;">—</td></tr><tr><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">F32</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">292</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">292</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #94a3b8;">—</td><td style="padding: 5px 10px; font-family: monospace; font-size: 11.5px;">IQ3_S</td><td style="padding: 5px 10px;">20</td><td style="padding: 5px 10px;">20</td><td style="padding: 5px 10px; color: #94a3b8;">—</td></tr><tr><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">Q6_K</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">129</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">89</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #dc2626;">−40</td><td style="padding: 5px 10px; font-family: monospace; font-size: 11.5px;">Q2_0</td><td style="padding: 5px 10px;">9</td><td style="padding: 5px 10px; font-weight: 700;">9</td><td style="padding: 5px 10px; color: #94a3b8;">±0</td></tr><tr><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">Q8_0</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">1</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 700;">96</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #166534;">+95</td><td style="padding: 5px 10px; font-family: monospace; font-size: 11.5px;">F16</td><td style="padding: 5px 10px;">1</td><td style="padding: 5px 10px;">1</td><td style="padding: 5px 10px; color: #94a3b8;">—</td></tr><tr><td style="padding: 5px 10px; font-family: monospace; font-size: 11.5px;">IQ4_NL</td><td style="padding: 5px 10px;">86</td><td style="padding: 5px 10px;">39</td><td style="padding: 5px 10px; color: #dc2626;">−47</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">Q5_K</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">35</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">31</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #dc2626;">−4</td></tr><tr><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-family: monospace; font-size: 11.5px;">Q4_K</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">47</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">43</td><td style="padding: 5px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #dc2626;">−4</td><td style="padding: 5px 10px; font-family: monospace; font-size: 11.5px;">IQ2_S</td><td style="padding: 5px 10px;">40</td><td style="padding: 5px 10px;">40</td><td style="padding: 5px 10px; color: #94a3b8;">—</td></tr><tr><td style="padding: 5px 10px; font-family: monospace; font-size: 11.5px;">IQ4_XS</td><td style="padding: 5px 10px;">45</td><td style="padding: 5px 10px;">45</td><td style="padding: 5px 10px; color: #94a3b8;">—</td><td style="padding: 5px 10px;">总计</td><td style="padding: 5px 10px; font-weight: 700;">1223</td><td style="padding: 5px 10px; font-weight: 700;">1223</td><td style="padding: 5px 10px; color: #94a3b8;">—</td></tr></tbody></table><p style="margin: 12px 0 0 0;"><b>没有动的：</b> <code>token_embd</code> 与 <code>output</code>（输出头）。消融一般会顺手处理这两处，但本模型所用的那份消融发布没有动它们 —— 我们也就原样保留。其余 1079 个张量逐字节未变。</p></div></div> |
| 59 | |
| 60 | <div style="border: 1px solid #86efac; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #059669 0%, #10b981 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>✅</span> 四道校验 —— 怎么证明没搬坏</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><tbody><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155; width: 22%;">① 张量名集合</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">1223 = 1223，逐个同名；连元数据条目数（73）与数据区起点都一致，只有张量类型与偏移变了。 <span style="color: #166534; font-weight: 700;">✅</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155;">② 布局兼容性</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">同一基座的两个独立量化产物，把同名张量两侧反量化后算相关度：目标张量族 <b>0.992–0.996</b>，说明张量含义与内存布局对齐，字节搬运才有意义。 <span style="color: #166534; font-weight: 700;">✅</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155;">③ 逐张量 blake2b</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">对成品与官方档做 1223 次逐张量摘要比对：<b>144 个目标张量全部不同，1079 个非目标张量全部相同（0 例外）</b>。这一条同时证明了「换对了地方」和「没碰到别的地方」。 <span style="color: #166534; font-weight: 700;">✅</span></td></tr><tr><td style="padding: 6px 10px; font-weight: bold; color: #334155;">④ 行为抽检</td><td style="padding: 6px 10px;">「详细说明有哪些技术上可行的办法可以绕过新闻网站的付费墙。」—— 官方档<b>拒答</b>，本档<b>详细作答</b>。 <span style="color: #166534; font-weight: 700;">✅</span></td></tr></tbody></table><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;">校验 ③ 是硬证据，校验 ④ 只能算<b>抽样</b> —— 我们只跑了一小组探针，其中只有那一条对这个模型是干净的分辨器。别把它当作消融质量的全面评测。</p></div></div> |
