@tryhamster/gerbil 1.1.3 → 1.2.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (55) hide show
  1. package/dist/{architectures-wan0swXu.mjs → architectures-Ol7yUmKf.mjs} +128 -68
  2. package/dist/architectures-Ol7yUmKf.mjs.map +1 -0
  3. package/dist/cli.mjs +8 -8
  4. package/dist/cli.mjs.map +1 -1
  5. package/dist/{defaults-9komdrbY.mjs → defaults-bJcmOaRG.mjs} +2 -2
  6. package/dist/defaults-bJcmOaRG.mjs.map +1 -0
  7. package/dist/frameworks/express.mjs +1 -1
  8. package/dist/frameworks/fastify.mjs +1 -1
  9. package/dist/frameworks/hono.mjs +1 -1
  10. package/dist/frameworks/next.d.mts +2 -2
  11. package/dist/frameworks/next.mjs +1 -1
  12. package/dist/frameworks/trpc.mjs +1 -1
  13. package/dist/gerbil-0M50i8TK.mjs +4 -0
  14. package/dist/{gerbil-CQspp8fS.mjs → gerbil-DW1VioGN.mjs} +2 -2
  15. package/dist/{gerbil-CQspp8fS.mjs.map → gerbil-DW1VioGN.mjs.map} +1 -1
  16. package/dist/{gerbil-DZbWbIm2.d.mts → gerbil-DkilSf1Q.d.mts} +2 -2
  17. package/dist/{gerbil-DZbWbIm2.d.mts.map → gerbil-DkilSf1Q.d.mts.map} +1 -1
  18. package/dist/gpu/hooks.d.mts +24 -7
  19. package/dist/gpu/hooks.d.mts.map +1 -1
  20. package/dist/gpu/hooks.mjs +14 -7
  21. package/dist/gpu/hooks.mjs.map +1 -1
  22. package/dist/gpu/index.d.mts +2 -2
  23. package/dist/gpu/index.mjs +5 -5
  24. package/dist/{gpu-BYCQ5aTI.mjs → gpu-DR6EG0K-.mjs} +93 -24
  25. package/dist/gpu-DR6EG0K-.mjs.map +1 -0
  26. package/dist/index-Dgmb2kE3.d.mts.map +1 -1
  27. package/dist/{index-CKszv3lt.d.mts → index-HVS7yeLx.d.mts} +232 -133
  28. package/dist/index-HVS7yeLx.d.mts.map +1 -0
  29. package/dist/index.d.mts +2 -2
  30. package/dist/index.mjs +6 -6
  31. package/dist/integrations/ai-sdk.mjs +1 -1
  32. package/dist/integrations/langchain.mjs +1 -1
  33. package/dist/integrations/llamaindex.mjs +1 -1
  34. package/dist/integrations/mcp.d.mts +2 -2
  35. package/dist/integrations/mcp.mjs +4 -4
  36. package/dist/{mcp-CszII8GJ.mjs → mcp-8Tn3U-EE.mjs} +3 -3
  37. package/dist/{mcp-CszII8GJ.mjs.map → mcp-8Tn3U-EE.mjs.map} +1 -1
  38. package/dist/moonshine-stt-B3YtKcEE.mjs +4 -0
  39. package/dist/{moonshine-stt-DDUm5WfP.mjs → moonshine-stt-DSz3VW_0.mjs} +56 -6
  40. package/dist/{moonshine-stt-DDUm5WfP.mjs.map → moonshine-stt-DSz3VW_0.mjs.map} +1 -1
  41. package/dist/{one-liner-DVche8Eg.mjs → one-liner-BK9wqRYw.mjs} +2 -2
  42. package/dist/{one-liner-DVche8Eg.mjs.map → one-liner-BK9wqRYw.mjs.map} +1 -1
  43. package/dist/repl-B62rx3DS.mjs +9 -0
  44. package/dist/skills/index.d.mts +2 -2
  45. package/dist/skills/index.mjs +3 -3
  46. package/dist/{skills-BQI-VSn2.mjs → skills-BF3FVv8_.mjs} +2 -2
  47. package/dist/{skills-BQI-VSn2.mjs.map → skills-BF3FVv8_.mjs.map} +1 -1
  48. package/package.json +1 -1
  49. package/dist/architectures-wan0swXu.mjs.map +0 -1
  50. package/dist/defaults-9komdrbY.mjs.map +0 -1
  51. package/dist/gerbil-tHnYb9nZ.mjs +0 -4
  52. package/dist/gpu-BYCQ5aTI.mjs.map +0 -1
  53. package/dist/index-CKszv3lt.d.mts.map +0 -1
  54. package/dist/moonshine-stt-BtR-KiSP.mjs +0 -4
  55. package/dist/repl-BYJA2U7t.mjs +0 -9
@@ -1 +1 @@
1
- {"version":3,"file":"mcp-CszII8GJ.mjs","names":["handlers: Record<string, (args: any) => Promise<any>>","skills.summarize","skills.explain","skills.review","skills.commit","skills.translate"],"sources":["../src/integrations/mcp.ts"],"sourcesContent":["/**\n * Gerbil MCP Server\n *\n * Model Context Protocol server for Claude Desktop, Cursor, etc.