@fugood/buttress-server 2.23.0-beta.45 → 2.23.0-beta.47

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 (2) hide show
  1. package/lib/index.mjs +12 -12
  2. package/package.json +3 -3
package/lib/index.mjs CHANGED
@@ -1,5 +1,5 @@
1
1
  #!/usr/bin/env node
2
- import{createRequire as s6}from"node:module";var l6=Object.defineProperty;var i6=($,X)=>{for(var Z in X)l6($,Z,{get:X[Z],enumerable:!0,configurable:!0,set:(W)=>X[Z]=()=>W})};var P=($,X)=>()=>($&&(X=$($=0)),X);var R1=s6(import.meta.url);var w1=($,X,Z)=>Math.min(Math.max($,X),Z),F1=($)=>$?40:0,E1=($=0)=>{if(!$)return 0;return w1($/12884901888*20,0,20)},M1=($=0)=>{if(!$)return 0;return w1($/34359738368*10,0,10)},x1=($)=>$?10:0,S1=($="default",X=null)=>{let Z=String($).toLowerCase();if(!Z)return 0;if(Z.includes("cuda"))return 20;if(Z.includes("vulkan"))return 10;if(Z.includes("default"))return X==="darwin"||X==="ios"?15:5;return 0},q0=({platform:$,variant:X,hasGpu:Z,gpuUsableBytes:W=0,cpuUsableBytes:j=0,ok:V=!0}={})=>{if(!V)return 0;let Y=F1(Z)+S1(X,$)+E1(W),J=M1(j),N=x1(V);return Math.min(100,Math.round(Y+J+N))},F4=({platform:$,variant:X,hasGpu:Z,gpuUsableBytes:W=0,cpuUsableBytes:j=0,ok:V=!0}={})=>({gpuPresence:F1(Z),variant:S1(X,$),gpuMemory:E1(W),cpuMemory:M1(j),availability:x1(V)});var D1,u0=0.85,f0=0.5,P1=($)=>{if(!$&&$!==0)return[];if(Array.isArray($))return $.filter((X)=>X!=null);return[$]},n6=($)=>{if(!$)return null;return String($).trim().toLowerCase()||null},r6=({variant:$,preferVariants:X=[],variantPreference:Z=[],defaultVariants:W=D1}={})=>{let j=[];if($)j.push($);j.push(...P1(X)),j.push(...P1(Z)),j.push(...W);let V=j.map(n6).filter(Boolean);return Array.from(new Set(V))},T1=($={})=>{let X=String($.type||$.deviceType||$.kind||"").toLowerCase();if(X.includes("gpu"))return!0;if(X.includes("cuda"))return!0;if(X.includes("metal"))return!0;if(X.includes("vulkan"))return!0;if(X.includes("snapdragon"))return!0;return!1},o6=($)=>{if(!Array.isArray($))return[];return $.map((X)=>({...X}))},a6=($,X)=>{if($==="snapdragon")return X.filter((Z)=>Z.deviceName!=="GPUOpenCL");return X},k1=({platform:$,totalMemoryInBytes:X,variant:Z,devices:W,gpuMemoryFraction:j,cpuMemoryFraction:V,ok:Y,error:J})=>{let N=o6(a6(Z,W)),H=N.some(T1),O=N.filter((R)=>T1(R)&&Number.isFinite(Number(R.maxMemorySize))).reduce((R,U)=>R+U.maxMemorySize,0),q=X,_=H?Math.floor(O*j):0,G=q?Math.floor(q*V):0,w={platform:$,variant:Z,hasGpu:H,gpuUsableBytes:_,cpuUsableBytes:G,ok:Y},z=q0(w),A=Y?F4(w):null;return{platform:$,ok:Y,variant:Z,hasGpu:H,devices:N,gpuTotalBytes:O,gpuUsableBytes:_,cpuTotalBytes:q,cpuUsableBytes:G,score:z,breakdown:A,error:J,timestamp:new Date().toISOString()}},E4=({device:$,modelBytes:X=0,kvCacheBytes:Z=0}={})=>{if(!$)return{totalRequiredBytes:X+Z,fitsInGpu:!1,fitsInCpu:!1,limiting:"unknown-device"};let W=Math.max(0,Number(X)||0)+Math.max(0,Number(Z)||0),j=$.hasGpu&&W>0&&W<=$.gpuUsableBytes,V=W>0&&W<=$.cpuUsableBytes,Y="ok";if(!j&&$.hasGpu)Y="gpu-memory";if(!V)Y=j?"cpu-memory":"insufficient-memory";return{totalRequiredBytes:W,fitsInGpu:j,fitsInCpu:V,limiting:Y}},E0=async({platform:$,variant:X=null,preferVariants:Z=[],variantPreference:W=[],gpuMemoryFraction:j=u0,cpuMemoryFraction:V=f0,includeBreakdown:Y=!1,totalMemoryInBytes:J,modelBytes:N=null,kvCacheBytes:H=null,limitedKvCacheBytes:O=null,dependencies:q={},defaultVariants:_=D1}={})=>{let{getBackendDevicesInfo:G,isLibVariantAvailable:w}=q;if(typeof G!=="function"||typeof w!=="function")throw TypeError("GGML capability detection requires getBackendDevicesInfo and isLibVariantAvailable functions");let z=r6({variant:X,preferVariants:Z,variantPreference:W,defaultVariants:_}),A=[];for(let B of z)try{if(!await w(B))throw Error(`Variant ${B} not available on this platform`);let F=await G(B);A.push(k1({platform:$,totalMemoryInBytes:J,variant:B,devices:F,gpuMemoryFraction:j,cpuMemoryFraction:V,ok:!0}))}catch(L){let F=L instanceof Error?L.message:String(L);A.push(k1({platform:$,totalMemoryInBytes:J,variant:B,devices:[],gpuMemoryFraction:j,cpuMemoryFraction:V,ok:!1,error:F}))}let U=A.filter((B)=>B.ok)[0]||null,Q={ok:Boolean(U),selected:U?{...U,breakdown:Y?U.breakdown:void 0}:null,attempts:A};if(!Y&&Q.selected)delete Q.selected.breakdown;if(!Q||!N&&!H)return Q;let K=(B)=>{if(!B)return B;let L=E4({device:B,modelBytes:N||0,kvCacheBytes:H||0}),F=null;if(O!=null&&O!==H)F=E4({device:B,modelBytes:N||0,kvCacheBytes:O});return{...B,fit:L,...F&&{limitedFit:F}}};return Q.selected=K(Q.selected),Q.attempts=Array.isArray(Q.attempts)?Q.attempts.map(K):Q.attempts,Q},g0="ggml-llm";var p0=P(()=>{D1=["cuda","vulkan","snapdragon","default"]});var M4="ggml-stt",b1,x4=async({platform:$,variant:X=null,preferVariants:Z=[],variantPreference:W=[],gpuMemoryFraction:j=u0,cpuMemoryFraction:V=f0,includeBreakdown:Y=!1,totalMemoryInBytes:J,modelBytes:N=null,processingBytes:H=null,kvCacheBytes:O=null,dependencies:q={}}={})=>{let _=W&&W.length>0?W:b1;return E0({platform:$,variant:X,preferVariants:Z,variantPreference:_,gpuMemoryFraction:j,cpuMemoryFraction:V,includeBreakdown:Y,totalMemoryInBytes:J,modelBytes:N,kvCacheBytes:H??O,dependencies:q,defaultVariants:b1})};var S4=P(()=>{p0();b1=["cuda","vulkan","default"]});var t6,_0=async({platform:$,totalMemoryInBytes:X,backend:Z=g0,dependencies:W,...j}={})=>{let V=t6.get(Z);if(!V)throw Error(`No capability detector registered for backend "${Z}"`);return await V({...j,dependencies:W,totalMemoryInBytes:X,platform:$})};var v1=P(()=>{p0();S4();t6=new Map([[g0,E0],[M4,x4]])});var h1,P4=($)=>{let X=$?String($).toLowerCase():"f16";return h1[X]||h1.f16},T4=($,X,Z,W,j,V={},{totalLayers:Y=null,swaLayers:J=0,swaContext:N=null,swaContextMultiplier:H=1,swaAdditionalTokens:O=0,swaFull:q=!1}={})=>{if(!$||!X||!Z||!W||!j)return 0;let _=Y!=null&&Y!==void 0?Number(Y):Number($),G=Math.max(0,Math.floor(_));if(!G)return 0;let w=P4(V.k),z=P4(V.v),A=Number(Z)*(Number(W)*w+Number(j)*z);if(!A)return 0;let R=Math.max(0,Number(X)||0),U=Math.min(G,Math.max(0,Math.floor(Number(J)||0))),Q=Math.max(0,G-U),K=N!=null&&Number.isFinite(Number(N))?Math.max(0,Number(N)):R,B=Math.max(1,Number(H)||1),L=Math.max(0,Number(O)||0),F=K*B+L,E=q?R:Math.min(R,F),M=Q*R+U*Math.max(0,Math.floor(E));return Math.round(A*M)},m0=({modelBytes:$=0,audioLengthSeconds:X=30,sampleRate:Z=16000,bytesPerSample:W=4}={})=>{let j=Math.max(0,Number($)||0),V=Math.max(0,Math.floor(Math.max(0,X)*Z*W)),Y=1048576,J=1073741824,N;if(j<209715200)N=125829120;else if(j<524288000)N=146800640;else if(j<2147483648)N=157286400;else N=167772160;let H;if(j<209715200)H=73400320;else if(j<524288000)H=141557760;else if(j<2147483648)H=230686720;else H=230686720;let O;if(j<104857600)O=20971520;else if(j<209715200)O=31457280;else if(j<524288000)O=89128960;else if(j<2147483648)O=225443840;else O=377487360;let q=N+H+O,_=j+q+V;return{modelBytes:j,audioBufferBytes:V,processingBufferBytes:q,totalBytes:_}};var k4=P(()=>{h1={f16:2,f32:4,q8_0:1,q6_k:0.75,q5_k:0.625,q5_k_m:0.625,q5_k_s:0.625,q5_1:0.625,q5_0:0.625,q4_k:0.5,q4_k_m:0.5,q4_k_s:0.5,q4_1:0.5,q4_0:0.5,iq4_nl:0.5}});var D4=($)=>$?String($).trim().toLowerCase():null,e6=($={},X=null)=>{if(!$)return null;let Z=D4(X),W=Z?`${Z}.attention.sliding_window`:null,j=(W&&$[W]!=null?$[W]:null)??$["llama.attention.sliding_window"];if(j==null)return null;let V=Number(j);return Number.isFinite(V)?V:null},C1=($=0,X=0,Z=!1)=>{let W=Math.max(0,Math.floor(Number($)||0)),j=Math.max(0,Math.floor(Number(X)||0));if(!W||j===1)return 0;if(j<=0)return W;let V=Math.max(0,j-1),Y=Math.floor(W/j),J=W%j,N=Z?Math.max(0,J-1):Math.min(J,V);return Y*V+N},I1=({arch:$,nLayer:X=0})=>({arch:D4($),enabled:!1,window:null,pattern:null,denseFirst:!1,type:null,kvLayers:Math.max(0,Math.floor(Number(X)||0)),swaLayers:0}),$3,c0=({arch:$,metadata:X={},nLayer:Z=0}={})=>{let W=D4($||X["general.architecture"]),j=Math.max(0,Math.floor(Number(Z)||0)),V=e6(X,W),Y=W?$3.get(W):null;if(!Y)return I1({arch:W,nLayer:Z});let J=Y({nLayer:j,nSwa:V,metadata:X});if(!J||!J.enabled||!J.window||J.window<=0)return I1({arch:W,nLayer:Z});let N=Math.max(0,Math.floor(Number(J.pattern)||0)),H=J.kvLayers!=null&&Number.isFinite(Number(J.kvLayers))?Number(J.kvLayers):j,O=Math.max(0,Math.floor(H)),q=C1(O,N,Boolean(J.denseFirst));return{arch:W,enabled:q>0,window:J.window,pattern:N,denseFirst:Boolean(J.denseFirst),type:J.type||"standard",kvLayers:O,swaLayers:q}};var y1=P(()=>{$3=new Map([["llama4",({nSwa:$})=>{if($===0)return{enabled:!1};return{enabled:!0,window:$&&$>0?$:8192,pattern:4,type:"chunked"}}],["afmoe",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["phi3",()=>({enabled:!1})],["gemma2",({nSwa:$})=>{let X=$&&$>0?$:4096;if(!X)return{enabled:!1};return{enabled:!0,window:X,pattern:2,type:"standard"}}],["gemma3",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:6,type:"standard"}}],["gemma3n",({nLayer:$,nSwa:X})=>{if(!X||X<=0)return{enabled:!1};return{enabled:!0,window:X,pattern:5,type:"standard",kvLayers:Math.min(20,$)}}],["gemma-embedding",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:6,type:"symmetric"}}],["cohere2",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["olmo2",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["exaone4",({nLayer:$,nSwa:X})=>{let Z=$>=64,W=null;if(X&&X>0)W=X;else if(Z)W=4096;if(!W)return{enabled:!1};return{enabled:!0,window:W,pattern:4,type:"standard"}}],["gpt-oss",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:2,type:"standard"}}],["smallthinker",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:4096,pattern:4,denseFirst:!0,type:"standard"}}]])});var X3,Z3,u1=($)=>$?String($).trim().toLowerCase():null,f1=($)=>{let X=u1($);return X?X3.has(X):!1},g1=($)=>{let X=u1($);return X?Z3.has(X):!1},b4=($)=>{if(f1($))return"recurrent";if(g1($))return"hybrid";return"transformer"},d0=($={})=>{let X=$["general.architecture"],Z=(F,E=null)=>{let M=$[F],x=Number(M);return Number.isFinite(x)?x:E},W=(F,E=null)=>{let M=$[F];if(Array.isArray(M))return M;let x=Number(M);return Number.isFinite(x)?x:E},j=X?Z(`${X}.context_length`,Z("llama.context_length")):null,V=X?Z(`${X}.block_count`,Z("llama.block_count")):null,Y=X?Z(`${X}.embedding_length`,Z("llama.embedding_length")):null,J=X?Z(`${X}.attention.head_count`,Z("llama.attention.head_count")):null,N=X?W(`${X}.attention.head_count_kv`,W("llama.attention.head_count_kv",J)):null,H=null,O=null;if(Array.isArray(N)){let F=N.filter((E)=>Number(E)>0);if(F.length>0)H=Math.max(...F.map(Number)),O=F.length;else H=0,O=0}else H=N;let q=X?Z(`${X}.attention.key_length`,Z("llama.attention.key_length")):null,_=X?Z(`${X}.attention.value_length`,Z("llama.attention.value_length")):null,G=$["general.quantization_version"]||null,w=$["general.file_type"]||null,z=X?Z(`${X}.ssm.conv_kernel`):null,A=X?Z(`${X}.ssm.state_size`):null,R=X?Z(`${X}.ssm.inner_size`):null,U=X?Z(`${X}.ssm.group_count`):null,Q=X?Z(`${X}.ssm.time_step_rank`):null,K=X?Z(`${X}.rwkv.head_size`):null,B=X?Z(`${X}.rwkv.token_shift_count`,2):null,L=O!=null&&V!=null?V-O:null;return{arch:X,nCtxTrain:j,nLayer:V,nEmbd:Y,nHead:J,nHeadKv:H,nEmbdHeadK:q,nEmbdHeadV:_,quantVersion:G,fileType:w,attentionLayerCount:O,recurrentLayerCount:L,ssmDConv:z,ssmDState:A,ssmDInner:R,ssmNGroup:U,ssmDtRank:Q,rwkvHeadSize:K,rwkvTokenShiftCount:B}},G0=({layerCount:$,headKvCount:X,embdHeadKCount:Z,embdHeadVCount:W,cacheTypes:j,swaConfig:V,kvUnified:Y=!1,nParallel:J=1,swaFull:N=!1,arch:H=null,attentionLayerCount:O=null})=>{let q=b4(H);if(q==="recurrent")return()=>0;let _=q==="hybrid"&&O!=null?Math.max(0,Math.floor(Number(O)||0)):$,G=V?.window&&Y?Math.max(1,Number(J)||1):1,w=Y?1:Math.max(1,Number(J)||1);return(z)=>T4(_,z,X,Z,W,j,{totalLayers:_,swaLayers:V?.swaLayers||0,swaContext:V?.window,swaFull:N,swaContextMultiplier:G})*w},l0=({nLayer:$,nEmbd:X,recurrentLayerCount:Z=null,nSeqMax:W=1,ssmDConv:j=null,ssmDState:V=null,ssmDInner:Y=null,ssmNGroup:J=null,ssmDtRank:N=null,rwkvHeadSize:H=null,rwkvTokenShiftCount:O=2,arch:q=null})=>{if(b4(q)==="transformer")return 0;let G=Z!=null?Math.max(0,Math.floor(Number(Z)||0)):Math.max(0,Math.floor(Number($)||0));if(G===0)return 0;let w=Math.max(1,Math.floor(Number(W)||1)),z=0,A=0;if(H!=null&&H>0&&X!=null&&X>0)z=Math.max(1,Number(O)||2)*X,A=X*H;else if(V!=null&&Y!=null){let Q=Math.max(0,Number(j)||0),K=Math.max(0,Number(V)||0),B=Math.max(0,Number(Y)||0),L=Math.max(1,Number(J)||1);if(Math.max(0,Number(N)||0)>0)z=Q>0?(Q-1)*2*L*K:0,A=Math.floor(K*B/2);else z=Q>0?