Build latest book artifacts / build (push) Canceled after 0s
dependency resolution / resolve (3.11) (push) Canceled after 0s
dependency resolution / resolve (3.13) (push) Canceled after 0s
deploy-pages / build (push) Canceled after 0s
deploy-pages / deploy (push) Canceled after 0s
i18n consistency check / check (push) Canceled after 0s
provider adoption tests / test (chapter2/context-compression) (push) Canceled after 0s
provider adoption tests / test (chapter2/prompt-injection) (push) Canceled after 0s
provider adoption tests / test (chapter2/system-hint) (push) Canceled after 0s
provider adoption tests / test (chapter3/log-sanitization) (push) Canceled after 0s
web-search-agent tests / test (push) Canceled after 0s
web-search-agent tests / agentbook (push) Canceled after 0s
35 lines
7.5 KiB
XML
35 lines
7.5 KiB
XML
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 40 780 450" width="780" height="450" style="background:#ffffff">
|
||
<defs><marker id="ah" markerWidth="12" markerHeight="8" refX="12" refY="4" orient="auto"><polygon points="0 0, 12 4, 0 8" fill="#333333"/></marker><marker id="ah-light" markerWidth="12" markerHeight="8" refX="12" refY="4" orient="auto"><polygon points="0 0, 12 4, 0 8" fill="#999999"/></marker></defs>
|
||
|
||
<rect x="300" y="80" width="180" height="120" rx="6" fill="#f0f0f0" stroke="#333333" stroke-width="2"/>
|
||
<text x="390" y="105" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="19.5" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Original weights W</text>
|
||
<text x="390" y="130" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">d × d</text>
|
||
<text x="390" y="152" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">Example: 4096 × 4096</text>
|
||
<text x="390" y="172" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="12.5" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">Parameters fixed (not trained)</text>
|
||
<text x="520" y="140" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="24" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">+</text>
|
||
<rect x="560" y="70" width="100" height="65" rx="6" fill="#d0d0d0" stroke="#333333" stroke-width="2"/>
|
||
<text x="610" y="88" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Matrix B</text>
|
||
<text x="610" y="108" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">d × r</text>
|
||
<text x="610" y="122" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">4096×8</text>
|
||
<rect x="560" y="145" width="100" height="65" rx="6" fill="#d0d0d0" stroke="#333333" stroke-width="2"/>
|
||
<text x="610" y="163" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Matrix A</text>
|
||
<text x="610" y="183" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">r × d</text>
|
||
<text x="610" y="197" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">8×4096</text>
|
||
<path d="M 665,70 C 685,70 685,135.0 690,140.0 C 685,145.0 685,210 665,210" fill="none" stroke="#333333" stroke-width="2"/>
|
||
<text x="695" y="140" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="bold">ΔW = αBA</text>
|
||
<text x="280" y="140" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="20" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="normal">x →</text>
|
||
<text x="695" y="165" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="start" dominant-baseline="central" font-weight="normal">rank r = 8</text>
|
||
<rect x="30" y="230" width="720" height="228" rx="4" fill="#f5f5f5" stroke="#333333" stroke-width="2"/>
|
||
<text x="390" y="252" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="20" fill="#333333" text-anchor="middle" dominant-baseline="central" font-weight="bold">Key Practical Findings (Thinking Machines Lab)</text>
|
||
<text x="50" y="282" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="bold">▸ Optimal learning rate</text>
|
||
<text x="50" y="300" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="normal">LoRA LR = Full fine-tuning LR × 10 (about 15× for short training)</text>
|
||
<text x="50" y="317" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="bold">▸ Scope of application</text>
|
||
<text x="50" y="335" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="13" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="normal">Must be applied to all layers (especially MLP/MoE; applying only to attention layers significantly degrades performance)</text>
|
||
<text x="50" y="352" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="bold">▸ RL scenario</text>
|
||
<text x="50" y="370" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="normal">Even rank=1 can achieve full fine-tuning performance (only O(1) bits absorbed per episode)</text>
|
||
<text x="50" y="387" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="bold">▸ Computational efficiency</text>
|
||
<text x="50" y="405" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="start" dominant-baseline="central" font-weight="normal">FLOPs ≈ 2/3 of full fine-tuning (2N² vs 3N²); LoRA is better when plotting performance against FLOPs</text>
|
||
<text x="50" y="422" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="16" fill="#333333" text-anchor="start" dominant-baseline="central" font-weight="bold">▸ Inference advantages</text>
|
||
<text x="50" y="440" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14" fill="#666666" text-anchor="start" dominant-baseline="central" font-weight="normal">Multi-tenant service: vLLM/SGLang support loading multiple adapters simultaneously</text>
|
||
<text x="390" y="472" font-family="Arial, 'Helvetica Neue', Helvetica, 'PingFang SC', 'Microsoft YaHei', sans-serif" font-size="14.5" fill="#666666" text-anchor="middle" dominant-baseline="central" font-weight="normal">SFT recommended rank 64-256 | RL recommended rank 8-32 | 1/r scaling makes LR approximately independent of rank</text>
|
||
</svg> |