kiku-agent 1.2.0 → 1.3.1

package/src/lib/vram/modelRegistry.js

@@ -0,0 +1,387 @@
+/**
+ * Built-in architecture registry for LM Studio models that advertise
+ * TOOL CALLING, VISION, or REASONING capabilities.
+ *
+ * Used as a fallback when neither GGUF metadata nor a HuggingFace config.json
+ * can be resolved. Geometry fields were verified against each model's published
+ * HuggingFace config.json (and HF safetensors totals where available) — see the
+ * `verified` tag per entry. All entries remain flagged `approx` because exact
+ * checkpoint metadata always wins when it can be read directly.
+ *
+ * paramCount semantics: total parameters loaded into VRAM (HF `safetensors.total`
+ * where shown, i.e. INCLUDING the vision encoder for multimodal models). For MoE
+ * models this is the TOTAL parameter count, not the active routing parameters.
+ *
+ * `capabilities` mirrors what LM Studio badges for the family:
+ *   - 'tools'     → native function/tool calling support
+ *   - 'vision'    → image input support in LM Studio
+ *   - 'reasoning' → thinking/reasoning channels (harmony, <think>, Z1, etc.)
+ */
+
+export const MODEL_REGISTRY = Object.freeze([
+  {
+    label: 'Gemma 3n E4B (incl. community "gemma-4-e4b")',
+    patterns: ['gemma-4-e4b', 'gemma4-e4b', 'gemma-3n-e4b', 'gemma3n-e4b'],
+    capabilities: ['vision'],
+    paramsB: 8.0,
+    layers: 35,
+    queryHeads: 8,
+    kvHeads: 2,
+    headDim: 256,
+    hiddenSize: 2048,
+    contextLength: 32768,
+    swa: { windowTokens: 512, patternLocal: 5, patternGlobal: 1 },
+    vision: { encoderParamsB: null, tokensPerImage: 576 },
+    approx: true,
+    verified: 'hf-config (mlx-community/gemma-3n-E4B-it-4bit)',
+    notes: 'Raw ~8B params loaded incl. MobileNet-V5 vision + audio encoders; MatFormer effective-4B class. Vision emits ~256 tokens/image — 576 here is the conservative floor.',
+  },
+  {
+    label: 'Gemma 3 4B IT',
+    patterns: ['gemma-3-4b', 'gemma3-4b'],
+    capabilities: ['vision'],
+    paramsB: 4.3,
+    layers: 34,
+    queryHeads: 8,
+    kvHeads: 4,
+    headDim: 256,
+    hiddenSize: 2560,
+    contextLength: 131072,
+    swa: { windowTokens: 1024, patternLocal: 5, patternGlobal: 1 },
+    vision: { encoderParamsB: null, tokensPerImage: 2048 },
+    approx: true,
+    verified: 'hf-config+api (unsloth/gemma-3-4b-it)',
+    notes: 'Total 4.30B incl. SigLIP-400M vision (paramsB already includes it). 5:1 local:global sliding-window hybrid; image tokens scale with resolution (896px ≈ 4k raw, clamped).',
+  },
+  {
+    label: 'Gemma 3 12B IT',
+    patterns: ['gemma-3-12b', 'gemma3-12b'],
+    capabilities: ['vision'],
+    paramsB: 12.19,
+    layers: 48,
+    queryHeads: 16,
+    kvHeads: 8,
+    headDim: 256,
+    hiddenSize: 3840,
+    contextLength: 131072,
+    swa: { windowTokens: 1024, patternLocal: 5, patternGlobal: 1 },
+    vision: { encoderParamsB: null, tokensPerImage: 2048 },
+    approx: true,
+    verified: 'hf-config+api (unsloth/gemma-3-12b-it)',
+    notes: 'Total 12.19B incl. vision encoder.',
+  },
+  {
+    label: 'Gemma 3 27B IT',
+    patterns: ['gemma-3-27b', 'gemma3-27b'],
+    capabilities: ['vision'],
+    paramsB: 27.43,
+    layers: 62,