(Q-1)*(B+2*L*K):0,A=K*B}else return 0;let R=4,U=(z+A)*w*G*R;return Math.max(0,U)},i0=({maxCtx:$,availableMemory:X,modelBytes:Z,kvBytesForCtx:W})=>{let j=Math.max(1,Math.floor(Number($)||0));if(!W||X<=Z)return j;let V=1,Y=j,J=j;while(V<=Y){let N=Math.floor((V+Y)/2);if(Z+W(N)<=X)J=N,V=N+1;else Y=N-1}return J};var p1=P(()=>{k4();X3=new Set(["mamba","mamba2","rwkv6","rwkv6qwen2","rwkv7","arwkv7"]),Z3=new Set(["jamba","falcon-h1","plamo2","granitehybrid","lfm2","lfm2moe","nemotron_h","nemotron_h_moe","qwen3next"])});var M0=P(()=>{v1();k4();p0();S4();y1();p1()});import{EventEmitter as W3}from"node:events";class h4{constructor($=j3){this.maxEntries=$,this.modelLoads=[],this.completions=[],this.transcriptions=[]}addModelLoad($){v4(this.modelLoads,$,this.maxEntries),e.emit("status:modelLoad",$),e.emit("status:change",{type:"modelLoad",entry:$})}addCompletion($){v4(this.completions,$,this.maxEntries),e.emit("status:completion",$),e.emit("status:change",{type:"completion",entry:$})}addTranscription($){v4(this.transcriptions,$,this.maxEntries),e.emit("status:transcription",$),e.emit("status:change",{type:"transcription",entry:$})}getModelLoadHistory(){return[...this.modelLoads].reverse()}getCompletionHistory(){return[...this.completions].reverse()}getTranscriptionHistory(){return[...this.transcriptions].reverse()}clear(){this.modelLoads=[],this.completions=[],this.transcriptions=[]}}function I4($){let X=(Z)=>$(Z);return e.on("status:change",X),()=>e.off("status:change",X)}function c1($){m1+=1;let X=m1,Z=I4($);return{subscriberId:X,unsubscribe:Z}}function C4($){let X=[];return{generators:Array.from($.entries()).filter(([,W])=>W.type==="ggml-llm").map(([W,j])=>{let{instance:V}=j,Y=[];if(V.contexts)Y=Array.from(V.contexts.entries()).map(([J,N])=>{let H={key:J,refCount:N.refCount,hasModel:Boolean(N.context)},O=N.context.parallel.getStatus();return H.parallelStatus=O,X.push({generatorId:W,contextKey:J,...O}),H});return{id:W,type:j.type,refCount:j.refCount,repoId:V.info?.model?.repoId||null,quantization:V.info?.model?.quantization||null,variant:V.info?.runtime?.variant||null,nCtx:V.info?.runtime?.n_ctx||null,nParallel:V.info?.runtime?.n_parallel||null,contexts:Y}}),parallelStatuses:X,history:{modelLoads:$0.getModelLoadHistory(),completions:$0.getCompletionHistory()}}}function y4($){return{generators:Array.from($.entries()).filter(([,Z])=>Z.type==="ggml-stt").map(([Z,W])=>{let{instance:j}=W,V=j.getStatus?.()||{},Y=V.queueStatus||{processing:!1,queuedCount:0};return{id:Z,type:W.type,refCount:W.refCount,repoId:j.info?.model?.repoId||null,quantization:j.info?.model?.quantization||null,modelType:j.info?.model?.modelType||null,variant:j.info?.runtime?.variant||null,hasContext:V.hasContext||!1,contextRefCount:V.contextRefCount||0,queueStatus:Y}}),history:{modelLoads:N0.getModelLoadHistory(),transcriptions:N0.getTranscriptionHistory()}}}function d1($){return{timestamp:new Date().toISOString(),ggmlLlm:C4($),ggmlStt:y4($)}}var j3=9999,e,v4=($,X,Z)=>{if($.push({...X,timestamp:X.timestamp||new Date().toISOString()}),$.length>Z)$.shift()},$0,N0,m1=0;var s0=P(()=>{e=new W3;e.setMaxListeners(100);$0=new h4,N0=new h4});import y from"node:path";import T0 from"node:os";import{stat as B0,mkdir as Y3,open as Q3,unlink as o,readFile as t1,writeFile as e1,rename as f4,readdir as J3}from"node:fs/promises";import{createHash as p4}from"node:crypto";import{gguf as V3}from"@huggingface/gguf";import{loadModel as N3,getBackendDevicesInfo as $2,isLibVariantAvailable as X2}from"@fugood/llama.node";import U3 from"bytes";import*as G3 from"node:stream/web";class L0{constructor($,X){this.config=$,this.plan=X,this.baseDir=$.runtime.cache_dir,this.enabled=$.runtime.session_cache?.enabled!==!1,this.maxSizeBytes=x3($.runtime.session_cache?.max_size_bytes,10737418240),this.maxEntries=$.runtime.session_cache?.max_entries||1000,this.metadata={variant:X.info?.runtime?.variant||null,n_gpu_layers:X.info?.runtime?.n_gpu_layers||0,n_ctx:X.info?.runtime?.n_ctx||0,modelPath:X.localPath,cacheTypeK:X.info?.runtime?.cache_type_k||"f16",cacheTypeV:X.info?.runtime?.cache_type_v||"f16",kvUnified:X.info?.runtime?.kv_unified??null,swaFull:X.info?.runtime?.swa_full??null,flashAttnType:X.info?.runtime?.flash_attn_type||"off",isRecurrent:!1,isHybrid:!1},this.cacheMap=null,this.initialized=!1}updateModelInfo($){if(!$)return;if(this.metadata.isRecurrent=Boolean($.is_recurrent),this.metadata.isHybrid=Boolean($.is_hybrid),this.metadata.isRecurrent||this.metadata.isHybrid)console.log(`[SessionCache] Model architecture: recurrent=${this.metadata.isRecurrent}, hybrid=${this.metadata.isHybrid}`)}requiresExactMatch(){return this.metadata.isRecurrent||this.metadata.isHybrid}static checkTokenPrefixMatch($,X){if($.length>X.length)return!1;for(let Z=0;Z<$.length;Z+=1)if($[Z]!==X[Z])return!1;return!0}static async tokenizeToArray($,X){let Z=await $.tokenize(X);return Array.from(Z?.tokens||[])}async findFormattedMatchForRecurrent($,X,Z){let W=await L0.tokenizeToArray(Z,X),j=$.map(async(_)=>{try{let G=await L0.tokenizeToArray(Z,_.fullText);if(L0.checkTokenPrefixMatch(G,W))return{entry:_,usePromptState:!1,tokenCount:G.length};if(_.promptStatePath&&_.promptText){let w=await L0.tokenizeToArray(Z,_.promptText);if(L0.checkTokenPrefixMatch(w,W))return{entry:_,usePromptState:!0,tokenCount:w.length}}return null}catch(G){return console.warn(`[SessionCache] Failed to check entry ${_.id}: ${G.message}`),null}}),Y=(await Promise.all(j)).find((_)=>_!==null);if(!Y)return console.log("[SessionCache] No token prefix match found for recurrent/hybrid model"),null;let{entry:J,usePromptState:N,tokenCount:H}=Y;console.log(`[SessionCache] Token prefix match: ${J.id} (${H} tokens, usePromptState=${N})`);let O=N?J.promptStatePath:J.stateFilePath;if(!await n1(O))return await this.removeStaleEntry(J),null;return J.lastAccessedAt=new Date().toISOString(),await x0(this.cacheMap,this.baseDir).catch(()=>{}),{entry:J,usePromptState:N}}async initialize(){if(!this.enabled||this.initialized)return;try{await H0($4(this.baseDir)),await H0(o0(this.baseDir)),await H0(U2(this.baseDir)),this.cacheMap=await A3(this.baseDir),this.initialized=!0,console.log(`[SessionCache] Initialized with ${Object.keys(this.cacheMap.entries).length} entries`)}catch($){console.warn(`[SessionCache] Failed to initialize: ${$.message}`),this.enabled=!1}}async removeStaleEntry($){if(console.log(`[SessionCache] Removing stale entry: ${$.id}`),$.stateFilePath)await o($.stateFilePath).catch(()=>{});if($.promptStatePath)await o($.promptStatePath).catch(()=>{});delete this.cacheMap.entries[$.id],this.cacheMap.totalSize-=($.stateFileSize||0)+($.promptStateSize||0),await x0(this.cacheMap,this.baseDir).catch(()=>{})}async findMatchingEntry($,X=null){if(!this.enabled||!this.cacheMap)return null;let Z=this.requiresExactMatch();if(Z&&X){let J=Object.values(this.cacheMap.entries).filter((N)=>m4(N.metadata,this.metadata)&&N.fullText);return this.findFormattedMatchForRecurrent(J,$,X)}let W=w3($,this.metadata,this.cacheMap,Z);if(!W)return null;let{entry:j}=W;if(!await n1(j.stateFilePath))return await this.removeStaleEntry(j),null;return j.lastAccessedAt=new Date().toISOString(),await x0(this.cacheMap,this.baseDir).catch(()=>{}),{entry:j,usePromptState:!1}}async prepareCompletionOptions($,X,Z=null){let W={options:$,cacheEntry:null,promptPrefix:null};if(!this.enabled)return W;let j=await this.findMatchingEntry(X,Z);if(!j)return W;let{entry:V,usePromptState:Y}=j,J=Y?V.promptStatePath:V.stateFilePath,N=Y?V.promptText:V.fullText;return console.log(`[SessionCache] Found matching entry: ${V.id} (${N.length} chars, usePromptState=${Y})`),{options:{...$,load_state_path:J},cacheEntry:V,promptPrefix:N}}async saveCompletionState($,X,Z,W=0,j=null){if(!this.enabled)return null;let V=$+X,Y=z3(V,this.metadata),J=()=>{if(Z)o(Z).catch(()=>{});if(j)o(j).catch(()=>{})};if(this.cacheMap.entries[Y]){console.log(`[SessionCache] Entry already exists for prompt: ${Y}, updating position`);let O=this.cacheMap.entries[Y];return O.lastAccessedAt=new Date().toISOString(),delete this.cacheMap.entries[Y],this.cacheMap.entries[Y]=O,await x0(this.cacheMap,this.baseDir).catch(()=>{}),J(),O}let N=s1(Y,this.baseDir),H=j?s1(`${Y}-prompt`,this.baseDir):null;try{await H0(y.dirname(N)),await f4(Z,N);let O=await B0(N),q=0;if(j&&H)try{await f4(j,H),q=(await B0(H)).size,console.log(`[SessionCache] Saved prompt state: ${H}`)}catch(G){console.warn(`[SessionCache] Failed to save prompt state: ${G.message}`)}let _={id:Y,promptText:$,completionText:X,fullText:V,promptTokenCount:W,stateFilePath:N,stateFileSize:O.size,promptStatePath:H||null,promptStateSize:q,metadata:{...this.metadata},createdAt:new Date().toISOString(),lastAccessedAt:new Date().toISOString()};if(this.cacheMap.entries[Y]=_,this.cacheMap.totalSize+=O.size+q,!this.requiresExactMatch())await E3(this.cacheMap,$,Y,this.metadata);return await F3(this.cacheMap,this.maxSizeBytes,this.maxEntries),await x0(this.cacheMap,this.baseDir),console.log(`[SessionCache] Saved entry: ${Y} (${O.size} bytes, ${V.length} chars)`),_}catch(O){return console.warn(`[SessionCache] Failed to save state: ${O.message}`),J(),null}}async generateTempStatePath(){return await H0(o0(this.baseDir)),B3(this.baseDir)}async cleanup(){await M3(this.baseDir)}}async function K2($,X,Z={}){let{globalDownloadManager:W=null}=Z,j=e0(X),V=await S3(j),Y=new L0(j,V);await Y.initialize();let J={id:$,type:"ggml-llm",config:j,plan:V,info:V.info,contexts:new Map,downloads:new Map,globalDownloadManager:W,sessionCache:Y,finalized:!1},N=async()=>{if(J.finalized)return;J.finalized=!0;let Q=Array.from(J.contexts.values()),K=Q.map((L)=>{if(L.released)return Promise.resolve(!1);if(L.releaseRequested||L.releaseTimer)return Promise.resolve(!1);if(L.refCount=Math.max(0,L.refCount-1),L.refCount>0)return Promise.resolve(!1);return r0(J,L)});if(await Promise.allSettled(K),Q.length===0||Q.every((L)=>L.released))await J.sessionCache.cleanup()},H=async(Q={})=>{let{onProgress:K}=Q,B=await D3(J,K);return J.sessionCache.updateModelInfo(B.modelInfo),{modelInfo:B.modelInfo?{...B.modelInfo}:null,runtime:{...J.plan.info.runtime},download:{...J.plan.info.download}}},O=async()=>{if(J.finalized)return!1;let Q=z0(J),K=J.contexts.get(Q);if(!K)return!1;return b3(J,K,!1)},q=async(Q={})=>{let{options:K={},useCache:B=!0}=Q,L=z0(J),F=J.contexts.get(L);if(!F)throw Error(`Context "${L}" not initialized`);await F.ready;let E=K.prompt||"",M=null,x=null;if(!E&&K.messages){({messages:M}=K),x={chatTemplate:K.chat_template||K.chatTemplate,jinja:K.jinja??!0,tools:K.tools,parallel_tool_calls:K.parallel_tool_calls,tool_choice:K.tool_choice,reasoning_format:K.reasoning_format,enable_thinking:K.enable_thinking,add_generation_prompt:K.add_generation_prompt,now:K.now,chat_template_kwargs:K.chat_template_kwargs};let I=await F.context.getFormattedChat(M,x.chatTemplate,x);E=I?.prompt||I||""}if(B&&J.sessionCache.enabled&&E){let{options:I}=await J.sessionCache.prepareCompletionOptions(K,E,F.context),b=await J.sessionCache.generateTempStatePath(),r=(await F.context.tokenize(E))?.tokens?.length||0,k={...I,save_state_path:b},d=J.sessionCache.requiresExactMatch(),v=Boolean(k.load_state_path),S=null;if(d&&!v)S=await J.sessionCache.generateTempStatePath(),k.save_prompt_state_path=S;let C={repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null};return T3(F.context,k,J.sessionCache,E,b,r,J.id,C,S)}let D={repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null};return P3(F.context,K,J.id,D)},_=async(Q={})=>{let{text:K="",params:B={}}=Q,L=z0(J),F=J.contexts.get(L);if(!F)throw Error(`Context "${L}" not initialized`);await F.ready;let E=await F.context.tokenize(K,B);if(!E)return{tokens:[]};let M=Array.from(E.tokens??