\n *\n * @example\n * ```bash\n * gerbil serve --mcp\n * ```\n *\n * @example Claude Desktop config\n * ```json\n * {\n * \"mcpServers\": {\n * \"gerbil\": {\n * \"command\": \"npx\",\n * \"args\": [\"-y\", \"gerbil\", \"serve\", \"--mcp\"]\n * }\n * }\n * }\n * ```\n */\n\nimport { Gerbil } from \"../core/gerbil.js\";\nimport { DEFAULT_MODEL } from \"../core/models.js\";\nimport type { GerbilConfig } from \"../core/types.js\";\nimport * as skills from \"../skills/index.js\";\n\nexport interface MCPServerOptions extends GerbilConfig {\n /** Port for HTTP transport (default: stdio) */\n port?: number;\n\n /** Tools to expose */\n tools?: string[];\n}\n\n// Tool definitions\nconst TOOL_DEFINITIONS = {\n generate: {\n name: \"gerbil_generate\",\n description: \"Generate text using a local LLM\",\n inputSchema: {\n type: \"object\",\n properties: {\n prompt: { type: \"string\", description: \"The prompt to generate from\" },\n maxTokens: { type: \"number\", description: \"Maximum tokens to generate\" },\n temperature: { type: \"number\", description: \"Sampling temperature (0-2)\" },\n system: { type: \"string\", description: \"System prompt\" },\n thinking: { type: \"boolean\", description: \"Enable thinking mode\" },\n },\n required: [\"prompt\"],\n },\n },\n summarize: {\n name: \"gerbil_summarize\",\n description: \"Summarize content\",\n inputSchema: {\n type: \"object\",\n properties: {\n content: { type: \"string\", description: \"Content to summarize\" },\n length: { type: \"string\", enum: [\"short\", \"medium\", \"long\"] },\n format: { type: \"string\", enum: [\"paragraph\", \"bullets\"] },\n },\n required: [\"content\"],\n },\n },\n explain: {\n name: \"gerbil_explain\",\n description: \"Explain code or concepts\",\n inputSchema: {\n type: \"object\",\n properties: {\n content: { type: \"string\", description: \"Code or concept to explain\" },\n level: { type: \"string\", enum: [\"beginner\", \"intermediate\", \"expert\"] },\n language: { type: \"string\", description: \"Programming language\" },\n },\n required: [\"content\"],\n },\n },\n review: {\n name: \"gerbil_review\",\n description: \"Review code for issues\",\n inputSchema: {\n type: \"object\",\n properties: {\n code: { type: \"string\", description: \"Code to review\" },\n focus: {\n type: \"array\",\n items: { type: \"string\", enum: [\"security\", \"performance\", \"style\", \"bugs\"] },\n },\n },\n required: [\"code\"],\n },\n },\n commit: {\n name: \"gerbil_commit\",\n description: \"Generate a commit message from a diff\",\n inputSchema: {\n type: \"object\",\n properties: {\n diff: { type: \"string\", description: \"Git diff\" },\n type: { type: \"string\", enum: [\"conventional\", \"simple\", \"detailed\"] },\n },\n required: [\"diff\"],\n },\n },\n translate: {\n name: \"gerbil_translate\",\n description: \"Translate text\",\n inputSchema: {\n type: \"object\",\n properties: {\n text: { type: \"string\", description: \"Text to translate\" },\n to: { type: \"string\", description: \"Target language\" },\n from: { type: \"string\", description: \"Source language (optional)\" },\n },\n required: [\"text\", \"to\"],\n },\n },\n embed: {\n name: \"gerbil_embed\",\n description: \"Generate embeddings for text\",\n inputSchema: {\n type: \"object\",\n properties: {\n text: { type: \"string\", description: \"Text to embed\" },\n },\n required: [\"text\"],\n },\n },\n};\n\n/**\n * Create MCP server\n */\nexport async function createMCPServer(options: MCPServerOptions = {}) {\n const g = new Gerbil(options);\n\n // Load model\n await g.loadModel(options.model || DEFAULT_MODEL);\n\n // Determine which tools to expose\n const enabledTools = options.tools || Object.keys(TOOL_DEFINITIONS);\n\n // Tool handlers\n const handlers: Record<string, (args: any) => Promise<any>> = {\n gerbil_generate: async (args) => {\n const result = await g.generate(args.prompt, {\n maxTokens: args.maxTokens,\n temperature: args.temperature,\n system: args.system,\n thinking: args.thinking,\n });\n return {\n content: [\n {\n type: \"text\",\n text: result.thinking\n ? `Thinking: ${result.thinking}\\n\\nAnswer: ${result.text}`\n : result.text,\n },\n ],\n };\n },\n\n gerbil_summarize: async (args) => {\n const summary = await skills.summarize({\n content: args.content,\n length: args.length,\n format: args.format,\n });\n return { content: [{ type: \"text\", text: summary }] };\n },\n\n gerbil_explain: async (args) => {\n const explanation = await skills.explain({\n content: args.content,\n level: args.level,\n language: args.language,\n } as Parameters<typeof