+    queryHeads: 32,
+    kvHeads: 16,
+    headDim: 128,
+    hiddenSize: 5376,
+    contextLength: 131072,
+    swa: { windowTokens: 1024, patternLocal: 5, patternGlobal: 1 },
+    vision: { encoderParamsB: null, tokensPerImage: 2048 },
+    approx: true,
+    verified: 'hf-config+api (unsloth/gemma-3-27b-it)',
+    notes: 'Total 27.43B incl. vision encoder.',
+  },
+  {
+    label: 'Qwen2.5 7B Instruct (+ Coder 7B)',
+    patterns: ['qwen2.5-7b', 'qwen25-7b', 'qwen2.5-coder-7b'],
+    capabilities: ['tools'],
+    paramsB: 7.62,
+    layers: 28,
+    queryHeads: 28,
+    kvHeads: 4,
+    headDim: 128,
+    hiddenSize: 3584,
+    contextLength: 32768,
+    approx: true,
+    verified: 'hf-config+api (Qwen/Qwen2.5-7B-Instruct)',
+    notes: 'Native LM Studio tool use. 7.6156B exact.',
+  },
+  {
+    label: 'Qwen2.5 14B Instruct (+ Coder 14B)',
+    patterns: ['qwen2.5-14b', 'qwen25-14b', 'qwen2.5-coder-14b'],
+    capabilities: ['tools'],
+    paramsB: 14.77,
+    layers: 48,
+    queryHeads: 40,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 5120,
+    contextLength: 32768,
+    approx: true,
+    verified: 'hf-config+api (Qwen/Qwen2.5-14B-Instruct)',
+    notes: 'Native tool use; 128k window reachable only via YaRN.',
+  },
+  {
+    label: 'Qwen2.5-Coder-32B-Instruct',
+    patterns: ['qwen2.5-coder-32b', 'qwen2.5-32b'],
+    capabilities: ['tools'],
+    paramsB: 32.76,
+    layers: 64,
+    queryHeads: 40,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 5120,
+    contextLength: 32768,
+    approx: true,
+    verified: 'hf-config+api (Qwen/Qwen2.5-Coder-32B-Instruct)',
+    notes: '32.7639B exact; config ctx 32,768 (sliding window to 131,072 via YaRN).',
+  },
+  {
+    label: 'Qwen2.5-VL 7B Instruct',
+    patterns: ['qwen2.5-vl-7b', 'qwen2-vl-7b'],
+    capabilities: ['vision'],
+    paramsB: 8.29,
+    layers: 28,
+    queryHeads: 28,
+    kvHeads: 4,
+    headDim: 128,
+    hiddenSize: 3584,
+    contextLength: 128000,
+    vision: { encoderParamsB: null, tokensPerImage: 1024 },
+    approx: true,
+    verified: 'hf-config+api (Qwen/Qwen2.5-VL-7B-Instruct)',
+    notes: 'Total 8.292B incl. dynamic-resolution ViT (paramsB already includes it). Image tokens vary heavily with resolution — 1024 is a mid-range placeholder.',
+  },
+  {
+    label: 'Qwen3 8B (hybrid thinking)',
+    patterns: ['qwen3-8b'],
+    capabilities: ['tools', 'reasoning'],
+    paramsB: 8.19,
+    layers: 36,
+    queryHeads: 32,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 4096,
+    contextLength: 40960,
+    approx: true,
+    verified: 'hf-config+api (Qwen/Qwen3-8B)',
+    notes: 'Single-model tool use + <think> reasoning; enable/disable thinking via chat template kwargs.',
+  },
+  {
+    label: 'Qwen3 30B-A3B 2507 (MoE, total params)',
+    patterns: ['qwen3-30b-a3b-instruct-2507', 'qwen3-30b-a3b-thinking-2507', 'qwen3-30b-a3b', 'qwen3-30b'],
+    capabilities: ['tools', 'reasoning'],
+    paramsB: 30.53,
+    layers: 48,
+    queryHeads: 32,
+    kvHeads: 4,
+    headDim: 128,
+    hiddenSize: 2048,
+    contextLength: 262144,
+    isMoE: true,
+    approx: true,
+    verified: 'hf-config+api (Qwen/Qwen3-30B-A3B-Instruct-2507)',