[]).map((x)=>Number(x));return{...E,tokens:M}},G=async(Q={})=>{let{tokens:K=[]}=Q,B=z0(J),L=J.contexts.get(B);if(!L)throw Error(`Context "${B}" not initialized`);await L.ready;let F=K.map((E)=>Number(E));return L.context.detokenize(F)},w=async(Q={})=>{let{messages:K=[],template:B,params:L}=Q,F=z0(J),E=J.contexts.get(F);if(!E)throw Error(`Context "${F}" not initialized`);return await E.ready,await E.context.getFormattedChat(K,B,L)},z=()=>Array.from(J.contexts.values()).some((Q)=>!Q.released&&(Q.releaseRequested||Q.releaseTimer||Q.refCount>0)),A=()=>{J.finalized=!1},R=()=>{let Q=[],K=Array.from(J.contexts.entries()).map(([B,L])=>{let F={key:B,refCount:L.refCount,hasModel:Boolean(L.context)},E=L.context.parallel.getStatus();return F.parallelStatus=E,Q.push({contextKey:B,...E}),F});return{id:J.id,type:J.type,repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null,nCtx:J.plan.info.runtime?.n_ctx||null,nParallel:J.plan.info.runtime?.n_parallel||null,contexts:K,parallelStatuses:Q}},U=(Q)=>{let K=Array.from(J.contexts.entries()).map(([B,L])=>L.context.parallel.subscribeToStatus((F)=>{Q({contextKey:B,...F})}));return{remove:()=>{K.forEach((B)=>{if(B?.remove)B.remove()})}}};return{id:$,type:"ggml-llm",info:V.info,contexts:J.contexts,initContext:H,completion:q,tokenize:_,detokenize:G,applyChatTemplate:w,releaseContext:O,finalize:N,getStatus:R,subscribeParallelStatus:U,hasPendingReleases:z,resetFinalized:A}}async function q2($,X,Z={}){let{onProgress:W,onComplete:j,onError:V}=Z;try{let Y=e0($),J=await c4(Y),N=H2(Y,J),{repoId:H}=J;if(await P0(N,J.size)){if(console.log(`[Download] Model already exists: ${H} at ${N}`),typeof j==="function")j({localPath:N,repoId:H,alreadyExists:!0});return{started:!1,localPath:N,repoId:H,alreadyExists:!0}}let q=X.getDownload(N);if(q)return console.log(`[Download] Already downloading: ${H}`),q.then(()=>{if(typeof j==="function")j({localPath:N,repoId:H,joinedExisting:!0})}).catch((G)=>{if(typeof V==="function")V(G)}),{started:!1,localPath:N,repoId:H,alreadyDownloading:!0};console.log(`[Download] Starting download: ${H}`);let _=(async()=>{try{if(J.isSplit&&J.splitCount>0){let G=/-(\d{5})-of-(\d{5})\.gguf$/,w=y.dirname(N),z=J.splitCount,A=0;for(let R=1;R<=z;R+=1){let U=String(R).padStart(5,"0"),Q=J.filename.replace(G,`-${U}-of-${String(z).padStart(5,"0")}.gguf`),K=`${Y.model.base_url.replace(/\/+$/,"")}/${J.repoId}/resolve/${J.revision}/${Q}`,B=y.join(w,Q);if(!await P0(B))await a0(K,J.headers,B,null,(F)=>{if(F>=0&&Number.isFinite(F)){let E=(A+F)/z;if(console.log(`[Download] ${H}: ${Math.round(E*100)}%`),typeof W==="function")W(E)}});A+=1}}else await a0(J.url,J.headers,N,J.size,(G)=>{if(G>=0&&Number.isFinite(G)){if(console.log(`[Download] ${H}: ${Math.round(G*100)}%`),typeof W==="function")W(G)}});if(console.log(`[Download] Completed: ${H}`),typeof j==="function")j({localPath:N,repoId:H})}catch(G){if(console.error(`[Download] Failed: ${H}`,G.message),typeof V==="function")V(G);throw G}finally{X.deleteDownload(N)}})();return X.setDownload(N,_),{started:!0,localPath:N,repoId:H}}catch(Y){if(console.error("[Download] Failed to start download:",Y.message),typeof V==="function")V(Y);return{started:!1,localPath:null,repoId:null,error:Y.message}}}async function v3($){let X=e0($),Z=await c4(X),W=await G2(Z.url,Z.headers,X.runtime.cache_dir),{arch:j,nCtxTrain:V,nLayer:Y,nEmbd:J,nHead:N,nHeadKv:H,nEmbdHeadK:O,nEmbdHeadV:q,quantVersion:_,fileType:G,attentionLayerCount:w,recurrentLayerCount:z,ssmDConv:A,ssmDState:R,ssmDInner:U,ssmNGroup:Q,ssmDtRank:K,rwkvHeadSize:B,rwkvTokenShiftCount:L}=d0(W),F=Number.isFinite(Number(Y))?Number(Y):0,E=Number.isFinite(Number(J))?Number(J):0,M=Number.isFinite(Number(N))?Number(N):0,x=Number.isFinite(Number(H))?Number(H):M,D=M>0&&E>0?E/M:128,I=O!=null&&Number.isFinite(Number(O))?Number(O):D,b=q!=null&&Number.isFinite(Number(q))?Number(q):D,u=c0({arch:j,metadata:W,nLayer:F}),r=u&&Number.isFinite(Number(u.kvLayers))?Number(u.kvLayers):F,k=Math.max(0,Math.floor(Number(r)||0)),v=(X.model.n_ctx?Number(X.model.n_ctx):null)||V||4096,S={k:X.model.cache_type_k,v:X.model.cache_type_v},C=Z.size>0?Z.size:0,h=X.model.n_parallel||4,p=G0({layerCount:k,headKvCount:x,embdHeadKCount:I,embdHeadVCount:b,cacheTypes:S,swaConfig:u,kvUnified:X.model.kv_unified,nParallel:h,swaFull:X.model.swa_full,arch:j,attentionLayerCount:w}),f=l0({nLayer:F,nEmbd:E,recurrentLayerCount:z,nSeqMax:h,ssmDConv:A,ssmDState:R,ssmDInner:U,ssmNGroup:Q,ssmDtRank:K,rwkvHeadSize:B,rwkvTokenShiftCount:L,arch:j}),s=X.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(X.backend.gpu_memory_fraction))):S0.backend.gpu_memory_fraction||1,W0=X.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(X.backend.cpu_memory_fraction))):t0,A0=p(v),l=await O2(X,{modelBytes:C,kvCacheBytes:A0}),F0=(l.selected.totalMemory||0)*s,A4=Math.max(0,T0.totalmem()*W0),h0=l.selected.hasGpu?F0:A4,I0=i0({maxCtx:v,availableMemory:h0,modelBytes:C,kvBytesForCtx:p}),U0=p(v),C0=p(I0);return{kvInfo:{nCtxTrain:V,nLayer:F,nEmbd:E,nHeadKv:x,nEmbdHeadK:I,nEmbdHeadV:b,nHeadCount:M,nHeadKvCount:x,kvLayerCount:k,swa:u?.enabled?{window:u.window,pattern:u.pattern,denseFirst:u.denseFirst,type:u.type,layers:u.swaLayers}:null},modelBytes:C,kvCacheBytes:U0,limitedKvCacheBytes:C0,memoryLimitedCtx:I0,recurrentMemoryBytes:f,quantization:{name:Z.quantization||null,fileType:G,version:_}}}async function _2($=null,X={}){let{threshold:Z=1.1,includeBreakdown:W=!1,config:j,...V}=X,Y=null,J=null,N=null,H=null,O=null,q=null,_=null;if(j)try{let{modelBytes:B,kvCacheBytes:L,limitedKvCacheBytes:F,memoryLimitedCtx:E,recurrentMemoryBytes:M,kvInfo:x,quantization:D}=await v3(j);Y=B,J=L,N=F,H=E,O=M,q=x,_=D}catch(B){}let G=j?.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.gpu_memory_fraction))):void 0,w=j?.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.cpu_memory_fraction))):void 0,z=await _0({...V,platform:process.platform,totalMemoryInBytes:T0.totalmem(),backend:"ggml-llm",includeBreakdown:W,gpuMemoryFraction:G,cpuMemoryFraction:w,dependencies:{getBackendDevicesInfo:$2,isLibVariantAvailable:X2},modelBytes:Y,kvCacheBytes:J,limitedKvCacheBytes:N}),A=z.selected,R=a1(A);A.modelBytes=Y||null,A.kvCacheBytes=J||null,A.memoryLimitedCtx=H||null,A.limitedKvCacheBytes=N||null,A.recurrentMemoryBytes=O||null,A.kvInfo=q||null,A.quantization=_||null;let U=null,Q=null;if($){let B=a1($);Q={...$,score:B};let L="buttress",F="buttress-higher-score";if(!z.ok)L="local",F="buttress-unavailable";else if(!B&&B!==0)L="buttress",F="missing-client-score";else{let{fit:E,limitedFit:M}=Q,x=A?.fit,D=A?.limitedFit,I=E?.fitsInGpu||E?.fitsInCpu||M?.fitsInGpu||M?.fitsInCpu,b=x?.fitsInGpu||x?.fitsInCpu||D?.fitsInGpu||D?.fitsInCpu;if(I&&!b)L="local",F="client-fits-in-memory";else if(b&&!I)L="buttress",F="buttress-fits-in-memory";else if(B>R*Z)L="local",F="client-better";else if(R>B*Z)L="buttress",F="buttress-better";else L="either",F="comparable-scores"}U={buttressScore:R,clientScore:B,threshold:Z,recommendation:L,reason:F}}if(!z.ok&&!U)U={buttressScore:R,clientScore:$?.score??null,threshold:Z,recommendation:"local",reason:"buttress-unavailable"};let K=null;if(j)K={repoId:j.model?.repo_id||null,quantization:j.model?.quantization||null,nCtx:j.model?.n_ctx||null,cacheKType:j.model?.cache_type_k||"f16",cacheVType:j.model?.cache_type_v||"f16"};return{type:"ggml-llm",timestamp:new Date().toISOString(),buttress:z,client:Q,comparison:U,modelConfig:K}}var H3=()=>{if(typeof globalThis<"u"&&globalThis.ReadableStream&&globalThis.WritableStream)return{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream};return G3},O3,Z2,K3,W2=($={},X={})=>{return Object.entries(X||{}).forEach(([Z,W])=>{if(W&&typeof W==="object"&&!Array.isArray(W)){if(!$[Z]||typeof $[Z]!=="object")$[Z]={};W2($[Z],W)}else $[Z]=W}),$},q3=".gguf",j2="https://huggingface.co",Y2="https://huggingface.co/api",n,Q2,t0=0.5,S0,u4=($,X=[])=>{if(!$&&$!==0)return[...X];if(Array.isArray($))return $.filter((Z)=>Z!=null);return[$]},n0=($)=>{if(!$)return null;let X=String($).toLowerCase();if(["cuda","vulkan","snapdragon","default"].includes(X))return X;return null},e0=($={})=>{let X=JSON.parse(JSON.stringify(S0));if(W2(X,$),X.backend.variant=n0(X.backend.variant),X.backend.variant_preference=Array.from(new Set(u4(X.backend.variant_preference).map(n0).filter(Boolean))),X.backend.variant_preference.length===0)X.backend.variant_preference=["cuda","vulkan","snapdragon","default"];if(X.runtime.prefer_variants=Array.from(new Set(u4(X.runtime.prefer_variants).map(n0).filter(Boolean))),X.model.preferred_quantizations=Array.from(new Set(u4(X.model.preferred_quantizations||X.model.quantizations).map((Z)=>Z?String(Z).toLowerCase():null).filter(Boolean))),X.model.quantization){let Z=String(X.model.quantization).toLowerCase();if(!X.model.preferred_quantizations.includes(Z))X.model.preferred_quantizations.unshift(Z)}return X.model.n_parallel=Math.max(1,Number(X.model.n_parallel)||4),X.model.n_batch=Math.max(1,Number(X.model.n_batch)||512),X.model.base_url=X.model.base_url||j2,X.model.api_base=X.model.api_base||Y2,X.runtime.cache_dir=X.runtime.cache_dir?y.resolve(X.runtime.cache_dir):n,X.runtime.session_cache={...S0.runtime.session_cache,...X.runtime.session_cache||{}},X.runtime.context_release_delay_ms=Math.max(0,Number(X.runtime.context_release_delay_ms)||S0.runtime.context_release_delay_ms),X},l1=($)=>{let X=$.toLowerCase();return Q2.find((W)=>X.includes(W))||null},_3=($)=>{let X=[];if($.backend.variant)X.push($.backend.variant);if($.runtime.prefer_variants.length>0)X.push(...$.runtime.prefer_variants);return X.push(...$.backend.variant_preference),X.push("default"),Array.from(new Set(X.map(n0).filter(Boolean)))},H0=async($)=>{await Y3($,{recursive:!0})},L3=($=n)=>y.join($,".metadata-cache"),J2=($,X,Z=n)=>{let W=p4("sha256").update($).digest("hex");return y.join(L3(Z),X,`${W}.json`)},V2=async($,X,Z=n)=>{try{let W=J2($,X,Z),j=await t1(W,"utf-8");return console.log(`[Cache] Hit ${X} cache:`,y.basename(W)),JSON.parse(j,(V,Y)=>{if(typeof Y==="string"&&Y.startsWith("__bigint__"))return BigInt(Y.slice(10));return Y})}catch(W){return null}},g4=async($,X,Z,W=n)=>{try{let j=J2($,X,W);await H0(y.dirname(j)),await e1(j,JSON.stringify(Z,(V,Y)=>{if(typeof Y==="bigint")return`__bigint__${Y.toString()}`;return Y}),"utf-8"),console.log(`[Cache] Wrote ${X} cache:`,y.basename(j))}catch(j){console.warn(`[Cache] Failed to write ${X} cache:`,j.message)}},$4=($=n)=>y.join($,".session-state-cache"),N2=($=n)=>y.join($4($),"cache-map.json"),o0=($=n)=>y.join($4($),"temp"),U2=($=n)=>y.join($4($),"states"),i1=()=>({version:1,entries:{},totalSize:0}),A3=async($=n)=>{try{let X=N2($),Z=await t1(X,"utf-8"),W=JSON.parse(Z);if(!W.entries||typeof W.entries!=="object")return i1();return W}catch{return i1()}},x0=async($,X=n)=>{let Z=N2(X),W=`${Z}.tmp.${Date.now()}`;try{await H0(y.dirname(Z)),await e1(W,JSON.stringify($,null,2),"utf-8"),await f4(W,Z)}catch(j){throw await o(W).catch(()=>{}),j}},z3=($,X)=>{let Z=JSON.stringify({text:$,model:X.modelPath,variant:X.variant,n_gpu_layers:X.n_gpu_layers,n_ctx:X.n_ctx,cacheTypeK:X.cacheTypeK,cacheTypeV:X.cacheTypeV,kvUnified:X.kvUnified,swaFull:X.swaFull,flashAttnType:X.flashAttnType});return p4("sha256").update(Z).digest("hex").slice(0,24)},s1=($,X=n)=>y.join(U2(X),`${$}.bin`),B3=($=n)=>{let X=`${Date.now()}-${Math.random().toString(36).slice(2,10)}`;return y.join(o0($),`${X}.bin`)},m4=($,X)=>$.modelPath===X.modelPath&&$.variant===X.variant&&$.n_gpu_layers===X.n_gpu_layers&&$.n_ctx>=X.n_ctx&&$.cacheTypeK===X.cacheTypeK&&$.cacheTypeV===X.cacheTypeV&&$.kvUnified===X.kvUnified&&$.swaFull===X.swaFull&&$.flashAttnType===X.flashAttnType&&Boolean($.isRecurrent)===Boolean(X.isRecurrent)&&Boolean($.isHybrid)===Boolean(X.isHybrid),R3=($,X)=>{let Z=Math.min($.length,X.length),W=0;while(W<Z&&$[W]===X[W])W+=1;return W},w3=($,X,Z,W=!1)=>{let j=Object.values(Z.entries);console.log(`[SessionCache] Finding match for promptText (${$.length} chars), exactMatch=${W}`),console.log(`[SessionCache] Checking ${j.length} cache entries`);let V=j.filter((J)=>m4(J.metadata,X));if(W){let J=V.find((N)=>N.fullText===$);if(J)return console.log(`[SessionCache] Exact match found: ${J.id} (${J.fullText.length} chars)`),{entry:J,prefixLength:J.fullText.length,exactMatch:!0};return null}let Y=V.reduce((J,N)=>{let H=R3($,N.fullText);if(H>J.prefixLen)return{entry:N,prefixLen:H};if(H===J.prefixLen&&N.fullText.length>(J.entry?.fullText?.length||0))return{entry:N,prefixLen:H};return J},{entry:null,prefixLen:0});if(Y.entry)return console.log(`[SessionCache] Prefix match found: ${Y.entry.id} (${Y.prefixLen}/${Y.entry.fullText.length} chars)`),{entry:Y.entry,prefixLength:Y.prefixLen};return console.log("[SessionCache] No match found"),null},F3=async($,X,Z)=>{let W=Object.values($.entries).sort((J,N)=>new Date(J.lastAccessedAt)-new Date(N.lastAccessedAt)),j=$.totalSize,V=Object.keys($.entries).length,Y=W.filter((J)=>{let N=j>X,H=V>Z;if(!N&&!H)return!1;return j-=(J.stateFileSize||0)+(J.promptStateSize||0),V-=1,!0});return await Promise.all(Y.map(async(J)=>{if(await o(J.stateFilePath).catch(()=>{}),J.promptStatePath)await o(J.promptStatePath).catch(()=>{});delete $.entries[J.id],console.log(`[SessionCache] Evicted entry: ${J.id}`)})),$.totalSize=Math.max(0,j),Y.map((J)=>J.id)},E3=async($,X,Z,W)=>{let j=Object.entries($.entries).filter(([V,Y])=>{if(V===Z)return!1;if(!m4(Y.metadata,W))return!1;return X.startsWith(Y.fullText)&&Y.fullText.length<X.length}).map(([,V])=>V);return await Promise.all(j.map(async(V)=>{if(await o(V.stateFilePath).catch(()=>{}),V.promptStatePath)await o(V.promptStatePath).catch(()=>{});$.totalSize-=(V.stateFileSize||0)+(V.promptStateSize||0),delete $.entries[V.id],console.log(`[SessionCache] Evicted superseded prefix entry: ${V.id} (${V.promptText.length} prompt chars)`)})),j.map((V)=>V.id)},M3=async($=n)=>{let X=o0($);try{let Z=await J3(X),W=Date.now(),j=3600000;await Promise.all(Z.map(async(V)=>{let Y=y.join(X,V),J=await B0(Y).catch(()=>null);if(J&&W-J.mtimeMs>3600000)await o(Y).catch(()=>{}),console.log(`[SessionCache] Cleaned up temp file: ${V}`)}))}catch{}},n1=async($)=>{try{return await B0($),!0}catch{return!1}},x3=($,X)=>{if($==null)return X;if(typeof $==="number")return $;if(typeof $==="string"){let Z=U3.parse($);return Z!=null?Z:X}return X},r1=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,X);if(!Z.ok){let W=await Z.text().catch(()=>"");throw Error(`Failed to fetch ${$}: ${Z.status} ${Z.statusText} ${W}`.trim())}return Z.json()},o1=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,{...X,method:"HEAD"});if(!Z.ok)throw Error(`Failed to fetch headers for ${$}: ${Z.status} ${Z.statusText}`);return Z},G2=async($,X,Z=n)=>{let W=JSON.stringify({url:$,headers:X}),j=await V2(W,"range-metadata",Z);if(j)return j;let V=!