File size: 5,041 Bytes
cba1825 23a6592 cba1825 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 | <svg xmlns="http://www.w3.org/2000/svg" width="1240" height="560" viewBox="0 0 1240 560" role="img">
<rect width="100%" height="100%" fill="#ffffff"/>
<style>
.title { font: 700 22px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #1f2937; }
.subtitle { font: 400 13px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #6b7280; }
.label { font: 600 12px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #1f2937; }
.small { font: 400 11px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #6b7280; }
.axis { stroke: #e5e7eb; stroke-width: 1; }
.tick { font: 400 10px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #6b7280; }
.value { font: 700 11px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #1f2937; }
.node-title { font: 700 13px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #1f2937; }
.node-text { font: 400 11px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; fill: #6b7280; }
</style>
<text x="32" y="38" class="title">End-to-end local Mac fine-tuning pipeline</text><text x="32" y="60" class="subtitle">Mixed int8-expert training creates the base behavioral change; LoRA repairs tool-call formatting and routing; GGUF export validates llama.cpp.</text><defs><marker id="arrow" markerWidth="10" markerHeight="10" refX="9" refY="3" orient="auto"><path d="M0,0 L0,6 L9,3 z" fill="#94a3b8"/></marker></defs><line x1="214" y1="172" x2="270" y2="172" stroke="#94a3b8" stroke-width="2" marker-end="url(#arrow)"/><line x1="454" y1="172" x2="510" y2="172" stroke="#94a3b8" stroke-width="2" marker-end="url(#arrow)"/><line x1="694" y1="172" x2="750" y2="172" stroke="#94a3b8" stroke-width="2" marker-end="url(#arrow)"/><line x1="934" y1="172" x2="990" y2="172" stroke="#94a3b8" stroke-width="2" marker-end="url(#arrow)"/><line x1="1084" y1="254" x2="420" y2="324" stroke="#94a3b8" stroke-width="2" marker-end="url(#arrow)"/><line x1="454" y1="404" x2="510" y2="404" stroke="#94a3b8" stroke-width="2" marker-end="url(#arrow)"/><line x1="694" y1="404" x2="750" y2="404" stroke="#94a3b8" stroke-width="2" marker-end="url(#arrow)"/><rect x="34" y="96" width="210" height="152" rx="8" fill="#f8fafc" stroke="#2563eb" stroke-width="2"/><circle cx="56" cy="121" r="7" fill="#2563eb"/><text x="74" y="127" class="node-title">Base model</text><text x="52" y="168" class="node-text">LiquidAI LFM2.5-8B-A1B</text><text x="52" y="192" class="node-text">BF16 MLX source</text><rect x="274" y="96" width="210" height="152" rx="8" fill="#f8fafc" stroke="#0891b2" stroke-width="2"/><circle cx="296" cy="121" r="7" fill="#0891b2"/><text x="314" y="127" class="node-title">Mixed core</text><text x="292" y="168" class="node-text">Quantize MoE experts to int8</text><text x="292" y="192" class="node-text">Routers/non-experts stay BF16</text><rect x="514" y="96" width="210" height="152" rx="8" fill="#f8fafc" stroke="#059669" stroke-width="2"/><circle cx="536" cy="121" r="7" fill="#059669"/><text x="554" y="127" class="node-title">Grouped training</text><text x="532" y="168" class="node-text">3 epochs, 1,746 steps</text><text x="532" y="192" class="node-text">overlapping groups g4/s3 at 10K cap</text><rect x="754" y="96" width="210" height="152" rx="8" fill="#f8fafc" stroke="#d97706" stroke-width="2"/><circle cx="776" cy="121" r="7" fill="#d97706"/><text x="794" y="127" class="node-title">Direct checkpoint</text><text x="772" y="168" class="node-text">step_01746_final</text><text x="772" y="192" class="node-text">experts + routers changed</text><rect x="994" y="96" width="210" height="152" rx="8" fill="#f8fafc" stroke="#7c3aed" stroke-width="2"/><circle cx="1016" cy="121" r="7" fill="#7c3aed"/><text x="1034" y="127" class="node-title">Repair adapters</text><text x="1012" y="168" class="node-text">Iter01 to Iter10 LoRA</text><text x="1012" y="192" class="node-text">structured tool-call targets</text><rect x="274" y="328" width="210" height="152" rx="8" fill="#f8fafc" stroke="#059669" stroke-width="2"/><circle cx="296" cy="353" r="7" fill="#059669"/><text x="314" y="359" class="node-title">Parser-disabled evals</text><text x="292" y="400" class="node-text">MLX fixed-Hermes 43/43</text><text x="292" y="424" class="node-text">0 text-tool leaks</text><rect x="514" y="328" width="210" height="152" rx="8" fill="#f8fafc" stroke="#2563eb" stroke-width="2"/><circle cx="536" cy="353" r="7" fill="#2563eb"/><text x="554" y="359" class="node-title">Export path</text><text x="532" y="400" class="node-text">Fuse + dequantize</text><text x="532" y="424" class="node-text">HF/safetensors source for GGUF</text><rect x="754" y="328" width="210" height="152" rx="8" fill="#f8fafc" stroke="#0891b2" stroke-width="2"/><circle cx="776" cy="353" r="7" fill="#0891b2"/><text x="794" y="359" class="node-title">XL GGUF quants</text><text x="772" y="400" class="node-text">Q4/Q5/Q6/Q8 KXL</text><text x="772" y="424" class="node-text">all pass 43-case llama.cpp eval</text>
</svg>
|