skills.explain>[0]);\n return { content: [{ type: \"text\", text: explanation }] };\n },\n\n gerbil_review: async (args) => {\n const reviewResult = await skills.review({\n code: args.code,\n focus: args.focus,\n } as Parameters<typeof skills.review>[0]);\n return { content: [{ type: \"text\", text: reviewResult }] };\n },\n\n gerbil_commit: async (args) => {\n const message = await skills.commit({\n diff: args.diff,\n type: args.type,\n } as Parameters<typeof skills.commit>[0]);\n return { content: [{ type: \"text\", text: message }] };\n },\n\n gerbil_translate: async (args) => {\n const translated = await skills.translate({\n text: args.text,\n to: args.to,\n from: args.from,\n } as Parameters<typeof skills.translate>[0]);\n return { content: [{ type: \"text\", text: translated }] };\n },\n\n gerbil_embed: async (args) => {\n const result = await g.embed(args.text);\n return {\n content: [\n {\n type: \"text\",\n text: JSON.stringify({\n vector: result.vector.slice(0, 10),\n dimensions: result.vector.length,\n }),\n },\n ],\n };\n },\n };\n\n return {\n // Server info\n info: {\n name: \"gerbil\",\n version: \"1.0.0\",\n description: \"Local LLM inference via Gerbil\",\n },\n\n // List available tools\n listTools: () =>\n enabledTools\n .filter((t) => TOOL_DEFINITIONS[t as keyof typeof TOOL_DEFINITIONS])\n .map((t) => TOOL_DEFINITIONS[t as keyof typeof TOOL_DEFINITIONS]),\n\n // Call a tool\n callTool: async (name: string, args: any) => {\n const handler = handlers[name];\n if (!handler) {\n throw new Error(`Unknown tool: ${name}`);\n }\n return handler(args);\n },\n\n // Get Gerbil instance\n gerbil: g,\n };\n}\n\n/**\n * Start MCP server with stdio transport\n */\nexport async function startMCPServer(options: MCPServerOptions = {}) {\n const server = await createMCPServer(options);\n\n // Simple stdio protocol handler\n const readline = await import(\"node:readline\");\n const rl = readline.createInterface({\n input: process.stdin,\n output: process.stdout,\n terminal: false,\n });\n\n rl.on(\"line\", async (line) => {\n try {\n const request = JSON.parse(line);\n\n let _response: any;\n\n switch (request.method) {\n case \"initialize\":\n _response = {\n protocolVersion: \"2024-11-05\",\n serverInfo: server.info,\n capabilities: {\n tools: {},\n },\n };\n break;\n\n case \"tools/list\":\n _response = { tools: server.listTools() };\n break;\n\n case \"tools/call\":\n _response = await server.callTool(request.params.name, request.params.arguments);\n break;\n\n default:\n _response = { error: { code: -32_601, message: \"Method not found\" } };\n }\n } catch (_error) {}\n });\n}\n\nexport default { createMCPServer, startMCPServer };\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,MAAM,mBAAmB;CACvB,UAAU;EACR,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,QAAQ;KAAE,MAAM;KAAU,aAAa;KAA+B;IACtE,WAAW;KAAE,MAAM;KAAU,aAAa;KAA8B;IACxE,aAAa;KAAE,MAAM;KAAU,aAAa;KAA8B;IAC1E,QAAQ;KAAE,MAAM;KAAU,aAAa;KAAiB;IACxD,UAAU;KAAE,MAAM;KAAW,aAAa;KAAwB;IACnE;GACD,UAAU,CAAC,SAAS;GACrB;EACF;CACD,WAAW;EACT,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,SAAS;KAAE,MAAM;KAAU,aAAa;KAAwB;IAChE,QAAQ;KAAE,MAAM;KAAU,MAAM;MAAC;MAAS;MAAU;MAAO;KAAE;IAC7D,QAAQ;KAAE,MAAM;KAAU,MAAM,CAAC,aAAa,UAAU;KAAE;IAC3D;GACD,UAAU,CAAC,UAAU;GACtB;EACF;CACD,SAAS;EACP,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,SAAS;KAAE,MAAM;KAAU,aAAa;KAA8B;IACtE,OAAO;KAAE,MAAM;KAAU,MAAM;MAAC;MAAY;MAAgB;MAAS;KAAE;IACvE,UAAU;KAAE,MAAM;KAAU,aAAa;KAAwB;IAClE;GACD,UAAU,CAAC,UAAU;GACtB;EACF;CACD,QAAQ;EACN,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAkB;IACvD,OAAO;KACL,MAAM;KACN,OAAO;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAe;OAAS;OAAO;MAAE;KAC9E;IACF;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACD,QAAQ;EACN,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAY;IACjD,MAAM;KAAE,MAAM;KAAU,MAAM;MAAC;MAAgB;MAAU;MAAW;KAAE;IACvE;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACD,WAAW;EACT,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAqB;IAC1D,IAAI;KAAE,MAAM;KAAU,aAAa;KAAmB;IACtD,MAAM;KAAE,MAAM;KAAU,aAAa;KAA8B;IACpE;GACD,UAAU,CAAC,QAAQ,KAAK;GACzB;EACF;CACD,OAAO;EACL,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY,EACV,MAAM;IAAE,MAAM;IAAU,aAAa;IAAiB,EACvD;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACF;;;;AAKD,eAAsB,gBAAgB,UAA4B,EAAE,EAAE;CACpE,MAAM,IAAI,IAAI,OAAO,QAAQ;AAG7B,OAAM,EAAE,UAAU,QAAQ,SAAS,cAAc;CAGjD,MAAM,eAAe,QAAQ,SAAS,OAAO,KAAK,iBAAiB;CAGnE,MAAMA,WAAwD;EAC5D,iBAAiB,OAAO,SAAS;GAC/B,MAAM,SAAS,MAAM,EAAE,SAAS,KAAK,QAAQ;IAC3C,WAAW,KAAK;IAChB,aAAa,KAAK;IAClB,QAAQ,KAAK;IACb,UAAU,KAAK;IAChB,CAAC;AACF,UAAO,EACL,SAAS,CACP;IACE,MAAM;IACN,MAAM,OAAO,WACT,aAAa,OAAO,SAAS,cAAc,OAAO,SAClD,OAAO;IACZ,CACF,EACF;;EAGH,kBAAkB,OAAO,SAAS;AAMhC,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALnB,MAAMC,UAAiB;KACrC,SAAS,KAAK;KACd,QAAQ,KAAK;KACb,QAAQ,KAAK;KACd,CAAC;IACgD,CAAC,EAAE;;EAGvD,gBAAgB,OAAO,SAAS;AAM9B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALf,MAAMC,QAAe;KACvC,SAAS,KAAK;KACd,OAAO,KAAK;KACZ,UAAU,KAAK;KAChB,CAAyC;IACY,CAAC,EAAE;;EAG3D,eAAe,OAAO,SAAS;AAK7B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MAJd,MAAMC,OAAc;KACvC,MAAM,KAAK;KACX,OAAO,KAAK;KACb,CAAwC;IACc,CAAC,EAAE;;EAG5D,eAAe,OAAO,SAAS;AAK7B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MAJnB,MAAMC,OAAc;KAClC,MAAM,KAAK;KACX,MAAM,KAAK;KACZ,CAAwC;IACS,CAAC,EAAE;;EAGvD,kBAAkB,OAAO,SAAS;AAMhC,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALhB,MAAMC,UAAiB;KACxC,MAAM,KAAK;KACX,IAAI,KAAK;KACT,MAAM,KAAK;KACZ,CAA2C;IACS,CAAC,EAAE;;EAG1D,cAAc,OAAO,SAAS;GAC5B,MAAM,SAAS,MAAM,EAAE,MAAM,KAAK,KAAK;AACvC,UAAO,EACL,SAAS,CACP;IACE,MAAM;IACN,MAAM,KAAK,UAAU;KACnB,QAAQ,OAAO,OAAO,MAAM,GAAG,GAAG;KAClC,YAAY,OAAO,OAAO;KAC3B,CAAC;IACH,CACF,EACF;;EAEJ;AAED,QAAO;EAEL,MAAM;GACJ,MAAM;GACN,SAAS;GACT,aAAa;GACd;EAGD,iBACE,aACG,QAAQ,MAAM,iBAAiB,GAAoC,CACnE,KAAK,MAAM,iBAAiB,GAAoC;EAGrE,UAAU,OAAO,MAAc,SAAc;GAC3C,MAAM,UAAU,SAAS;AACzB,OAAI,CAAC,QACH,OAAM,IAAI,MAAM,iBAAiB,OAAO;AAE1C,UAAO,QAAQ,KAAK;;EAItB,QAAQ;EACT;;;;;AAMH,eAAsB,eAAe,UAA4B,EAAE,EAAE;CACnE,MAAM,SAAS,MAAM,gBAAgB,QAAQ;AAU7C,EAPiB,MAAM,OAAO,kBACV,gBAAgB;EAClC,OAAO,QAAQ;EACf,QAAQ,QAAQ;EAChB,UAAU;EACX,CAAC,CAEC,GAAG,QAAQ,OAAO,SAAS;AAC5B,MAAI;GACF,MAAM,UAAU,KAAK,MAAM,KAAK;AAIhC,WAAQ,QAAQ,QAAhB;IACE,KAAK;AACH,KAEc,OAAO;AAKrB;IAEF,KAAK;AACH,KAAqB,OAAO,WAAW;AACvC;IAEF,KAAK;AACH,KAAY,MAAM,OAAO,SAAS,QAAQ,OAAO,MAAM,QAAQ,OAAO,UAAU;AAChF;IAEF;;WAGK,QAAQ;GACjB;;AAGJ,kBAAe;CAAE;CAAiB;CAAgB"}
1
+ {"version":3,"file":"mcp-8Tn3U-EE.mjs","names":["handlers: Record<string, (args: any) => Promise<any>>","skills.summarize","skills.explain","skills.review","skills.commit","skills.translate"],"sources":["../src/integrations/mcp.ts"],"sourcesContent":["/**\n * Gerbil MCP Server\n *\n * Model Context Protocol server for Claude Desktop, Cursor, etc.\n *\n * @example\n * ```bash\n * gerbil serve --mcp\n * ```\n *\n * @example Claude Desktop config\n * ```json\n * {\n * \"mcpServers\": {\n * \"gerbil\": {\n * \"command\": \"npx\",\n * \"args\": [\"-y\", \"gerbil\", \"serve\", \"--mcp\"]\n * }\n * }\n * }\n * ```\n */\n\nimport { Gerbil } from \"../core/gerbil.js\";\nimport { DEFAULT_MODEL } from \"../core/models.js\";\nimport type { GerbilConfig } from \"../core/types.js\";\nimport * as skills from \"../skills/index.js\";\n\nexport interface MCPServerOptions extends GerbilConfig {\n /** Port for HTTP transport (default: stdio) */\n port?: number;\n\n /** Tools to expose */\n tools?: string[];\n}\n\n// Tool definitions\nconst TOOL_DEFINITIONS = {\n generate: {\n name: \"gerbil_generate\",\n description: \"Generate text using a local LLM\",\n inputSchema: {\n type: \"object\",\n properties: {\n prompt: { type: \"string\", description: \"The prompt to generate from\" },\n maxTokens: { type: \"number\", description: \"Maximum tokens to generate\" },\n temperature: { type: \"number\", description: \"Sampling temperature (0-2)\" },\n system: { type: \"string\", description: \"System prompt\" },\n thinking: { type: \"boolean\", description: \"Enable thinking mode\" },\n },\n