/^https?:/i.test($),{metadata:Y}=await V3($,{fetch,additionalFetchHeaders:X,allowLocalFile:V});return await g4(W,"range-metadata",Y,Z),Y},H2=($,X)=>{if($.model.local_path)return y.resolve($.model.local_path);let Z=X.repoId.split("/"),W=y.join($.runtime.cache_dir,...Z,X.revision);return y.join(W,X.filename)},P0=async($,X)=>{try{let Z=await B0($);if(!X)return!0;return Z.size===X}catch(Z){return!1}},a0=async($,X,Z,W,j)=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");await H0(y.dirname(Z));let V=await fetch($,{headers:X});if(!V.ok||!V.body)throw Error(`Failed to download ${$}: ${V.status} ${V.statusText}`);let Y=await Q3(Z,"w"),J=Number(V.headers.get("content-length"))||W||0,N=0,H=0.05;try{await V.body.pipeTo(new K3({async write(O){if(await Y.write(O),N+=O.byteLength,typeof j==="function"&&J>0){let q=Math.min(1,N/J);while(q>=H)j(H),H+=0.05}},async close(){if(await Y.close(),typeof j==="function")j(1)},async abort(O){throw await Y.close().catch(()=>{}),await o(Z).catch(()=>{}),O}}))}catch(O){throw await Y.close().catch(()=>{}),await o(Z).catch(()=>{}),O}if(W){let O=await B0(Z);if(O.size!==W)throw await o(Z).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${W} got ${O.size}`)}},c4=async($)=>{let X=$.model.repo_id||$.model.repository||$.model.model;if(!X)throw Error("`model.repo_id` is required in Buttress backend config");let Z=$.model.revision||"main",W=$.runtime.cache_dir,j=JSON.stringify({repoId:X,revision:Z,filename:$.model.filename,url:$.model.url,quantization:$.model.quantization,preferred_quantizations:$.model.preferred_quantizations}),V=await V2(j,"artifact-info",W);if(V)return V;let Y={...$.runtime.http_headers||{}};if($.runtime.huggingface_token)Y.Authorization=`Bearer ${$.runtime.huggingface_token}`;if($.model.url){let Q=await o1($.model.url,{headers:Y}),K=Number(Q.headers.get("content-length"))||null,B=$.model.filename||$.model.url.split("/").pop(),L={repoId:X,revision:Z,filename:B,url:$.model.url,size:K,headers:Y};return await g4(j,"artifact-info",L,W),L}let{filename:J}=$.model,N=$.model.quantization&&String($.model.quantization).toLowerCase(),H=await r1(`${$.model.api_base}/models/${X}?revision=${Z}&blobs=true`,{headers:Y}),q=(H?.siblings||H?.files||[]).map((Q)=>Q.rfilename||Q.path||Q.filename).filter((Q)=>typeof Q==="string"&&Q.endsWith(q3));if(q.length===0)throw Error(`No GGUF artifacts found in repo ${X}`);let _=$.model.preferred_quantizations.length>0?$.model.preferred_quantizations:Q2,G=()=>{let Q=_.find((K)=>{return q.find((L)=>L.toLowerCase().includes(K))});if(Q)return{filename:q.find((B)=>B.toLowerCase().includes(Q)),quantization:Q};return null};if(!J){let Q=G()||{filename:q[0],quantization:null},{filename:K,quantization:B}=Q;J=K,N=B||l1(J)}else if(!N)N=l1(J);let w=`${$.model.base_url.replace(/\/+$/,"")}/${X}/resolve/${Z}/${J}`,z=/-(\d{5})-of-(\d{5})\.gguf$/,A=J.match(z),R=null;if(A){let[,,Q]=A,K=await r1(`${$.model.api_base}/models/${X}?revision=${Z}&blobs=true`,{headers:Y}),B=K?.siblings||K?.files||[],L=Number(Q);R=0;for(let F=1;F<=L;F+=1){let E=String(F).padStart(5,"0"),M=J.replace(z,`-${E}-of-${Q}.gguf`),x=B.find((I)=>(I.rfilename||I.path||I.filename)===M),D=Number(x?.size);if(Number.isFinite(D)&&D>0)R+=D}}else{let Q=await o1(w,{headers:Y});R=Number(Q.headers.get("content-length"))||null}let U={repoId:X,revision:Z,filename:J,url:w,size:R,quantization:N,headers:Y,isSplit:Boolean(A),splitCount:A?Number(A[2]):0};return await g4(j,"artifact-info",U,W),U},O2=async($,{modelBytes:X=null,kvCacheBytes:Z=null}={})=>{let W=_3($),[j,...V]=W,Y=$.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.gpu_memory_fraction))):S0.backend.gpu_memory_fraction||1,J=$.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):t0,N=await _0({platform:process.platform,totalMemoryInBytes:T0.totalmem(),backend:"ggml-llm",variant:j||null,preferVariants:V,gpuMemoryFraction:Y,cpuMemoryFraction:J,dependencies:{getBackendDevicesInfo:$2,isLibVariantAvailable:X2},modelBytes:X,kvCacheBytes:Z}),H=(_)=>({..._,devices:Array.isArray(_.devices)?_.devices:[],ok:_.ok,hasGpu:Boolean(_.hasGpu),totalMemory:_.gpuTotalBytes||_.totalMemory||0,error:_.ok?null:Error(_.error||`Variant ${_.variant} not available on this platform`)});if(!N.ok||!N.selected){let _=(N.attempts||[]).map((G)=>`${G.variant}: ${G.error||"unknown error"}`).join("; ");throw Error(`Unable to initialize any backend variant (${W.join(", ")}). Errors: ${_}`)}let O=(N.attempts||[]).map(H);return{selected:H(N.selected),attempts:O}},S3=async($)=>{let X=await c4($),Z=await G2(X.url,X.headers,$.runtime.cache_dir),{arch:W,nCtxTrain:j,nLayer:V,nEmbd:Y,nHead:J,nHeadKv:N,nEmbdHeadK:H,nEmbdHeadV:O,quantVersion:q,fileType:_,attentionLayerCount:G,recurrentLayerCount:w,ssmDConv:z,ssmDState:A,ssmDInner:R,ssmNGroup:U,ssmDtRank:Q,rwkvHeadSize:K,rwkvTokenShiftCount:B}=d0(Z),L=Number.isFinite(Number(V))?Number(V):0,F=Number.isFinite(Number(Y))?Number(Y):0,E=Number.isFinite(Number(J))?Number(J):0,M=Number.isFinite(Number(N))?Number(N):E,x=E>0&&F>0?F/E:128,D=H!=null&&Number.isFinite(Number(H))?Number(H):x,I=O!=null&&Number.isFinite(Number(O))?Number(O):x,b=c0({arch:W,metadata:Z,nLayer:L}),u=b&&Number.isFinite(Number(b.kvLayers))?Number(b.kvLayers):L,r=Math.max(0,Math.floor(Number(u)||0)),k={use_mmap:$.model.use_mmap??$.runtime.use_mmap,use_mlock:$.model.use_mlock??$.runtime.use_mlock,n_threads:$.model.n_threads??$.runtime.n_threads,n_ctx:$.model.n_ctx??$.runtime.n_ctx,n_batch:$.model.n_batch??$.runtime.n_batch,n_ubatch:$.model.n_ubatch??$.runtime.n_ubatch,n_cpu_moe:$.model.n_cpu_moe??$.runtime.n_cpu_moe,n_parallel:$.model.n_parallel??$.runtime.n_parallel,cpu_mask:$.model.cpu_mask??$.runtime.cpu_mask,cpu_strict:$.model.cpu_strict??$.runtime.cpu_strict,devices:$.model.devices??$.runtime.devices,n_gpu_layers:$.model.n_gpu_layers??$.runtime.n_gpu_layers,flash_attn_type:$.model.flash_attn_type??$.runtime.flash_attn_type,cache_type_k:$.model.cache_type_k??$.runtime.cache_type_k,cache_type_v:$.model.cache_type_v??$.runtime.cache_type_v,kv_unified:$.model.kv_unified??$.runtime.kv_unified,swa_full:$.model.swa_full??$.runtime.swa_full,ctx_shift:$.model.ctx_shift??$.runtime.ctx_shift},d=k.n_ctx?Number(k.n_ctx):null,v=d||j||4096,S=[],C=[],h=!0;if(d&&j&&d>j){h=!1;let V0=`Requested context length (${d}) exceeds model training context (${j})`;S.push(V0),C.push(V0),v=j}if(d&&!j)S.push("Model metadata missing training context length, using requested value");let p={k:k.cache_type_k,v:k.cache_type_v},f=X.size>0?X.size:0,s=G0({layerCount:r,headKvCount:M,embdHeadKCount:D,embdHeadVCount:I,cacheTypes:p,swaConfig:b,kvUnified:k.kv_unified,nParallel:k.n_parallel,swaFull:k.swa_full,arch:W,attentionLayerCount:G}),W0=l0({nLayer:L,nEmbd:F,recurrentLayerCount:w,nSeqMax:k.n_parallel||4,ssmDConv:z,ssmDState:A,ssmDInner:R,ssmNGroup:U,ssmDtRank:Q,rwkvHeadSize:K,rwkvTokenShiftCount:B,arch:W}),A0=s(v),l=await O2($,{modelBytes:f,kvCacheBytes:A0+W0}),L4=l.selected.totalMemory||0,F0=L4*($.backend.gpu_memory_fraction||1),A4=$.backend.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):t0,h0=Math.max(0,T0.totalmem()*A4),I0=l.selected.hasGpu?F0:h0,U0=i0({maxCtx:v,availableMemory:I0,modelBytes:f,kvBytesForCtx:s});if(!d&&U0){let V0=j?Math.min(U0,j):U0,R4=Math.max(32,V0);if(R4<v)S.push(`Context length capped to ${R4} by memory limits`);v=R4}if(v>U0)v=U0;let C0=Math.floor(U0);console.log(`[buttress] Memory-limited context length: ${C0}`);let A1=s(v),p6=f+A1+W0,z1=L?f/(L+1):f,y0=0;if(l.selected.hasGpu&&z1>0)y0=Math.min(L+1,Math.max(0,Math.floor(F0/z1)));console.log(`[buttress] Auto GPU layer capacity (${l.selected.variant}): ${y0}/${L+1}`);let z4;if(k.n_gpu_layers==="auto"||k.n_gpu_layers==null)z4=y0;else z4=Math.max(0,Math.min(Number(k.n_gpu_layers)||0,L+1));let m6=(()=>{let V0=k.flash_attn_type&&String(k.flash_attn_type).toLowerCase();if(V0==="on"||V0==="off")return V0;if(V0==="auto")return l.selected.hasGpu?"auto":"off";return l.selected.hasGpu?"auto":"off"})(),c6=$.runtime.cache_dir,B4=H2($,X),B1=await P0(B4,X.size),d6={ok:h,backend:"ggml-llm",warnings:S,errors:C,model:{repoId:X.repoId,revision:X.revision,filename:X.filename,quantization:X.quantization,url:X.url,sizeBytes:X.size,metadata:{architecture:W,n_ctx_train:j,n_layer:L,n_embd:F,quantization_version:q,file_type:_,kv_layer_count:r,swa:b?.enabled?{window:b.window,pattern:b.pattern,dense_first:b.denseFirst,type:b.type,layers:b.swaLayers}:null}},runtime:{...k,variant:l.selected.variant,n_ctx:v,requested_ctx:d,n_gpu_layers:z4,auto_gpu_layers:y0,flash_attn_type:m6,cache_type_k:p.k,cache_type_v:p.v,estimated_max_n_ctx:C0},resources:{modelBytes:f,kvCacheBytes:A1,recurrentMemoryBytes:W0,totalEstimatedBytes:p6,gpuCapacityBytes:L4,gpuUsableBytes:F0,cpuUsableBytes:h0,fit:l.selected.fit},devices:{selected:l.selected,attempts:l.attempts},download:{cacheDir:c6,localPath:B4,exists:B1},timestamp:new Date().toISOString()};return{config:$,info:d6,artifact:X,metadata:{arch:W,nCtxTrain:j,nLayer:L,nEmbd:F},devices:l,cacheTypes:p,localPath:B4,localExists:B1}},P3=($,X,Z=null,W=null)=>{let j,V=Date.now(),Y=0;return new Z2({async start(J){try{let N=await $.parallel.completion(X,(G,w)=>{if(!w)return;if(w.token)Y+=1;J.enqueue({event:"token",data:{requestId:G,...w}})}),{requestId:H}=N;j=N.stop;let O=await N.promise;console.log("[Completion] Result:",O),J.enqueue({event:"result",data:{requestId:H,...O}}),J.close();let q=Date.now()-V,_=O.timings||{};$0.addCompletion({id:`completion-${H}`,generatorId:Z,requestId:H,repoId:W?.repoId||null,quantization:W?.quantization||null,variant:W?.variant||null,cacheTokens:_.cache_n??0,promptTokens:_.prompt_n??0,tokensGenerated:_.predicted_n??Y,tokensPerSecond:_.predicted_per_second??0,promptPerSecond:_.prompt_per_second??0,durationMs:q,success:!0,interrupted:O.interrupted||!1,contextFull:O.context_full||O.contextFull||!1})}catch(N){J.enqueue({event:"error",data:{message:N?.message||String(N)}}),J.error(N),$0.addCompletion({id:`completion-${Date.now()}`,generatorId:Z,repoId:W?.repoId||null,quantization:W?.quantization||null,variant:W?.variant||null,durationMs:Date.now()-V,tokensGenerated:Y,success:!1,error:N?.message||String(N)})}},cancel(){if(j)j()}})},T3=($,X,Z,W,j,V,Y=null,J=null,N=null)=>{let H,O="",q=!1,_=Date.now(),G=0,w=()=>{if(j)o(j).catch(()=>{});if(N)o(N).catch(()=>{})};return new Z2({async start(z){try{let A=await $.parallel.completion(X,(B,L)=>{if(!L)return;if(L.token)O+=L.token,G+=1;z.enqueue({event:"token",data:{requestId:B,...L}})}),{requestId:R}=A;H=A.stop;let U=await A.promise;if(U.text)O=U.text;else if(U.content)O=U.content;q=!U.interrupted&&!U.context_full,console.log("[Completion] Result:",U),z.enqueue({event:"result",data:{requestId:R,...U}}),z.close();let Q=Date.now()-_,K=U.timings||{};if($0.addCompletion({id:`completion-${R}`,generatorId:Y,requestId:R,repoId:J?.repoId||null,quantization:J?.quantization||null,variant:J?.variant||null,cacheTokens:K.cache_n??0,promptTokens:K.prompt_n??V??0,tokensGenerated:K.predicted_n??G,tokensPerSecond:K.predicted_per_second??0,promptPerSecond:K.prompt_per_second??0,durationMs:Q,success:!0,interrupted:U.interrupted||!1,contextFull:U.context_full||U.contextFull||!1,usedCache:Boolean(X.load_state_path)}),q&&Z.enabled&&O)Z.saveCompletionState(W,O,j,V,N).catch((B)=>{console.warn("[SessionCache] Save failed:",B.message)});else w()}catch(A){z.enqueue({event:"error",data:{message:A?.message||String(A)}}),z.error(A),$0.addCompletion({id:`completion-${Date.now()}`,generatorId:Y,repoId:J?.repoId||null,quantization:J?.quantization||null,variant:J?.variant||null,durationMs:Date.now()-_,tokensGenerated:G,success:!1,error:A?.message||String(A)}),w()}},cancel(){if(H)H();w()}})},z0=($)=>{let X={model:$.plan.localPath,runtime:$.plan.info.runtime};return p4("sha256").update(JSON.stringify(X)).digest("hex").slice(0,24)},k3=async($,X,Z,W=null)=>{let{config:j,localPath:V,artifact:Y}=$;if($.localExists&&!X.has(V)){if($.info.download.exists=!0,typeof Z==="function")Z(0.5);return V}if(j.model.local_path&&!j.model.allow_local_file)throw Error("Local model path provided but `model.allow_local_file` is not enabled");let J=V;if(W){let N=W.getDownload(J);if(N){console.log(`[ensureModelFile] Waiting for global download: ${Y.repoId}`);try{if(await N,await P0(V,Y.size)){if($.localExists=!0,$.info.download.exists=!0,typeof Z==="function")Z(0.5);return V}}catch(H){console.warn(`[ensureModelFile] Global download failed, will retry: ${H.message}`)}}}if(!X.has(J))X.set(J,(async()=>{if(Y.isSplit&&Y.splitCount>0){let N=/-(\d{5})-of-(\d{5})\.gguf$/,H=y.dirname(V),O=Y.splitCount,q=0;for(let _=1;_<=O;_+=1){let G=String(_).padStart(5,"0"),w=Y.filename.replace(N,`-${G}-of-${String(O).padStart(5,"0")}.gguf`),z=`${j.model.base_url.replace(/\/+$/,"")}/${Y.repoId}/resolve/${Y.revision}/${w}`,A=y.join(H,w);if(!await P0(A))await a0(z,Y.headers,A,null,(U)=>{if(U>=0&&Number.isFinite(U)){let Q=(q+U)/O,K=Math.round(Q*100);if(console.log(`Downloading model splits: ${Math.min(100,K)}%`),typeof Z==="function")Z(Q*0.5)}});q+=1}}else console.log("Downloading model: 0%"),await a0(Y.url,Y.headers,V,Y.size,(N)=>{if(N>=0&&Number.isFinite(N)){let H=Math.round(N*100);if(console.log(`Downloading model: ${Math.min(100,H)}%`),typeof Z==="function")Z(N*0.5)}});$.localExists=!0,$.info.download.exists=!0})());try{await X.get(J)}finally{X.delete(J)}return V},D3=async($,X)=>{let Z=z0($),W=$.contexts.get(Z);if(W&&!W.released){if(W.releaseTimer)clearTimeout(W.releaseTimer),W.releaseTimer=null,console.log(`[Context] Cancelled pending release for context "${Z}"`);if(W.releaseRequested=!1,W.refCount+=1,console.log(`[Context] Reusing existing context "${Z}", refCount=${W.refCount}`),typeof X==="function")X(0);if(!W.context)await W.ready;if(typeof X==="function")X(1);return W}if(W)console.log(`[Context] Record exists but released=${W.released}, creating new context`);else console.log(`[Context] No existing record for "${Z}", creating new context`);W={key:Z,refCount:1,ready:null,released:!1},$.contexts.set(Z,W),W.ready=(async()=>{let j=Date.now(),V=await k3($.plan,$.downloads,X,$.globalDownloadManager);if(typeof X==="function")X(0.5);let Y={model:V,n_threads:$.plan.info.runtime.n_threads,use_mmap:$.plan.info.runtime.use_mmap,use_mlock:$.plan.info.runtime.use_mlock,cpu_mask:$.plan.info.runtime.cpu_mask,cpu_strict:$.plan.info.runtime.cpu_strict,devices:$.plan.info.runtime.devices,n_ctx:$.plan.info.runtime.n_ctx,n_gpu_layers:$.plan.info.runtime.n_gpu_layers,n_parallel:$.plan.info.runtime.n_parallel,n_batch:$.plan.info.runtime.n_batch,n_ubatch:$.plan.info.runtime.n_ubatch,n_cpu_moe:$.plan.info.runtime.n_cpu_moe,flash_attn_type:$.plan.info.runtime.flash_attn_type,ctx_shift:$.plan.info.runtime.ctx_shift,kv_unified:$.plan.info.runtime.kv_unified,swa_full:$.plan.info.runtime.swa_full,lib_variant:$.plan.info.runtime.variant};if($.plan.info.runtime.flash_attn_type!