+    notes: 'MoE: all 30.53B params sit in VRAM, ~3.3B active per token. Thinking-2507 twin has the same geometry (reasoning); Instruct-2507 drops the thinking channel. 262,144 native context.',
+  },
+  {
+    label: 'Llama 3.1 8B Instruct',
+    patterns: ['llama-3.1-8b', 'llama3.1-8b', 'meta-llama-3.1-8b'],
+    capabilities: ['tools'],
+    paramsB: 8.03,
+    layers: 32,
+    queryHeads: 32,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 4096,
+    contextLength: 131072,
+    approx: true,
+    verified: 'hf-config+api (NousResearch/Meta-Llama-3.1-8B-Instruct)',
+    notes: 'Baseline GQA tool-caller supported by LM Studio.',
+  },
+  {
+    label: 'Llama 3.3 70B Instruct',
+    patterns: ['llama-3.3-70b', 'llama3.3-70b'],
+    capabilities: ['tools'],
+    paramsB: 70.6,
+    layers: 80,
+    queryHeads: 64,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 8192,
+    contextLength: 131072,
+    approx: true,
+    verified: 'documented (shares Llama-3.1-70B architecture)',
+    notes: 'Multi-GPU territory: ~40 GB at Q4_K_M before KV cache.',
+  },
+  {
+    label: 'Hermes 3 Llama 3.1 8B',
+    patterns: ['hermes-3-llama-3.1-8b', 'hermes-3-8b', 'hermes-3'],
+    capabilities: ['tools'],
+    paramsB: 8.03,
+    layers: 32,
+    queryHeads: 32,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 4096,
+    contextLength: 131072,
+    approx: true,
+    verified: 'hf-config+api (NousResearch/Hermes-3-Llama-3.1-8B)',
+    notes: 'Strong agentic/function-calling fine-tune on the Llama 3.1 8B base.',
+  },
+  {
+    label: 'DeepSeek-R1-Distill-Qwen-14B',
+    patterns: ['r1-distill-qwen-14b', 'deepseek-r1-distill-qwen-14b'],
+    capabilities: ['reasoning'],
+    paramsB: 14.77,
+    layers: 48,
+    queryHeads: 40,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 5120,
+    contextLength: 131072,
+    approx: true,
+    verified: 'hf-config+api (deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)',
+    notes: 'Qwen2.5-14B geometry fine-tuned for <think> reasoning; no native tool calling.',
+  },
+  {
+    label: 'DeepSeek-R1-Distill-Llama-8B',
+    patterns: ['r1-distill-llama-8b', 'deepseek-r1-distill-llama-8b'],
+    capabilities: ['reasoning'],
+    paramsB: 8.03,
+    layers: 32,
+    queryHeads: 32,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 4096,
+    contextLength: 131072,
+    approx: true,
+    verified: 'hf-config (deepseek-ai/DeepSeek-R1-Distill-Llama-8B, Llama-3.1 geometry)',
+    notes: 'Llama-3.1-8B geometry; reasoning-only.',
+  },
+  {
+    label: 'GPT-OSS 20B (MoE, total params)',
+    patterns: ['gpt-oss-20b'],
+    capabilities: ['tools', 'reasoning'],
+    paramsB: 20.91,
+    layers: 24,
+    queryHeads: 64,
+    kvHeads: 8,
+    headDim: 64,
+    hiddenSize: 2880,
+    contextLength: 131072,
+    isMoE: true,
+    defaultQuant: 'mxfp4',
+    approx: true,
+    verified: 'hf-config+api (openai/gpt-oss-20b)',
+    notes: 'Native MXFP4 checkpoint (~4.25 bpw) — all 20.91B params in VRAM, ~3.6B active. Harmony format with reasoning + tool channels.',
+  },
+  {
+    label: 'GLM-4 9B 0414 / GLM-Z1 9B 0414',
+    patterns: ['glm-4-9b-0414', 'glm-z1-9b-0414', 'glm4-9b'],
+    capabilities: ['tools', 'reasoning'],