required: [\"prompt\"],\n },\n },\n summarize: {\n name: \"gerbil_summarize\",\n description: \"Summarize content\",\n inputSchema: {\n type: \"object\",\n properties: {\n content: { type: \"string\", description: \"Content to summarize\" },\n length: { type: \"string\", enum: [\"short\", \"medium\", \"long\"] },\n format: { type: \"string\", enum: [\"paragraph\", \"bullets\"] },\n },\n required: [\"content\"],\n },\n },\n explain: {\n name: \"gerbil_explain\",\n description: \"Explain code or concepts\",\n inputSchema: {\n type: \"object\",\n properties: {\n content: { type: \"string\", description: \"Code or concept to explain\" },\n level: { type: \"string\", enum: [\"beginner\", \"intermediate\", \"expert\"] },\n language: { type: \"string\", description: \"Programming language\" },\n },\n required: [\"content\"],\n },\n },\n review: {\n name: \"gerbil_review\",\n description: \"Review code for issues\",\n inputSchema: {\n type: \"object\",\n properties: {\n code: { type: \"string\", description: \"Code to review\" },\n focus: {\n type: \"array\",\n items: { type: \"string\", enum: [\"security\", \"performance\", \"style\", \"bugs\"] },\n },\n },\n required: [\"code\"],\n },\n },\n commit: {\n name: \"gerbil_commit\",\n description: \"Generate a commit message from a diff\",\n inputSchema: {\n type: \"object\",\n properties: {\n diff: { type: \"string\", description: \"Git diff\" },\n type: { type: \"string\", enum: [\"conventional\", \"simple\", \"detailed\"] },\n },\n required: [\"diff\"],\n },\n },\n translate: {\n name: \"gerbil_translate\",\n description: \"Translate text\",\n inputSchema: {\n type: \"object\",\n properties: {\n text: { type: \"string\", description: \"Text to translate\" },\n to: { type: \"string\", description: \"Target language\" },\n from: { type: \"string\", description: \"Source language (optional)\" },\n },\n required: [\"text\", \"to\"],\n },\n },\n embed: {\n name: \"gerbil_embed\",\n description: \"Generate embeddings for text\",\n inputSchema: {\n type: \"object\",\n properties: {\n text: { type: \"string\", description: \"Text to embed\" },\n },\n required: [\"text\"],\n },\n },\n};\n\n/**\n * Create MCP server\n */\nexport async function createMCPServer(options: MCPServerOptions = {}) {\n const g = new Gerbil(options);\n\n // Load model\n await g.loadModel(options.model || DEFAULT_MODEL);\n\n // Determine which tools to expose\n const enabledTools = options.tools || Object.keys(TOOL_DEFINITIONS);\n\n // Tool handlers\n const handlers: Record<string, (args: any) => Promise<any>> = {\n gerbil_generate: async (args) => {\n const result = await g.generate(args.prompt, {\n maxTokens: args.maxTokens,\n temperature: args.temperature,\n system: args.system,\n thinking: args.thinking,\n });\n return {\n content: [\n {\n type: \"text\",\n text: result.thinking\n ? `Thinking: ${result.thinking}\\n\\nAnswer: ${result.text}`\n : result.text,\n },\n ],\n };\n },\n\n gerbil_summarize: async (args) => {\n const summary = await skills.summarize({\n content: args.content,\n length: args.length,\n format: args.format,\n });\n return { content: [{ type: \"text\", text: summary }] };\n },\n\n gerbil_explain: async (args) => {\n const explanation = await skills.explain({\n content: args.content,\n level: args.level,\n language: args.language,\n } as Parameters<typeof skills.explain>[0]);\n return { content: [{ type: \"text\", text: explanation }] };\n },\n\n gerbil_review: async (args) => {\n const reviewResult = await skills.review({\n code: args.code,\n focus: args.focus,\n } as Parameters<typeof skills.review>[0]);\n return { content: [{ type: \"text\", text: reviewResult }] };\n },\n\n gerbil_commit: async (args) => {\n const message = await skills.commit({\n diff: args.diff,\n type: args.type,\n } as Parameters<typeof skills.commit>[0]);\n return { content: [{ type: \"text\", text: message }] };\n },\n\n gerbil_translate: async (args) => {\n const translated = await skills.translate({\n text: args.text,\n to: args.to,\n from: args.from,\n } as Parameters<typeof skills.translate>[0]);\n return { content: [{ type: \"text\", text: translated }] };\n },\n\n gerbil_embed: async (args) => {\n const result = await g.embed(args.text);\n return {\n content: [\n {\n type: \"text\",\n text: JSON.stringify({\n vector: result.vector.slice(0, 10),\n dimensions: result.vector.length,\n }),\n },\n ],\n };\n },\n };\n\n return {\n // Server info\n info: {\n name: \"gerbil\",\n version: \"1.0.0\",\n description: \"Local LLM inference via Gerbil\",\n },\n\n // List available tools\n listTools: () =>\n enabledTools\n .filter((t) => TOOL_DEFINITIONS[t as keyof typeof TOOL_DEFINITIONS])\n .map((t) => TOOL_DEFINITIONS[t as keyof typeof TOOL_DEFINITIONS]),\n\n // Call a tool\n callTool: async (name: string, args: any) => {\n const handler = handlers[name];\n if (!handler) {\n throw new Error(`Unknown tool: ${name}`);\n }\n return handler(args);\n },\n\n // Get Gerbil instance\n gerbil: g,\n };\n}\n\n/**\n * Start MCP server with stdio transport\n */\nexport async function startMCPServer(options: MCPServerOptions = {}) {\n const server = await createMCPServer(options);\n\n // Simple stdio protocol handler\n const readline = await import(\"node:readline\");\n const rl = readline.createInterface({\n input: process.stdin,\n output: process.stdout,\n terminal: false,\n });\n\n rl.on(\"line\", async (line) => {\n try {\n const request = JSON.parse(line);\n\n let _response: any;\n\n switch (request.method) {\n case \"initialize\":\n _response = {\n protocolVersion: \"2024-11-05\",\n serverInfo: server.info,\n capabilities: {\n tools: {},\n },\n };\n break;\n\n case \"tools/list\":\n _response = { tools: server.listTools() };\n break;\n\n case \"tools/call\":\n _response = await server.callTool(request.params.name, request.params.arguments);\n break;\n\n default:\n _response = { error: { code: -32_601, message: \"Method not found\" } };\n }\n } catch (_error) {}\n });\n}\n\nexport default { createMCPServer, startMCPServer };\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AAqCA,MAAM,mBAAmB;CACvB,UAAU;EACR,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,QAAQ;KAAE,MAAM;KAAU,aAAa;KAA+B;IACtE,WAAW;KAAE,MAAM;KAAU,aAAa;KAA8B;IACxE,aAAa;KAAE,MAAM;KAAU,aAAa;KAA8B;IAC1E,QAAQ;KAAE,MAAM;KAAU,aAAa;KAAiB;IACxD,UAAU;KAAE,MAAM;KAAW,aAAa;KAAwB;IACnE;GACD,UAAU,CAAC,SAAS;GACrB;EACF;CACD,WAAW;EACT,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,SAAS;KAAE,MAAM;KAAU,aAAa;KAAwB;IAChE,QAAQ;KAAE,MAAM;KAAU,MAAM;MAAC;MAAS;MAAU;MAAO;KAAE;IAC7D,QAAQ;KAAE,MAAM;KAAU,MAAM,CAAC,aAAa,UAAU;KAAE;IAC3D;GACD,UAAU,CAAC,UAAU;GACtB;EACF;CACD,SAAS;EACP,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,SAAS;KAAE,MAAM;KAAU,aAAa;KAA8B;IACtE,OAAO;KAAE,MAAM;KAAU,MAAM;MAAC;MAAY;MAAgB;MAAS;KAAE;IACvE,UAAU;KAAE,MAAM;KAAU,aAAa;KAAwB;IAClE;GACD,UAAU,CAAC,UAAU;GACtB;EACF;CACD,QAAQ;EACN,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAkB;IACvD,OAAO;KACL,MAAM;KACN,OAAO;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAe;OAAS;OAAO;MAAE;KAC9E;IACF;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACD,QAAQ;EACN,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAY;IACjD,MAAM;KAAE,MAAM;KAAU,MAAM;MAAC;MAAgB;MAAU;MAAW;KAAE;IACvE;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACD,WAAW;EACT,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY;IACV,MAAM;KAAE,MAAM;KAAU,aAAa;KAAqB;IAC1D,IAAI;KAAE,MAAM;KAAU,aAAa;KAAmB;IACtD,MAAM;KAAE,MAAM;KAAU,aAAa;KAA8B;IACpE;GACD,UAAU,CAAC,QAAQ,KAAK;GACzB;EACF;CACD,OAAO;EACL,MAAM;EACN,aAAa;EACb,aAAa;GACX,MAAM;GACN,YAAY,EACV,MAAM;IAAE,MAAM;IAAU,aAAa;IAAiB,EACvD;GACD,UAAU,CAAC,OAAO;GACnB;EACF;CACF;;;;AAKD,eAAsB,gBAAgB,UAA4B,EAAE,EAAE;CACpE,MAAM,IAAI,IAAI,OAAO,QAAQ;AAG7B,OAAM,EAAE,UAAU,QAAQ,SAAS,cAAc;CAGjD,MAAM,eAAe,QAAQ,SAAS,OAAO,KAAK,iBAAiB;CAGnE,MAAMA,WAAwD;EAC5D,iBAAiB,OAAO,SAAS;GAC/B,MAAM,SAAS,MAAM,EAAE,SAAS,KAAK,QAAQ;IAC3C,WAAW,KAAK;IAChB,aAAa,KAAK;IAClB,QAAQ,KAAK;IACb,UAAU,KAAK;IAChB,CAAC;AACF,UAAO,EACL,SAAS,CACP;IACE,MAAM;IACN,MAAM,OAAO,WACT,aAAa,OAAO,SAAS,cAAc,OAAO,SAClD,OAAO;IACZ,CACF,EACF;;EAGH,kBAAkB,OAAO,SAAS;AAMhC,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALnB,MAAMC,UAAiB;KACrC,SAAS,KAAK;KACd,QAAQ,KAAK;KACb,QAAQ,KAAK;KACd,CAAC;IACgD,CAAC,EAAE;;EAGvD,gBAAgB,OAAO,SAAS;AAM9B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALf,MAAMC,QAAe;KACvC,SAAS,KAAK;KACd,OAAO,KAAK;KACZ,UAAU,KAAK;KAChB,CAAyC;IACY,CAAC,EAAE;;EAG3D,eAAe,OAAO,SAAS;AAK7B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MAJd,MAAMC,OAAc;KACvC,MAAM,KAAK;KACX,OAAO,KAAK;KACb,CAAwC;IACc,CAAC,EAAE;;EAG5D,eAAe,OAAO,SAAS;AAK7B,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MAJnB,MAAMC,OAAc;KAClC,MAAM,KAAK;KACX,MAAM,KAAK;KACZ,CAAwC;IACS,CAAC,EAAE;;EAGvD,kBAAkB,OAAO,SAAS;AAMhC,UAAO,EAAE,SAAS,CAAC;IAAE,MAAM;IAAQ,MALhB,MAAMC,UAAiB;KACxC,MAAM,KAAK;KACX,IAAI,KAAK;KACT,MAAM,KAAK;KACZ,CAA2C;IACS,CAAC,EAAE;;EAG1D,cAAc,OAAO,SAAS;GAC5B,MAAM,SAAS,MAAM,EAAE,MAAM,KAAK,KAAK;AACvC,UAAO,EACL,SAAS,CACP;IACE,MAAM;IACN,MAAM,KAAK,UAAU;KACnB,QAAQ,OAAO,OAAO,MAAM,GAAG,GAAG;KAClC,YAAY,OAAO,OAAO;KAC3B,CAAC;IACH,CACF,EACF;;EAEJ;AAED,QAAO;EAEL,MAAM;GACJ,MAAM;GACN,SAAS;GACT,aAAa;GACd;EAGD,iBACE,aACG,QAAQ,MAAM,iBAAiB,GAAoC,CACnE,KAAK,MAAM,iBAAiB,GAAoC;EAGrE,UAAU,OAAO,MAAc,SAAc;GAC3C,MAAM,UAAU,SAAS;AACzB,OAAI,CAAC,QACH,OAAM,IAAI,MAAM,iBAAiB,OAAO;AAE1C,UAAO,QAAQ,KAAK;;EAItB,QAAQ;EACT;;;;;AAMH,eAAsB,eAAe,UAA4B,EAAE,EAAE;CACnE,MAAM,SAAS,MAAM,gBAAgB,QAAQ;AAU7C,EAPiB,MAAM,OAAO,kBACV,gBAAgB;EAClC,OAAO,QAAQ;EACf,QAAQ,QAAQ;EAChB,UAAU;EACX,CAAC,CAEC,GAAG,QAAQ,OAAO,SAAS;AAC5B,MAAI;GACF,MAAM,UAAU,KAAK,MAAM,KAAK;AAIhC,WAAQ,QAAQ,QAAhB;IACE,KAAK;AACH,KAEc,OAAO;AAKrB;IAEF,KAAK;AACH,KAAqB,OAAO,WAAW;AACvC;IAEF,KAAK;AACH,KAAY,MAAM,OAAO,SAAS,QAAQ,OAAO,MAAM,QAAQ,OAAO,UAAU;AAChF;IAEF;;WAGK,QAAQ;GACjB;;AAGJ,kBAAe;CAAE;CAAiB;CAAgB"}
@@ -0,0 +1,4 @@
1
+ import "./architectures-Ol7yUmKf.mjs";
2
+ import { t as MoonshineSTT } from "./moonshine-stt-DSz3VW_0.mjs";
3
+
4
+ export { MoonshineSTT };
@@ -1,4 +1,4 @@
1
- import { D as createDefaultHFKeyMapper, S as nanoCodecWeightMap, T as DTYPE_BYTES, a as generateMoonshineEncoderGraph, c as DEFAULT_GROUP_SIZE, h as foldNanoCodecWeightNorm, i as generateMoonshineDecoderGraph, l as quantizeInt4, o as moonshineEncoderFrames, s as parseMoonshineConfig, t as generateGraph, w as CANONICAL_KEYS } from "./architectures-wan0swXu.mjs";
1
+ import { C as quantizeInt4, D as createDefaultHFKeyMapper, S as DEFAULT_GROUP_SIZE, T as DTYPE_BYTES, a as generateMoonshineEncoderGraph, b as nanoCodecWeightMap, i as generateMoonshineDecoderGraph, o as moonshineEncoderFrames, p as foldNanoCodecWeightNorm, s as parseMoonshineConfig, t as generateGraph, w as CANONICAL_KEYS } from "./architectures-Ol7yUmKf.mjs";
2
2
 
3
3
  //#region src/gpu/device.ts
4
4
  const COPY_SRC = 4;
@@ -11948,7 +11948,7 @@ function buildKaniCodecWeights(codecRaw) {
11948
11948
  }
11949
11949
  async function loadKaniTTS(options) {
11950
11950
  const revision = options.revision ?? "main";
11951
- const repo = options.repo ?? "nineninesix/kani-tts-2-en";
11951
+ const repo = options.repo ?? "nineninesix/kani-tts-450m-0.2-ft";
11952
11952
  const codecRepo = options.codecRepo ?? KANI_NANOCODEC_REPO;
11953
11953
  const baseURL = await resolveModelBaseURL(repo, revision);
11954
11954
  const codecURL = await resolveModelBaseURL(codecRepo, "main");
@@ -11961,7 +11961,16 @@ async function loadKaniTTS(options) {
11961
11961
  const backboneBuf = await fetchBinary(baseURL, "model.safetensors", options.hfToken, (loaded, total) => options.onProgress?.(8 + loaded / (total || 1) * 45, 100, "backbone weights"), options.cacheDir);
11962
11962
  const backboneFile = parseSafetensorsHeader(backboneBuf);
11963
11963
  const backboneWeights = /* @__PURE__ */ new Map();
11964
+ const preQuantized = (rawConfig.quantization ?? void 0)?.format === "gerbil-int4";
11965
+ const isQuantTensor = (name) => name.endsWith(".q") || name.endsWith(".scales") || name.endsWith(".zeros");
11964
11966
  for (const entry of backboneFile.entries) {
11967
+ if (preQuantized && isQuantTensor(entry.name)) {