=="off")Y.cache_type_k=$.plan.info.runtime.cache_type_k,Y.cache_type_v=$.plan.info.runtime.cache_type_v;console.log("[Context] Load Options:",Y);let J;try{if(J=await N3(Y,(N)=>{if(typeof X==="function"){if(X(0.5+N*0.25),N%5===0)console.log("[Context] Load Model Progress:",N)}}),$.plan.info.runtime.n_parallel){if(!await J.parallel.enable({n_parallel:$.plan.info.runtime.n_parallel,n_batch:$.plan.info.runtime.n_batch}))throw Error("Failed to enable parallel decoding mode for context")}if(typeof X==="function")X(1);return W.context=J,W.modelInfo=J.getModelInfo(),$0.addModelLoad({id:`${$.id}-${Z}`,generatorId:$.id,contextKey:Z,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,nCtx:$.plan.info.runtime?.n_ctx||null,nGpuLayers:$.plan.info.runtime?.n_gpu_layers||null,durationMs:Date.now()-j,success:!0}),W}catch(N){if($0.addModelLoad({id:`${$.id}-${Z}`,generatorId:$.id,contextKey:Z,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,durationMs:Date.now()-j,success:!1,error:N?.message||String(N)}),J)try{J.release()}catch(H){}throw N}})();try{return await W.ready,W}catch(j){throw $.contexts.delete(Z),j}},r0=async($,X,Z=!1)=>{if(X.released)return!1;if(!Z&&X.refCount>0)return!1;X.released=!0,$.contexts.delete(X.key);try{X.context?.parallel?.disable?.()}catch(W){}return await X.context?.release?.(),!0},b3=async($,X,Z=!1)=>{if(X.releaseRequested=!0,X.releaseTimer)clearTimeout(X.releaseTimer),X.releaseTimer=null;if(Z)X.refCount=0;else if(X.refCount=Math.max(0,X.refCount-1),X.refCount>0)return X.releaseRequested=!1,!1;let W=$.config.runtime.context_release_delay_ms;if(typeof W!=="number"||!Number.isFinite(W))return r0($,X);let j=Math.max(0,Math.floor(W));if(Z||j<=0)return r0($,X);return console.log(`[Context] Scheduling release in ${j}ms for context "${X.key}"`),X.releaseTimer=setTimeout(async()=>{if(X.releaseTimer=null,X.refCount>0){console.log(`[Context] Release cancelled, refCount=${X.refCount} for context "${X.key}"`),X.releaseRequested=!1;return}console.log(`[Context] Releasing context "${X.key}" after ${j}ms delay`),await r0($,X)},j),!0},d4=($)=>{let X=e0($);return X.model.repo_id||X.model.repository||X.model.model||null},a1=($)=>{if(!$)return 0;if(typeof $.score==="number"&&Number.isFinite($.score))return Number($.score);return q0($)};var l4=P(()=>{M0();s0();O3=H3(),{ReadableStream:Z2,WritableStream:K3}=O3,n=y.join(T0.homedir(),".buttress","models"),Q2=["mxfp4","q8_0","q6_k","q6","q5_k_m","q5_k_s","q5_k","q5_1","q5_0","q4_k_m","q4_k_s","q4_k","q4_1","q4_0","q3","q2"],S0={backend:{type:"ggml-llm",variant:null,variant_preference:["cuda","vulkan","snapdragon","default"],gpu_memory_fraction:0.85,cpu_memory_fraction:t0},model:{repo_id:null,revision:"main",filename:null,url:null,quantization:null,preferred_quantizations:[],n_ctx:null,n_gpu_layers:"auto",allow_local_file:!1,local_path:null,api_base:Y2,base_url:j2},runtime:{cache_dir:n,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},session_cache:{enabled:!0,max_size_bytes:10737418240,max_entries:1000},context_release_delay_ms:1e4}}});import j0 from"node:path";import t4 from"node:os";import{stat as B2,mkdir as h3,open as I3,unlink as i4,readFile as C3,writeFile as y3}from"node:fs/promises";import{createHash as u3}from"node:crypto";import{initWhisper as f3}from"@fugood/whisper.node";import{getBackendDevicesInfo as R2,isLibVariantAvailable as w2}from"@fugood/llama.node";import*as g3 from"node:stream/web";class v2{constructor(){this.queue=[],this.processing=!1,this.currentTaskId=null}async enqueue($,X=null){return new Promise((Z,W)=>{this.queue.push({task:$,resolve:Z,reject:W,taskId:X}),this.processNext()})}async processNext(){if(this.processing||this.queue.length===0)return;this.processing=!0;let{task:$,resolve:X,reject:Z,taskId:W}=this.queue.shift();this.currentTaskId=W;try{let j=await $();X(j)}catch(j){Z(j)}finally{this.processing=!1,this.currentTaskId=null,this.processNext()}}getStatus(){return{processing:this.processing,queuedCount:this.queue.length,currentTaskId:this.currentTaskId}}}async function h2($,X,Z={}){let{globalDownloadManager:W=null}=Z,j=j4(X),V=await o3(j),Y={id:$,type:"ggml-stt",config:j,plan:V,info:V.info,contextRecord:null,downloads:new Map,globalDownloadManager:W,queue:new v2,finalized:!1},J=async()=>{if(Y.finalized)return;Y.finalized=!0;let A=Y.contextRecord;if(!A)return;if(A.released)return;if(A.releaseRequested||A.releaseTimer)return;if(A.refCount=Math.max(0,A.refCount-1),A.refCount>0)return;await Z4(Y,A)},N=async(A={})=>{let{onProgress:R}=A;try{let U=await e3(Y,R);return{modelInfo:U.modelInfo&&typeof U.modelInfo==="object"?{...U.modelInfo}:null,runtime:{...Y.plan.info.runtime},download:{...Y.plan.info.download}}}catch(U){throw console.error("[Context] Error initializing context:",U),U}},H=async()=>{if(Y.finalized)return!1;let A=Y.contextRecord;if(!A)return!1;return $9(Y,A)},O=async(A={})=>{let{audioPath:R,audioData:U,options:Q={}}=A,K=Y.contextRecord;if(!K)throw Error("Context not initialized");let B={...Q};if(Y.plan.info.runtime.max_threads&&B.maxThreads==null)B.maxThreads=Y.plan.info.runtime.max_threads;let L=`transcription-${Date.now()}-${Math.random().toString(36).slice(2,8)}`,F=Date.now();return Y.queue.enqueue(async()=>{await K.ready;try{let E;if(U){let M=t3(U),{promise:x}=K.context.transcribeData(M,B);E=await x}else{if(!R)throw Error("audioPath or audioData is required for transcription");let M=j0.resolve(R),{promise:x}=K.context.transcribe(M,B);E=await x}return N0.addTranscription({id:L,generatorId:Y.id,repoId:Y.plan.info.model?.repoId||null,quantization:Y.plan.info.model?.quantization||null,modelType:Y.plan.info.model?.modelType||null,variant:Y.plan.info.runtime?.variant||null,durationMs:Date.now()-F,segmentCount:E?.segments?.length||0,textLength:E?.text?.length||0,success:!0}),E}catch(E){throw N0.addTranscription({id:L,generatorId:Y.id,repoId:Y.plan.info.model?.repoId||null,quantization:Y.plan.info.model?.quantization||null,modelType:Y.plan.info.model?.modelType||null,variant:Y.plan.info.runtime?.variant||null,durationMs:Date.now()-F,success:!1,error:E?.message||String(E)}),E}},L)},q=async(A={})=>O(A),_=async(A={})=>O(A),G=()=>{let A=Y.contextRecord;if(!A)return!1;return!A.released&&(A.releaseRequested||A.releaseTimer||A.refCount>0)},w=()=>{Y.finalized=!1},z=()=>({id:Y.id,type:Y.type,repoId:Y.plan.info.model?.repoId||null,quantization:Y.plan.info.model?.quantization||null,modelType:Y.plan.info.model?.modelType||null,variant:Y.plan.info.runtime?.variant||null,hasContext:Boolean(Y.contextRecord?.context),contextRefCount:Y.contextRecord?.refCount||0,queueStatus:Y.queue.getStatus()});return{id:$,type:"ggml-stt",info:V.info,queue:Y.queue,initContext:N,transcribe:q,transcribeData:_,releaseContext:H,finalize:J,getStatus:z,hasPendingReleases:G,resetFinalized:w}}async function I2($,X,Z={}){let{onProgress:W,onComplete:j,onError:V}=Z;try{let Y=j4($),J=await e4(Y),N=D2(Y,J),{repoId:H}=J;if(await W4(N,J.size)){if(console.log(`[Download] STT model already exists: ${H} at ${N}`),typeof j==="function")j({localPath:N,repoId:H,alreadyExists:!0});return{started:!1,localPath:N,repoId:H,alreadyExists:!0}}let q=X.getDownload(N);if(q)return console.log(`[Download] Already downloading STT model: ${H}`),q.then(()=>{if(typeof j==="function")j({localPath:N,repoId:H,joinedExisting:!0})}).catch((G)=>{if(typeof V==="function")V(G)}),{started:!1,localPath:N,repoId:H,alreadyDownloading:!0};console.log(`[Download] Starting STT model download: ${H}`);let _=(async()=>{try{if(await b2(J.url,J.headers,N,J.size,(G)=>{if(G>=0&&Number.isFinite(G)){if(console.log(`[Download] ${H}: ${Math.round(G*100)}%`),typeof W==="function")W(G)}}),console.log(`[Download] Completed STT model: ${H}`),typeof j==="function")j({localPath:N,repoId:H})}catch(G){if(console.error(`[Download] Failed STT model: ${H}`,G.message),typeof V==="function")V(G);throw G}finally{X.deleteDownload(N)}})();return X.setDownload(N,_),{started:!0,localPath:N,repoId:H}}catch(Y){if(console.error("[Download] Failed to start STT download:",Y.message),typeof V==="function")V(Y);return{started:!1,localPath:null,repoId:null,error:Y.message}}}async function C2($=null,X={}){let{threshold:Z=1.1,includeBreakdown:W=!1,config:j,...V}=X,Y=null,J=null,N=null;if(j)try{let R=j4(j),U=await e4(R);Y=U.size??null,{processingBufferBytes:J}=m0({modelBytes:Y}),N=U.quantization||null}catch(R){}let H=j?.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.gpu_memory_fraction))):void 0,O=j?.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.cpu_memory_fraction))):void 0,q=await _0({...V,platform:process.platform,totalMemoryInBytes:t4.totalmem(),backend:"ggml-stt",includeBreakdown:W,gpuMemoryFraction:H,cpuMemoryFraction:O,dependencies:{getBackendDevicesInfo:R2,isLibVariantAvailable:w2},modelBytes:Y,kvCacheBytes:J}),_=q.selected,G=z2(_);if(_)_.modelBytes=Y||null,_.processingBytes=J||null,_.quantization=N||null;let w=null,z=null;if($){let R=z2($);z={...$,score:R};let U="buttress",Q="buttress-higher-score";if(!q.ok)U="local",Q="buttress-unavailable";else if(!R&&R!==0)U="buttress",Q="missing-client-score";else if($.fit&&_?.fit){let K=$.fit.fitsInGpu||$.fit.fitsInCpu,B=_.fit.fitsInGpu||_.fit.fitsInCpu;if(K&&!B)U="local",Q="client-fits-in-memory";else if(B&&!K)U="buttress",Q="buttress-fits-in-memory";else if(R>G*Z)U="local",Q="client-better";else if(G>R*Z)U="buttress",Q="buttress-better";else U="either",Q="comparable-scores"}else if(R>G*Z)U="local",Q="client-better";else if(G>R*Z)U="buttress",Q="buttress-better";else U="either",Q="comparable-scores";w={buttressScore:G,clientScore:R,threshold:Z,recommendation:U,reason:Q}}if(!q.ok&&!w)w={buttressScore:G,clientScore:$?.score??null,threshold:Z,recommendation:"local",reason:"buttress-unavailable"};let A=null;if(j)A={repoId:j.model?.repo_id||null,quantization:j.model?.quantization||null,filename:j.model?.filename||null};return{type:"ggml-stt",timestamp:new Date().toISOString(),buttress:q,client:z,comparison:w,modelConfig:A}}var p3=()=>{if(typeof globalThis<"u"&&globalThis.ReadableStream&&globalThis.WritableStream)return{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream};return g3},m3,F2=($={},X={})=>{return Object.entries(X||{}).forEach(([Z,W])=>{if(W&&typeof W==="object"&&!Array.isArray(W)){if(!$[Z]||typeof $[Z]!=="object")$[Z]={};F2($[Z],W)}else $[Z]=W}),$},c3=".bin",E2="https://huggingface.co",M2="https://huggingface.co/api",R0,r4,o4,x2="fp16",S2=0.5,d3,P2=($)=>{if(!$)return null;let X=$.toLowerCase();return d3.find((Z)=>X.includes(Z))||null},a4,s4=($,X=[])=>{if(!$&&$!==0)return[...X];if(Array.isArray($))return $.filter((Z)=>Z!=null);return[$]},X4=($)=>{if(!