+    paramsB: 9.4,
+    layers: 40,
+    queryHeads: 32,
+    kvHeads: 2,
+    headDim: 128,
+    hiddenSize: 4096,
+    contextLength: 32768,
+    approx: true,
+    verified: 'hf-config+api (zai-org/GLM-4-9B-0414)',
+    notes: 'GLM-4-9B-0414 = tool calling; GLM-Z1-9B-0414 twin = thinking/reasoning, same geometry. Only 2 KV heads — cheap cache.',
+  },
+  {
+    label: 'GLM-4.1V 9B Thinking (VLM)',
+    patterns: ['glm-4.1v-9b-thinking', 'glm-4.1v-9b', 'glm4.1v'],
+    capabilities: ['vision', 'reasoning'],
+    paramsB: 10.29,
+    layers: 40,
+    queryHeads: 32,
+    kvHeads: 2,
+    headDim: 128,
+    hiddenSize: 4096,
+    contextLength: 65536,
+    vision: { encoderParamsB: null, tokensPerImage: 576 },
+    approx: true,
+    verified: 'hf-config+api (zai-org/GLM-4.1V-9B-Thinking)',
+    notes: 'Total 10.293B incl. AIMv2 vision encoder (paramsB already includes it); <think> reasoning over images.',
+  },
+  {
+    label: 'Mistral Small 3.2 24B Instruct',
+    patterns: ['mistral-small-3.2-24b', 'mistral-small-3.1-24b', 'mistral-small-24b'],
+    capabilities: ['tools', 'vision'],
+    paramsB: 23.57,
+    layers: 40,
+    queryHeads: 32,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 5120,
+    contextLength: 131072,
+    vision: { encoderParamsB: null, tokensPerImage: 2048 },
+    approx: true,
+    verified: 'hf-config+api (mlx-community/Mistral-Small-3.2-24B-Instruct-2506-4bit)',
+    notes: 'Total 23.57B incl. SigLIP vision encoder (paramsB already includes it). Tool use + image input; no thinking channel.',
+  },
+  {
+    label: 'Pixtral 12B (VLM)',
+    patterns: ['pixtral-12b', 'pixtral'],
+    capabilities: ['vision'],
+    paramsB: 12.68,
+    layers: 40,
+    queryHeads: 32,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 5120,
+    contextLength: 131072,
+    vision: { encoderParamsB: null, tokensPerImage: 2048 },
+    approx: true,
+    verified: 'hf-config+api (mlx-community/Pixtral-12B-4bit)',
+    notes: 'Total 12.683B incl. vision encoder. config max_position_embeddings is 1,024,000 — practical LM Studio default used here; image tokens scale with resolution.',
+  },
+  {
+    label: 'Phi-4-reasoning-plus 14B',
+    patterns: ['phi-4-reasoning-plus', 'phi-4-reasoning', 'phi4-reasoning'],
+    capabilities: ['reasoning'],
+    paramsB: 14.66,
+    layers: 40,
+    queryHeads: 40,
+    kvHeads: 10,
+    headDim: 128,
+    hiddenSize: 5120,
+    contextLength: 32768,
+    approx: true,
+    verified: 'hf-config+api (microsoft/phi-4-reasoning-plus)',
+    notes: 'Chain-of-thought fine-tune of Phi-4 (14.659B); no vision, tool use is not native.',
+  },
+  {
+    label: 'LLaVA 1.6 Mistral 7B (VLM)',
+    patterns: ['llava-v1.6-mistral-7b', 'llava-mistral-7b'],
+    capabilities: ['vision'],
+    paramsB: 7.57,
+    layers: 32,
+    queryHeads: 32,
+    kvHeads: 8,
+    headDim: 128,
+    hiddenSize: 4096,
+    contextLength: 32768,
+    vision: { encoderParamsB: null, tokensPerImage: 576 },
+    approx: true,
+    verified: 'hf-config+api (llava-hf/llava-v1.6-mistral-7b-hf)',
+    notes: 'Total 7.567B incl. CLIP ViT-L/14-336px (577 tokens/image; paramsB already includes it). Vision only — no tool calling.',