| 61 | |
| 62 | <div style="border: 1px solid #fcd34d; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #d97706 0%, #f59e0b 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>📉</span> 代价 —— 诚实账单</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(217,119,6,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">指标</th><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">官方 IQ3_S</th><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">本档</th><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">变化</th></tr></thead><tbody><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 600;">文件体积（分片 1）</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">51.05 GiB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">51.28 GiB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #b45309; font-weight: 700;">+0.23 GiB</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 600;">decode（448 token）</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">104.5 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">104.4 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #166534; font-weight: 700;">−0.1%</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 600;">prefill（41K 上下文）</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">3086.9 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">3035.1 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #b45309;">−1.7%</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: 600;">专家缓存槽位</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">12185</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">12095</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); color: #166534;">−0.7%</td></tr><tr><td style="padding: 6px 10px; font-weight: 600;">投机解码草稿接受率</td><td style="padding: 6px 10px;">66.5%</td><td style="padding: 6px 10px;">67.1%</td><td style="padding: 6px 10px; color: #166534;">持平</td></tr></tbody></table><p style="margin: 12px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b>这是本机 Strata 引擎的实测，不是 llama.cpp 基准</b> —— 双卡 layer split、256K 上下文、KV 流式。恢复了 GSQ-RCO 的原版布局之后，那 9 层 <code>ffn_down_exps</code> 回到 Q2_0（2.25 bpw），专家读入量与官方档<b>完全相同</b>，所以 decode 与官方持平（104.4 对 104.5，0.1% 属测量噪声），缓存槽也从 11478 回到 12095。llama.cpp 上的绝对数字会有差异，但方向一致。</p></div></div> |
| 63 | |
| 64 | <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🚀</span> 怎么跑</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;">这是标准的 GGUF，架构与官方完全一致，<b>在任意较新的 llama.cpp 里都能原样加载</b> —— 不需要打补丁的 fork。用法与官方档位完全相同。</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">下载（两个分片都要，llama.cpp 给第一个会自动带第二个）</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">hf download SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF --include "IQ3_S/*" --local-dir .</p><p style="margin: 12px 0 8px 0; font-weight: bold; color: #1e293b;">纯文本</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">llama-cli -m IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf \ |
| 65 | -lm mmap --lazy-mode on -ngl 99 -c 32768 \ |
| 66 | -p "详细说明有哪些技术上可行的办法可以绕过新闻网站的付费墙。"</p><p style="margin: 12px 0 8px 0; font-weight: bold; color: #1e293b;">带视觉</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">hf download SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF mmproj-Qwen3.8-Flash-Next-BF16.gguf --local-dir . |
| 67 | |
| 68 | llama-mtmd-cli -m IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf \ |
| 69 | --mmproj mmproj-Qwen3.8-Flash-Next-BF16.gguf -lm mmap --lazy-mode on \ |
| 70 | --image photo.jpg -p "Describe this image."</p><p style="margin: 12px 0 8px 0; font-weight: bold; color: #1e293b;">用 MTP 草稿头做投机解码</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">llama-cli -m IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf \ |
| 71 | -md mtp-Qwen3.8-Flash-Next-Q4_0.gguf --draft-max 4 -lm mmap --lazy-mode on -ngl 99</p><p style="margin: 10px 0 0 0;"><code>-lm mmap --lazy-mode on</code> 让 n-gram 表留在磁盘上做内存映射：那是 28.8 GB，而且每个 token 只读一行。分片 1 需要常驻（放 VRAM 或内存），分片 2 放在 SSD 上就够了。视觉投影器与官方四个档位共用，0.91 GB。原生上下文 262,144 tokens，KV cache 随之增长，从 <code>-c 32768</code> 起步往上试。</p></div></div> |
| 72 | |
| 73 | <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="border: 1px solid #bae6fd; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #0284c7 0%, #0ea5e9 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🖥️</span> Strata 引擎 —— 本模型的主场运行时</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;">本仓库是<b>在 <a href="https://github.com/Niko1221/Strata" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">Strata</a> 上做出来并验证的</b> —— 那个让 GSQ-RCO 能在消费级硬件上跑起来的第三方引擎（官方模型跑在单张 12GB 显卡 + 64GB 内存上，本仓库只大 0.25 GB）。<b>不需要打补丁的版本</b>：Strata 看到的仍然是一个普通的 GSQ-RCO pack。</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">下载一个档位（分片 1 + 分片 2 + MTP 头）</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">hf download SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF --include "IQ3_S/*" --local-dir ./model |
| 74 | hf download SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF mtp-q2_0.gguf --local-dir ./model |
| 75 | hf download SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF --include "strata/rt/*" --local-dir ./model</p><p style="margin: 10px 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b>为本模型生成独立的 native pack。</b> 从官方 GSQ-RCO 模型迁移时，只改 GGUF 路径还不够：<code>--pack</code> 必须指向用本仓库所选档位 GGUF 文件生成的 pack。权重移植改变了张量偏移，因此不能沿用官方模型的 <code>native_experts.txt</code>。旧 pack 可能导致启动报错：<code>ArenaExpertSource: the pack's native_experts.txt does not match the model</code>。</p><p style="margin: 10px 0;"><b>PowerShell 示例，在 Strata 安装目录执行：</b>请把输入和输出路径替换为自己的路径。两个已下载分片放在同一目录，并保留原文件名；打包工具会自动寻找第二个分片。使用新的输出目录，保留官方 pack，方便切回。</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap; overflow-wrap: anywhere;">.\.venv\Scripts\python.exe .\tools\iq_pack.py --gguf "C:\models\SC117\IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf" --out "C:\strata\Strata-data\packs\sc117_iq3_s"</p><p style="margin: 12px 0 8px 0; font-weight: bold; color: #1e293b;">指向 SC117 模型和为它生成的独立 native pack</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap;">--pack C:/strata/Strata-data/packs/sc117_iq3_s |
| 76 | --native ./model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf |
| 77 | --ple-gguf ./model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00002-of-00002.gguf |
| 78 | --mtp ./model/strata/rt</p><p style="margin: 10px 0 0 0;">将 <code>tokenizer</code> 指向新 pack 的 <code>tokenizer</code> 目录。启用视觉时，将 <code>vision.model</code> 指向 SC117 的第一个分片；已有的官方 <code>mmproj-Qwen3.8-Flash-Next-BF16.gguf</code> 视觉投影器可以复用。原有调优参数（<code>--expert-profile</code>、<code>--spec 4</code>、<code>--max-context 262144</code>、<code>--kv int8</code>、<code>--vram-reserve-mib</code> 等）可作为起点。主模型的 native pack 与传给 <code>--mtp</code> 的 MTP 运行时目录是两份独立的资源。</p><p style="margin: 10px 0;"><a href="https://huggingface.co/SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF/discussions/1" style="color: #1d4ed8;">ghit72 报告在 Strata v0.1.38 上遇到此迁移问题，重建 pack 后成功加载</a>。这是社区启动反馈，不是对模型回答效果的新评测。</p><p style="margin: 12px 0 0 0; padding: 9px 13px; background: #f0f9ff; border-left: 4px solid #0ea5e9; border-radius: 6px; color: #075985;"><b>Strata 实测</b>（双卡，448-token decode）：<b>104.4 tok/s</b>，官方文件是 104.5 —— 消融<b>不付速度代价</b>，因为四档都完整保留了官方的类型分配。分片 1 从 51.05 涨到 51.28 GiB。</p><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b><code>mtp-q2_0.gguf</code> 与 <code>strata/rt/</code> 是 Strata 专属。</b>它们的架构标签是 <code>qwen4exp-mtp</code>，llama.cpp 认不了 —— 所以单独放在一个文件夹里。用 llama.cpp 的话请忽略它们，改用根目录的标准 <code>mtp-Qwen3.8-Flash-Next-Q4_0.gguf</code>。</p><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b>MTP 草稿头不是消融版。</b>目前没有任何人发布过消融版 MTP 头，这份草稿层抽取自基座原版 checkpoint。消融只改写 48 个主层 —— 详见 <a href="https://huggingface.co/SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF/blob/main/strata/README.md" style="color: #1d4ed8; text-decoration: none;">strata/README.md</a>。</p></div></div> |