$)return null;let X=String($).toLowerCase();if(["cuda","vulkan","default"].includes(X))return X;return null},j4=($={})=>{let X=JSON.parse(JSON.stringify(a4));if(F2(X,$),X.backend.variant=X4(X.backend.variant),X.backend.variant_preference=Array.from(new Set(s4(X.backend.variant_preference||r4).map(X4).filter(Boolean))),X.backend.variant_preference.length===0)X.backend.variant_preference=[...r4];if(X.runtime.prefer_variants=Array.from(new Set(s4(X.runtime.prefer_variants).map(X4).filter(Boolean))),X.model.preferred_quantizations=Array.from(new Set(s4(X.model.preferred_quantizations||X.model.quantizations).map((Z)=>Z?String(Z).toLowerCase():null).filter(Boolean))),X.model.quantization){let Z=String(X.model.quantization).toLowerCase();if(!X.model.preferred_quantizations.includes(Z))X.model.preferred_quantizations.unshift(Z)}return X.model.base_url=X.model.base_url||E2,X.model.api_base=X.model.api_base||M2,X.runtime.cache_dir=X.runtime.cache_dir?j0.resolve(X.runtime.cache_dir):R0,X.runtime.context_release_delay_ms=Math.max(0,Number(X.runtime.context_release_delay_ms)||a4.runtime.context_release_delay_ms),X},n4=($)=>{let X=$.toLowerCase();return o4.find((W)=>X.includes(W))||null},l3=($)=>{let X=[];if($.backend.variant)X.push($.backend.variant);if($.runtime.prefer_variants.length>0)X.push(...$.runtime.prefer_variants);return X.push(...$.backend.variant_preference),X.push("default"),Array.from(new Set(X.map(X4).filter(Boolean)))},T2=async($)=>{await h3($,{recursive:!0})},i3=($=R0)=>j0.join($,".metadata-cache"),k2=($,X,Z=R0)=>{let W=u3("sha256").update($).digest("hex");return j0.join(i3(Z),X,`${W}.json`)},s3=async($,X,Z=R0)=>{try{let W=k2($,X,Z),j=await C3(W,"utf-8");return JSON.parse(j)}catch(W){return null}},L2=async($,X,Z,W=R0)=>{try{let j=k2($,X,W);await T2(j0.dirname(j)),await y3(j,JSON.stringify(Z),"utf-8")}catch(j){}},n3=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,X);if(!Z.ok){let W=await Z.text().catch(()=>"");throw Error(`Failed to fetch ${$}: ${Z.status} ${Z.statusText} ${W}`.trim())}return Z.json()},A2=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,{...X,method:"HEAD"});if(!Z.ok)throw Error(`Failed to fetch headers for ${$}: ${Z.status} ${Z.statusText}`);return Z},D2=($,X)=>{if($.model.local_path)return j0.resolve($.model.local_path);let Z=X.repoId.split("/"),W=j0.join($.runtime.cache_dir,...Z,X.revision);return j0.join(W,X.filename)},W4=async($,X)=>{try{let Z=await B2($);if(!X)return!0;return Z.size===X}catch(Z){return!1}},b2=async($,X,Z,W,j)=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");await T2(j0.dirname(Z));let V=await fetch($,{headers:X});if(!V.ok||!V.body)throw Error(`Failed to download ${$}: ${V.status} ${V.statusText}`);let Y=await I3(Z,"w"),J=Number(V.headers.get("content-length"))||W||0,N=0,H=0.05;try{await V.body.pipeTo(new m3({async write(O){if(await Y.write(O),N+=O.byteLength,typeof j==="function"&&J>0){let q=Math.min(1,N/J);while(q>=H)j(H),H+=0.05}},async close(){if(await Y.close(),typeof j==="function")j(1)},async abort(O){throw await Y.close().catch(()=>{}),await i4(Z).catch(()=>{}),O}}))}catch(O){throw await Y.close().catch(()=>{}),await i4(Z).catch(()=>{}),O}if(W){let O=await B2(Z);if(O.size!==W)throw await i4(Z).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${W} got ${O.size}`)}},e4=async($)=>{let X=$.model.repo_id||$.model.repository||$.model.model;if(!X)throw Error("`model.repo_id` is required in Buttress backend config");let Z=$.model.revision||"main",W=$.runtime.cache_dir,j=JSON.stringify({repoId:X,revision:Z,filename:$.model.filename,url:$.model.url,quantization:$.model.quantization,preferred_quantizations:$.model.preferred_quantizations}),V=await s3(j,"artifact-info",W);if(V)return V;let Y={...$.runtime.http_headers||{}};if($.runtime.huggingface_token)Y.Authorization=`Bearer ${$.runtime.huggingface_token}`;if($.model.url){let U=await A2($.model.url,{headers:Y}),Q=Number(U.headers.get("content-length"))||null,K=$.model.filename||$.model.url.split("/").pop(),B={repoId:X,revision:Z,filename:K,url:$.model.url,size:Q,quantization:n4(K||""),headers:Y};return await L2(j,"artifact-info",B,W),B}let{filename:J}=$.model,N=$.model.quantization&&String($.model.quantization).toLowerCase(),H=await n3(`${$.model.api_base}/models/${X}?revision=${Z}&blobs=true`,{headers:Y}),q=(H?.siblings||H?.files||[]).map((U)=>U.rfilename||U.path||U.filename).filter((U)=>typeof U==="string"&&U.endsWith(c3));if(q.length===0)throw Error(`No model artifacts found in repo ${X}`);let _=$.model.preferred_quantizations.length>0?$.model.preferred_quantizations:o4,G=()=>{for(let U of _)if(U===x2){let Q=q.find((K)=>{let B=K.toLowerCase();return!o4.some((L)=>B.includes(L))});if(Q)return{filename:Q,quantization:null}}else{let Q=q.find((K)=>K.toLowerCase().includes(U));if(Q)return{filename:Q,quantization:U}}return null};if(!J){let U=G()||{filename:q[0],quantization:null},{filename:Q,quantization:K}=U;J=Q,N=K||n4(J)}else if(!N)N=n4(J);let w=`${$.model.base_url.replace(/\/+$/,"")}/${X}/resolve/${Z}/${J}`,z=await A2(w,{headers:Y}),A=Number(z.headers.get("content-length"))||null,R={repoId:X,revision:Z,filename:J,url:w,size:A,quantization:N,headers:Y,isSplit:!1,splitCount:0};return await L2(j,"artifact-info",R,W),R},r3=async($,{modelBytes:X=null,processingBytes:Z=null}={})=>{let W=l3($),[j,...V]=W,Y=$.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.gpu_memory_fraction))):a4.backend.gpu_memory_fraction||1,J=$.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):S2,N=await _0({platform:process.platform,totalMemoryInBytes:t4.totalmem(),backend:"ggml-stt",variant:j||null,preferVariants:V,variantPreference:$.backend.variant_preference,gpuMemoryFraction:Y,cpuMemoryFraction:J,dependencies:{getBackendDevicesInfo:R2,isLibVariantAvailable:w2},modelBytes:X,kvCacheBytes:Z}),H=(_)=>({..._,devices:Array.isArray(_.devices)?_.devices:[],ok:_.ok,hasGpu:Boolean(_.hasGpu),totalMemory:_.gpuTotalBytes||_.totalMemory||0,error:_.ok?null:Error(_.error||`Variant ${_.variant} not available on this platform`)});if(!N.ok||!N.selected){let _=(N.attempts||[]).map((G)=>`${G.variant}: ${G.error||"unknown error"}`).join("; ");throw Error(`Unable to initialize any backend variant (${W.join(", ")}). Errors: ${_}`)}let O=(N.attempts||[]).map(H);return{selected:H(N.selected),attempts:O}},o3=async($)=>{let X=await e4($),Z=m0({modelBytes:X.size>0?X.size:0}),W=await r3($,{modelBytes:Z.modelBytes,processingBytes:Z.processingBufferBytes}),j=W.selected.hasGpu&&(W.selected.fit?.fitsInGpu!==void 0?W.selected.fit.fitsInGpu:!0);if($.model.use_gpu===!1)j=!1;let V=$.model.use_flash_attn&&String($.model.use_flash_attn).toLowerCase(),Y;if(V==="on"||V==="true")Y=!0;else if(V==="off"||V==="false")Y=!1;else Y=j;let J=$.runtime.cache_dir,N=D2($,X),H=await W4(N,X.size),O={ok:!0,backend:"ggml-stt",model:{repoId:X.repoId,revision:X.revision,filename:X.filename,quantization:X.quantization,modelType:P2(X.filename),url:X.url,sizeBytes:X.size},runtime:{variant:W.selected.variant,use_gpu:j,use_flash_attn:Y,max_threads:$.runtime.max_threads?Number($.runtime.max_threads):null},resources:{...Z,gpuCapacityBytes:W.selected.gpuTotalBytes,gpuUsableBytes:W.selected.gpuUsableBytes,cpuUsableBytes:W.selected.cpuUsableBytes,fit:W.selected.fit},devices:{selected:W.selected,attempts:W.attempts},download:{cacheDir:J,localPath:N,exists:H},timestamp:new Date().toISOString()};return{config:$,info:O,artifact:X,memory:Z,devices:W,localPath:N,localExists:H}},a3=async($,X,Z,W=null)=>{let{localPath:j,artifact:V,config:Y}=$;if($.localExists){if(typeof Z==="function")Z(1);return j}if(W){let H=W.getDownload(j);if(H){console.log(`[ensureModelFile] Waiting for global STT download: ${V.repoId}`);try{if(await H,await W4(j,V.size)){if($.localExists=!0,$.info.download.exists=!0,typeof Z==="function")Z(1);return j}}catch(O){console.warn(`[ensureModelFile] Global STT download failed, will retry: ${O.message}`)}}}let J=X.get(j);if(J){if(await J,typeof Z==="function")Z(1);return j}let N=(async()=>{if(Y.model.allow_local_file){if(!await W4(j,V.size))throw Error(`Local model file not found: ${j}`);return j}return await b2(V.url,V.headers,j,V.size,Z),j})();X.set(j,N);try{return await N,j}finally{X.delete(j)}},t3=($)=>{if(!$)return null;if($ instanceof ArrayBuffer)return $;if(ArrayBuffer.isView($))return $.buffer;if(typeof $==="string"){let X=$.startsWith("data:")?$.split(",")[1]||"":$,Z=Buffer.from(X,"base64");return Z.buffer.slice(Z.byteOffset,Z.byteOffset+Z.byteLength)}throw Error("Unsupported audioData format, expected base64 string or ArrayBuffer")},e3=async($,X)=>{if($.contextRecord&&!$.contextRecord.released){if($.contextRecord.releaseTimer)clearTimeout($.contextRecord.releaseTimer),$.contextRecord.releaseTimer=null,console.log("[Context] Cancelled pending STT release");if($.contextRecord.releaseRequested=!1,$.contextRecord.refCount+=1,console.log(`[Context] Reusing existing STT context, refCount=${$.contextRecord.refCount}`),typeof X==="function")X(0);if(!$.contextRecord.context)await $.contextRecord.ready;if(typeof X==="function")X(1);return $.contextRecord}if($.contextRecord)console.log(`[Context] STT record exists but released=${$.contextRecord.released}, creating new context`);else console.log("[Context] No existing STT record, creating new context");let Z={refCount:1,ready:null,released:!1};$.contextRecord=Z,Z.ready=(async()=>{let W=Date.now();try{if(typeof X==="function")X(0);let j=await a3($.plan,$.downloads,X,$.globalDownloadManager);if(typeof X==="function")X(0.5);let V=await f3({filePath:j,useFlashAttn:$.plan.info.runtime.flash_attn_type==="on",useGpu:$.plan.info.runtime.n_gpu_layers>0,nThreads:$.plan.info.runtime.n_threads},$.plan.info.runtime.variant);if(typeof X==="function")X(1);Z.context=V;try{Z.modelInfo=V.getModelInfo()}catch(Y){Z.modelInfo=null}return N0.addModelLoad({id:$.id,generatorId:$.id,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,modelType:$.plan.info.model?.modelType||null,variant:$.plan.info.runtime?.variant||null,useGpu:$.plan.info.runtime?.use_gpu||!1,durationMs:Date.now()-W,success:!0}),Z}catch(j){throw N0.addModelLoad({id:$.id,generatorId:$.id,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,modelType:$.plan.info.model?.modelType||null,variant:$.plan.info.runtime?.variant||null,durationMs:Date.now()-W,success:!1,error:j?.message||String(j)}),j}})();try{if(await Z.ready,typeof X==="function")X(1);return Z}catch(W){throw $.contextRecord=null,W}},Z4=async($,X,Z=!1)=>{if(X.released)return!1;if(!Z&&X.refCount>0)return!1;return X.released=!0,$.contextRecord=null,await X.context?.release?.(),!0},$9=async($,X,Z=!1)=>{if(X.releaseRequested=!0,X.releaseTimer)clearTimeout(X.releaseTimer),X.releaseTimer=null;if(Z)X.refCount=0;else if(X.refCount=Math.max(0,X.refCount-1),X.refCount>0)return X.releaseRequested=!1,!1;let W=$.config.runtime.context_release_delay_ms;if(typeof W!=="number"||!Number.isFinite(W))return Z4($,X);let j=Math.max(0,Math.floor(W));if(Z||j<=0)return Z4($,X);return console.log(`[Context] Scheduling STT release in ${j}ms`),X.releaseTimer=setTimeout(async()=>{if(X.releaseTimer=null,X.refCount>0){console.log(`[Context] STT release cancelled, refCount=${X.refCount}`),X.releaseRequested=!1;return}console.log(`[Context] Releasing STT context after ${j}ms delay`),await Z4($,X)},j),!0},$1=($)=>{let X=j4($),Z=X.model.repo_id||X.model.repository||X.model.model||null;if(!Z)return null;let W=P2(X.model.filename);if(W)return`${Z}:${W}`;return Z},z2=($)=>{if(!$)return 0;if(typeof $.score==="number"&&Number.isFinite($.score))return Number($.score);return q0($)};var X1=P(()=>{M0();s0();({WritableStream:m3}=p3()),R0=j0.join(t4.homedir(),".buttress","models"),r4=["cuda","vulkan","default"],o4=["q8_0","q5_1","q5_0","q4_1","q4_0"],d3=["large-v3-turbo","distil-large-v3","large-v3","large-v2","large-v1","large","distil-medium","medium.en","medium","small.en-tdrz","distil-small.en","small.en","small","base.en","base","tiny.en","tiny"],a4={backend:{type:"ggml-stt",variant:null,variant_preference:r4,gpu_memory_fraction:0.85,cpu_memory_fraction:S2},model:{repo_id:"BricksDisplay/whisper-ggml",revision:"main",filename:null,url:null,quantization:null,preferred_quantizations:["q8_0",x2,"q5_1"],allow_local_file:!1,local_path:null,api_base:M2,base_url:E2,use_gpu:!0,use_flash_attn:"auto"},runtime:{cache_dir:R0,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},max_threads:null,context_release_delay_ms:1e4}}});async function t($,X=null,Z={}){if($==="ggml-llm")return _2(X,Z);if($==="ggml-stt")return C2(X,Z);throw Error(`Unknown backend type: ${$}`)}var Z1=P(()=>{l4();X1()});var g;var y2=P(()=>{g={name:"@fugood/buttress-backend-core",private:!0,type:"module",version:"2.23.0-beta.44",main:"src/index.js",types:"lib/types/index.d.ts",scripts:{build:"tsc --noResolve --noCheck --declaration --emitDeclarationOnly --allowJs --outDir lib/types src/index.js"},dependencies:{"@fugood/buttress-hardware-guardrails":"^2.23.0-beta.44","@fugood/llama.node":"^1.4.14","@fugood/whisper.node":"^1.0.13","@huggingface/gguf":"^0.3.2","@iarna/toml":"^3.0.0",bytes:"^3.1.0"}}});import m from"node:os";import u2 from"node:fs";import f2 from"node:path";import{execSync as Y4}from"node:child_process";import g2 from"@iarna/toml";async function m2({modelIds:$=[],defaultConfig:X=null}={}){let Z=[];if(console.log(`${g.name} v${g.version}`),console.log(`Generating model capabilities comparison...