+  },
+]);
+
+const SORTED_REGISTRY = [...MODEL_REGISTRY].sort(
+  (a, b) => Math.max(...b.patterns.map(p => p.length)) - Math.max(...a.patterns.map(p => p.length))
+);
+
+function normalizeForMatch(name) {
+  return String(name || '').toLowerCase().replace(/[^a-z0-9.\-]/g, '');
+}
+
+export function findRegistryEntry(name) {
+  const normalized = normalizeForMatch(name);
+  if (!normalized) return null;
+  for (const entry of SORTED_REGISTRY) {
+    for (const pattern of entry.patterns) {
+      if (normalized.includes(pattern)) return entry;
+    }
+  }
+  return null;
+}

package/tokens.css

@@ -23,6 +23,21 @@
   --color-agent-status-done: oklch(76% 0.15 158);
   --color-agent-status-done-ink: oklch(20% 0.035 158);
   --color-agent-status-error: oklch(70% 0.19 24);
+  --color-settings-canvas: oklch(15% 0.018 270);
+  --color-settings-surface: oklch(19% 0.02 270);
+  --color-settings-surface-raised: oklch(23% 0.022 270);
+  --color-settings-rule: oklch(34% 0.025 270);
+  --color-settings-rule-strong: oklch(50% 0.035 270);
+  --color-settings-ink: oklch(95% 0.008 270);
+  --color-settings-muted: oklch(73% 0.02 270);
+  --color-settings-accent: oklch(72% 0.16 248);
+  --color-settings-accent-soft: oklch(26% 0.055 248);
+  --color-settings-accent-ink: oklch(16% 0.025 248);
+  --color-settings-focus: oklch(78% 0.16 248);
+  --color-settings-success: oklch(75% 0.14 158);
+  --color-settings-warning: oklch(80% 0.14 78);
+  --color-settings-error: oklch(70% 0.19 24);
+  --color-settings-overlay: oklch(8% 0.015 270 / 0.82);
   --user-msg: linear-gradient(135deg, #e11d48 0%, #be123c 60%, #9f1239 100%);
 
   --font-sans: 'Google Sans', 'Google Sans Text', 'Product Sans', 'Noto Sans Thai', 'Noto Sans SC', sans-serif;
@@ -59,7 +74,8 @@
   --radius-sm: 0.625rem;
   --radius-card: 1.25rem;
   --radius-input: 1.75rem;
-  --radius-pill: 999px;
+  --radius-pill: 999px;
+  --radius-settings: 0.75rem;
   --rule-hair: 1px;
 
   --z-base: 1;
@@ -71,7 +87,7 @@
   --z-tooltip: 600;
 }
 
-[data-theme='light'] {
+[data-theme='light'] {
   --color-paper: #FFFDF9;
   --color-paper-2: rgba(255, 255, 255, 0.94);
   --color-paper-3: rgba(255, 250, 252, 0.96);
@@ -93,5 +109,20 @@
   --color-agent-status-done: oklch(48% 0.14 158);
   --color-agent-status-done-ink: oklch(98% 0.008 158);
   --color-agent-status-error: oklch(52% 0.2 24);
+  --color-settings-canvas: oklch(98% 0.008 255);
+  --color-settings-surface: oklch(95% 0.012 255);
+  --color-settings-surface-raised: oklch(100% 0.004 255);
+  --color-settings-rule: oklch(84% 0.025 255);
+  --color-settings-rule-strong: oklch(66% 0.04 255);
+  --color-settings-ink: oklch(22% 0.035 270);
+  --color-settings-muted: oklch(46% 0.035 270);
+  --color-settings-accent: oklch(48% 0.18 248);
+  --color-settings-accent-soft: oklch(92% 0.045 248);
+  --color-settings-accent-ink: oklch(98% 0.008 255);
+  --color-settings-focus: oklch(44% 0.2 248);
+  --color-settings-success: oklch(48% 0.14 158);
+  --color-settings-warning: oklch(52% 0.14 70);
+  --color-settings-error: oklch(52% 0.2 24);
+  --color-settings-overlay: oklch(18% 0.02 270 / 0.52);
   --user-msg: linear-gradient(135deg, #f43f5e 0%, #e11d48 100%);
 }