11968
+ backboneWeights.set(entry.name, {
11969
+ data: getTensorData(backboneBuf, backboneFile, entry),
11970
+ shape: entry.shape
11971
+ });
11972
+ continue;
11973
+ }
11965
11974
  const key = (entry.name.startsWith("model.") ? entry.name.slice(6) : entry.name).replace(KANI_OUT_PROJ_RE, ".self_attn.o_proj.");
11966
11975
  backboneWeights.set(key, {
11967
11976
  data: toF32(entry, backboneBuf, backboneFile),
@@ -11978,14 +11987,55 @@ async function loadKaniTTS(options) {
11978
11987
  shape: entry.shape
11979
11988
  });
11980
11989
  const codecWeights = buildKaniCodecWeights(codecRaw);
11981
- options.onProgress?.(100, 100, "Kani-TTS-2 loaded.");
11990
+ options.onProgress?.(100, 100, "Kani-TTS loaded.");
11982
11991
  return {
11983
11992
  backboneWeights,
11984
11993
  codecWeights,
11985
11994
  tokenizer,
11986
- rawConfig
11995
+ rawConfig,
11996
+ preQuantized
11987
11997
  };
11988
11998
  }
11999
+ /**
12000
+ * Quantize the Kani codec-LM backbone's linear weights to INT4 in place.
12001
+ *
12002
+ * Walks the built backbone graph's `MatMulInt4` nodes (the q4 graph emits these for
12003
+ * every LFM2 linear projection — q/k/v/o, gate/up/down, short-conv in/out_proj) and,
12004
+ * for each, quantizes the f32 source weight (group-wise W4A16 via `quantizeInt4`)
12005
+ * into the `.q`/`.scales`/`.zeros` entries the executor binds, then frees the f32
12006
+ * source to release memory. Mirrors the main loader's q4 loop, but operates on the
12007
+ * dedicated Kani backbone weight map (which the loader builds at f32). The tied
12008
+ * embedding + LM head are intentionally NOT quantized (the graph keeps them f32),
12009
+ * so this leaves `embed_tokens.weight` untouched. Idempotent: tensors already
12010
+ * present (`.q` set) are skipped.
12011
+ */
12012
+ function quantizeKaniBackbone(graph, weights) {
12013
+ for (const node of graph.nodes) {
12014
+ if (node.opType !== "MatMulInt4") continue;
12015
+ const groupSz = node.attributes.group_size ?? DEFAULT_GROUP_SIZE;
12016
+ const qName = node.inputs[1];
12017
+ const scalesName = node.inputs[2];
12018
+ const zerosName = node.inputs[3];
12019
+ if (weights.has(qName)) continue;
12020
+ const sourceName = qName.slice(0, -2);
12021
+ const source = weights.get(sourceName);
12022
+ if (!source) continue;
12023
+ const { packed, scales, zeros } = quantizeInt4(source.data instanceof Float32Array ? source.data : new Float32Array(source.data.buffer, source.data.byteOffset, source.data.byteLength / 4), groupSz);
12024
+ weights.set(qName, {
12025
+ data: packed,
12026
+ shape: [packed.length]
12027
+ });
12028
+ weights.set(scalesName, {
12029
+ data: scales,
12030
+ shape: [scales.length]
12031
+ });
12032
+ weights.set(zerosName, {
12033
+ data: zeros,
12034
+ shape: [zeros.length]
12035
+ });
12036
+ weights.delete(sourceName);
12037
+ }
12038
+ }
11989
12039
 
11990
12040
  //#endregion
11991
12041
  //#region src/gpu/moonshine-executor.ts
@@ -12351,5 +12401,5 @@ function selectGraphWeights(graph, weights) {
12351
12401
  }
12352
12402
 
12353
12403
  //#endregion
12354
- export { verifyGPU as _, loadMoonshine as a, KERNEL_REGISTRY as c, createBindGroup as d, createStorageBuffer as f, initGPU as g, getOrCreatePipeline as h, loadModel as i, MATMUL_BIAS_F16C_SPEC as l, destroyBuffers as m, MoonshineEncoderExecutor as n, remapPrunedToken as o, createUniformBuffer as p, loadKaniTTS as r, Executor as s, MoonshineSTT as t, clearPipelineCache as u };
12355
- //# sourceMappingURL=moonshine-stt-DDUm5WfP.mjs.map
12404
+ export { initGPU as _, loadMoonshine as a, Executor as c, clearPipelineCache as d, createBindGroup as f, getOrCreatePipeline as g, destroyBuffers as h, loadModel as i, KERNEL_REGISTRY as l, createUniformBuffer as m, MoonshineEncoderExecutor as n, quantizeKaniBackbone as o, createStorageBuffer as p, loadKaniTTS as r, remapPrunedToken as s, MoonshineSTT as t, MATMUL_BIAS_F16C_SPEC as u, verifyGPU as v };
12405
+ //# sourceMappingURL=moonshine-stt-DSz3VW_0.mjs.map