2
+ import{createRequire as s6}from"node:module";var l6=Object.defineProperty;var i6=($,X)=>{for(var Z in X)l6($,Z,{get:X[Z],enumerable:!0,configurable:!0,set:(W)=>X[Z]=()=>W})};var P=($,X)=>()=>($&&(X=$($=0)),X);var R1=s6(import.meta.url);var w1=($,X,Z)=>Math.min(Math.max($,X),Z),F1=($)=>$?40:0,E1=($=0)=>{if(!$)return 0;return w1($/12884901888*20,0,20)},M1=($=0)=>{if(!$)return 0;return w1($/34359738368*10,0,10)},x1=($)=>$?10:0,S1=($="default",X=null)=>{let Z=String($).toLowerCase();if(!Z)return 0;if(Z.includes("cuda"))return 20;if(Z.includes("vulkan"))return 10;if(Z.includes("default"))return X==="darwin"||X==="ios"?15:5;return 0},q0=({platform:$,variant:X,hasGpu:Z,gpuUsableBytes:W=0,cpuUsableBytes:j=0,ok:V=!0}={})=>{if(!V)return 0;let Y=F1(Z)+S1(X,$)+E1(W),J=M1(j),N=x1(V);return Math.min(100,Math.round(Y+J+N))},F4=({platform:$,variant:X,hasGpu:Z,gpuUsableBytes:W=0,cpuUsableBytes:j=0,ok:V=!0}={})=>({gpuPresence:F1(Z),variant:S1(X,$),gpuMemory:E1(W),cpuMemory:M1(j),availability:x1(V)});var D1,u0=0.85,f0=0.5,P1=($)=>{if(!$&&$!==0)return[];if(Array.isArray($))return $.filter((X)=>X!=null);return[$]},n6=($)=>{if(!$)return null;return String($).trim().toLowerCase()||null},r6=({variant:$,preferVariants:X=[],variantPreference:Z=[],defaultVariants:W=D1}={})=>{let j=[];if($)j.push($);j.push(...P1(X)),j.push(...P1(Z)),j.push(...W);let V=j.map(n6).filter(Boolean);return Array.from(new Set(V))},T1=($={})=>{let X=String($.type||$.deviceType||$.kind||"").toLowerCase();if(X.includes("gpu"))return!0;if(X.includes("cuda"))return!0;if(X.includes("metal"))return!0;if(X.includes("vulkan"))return!0;if(X.includes("snapdragon"))return!0;return!1},o6=($)=>{if(!Array.isArray($))return[];return $.map((X)=>({...X}))},a6=($,X)=>{if($==="snapdragon")return X.filter((Z)=>Z.deviceName!=="GPUOpenCL");return X},k1=({platform:$,totalMemoryInBytes:X,variant:Z,devices:W,gpuMemoryFraction:j,cpuMemoryFraction:V,ok:Y,error:J})=>{let N=o6(a6(Z,W)),H=N.some(T1),O=N.filter((R)=>T1(R)&&Number.isFinite(Number(R.maxMemorySize))).reduce((R,U)=>R+U.maxMemorySize,0),q=X,_=H?Math.floor(O*j):0,G=q?Math.floor(q*V):0,w={platform:$,variant:Z,hasGpu:H,gpuUsableBytes:_,cpuUsableBytes:G,ok:Y},z=q0(w),A=Y?F4(w):null;return{platform:$,ok:Y,variant:Z,hasGpu:H,devices:N,gpuTotalBytes:O,gpuUsableBytes:_,cpuTotalBytes:q,cpuUsableBytes:G,score:z,breakdown:A,error:J,timestamp:new Date().toISOString()}},E4=({device:$,modelBytes:X=0,kvCacheBytes:Z=0}={})=>{if(!$)return{totalRequiredBytes:X+Z,fitsInGpu:!1,fitsInCpu:!1,limiting:"unknown-device"};let W=Math.max(0,Number(X)||0)+Math.max(0,Number(Z)||0),j=$.hasGpu&&W>0&&W<=$.gpuUsableBytes,V=W>0&&W<=$.cpuUsableBytes,Y="ok";if(!j&&$.hasGpu)Y="gpu-memory";if(!V)Y=j?"cpu-memory":"insufficient-memory";return{totalRequiredBytes:W,fitsInGpu:j,fitsInCpu:V,limiting:Y}},E0=async({platform:$,variant:X=null,preferVariants:Z=[],variantPreference:W=[],gpuMemoryFraction:j=u0,cpuMemoryFraction:V=f0,includeBreakdown:Y=!1,totalMemoryInBytes:J,modelBytes:N=null,kvCacheBytes:H=null,limitedKvCacheBytes:O=null,dependencies:q={},defaultVariants:_=D1}={})=>{let{getBackendDevicesInfo:G,isLibVariantAvailable:w}=q;if(typeof G!=="function"||typeof w!=="function")throw TypeError("GGML capability detection requires getBackendDevicesInfo and isLibVariantAvailable functions");let z=r6({variant:X,preferVariants:Z,variantPreference:W,defaultVariants:_}),A=[];for(let B of z)try{if(!await w(B))throw Error(`Variant ${B} not available on this platform`);let F=await G(B);A.push(k1({platform:$,totalMemoryInBytes:J,variant:B,devices:F,gpuMemoryFraction:j,cpuMemoryFraction:V,ok:!0}))}catch(L){let F=L instanceof Error?L.message:String(L);A.push(k1({platform:$,totalMemoryInBytes:J,variant:B,devices:[],gpuMemoryFraction:j,cpuMemoryFraction:V,ok:!1,error:F}))}let U=A.filter((B)=>B.ok)[0]||null,Q={ok:Boolean(U),selected:U?{...U,breakdown:Y?U.breakdown:void 0}:null,attempts:A};if(!Y&&Q.selected)delete Q.selected.breakdown;if(!Q||!N&&!H)return Q;let K=(B)=>{if(!B)return B;let L=E4({device:B,modelBytes:N||0,kvCacheBytes:H||0}),F=null;if(O!=null&&O!==H)F=E4({device:B,modelBytes:N||0,kvCacheBytes:O});return{...B,fit:L,...F&&{limitedFit:F}}};return Q.selected=K(Q.selected),Q.attempts=Array.isArray(Q.attempts)?Q.attempts.map(K):Q.attempts,Q},g0="ggml-llm";var p0=P(()=>{D1=["cuda","vulkan","snapdragon","default"]});var M4="ggml-stt",b1,x4=async({platform:$,variant:X=null,preferVariants:Z=[],variantPreference:W=[],gpuMemoryFraction:j=u0,cpuMemoryFraction:V=f0,includeBreakdown:Y=!1,totalMemoryInBytes:J,modelBytes:N=null,processingBytes:H=null,kvCacheBytes:O=null,dependencies:q={}}={})=>{let _=W&&W.length>0?W:b1;return E0({platform:$,variant:X,preferVariants:Z,variantPreference:_,gpuMemoryFraction:j,cpuMemoryFraction:V,includeBreakdown:Y,totalMemoryInBytes:J,modelBytes:N,kvCacheBytes:H??O,dependencies:q,defaultVariants:b1})};var S4=P(()=>{p0();b1=["cuda","vulkan","default"]});var t6,_0=async({platform:$,totalMemoryInBytes:X,backend:Z=g0,dependencies:W,...j}={})=>{let V=t6.get(Z);if(!V)throw Error(`No capability detector registered for backend "${Z}"`);return await V({...j,dependencies:W,totalMemoryInBytes:X,platform:$})};var v1=P(()=>{p0();S4();t6=new Map([[g0,E0],[M4,x4]])});var h1,P4=($)=>{let X=$?String($).toLowerCase():"f16";return h1[X]||h1.f16},T4=($,X,Z,W,j,V={},{totalLayers:Y=null,swaLayers:J=0,swaContext:N=null,swaContextMultiplier:H=1,swaAdditionalTokens:O=0,swaFull:q=!1}={})=>{if(!$||!X||!Z||!W||!j)return 0;let _=Y!=null&&Y!==void 0?Number(Y):Number($),G=Math.max(0,Math.floor(_));if(!G)return 0;let w=P4(V.k),z=P4(V.v),A=Number(Z)*(Number(W)*w+Number(j)*z);if(!A)return 0;let R=Math.max(0,Number(X)||0),U=Math.min(G,Math.max(0,Math.floor(Number(J)||0))),Q=Math.max(0,G-U),K=N!=null&&Number.isFinite(Number(N))?Math.max(0,Number(N)):R,B=Math.max(1,Number(H)||1),L=Math.max(0,Number(O)||0),F=K*B+L,E=q?R:Math.min(R,F),M=Q*R+U*Math.max(0,Math.floor(E));return Math.round(A*M)},m0=({modelBytes:$=0,audioLengthSeconds:X=30,sampleRate:Z=16000,bytesPerSample:W=4}={})=>{let j=Math.max(0,Number($)||0),V=Math.max(0,Math.floor(Math.max(0,X)*Z*W)),Y=1048576,J=1073741824,N;if(j<209715200)N=125829120;else if(j<524288000)N=146800640;else if(j<2147483648)N=157286400;else N=167772160;let H;if(j<209715200)H=73400320;else if(j<524288000)H=141557760;else if(j<2147483648)H=230686720;else H=230686720;let O;if(j<104857600)O=20971520;else if(j<209715200)O=31457280;else if(j<524288000)O=89128960;else if(j<2147483648)O=225443840;else O=377487360;let q=N+H+O,_=j+q+V;return{modelBytes:j,audioBufferBytes:V,processingBufferBytes:q,totalBytes:_}};var k4=P(()=>{h1={f16:2,f32:4,q8_0:1,q6_k:0.75,q5_k:0.625,q5_k_m:0.625,q5_k_s:0.625,q5_1:0.625,q5_0:0.625,q4_k:0.5,q4_k_m:0.5,q4_k_s:0.5,q4_1:0.5,q4_0:0.5,iq4_nl:0.5}});var D4=($)=>$?String($).trim().toLowerCase():null,e6=($={},X=null)=>{if(!$)return null;let Z=D4(X),W=Z?`${Z}.attention.sliding_window`:null,j=(W&&$[W]!=null?$[W]:null)??$["llama.attention.sliding_window"];if(j==null)return null;let V=Number(j);return Number.isFinite(V)?V:null},C1=($=0,X=0,Z=!1)=>{let W=Math.max(0,Math.floor(Number($)||0)),j=Math.max(0,Math.floor(Number(X)||0));if(!W||j===1)return 0;if(j<=0)return W;let V=Math.max(0,j-1),Y=Math.floor(W/j),J=W%j,N=Z?Math.max(0,J-1):Math.min(J,V);return Y*V+N},I1=({arch:$,nLayer:X=0})=>({arch:D4($),enabled:!1,window:null,pattern:null,denseFirst:!1,type:null,kvLayers:Math.max(0,Math.floor(Number(X)||0)),swaLayers:0}),$3,c0=({arch:$,metadata:X={},nLayer:Z=0}={})=>{let W=D4($||X["general.architecture"]),j=Math.max(0,Math.floor(Number(Z)||0)),V=e6(X,W),Y=W?$3.get(W):null;if(!Y)return I1({arch:W,nLayer:Z});let J=Y({nLayer:j,nSwa:V,metadata:X});if(!J||!J.enabled||!J.window||J.window<=0)return I1({arch:W,nLayer:Z});let N=Math.max(0,Math.floor(Number(J.pattern)||0)),H=J.kvLayers!=null&&Number.isFinite(Number(J.kvLayers))?Number(J.kvLayers):j,O=Math.max(0,Math.floor(H)),q=C1(O,N,Boolean(J.denseFirst));return{arch:W,enabled:q>0,window:J.window,pattern:N,denseFirst:Boolean(J.denseFirst),type:J.type||"standard",kvLayers:O,swaLayers:q}};var y1=P(()=>{$3=new Map([["llama4",({nSwa:$})=>{if($===0)return{enabled:!1};return{enabled:!0,window:$&&$>0?$:8192,pattern:4,type:"chunked"}}],["afmoe",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["phi3",()=>({enabled:!1})],["gemma2",({nSwa:$})=>{let X=$&&$>0?$:4096;if(!X)return{enabled:!1};return{enabled:!0,window:X,pattern:2,type:"standard"}}],["gemma3",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:6,type:"standard"}}],["gemma3n",({nLayer:$,nSwa:X})=>{if(!X||X<=0)return{enabled:!1};return{enabled:!0,window:X,pattern:5,type:"standard",kvLayers:Math.min(20,$)}}],["gemma-embedding",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:6,type:"symmetric"}}],["cohere2",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["olmo2",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["exaone4",({nLayer:$,nSwa:X})=>{let Z=$>=64,W=null;if(X&&X>0)W=X;else if(Z)W=4096;if(!W)return{enabled:!1};return{enabled:!0,window:W,pattern:4,type:"standard"}}],["gpt-oss",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:2,type:"standard"}}],["smallthinker",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:4096,pattern:4,denseFirst:!0,type:"standard"}}]])});var X3,Z3,u1=($)=>$?String($).trim().toLowerCase():null,f1=($)=>{let X=u1($);return X?X3.has(X):!1},g1=($)=>{let X=u1($);return X?Z3.has(X):!1},b4=($)=>{if(f1($))return"recurrent";if(g1($))return"hybrid";return"transformer"},d0=($={})=>{let X=$["general.architecture"],Z=(F,E=null)=>{let M=$[F],x=Number(M);return Number.isFinite(x)?x:E},W=(F,E=null)=>{let M=$[F];if(Array.isArray(M))return M;let x=Number(M);return Number.isFinite(x)?x:E},j=X?Z(`${X}.context_length`,Z("llama.context_length")):null,V=X?Z(`${X}.block_count`,Z("llama.block_count")):null,Y=X?Z(`${X}.embedding_length`,Z("llama.embedding_length")):null,J=X?Z(`${X}.attention.head_count`,Z("llama.attention.head_count")):null,N=X?W(`${X}.attention.head_count_kv`,W("llama.attention.head_count_kv",J)):null,H=null,O=null;if(Array.isArray(N)){let F=N.filter((E)=>Number(E)>0);if(F.length>0)H=Math.max(...F.map(Number)),O=F.length;else H=0,O=0}else H=N;let q=X?Z(`${X}.attention.key_length`,Z("llama.attention.key_length")):null,_=X?Z(`${X}.attention.value_length`,Z("llama.attention.value_length")):null,G=$["general.quantization_version"]||null,w=$["general.file_type"]||null,z=X?Z(`${X}.ssm.conv_kernel`):null,A=X?Z(`${X}.ssm.state_size`):null,R=X?Z(`${X}.ssm.inner_size`):null,U=X?Z(`${X}.ssm.group_count`):null,Q=X?Z(`${X}.ssm.time_step_rank`):null,K=X?Z(`${X}.rwkv.head_size`):null,B=X?Z(`${X}.rwkv.token_shift_count`,2):null,L=O!=null&&V!=null?V-O:null;return{arch:X,nCtxTrain:j,nLayer:V,nEmbd:Y,nHead:J,nHeadKv:H,nEmbdHeadK:q,nEmbdHeadV:_,quantVersion:G,fileType:w,attentionLayerCount:O,recurrentLayerCount:L,ssmDConv:z,ssmDState:A,ssmDInner:R,ssmNGroup:U,ssmDtRank:Q,rwkvHeadSize:K,rwkvTokenShiftCount:B}},G0=({layerCount:$,headKvCount:X,embdHeadKCount:Z,embdHeadVCount:W,cacheTypes:j,swaConfig:V,kvUnified:Y=!1,nParallel:J=1,swaFull:N=!1,arch:H=null,attentionLayerCount:O=null})=>{let q=b4(H);if(q==="recurrent")return()=>0;let _=q==="hybrid"&&O!=null?Math.max(0,Math.floor(Number(O)||0)):$,G=V?.window&&Y?Math.max(1,Number(J)||1):1,w=Y?1:Math.max(1,Number(J)||1);return(z)=>T4(_,z,X,Z,W,j,{totalLayers:_,swaLayers:V?.swaLayers||0,swaContext:V?.window,swaFull:N,swaContextMultiplier:G})*w},l0=({nLayer:$,nEmbd:X,recurrentLayerCount:Z=null,nSeqMax:W=1,ssmDConv:j=null,ssmDState:V=null,ssmDInner:Y=null,ssmNGroup:J=null,ssmDtRank:N=null,rwkvHeadSize:H=null,rwkvTokenShiftCount:O=2,arch:q=null})=>{if(b4(q)==="transformer")return 0;let G=Z!=null?Math.max(0,Math.floor(Number(Z)||0)):Math.max(0,Math.floor(Number($)||0));if(G===0)return 0;let w=Math.max(1,Math.floor(Number(W)||1)),z=0,A=0;if(H!=null&&H>0&&X!=null&&X>0)z=Math.max(1,Number(O)||2)*X,A=X*H;else if(V!=null&&Y!=null){let Q=Math.max(0,Number(j)||0),K=Math.max(0,Number(V)||0),B=Math.max(0,Number(Y)||0),L=Math.max(1,Number(J)||1);if(Math.max(0,Number(N)||0)>0)z=Q>0?(Q-1)*2*L*K:0,A=Math.floor(K*B/2);else z=Q>0?(Q-1)*(B+2*L*K):0,A=K*B}else return 0;let R=4,U=(z+A)*w*G*R;return Math.max(0,U)},i0=({maxCtx:$,availableMemory:X,modelBytes:Z,kvBytesForCtx:W})=>{let j=Math.max(1,Math.floor(Number($)||0));if(!W||X<=Z)return j;let V=1,Y=j,J=j;while(V<=Y){let N=Math.floor((V+Y)/2);if(Z+W(N)<=X)J=N,V=N+1;else Y=N-1}return J};var p1=P(()=>{k4();X3=new Set(["mamba","mamba2","rwkv6","rwkv6qwen2","rwkv7","arwkv7"]),Z3=new Set(["jamba","falcon-h1","plamo2","granitehybrid","lfm2","lfm2moe","nemotron_h","nemotron_h_moe","qwen3next"])});var M0=P(()=>{v1();k4();p0();S4();y1();p1()});import{EventEmitter as W3}from"node:events";class h4{constructor($=j3){this.maxEntries=$,this.modelLoads=[],this.completions=[],this.transcriptions=[]}addModelLoad($){v4(this.modelLoads,$,this.maxEntries),e.emit("status:modelLoad",$),e.emit("status:change",{type:"modelLoad",entry:$})}addCompletion($){v4(this.completions,$,this.maxEntries),e.emit("status:completion",$),e.emit("status:change",{type:"completion",entry:$})}addTranscription($){v4(this.transcriptions,$,this.maxEntries),e.emit("status:transcription",$),e.emit("status:change",{type:"transcription",entry:$})}getModelLoadHistory(){return[...this.modelLoads].reverse()}getCompletionHistory(){return[...this.completions].reverse()}getTranscriptionHistory(){return[...this.transcriptions].reverse()}clear(){this.modelLoads=[],this.completions=[],this.transcriptions=[]}}function I4($){let X=(Z)=>$(Z);return e.on("status:change",X),()=>e.off("status:change",X)}function c1($){m1+=1;let X=m1,Z=I4($);return{subscriberId:X,unsubscribe:Z}}function C4($){let X=[];return{generators:Array.from($.entries()).filter(([,W])=>W.type==="ggml-llm").map(([W,j])=>{let{instance:V}=j,Y=[];if(V.contexts)Y=Array.from(V.contexts.entries()).map(([J,N])=>{let H={key:J,refCount:N.refCount,hasModel:Boolean(N.context)},O=N.context.parallel.getStatus();return H.parallelStatus=O,X.push({generatorId:W,contextKey:J,...O}),H});return{id:W,type:j.type,refCount:j.refCount,repoId:V.info?.model?.repoId||null,quantization:V.info?.model?.quantization||null,variant:V.info?.runtime?.variant||null,nCtx:V.info?.runtime?.n_ctx||null,nParallel:V.info?.runtime?.n_parallel||null,contexts:Y}}),parallelStatuses:X,history:{modelLoads:$0.getModelLoadHistory(),completions:$0.getCompletionHistory()}}}function y4($){return{generators:Array.from($.entries()).filter(([,Z])=>Z.type==="ggml-stt").map(([Z,W])=>{let{instance:j}=W,V=j.getStatus?.