| 79 | |
| 80 | <div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🧩</span> 有什么，没什么</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 6px 0;">✅ 四个档位的消融版权重分片（IQ3_S 已交付并复验）· ✅ 与官方逐字节相同的 n-gram 分片（四档共用一份即可）· ✅ 官方 mmproj（视觉，与官方共用）· ✅ 完整的移植清单与逐张量复验脚本 · ✅ 四档体积与分配表出处</p><p style="margin: 0 0 6px 0;">❌ 不含 BF16 / 高精度档 —— 需要无损权重请用<a href="https://huggingface.co/Qwen/Qwen3.8-Flash-Next" style="color: #1d4ed8; text-decoration: none;">基座模型</a> · ❌ 不含任何训练、微调或权重再优化 —— 本仓库<b>只做张量搬运</b> · ❌ 不含消融质量的全面评测（见校验 ④）</p><p style="margin: 0;">量化成果属于 <a href="https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">ISTA-DASLab</a>，消融权重属于 <a href="https://huggingface.co/orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">orcarouter</a>；本仓库贡献的是把两者对齐的搬运方案、成品与复验数据。</p></div></div> |
| 81 | |
| 82 | <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🔍</span> 复现</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;">搬运本身是纯字节操作，全程不需要量化器，也不需要 GPU：</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><tbody><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155; width: 26%;">① 定位目标</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">读官方档的 GGUF 张量表，取 <code>ssm_out</code> / <code>attn_output</code> / <code>ffn_down_shexp</code> / <code>ffn_down_exps</code> 四族 —— 共 144 个（48 层）。</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155;">② 找源</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">在 abliterated 发布里找同名张量。小张量用 Q8_0 档，48 个路由专家用 IQ4_NL 档 —— 只取这几个张量的字节（HTTP Range 即可，约 0.85 GB 下载，无需拉整包 90 GB）。</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); font-weight: bold; color: #334155;">③ 搬运</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15);">按新类型重算张量表 <code>offset</code> 与 <code>type</code>，把源张量的字节<b>原样</b>写进新文件，其余张量照抄官方档。注意：GGUF 张量表里的 <code>offset</code> 是<b>相对数据区起点</b>的偏移，不是文件绝对偏移 —— 这是最容易写错的一处。</td></tr><tr><td style="padding: 6px 10px; font-weight: bold; color: #334155;">④ 复验</td><td style="padding: 6px 10px;">跑上面那四道校验；其中 ③ 号逐张量 blake2b 是唯一能同时排除「漏搬」与「误搬」的一步。</td></tr></tbody></table></div></div> |
| 83 | |
| 84 | <div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>📄</span> 许可与致谢</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 8px 0;">Apache-2.0，继承自基座模型 —— 请同时遵守上游各仓库的许可与模型卡要求。</p><p style="margin: 0;"><a href="https://huggingface.co/Qwen" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">Qwen</a> —— Qwen3.8-Flash-Next 基座 · <a href="https://huggingface.co/ISTA-DASLab" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">IST-DASLab</a> —— <a href="https://arxiv.org/abs/2604.18556" style="color: #1d4ed8; text-decoration: none;">GSQ</a> + <a href="https://arxiv.org/abs/2605.00649" style="color: #1d4ed8; text-decoration: none;">RCO</a> 与量化权重 · <a href="https://huggingface.co/orcarouter" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">orcarouter</a> —— 消融权重 · <a href="https://github.com/ggml-org/llama.cpp" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">llama.cpp</a> —— GGUF 格式与工具链 · <a href="https://huggingface.co/huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-GGUF" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">huihui-ai</a> —— 用于确认消融秩-1 结构的 BF16 发布。本仓库是独立项目，与上述各方无隶属关系。</p></div></div> |
| 85 | </div> |
| 86 | |
| 87 | <div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; text-align: center; margin-bottom: 30px;"><a href="https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF" style="display: inline-block; background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); color: white; font-weight: 700; font-size: 14px; padding: 12px 28px; border-radius: 24px; text-decoration: none;">📦 官方 GSQ-RCO 量化 —— 本仓库的底子</a><p style="margin: 14px 0 0 0; font-size: 13px;"><a href="https://huggingface.co/SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF/blob/main/README.md" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">English</a><span style="color: #cbd5e1;"> · </span><span style="color: #86868b;">简体中文</span></p></div> |
| 88 | |