()||{},Y=V.queueStatus||{processing:!1,queuedCount:0};return{id:Z,type:W.type,refCount:W.refCount,repoId:j.info?.model?.repoId||null,quantization:j.info?.model?.quantization||null,modelType:j.info?.model?.modelType||null,variant:j.info?.runtime?.variant||null,hasContext:V.hasContext||!1,contextRefCount:V.contextRefCount||0,queueStatus:Y}}),history:{modelLoads:N0.getModelLoadHistory(),transcriptions:N0.getTranscriptionHistory()}}}function d1($){return{timestamp:new Date().toISOString(),ggmlLlm:C4($),ggmlStt:y4($)}}var j3=9999,e,v4=($,X,Z)=>{if($.push({...X,timestamp:X.timestamp||new Date().toISOString()}),$.length>Z)$.shift()},$0,N0,m1=0;var s0=P(()=>{e=new W3;e.setMaxListeners(100);$0=new h4,N0=new h4});import y from"node:path";import T0 from"node:os";import{stat as B0,mkdir as Y3,open as Q3,unlink as o,readFile as t1,writeFile as e1,rename as f4,readdir as J3}from"node:fs/promises";import{createHash as p4}from"node:crypto";import{gguf as V3}from"@huggingface/gguf";import{loadModel as N3,getBackendDevicesInfo as $2,isLibVariantAvailable as X2}from"@fugood/llama.node";import U3 from"bytes";import*as G3 from"node:stream/web";class L0{constructor($,X){this.config=$,this.plan=X,this.baseDir=$.runtime.cache_dir,this.enabled=$.runtime.session_cache?.enabled!==!1,this.maxSizeBytes=x3($.runtime.session_cache?.max_size_bytes,10737418240),this.maxEntries=$.runtime.session_cache?.max_entries||1000,this.metadata={variant:X.info?.runtime?.variant||null,n_gpu_layers:X.info?.runtime?.n_gpu_layers||0,n_ctx:X.info?.runtime?.n_ctx||0,modelPath:X.localPath,cacheTypeK:X.info?.runtime?.cache_type_k||"f16",cacheTypeV:X.info?.runtime?.cache_type_v||"f16",kvUnified:X.info?.runtime?.kv_unified??null,swaFull:X.info?.runtime?.swa_full??null,flashAttnType:X.info?.runtime?.flash_attn_type||"off",isRecurrent:!1,isHybrid:!1},this.cacheMap=null,this.initialized=!1}updateModelInfo($){if(!$)return;if(this.metadata.isRecurrent=Boolean($.is_recurrent),this.metadata.isHybrid=Boolean($.is_hybrid),this.metadata.isRecurrent||this.metadata.isHybrid)console.log(`[SessionCache] Model architecture: recurrent=${this.metadata.isRecurrent}, hybrid=${this.metadata.isHybrid}`)}requiresExactMatch(){return this.metadata.isRecurrent||this.metadata.isHybrid}static checkTokenPrefixMatch($,X){if($.length>X.length)return!1;for(let Z=0;Z<$.length;Z+=1)if($[Z]!==X[Z])return!1;return!0}static async tokenizeToArray($,X){let Z=await $.tokenize(X);return Array.from(Z?.tokens||[])}async findFormattedMatchForRecurrent($,X,Z){let W=await L0.tokenizeToArray(Z,X),j=$.map(async(_)=>{try{let G=await L0.tokenizeToArray(Z,_.fullText);if(L0.checkTokenPrefixMatch(G,W))return{entry:_,usePromptState:!1,tokenCount:G.length};if(_.promptStatePath&&_.promptText){let w=await L0.tokenizeToArray(Z,_.promptText);if(L0.checkTokenPrefixMatch(w,W))return{entry:_,usePromptState:!0,tokenCount:w.length}}return null}catch(G){return console.warn(`[SessionCache] Failed to check entry ${_.id}: ${G.message}`),null}}),Y=(await Promise.all(j)).find((_)=>_!==null);if(!Y)return console.log("[SessionCache] No token prefix match found for recurrent/hybrid model"),null;let{entry:J,usePromptState:N,tokenCount:H}=Y;console.log(`[SessionCache] Token prefix match: ${J.id} (${H} tokens, usePromptState=${N})`);let O=N?J.promptStatePath:J.stateFilePath;if(!await n1(O))return await this.removeStaleEntry(J),null;return J.lastAccessedAt=new Date().toISOString(),await x0(this.cacheMap,this.baseDir).catch(()=>{}),{entry:J,usePromptState:N}}async initialize(){if(!this.enabled||this.initialized)return;try{await H0($4(this.baseDir)),await H0(o0(this.baseDir)),await H0(U2(this.baseDir)),this.cacheMap=await A3(this.baseDir),this.initialized=!0,console.log(`[SessionCache] Initialized with ${Object.keys(this.cacheMap.entries).length} entries`)}catch($){console.warn(`[SessionCache] Failed to initialize: ${$.message}`),this.enabled=!1}}async removeStaleEntry($){if(console.log(`[SessionCache] Removing stale entry: ${$.id}`),$.stateFilePath)await o($.stateFilePath).catch(()=>{});if($.promptStatePath)await o($.promptStatePath).catch(()=>{});delete this.cacheMap.entries[$.id],this.cacheMap.totalSize-=($.stateFileSize||0)+($.promptStateSize||0),await x0(this.cacheMap,this.baseDir).catch(()=>{})}async findMatchingEntry($,X=null){if(!this.enabled||!this.cacheMap)return null;let Z=this.requiresExactMatch();if(Z&&X){let J=Object.values(this.cacheMap.entries).filter((N)=>m4(N.metadata,this.metadata)&&N.fullText);return this.findFormattedMatchForRecurrent(J,$,X)}let W=w3($,this.metadata,this.cacheMap,Z);if(!W)return null;let{entry:j}=W;if(!await n1(j.stateFilePath))return await this.removeStaleEntry(j),null;return j.lastAccessedAt=new Date().toISOString(),await x0(this.cacheMap,this.baseDir).catch(()=>{}),{entry:j,usePromptState:!1}}async prepareCompletionOptions($,X,Z=null){let W={options:$,cacheEntry:null,promptPrefix:null};if(!this.enabled)return W;let j=await this.findMatchingEntry(X,Z);if(!j)return W;let{entry:V,usePromptState:Y}=j,J=Y?V.promptStatePath:V.stateFilePath,N=Y?V.promptText:V.fullText;return console.log(`[SessionCache] Found matching entry: ${V.id} (${N.length} chars, usePromptState=${Y})`),{options:{...$,load_state_path:J},cacheEntry:V,promptPrefix:N}}async saveCompletionState($,X,Z,W=0,j=null){if(!this.enabled)return null;let V=$+X,Y=z3(V,this.metadata),J=()=>{if(Z)o(Z).catch(()=>{});if(j)o(j).catch(()=>{})};if(this.cacheMap.entries[Y]){console.log(`[SessionCache] Entry already exists for prompt: ${Y}, updating position`);let O=this.cacheMap.entries[Y];return O.lastAccessedAt=new Date().toISOString(),delete this.cacheMap.entries[Y],this.cacheMap.entries[Y]=O,await x0(this.cacheMap,this.baseDir).catch(()=>{}),J(),O}let N=s1(Y,this.baseDir),H=j?s1(`${Y}-prompt`,this.baseDir):null;try{await H0(y.dirname(N)),await f4(Z,N);let O=await B0(N),q=0;if(j&&H)try{await f4(j,H),q=(await B0(H)).size,console.log(`[SessionCache] Saved prompt state: ${H}`)}catch(G){console.warn(`[SessionCache] Failed to save prompt state: ${G.message}`)}let _={id:Y,promptText:$,completionText:X,fullText:V,promptTokenCount:W,stateFilePath:N,stateFileSize:O.size,promptStatePath:H||null,promptStateSize:q,metadata:{...this.metadata},createdAt:new Date().toISOString(),lastAccessedAt:new Date().toISOString()};if(this.cacheMap.entries[Y]=_,this.cacheMap.totalSize+=O.size+q,!this.requiresExactMatch())await E3(this.cacheMap,$,Y,this.metadata);return await F3(this.cacheMap,this.maxSizeBytes,this.maxEntries),await x0(this.cacheMap,this.baseDir),console.log(`[SessionCache] Saved entry: ${Y} (${O.size} bytes, ${V.length} chars)`),_}catch(O){return console.warn(`[SessionCache] Failed to save state: ${O.message}`),J(),null}}async generateTempStatePath(){return await H0(o0(this.baseDir)),B3(this.baseDir)}async cleanup(){await M3(this.baseDir)}}async function K2($,X,Z={}){let{globalDownloadManager:W=null}=Z,j=e0(X),V=await S3(j),Y=new L0(j,V);await Y.initialize();let J={id:$,type:"ggml-llm",config:j,plan:V,info:V.info,contexts:new Map,downloads:new Map,globalDownloadManager:W,sessionCache:Y,finalized:!1},N=async()=>{if(J.finalized)return;J.finalized=!0;let Q=Array.from(J.contexts.values()),K=Q.map((L)=>{if(L.released)return Promise.resolve(!1);if(L.releaseRequested||L.releaseTimer)return Promise.resolve(!1);if(L.refCount=Math.max(0,L.refCount-1),L.refCount>0)return Promise.resolve(!1);return r0(J,L)});if(await Promise.allSettled(K),Q.length===0||Q.every((L)=>L.released))await J.sessionCache.cleanup()},H=async(Q={})=>{let{onProgress:K}=Q,B=await D3(J,K);return J.sessionCache.updateModelInfo(B.modelInfo),{modelInfo:B.modelInfo?{...B.modelInfo}:null,runtime:{...J.plan.info.runtime},download:{...J.plan.info.download}}},O=async()=>{if(J.finalized)return!1;let Q=z0(J),K=J.contexts.get(Q);if(!K)return!1;return b3(J,K,!1)},q=async(Q={})=>{let{options:K={},useCache:B=!0}=Q,L=z0(J),F=J.contexts.get(L);if(!F)throw Error(`Context "${L}" not initialized`);await F.ready;let E=K.prompt||"",M=null,x=null;if(!E&&K.messages){({messages:M}=K),x={chatTemplate:K.chat_template||K.chatTemplate,jinja:K.jinja??!0,tools:K.tools,parallel_tool_calls:K.parallel_tool_calls,tool_choice:K.tool_choice,reasoning_format:K.reasoning_format,enable_thinking:K.enable_thinking,add_generation_prompt:K.add_generation_prompt,now:K.now,chat_template_kwargs:K.chat_template_kwargs};let I=await F.context.getFormattedChat(M,x.chatTemplate,x);E=I?.prompt||I||""}if(B&&J.sessionCache.enabled&&E){let{options:I}=await J.sessionCache.prepareCompletionOptions(K,E,F.context),b=await J.sessionCache.generateTempStatePath(),r=(await F.context.tokenize(E))?.tokens?.length||0,k={...I,save_state_path:b},d=J.sessionCache.requiresExactMatch(),v=Boolean(k.load_state_path),S=null;if(d&&!v)S=await J.sessionCache.generateTempStatePath(),k.save_prompt_state_path=S;let C={repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null};return T3(F.context,k,J.sessionCache,E,b,r,J.id,C,S)}let D={repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null};return P3(F.context,K,J.id,D)},_=async(Q={})=>{let{text:K="",params:B={}}=Q,L=z0(J),F=J.contexts.get(L);if(!F)throw Error(`Context "${L}" not initialized`);await F.ready;let E=await F.context.tokenize(K,B);if(!E)return{tokens:[]};let M=Array.from(E.tokens??[]).map((x)=>Number(x));return{...E,tokens:M}},G=async(Q={})=>{let{tokens:K=[]}=Q,B=z0(J),L=J.contexts.get(B);if(!L)throw Error(`Context "${B}" not initialized`);await L.ready;let F=K.map((E)=>Number(E));return L.context.detokenize(F)},w=async(Q={})=>{let{messages:K=[],template:B,params:L}=Q,F=z0(J),E=J.contexts.get(F);if(!E)throw Error(`Context "${F}" not initialized`);return await E.ready,await E.context.getFormattedChat(K,B,L)},z=()=>Array.from(J.contexts.values()).some((Q)=>!Q.released&&(Q.releaseRequested||Q.releaseTimer||Q.refCount>0)),A=()=>{J.finalized=!1},R=()=>{let Q=[],K=Array.from(J.contexts.entries()).map(([B,L])=>{let F={key:B,refCount:L.refCount,hasModel:Boolean(L.context)},E=L.context.parallel.getStatus();return F.parallelStatus=E,Q.push({contextKey:B,...E}),F});return{id:J.id,type:J.type,repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null,nCtx:J.plan.info.runtime?.n_ctx||null,nParallel:J.plan.info.runtime?.n_parallel||null,contexts:K,parallelStatuses:Q}},U=(Q)=>{let K=Array.from(J.contexts.entries()).map(([B,L])=>L.context.parallel.subscribeToStatus((F)=>{Q({contextKey:B,...F})}));return{remove:()=>{K.forEach((B)=>{if(B?.remove)B.remove()})}}};return{id:$,type:"ggml-llm",info:V.info,contexts:J.contexts,initContext:H,completion:q,tokenize:_,detokenize:G,applyChatTemplate:w,releaseContext:O,finalize:N,getStatus:R,subscribeParallelStatus:U,hasPendingReleases:z,resetFinalized:A}}async function q2($,X,Z={}){let{onProgress:W,onComplete:j,onError:V}=Z;try{let Y=e0($),J=await c4(Y),N=H2(Y,J),{repoId:H}=J;if(await P0(N,J.size)){if(console.log(`[Download] Model already exists: ${H} at ${N}`),typeof j==="function")j({localPath:N,repoId:H,alreadyExists:!0});return{started:!1,localPath:N,repoId:H,alreadyExists:!0}}let q=X.getDownload(N);if(q)return console.log(`[Download] Already downloading: ${H}`),q.then(()=>{if(typeof j==="function")j({localPath:N,repoId:H,joinedExisting:!0})}).catch((G)=>{if(typeof V==="function")V(G)}),{started:!1,localPath:N,repoId:H,alreadyDownloading:!0};console.log(`[Download] Starting download: ${H}`);let _=(async()=>{try{if(J.isSplit&&J.splitCount>0){let G=/-(\d{5})-of-(\d{5})\.gguf$/,w=y.dirname(N),z=J.splitCount,A=0;for(let R=1;R<=z;R+=1){let U=String(R).padStart(5,"0"),Q=J.filename.replace(G,`-${U}-of-${String(z).padStart(5,"0")}.gguf`),K=`${Y.model.base_url.replace(/\/+$/,"")}/${J.repoId}/resolve/${J.revision}/${Q}`,B=y.join(w,Q);if(!await P0(B))await a0(K,J.headers,B,null,(F)=>{if(F>=0&&Number.isFinite(F)){let E=(A+F)/z;if(console.log(`[Download] ${H}: ${Math.round(E*100)}%`),typeof W==="function")W(E)}});A+=1}}else await a0(J.url,J.headers,N,J.size,(G)=>{if(G>=0&&Number.isFinite(G)){if(console.log(`[Download] ${H}: ${Math.round(G*100)}%`),typeof W==="function")W(G)}});if(console.log(`[Download] Completed: ${H}`),typeof j==="function")j({localPath:N,repoId:H})}catch(G){if(console.error(`[Download] Failed: ${H}`,G.message),typeof V==="function")V(G);throw G}finally{X.deleteDownload(N)}})();return X.setDownload(N,_),{started:!0,localPath:N,repoId:H}}catch(Y){if(console.error("[Download] Failed to start download:",Y.message),typeof V==="function")V(Y);return{started:!1,localPath:null,repoId:null,error:Y.message}}}async function v3($){let X=e0($),Z=await c4(X),W=await G2(Z.url,Z.headers,X.runtime.cache_dir),{arch:j,nCtxTrain:V,nLayer:Y,nEmbd:J,nHead:N,nHeadKv:H,nEmbdHeadK:O,nEmbdHeadV:q,quantVersion:_,fileType:G,attentionLayerCount:w,recurrentLayerCount:z,ssmDConv:A,ssmDState:R,ssmDInner:U,ssmNGroup:Q,ssmDtRank:K,rwkvHeadSize:B,rwkvTokenShiftCount:L}=d0(W),F=Number.isFinite(Number(Y))?Number(Y):0,E=Number.isFinite(Number(J))?Number(J):0,M=Number.isFinite(Number(N))?Number(N):0,x=Number.isFinite(Number(H))?Number(H):M,D=M>0&&E>0?E/M:128,I=O!=null&&Number.isFinite(Number(O))?Number(O):D,b=q!=null&&Number.isFinite(Number(q))?Number(q):D,u=c0({arch:j,metadata:W,nLayer:F}),r=u&&Number.isFinite(Number(u.kvLayers))?Number(u.kvLayers):F,k=Math.max(0,Math.floor(Number(r)||0)),v=(X.model.n_ctx?Number(X.model.n_ctx):null)||V||4096,S={k:X.model.cache_type_k,v:X.model.cache_type_v},C=Z.size>0?Z.size:0,h=X.model.n_parallel||4,p=G0({layerCount:k,headKvCount:x,embdHeadKCount:I,embdHeadVCount:b,cacheTypes:S,swaConfig:u,kvUnified:X.model.kv_unified,nParallel:h,swaFull:X.model.swa_full,arch:j,attentionLayerCount:w}),f=l0({nLayer:F,nEmbd:E,recurrentLayerCount:z,nSeqMax:h,ssmDConv:A,ssmDState:R,ssmDInner:U,ssmNGroup:Q,ssmDtRank:K,rwkvHeadSize:B,rwkvTokenShiftCount:L,arch:j}),s=X.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(X.backend.gpu_memory_fraction))):S0.backend.gpu_memory_fraction||1,W0=X.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(X.backend.cpu_memory_fraction))):t0,A0=p(v),l=await O2(X,{modelBytes:C,kvCacheBytes:A0}),F0=(l.selected.totalMemory||0)*s,A4=Math.max(0,T0.totalmem()*W0),h0=l.selected.hasGpu?F0:A4,I0=i0({maxCtx:v,availableMemory:h0,modelBytes:C,kvBytesForCtx:p}),U0=p(v),C0=p(I0);return{kvInfo:{nCtxTrain:V,nLayer:F,nEmbd:E,nHeadKv:x,nEmbdHeadK:I,nEmbdHeadV:b,nHeadCount:M,nHeadKvCount:x,kvLayerCount:k,swa:u?.enabled?{window:u.window,pattern:u.pattern,denseFirst:u.denseFirst,type:u.type,layers:u.swaLayers}:null},modelBytes:C,kvCacheBytes:U0,limitedKvCacheBytes:C0,memoryLimitedCtx:I0,recurrentMemoryBytes:f,quantization:{name:Z.quantization||null,fileType:G,version:_}}}async function _2($=null,X={}){let{threshold:Z=1.1,includeBreakdown:W=!1,config:j,...V}=X,Y=null,J=null,N=null,H=null,O=null,q=null,_=null;if(j)try{let{modelBytes:B,kvCacheBytes:L,limitedKvCacheBytes:F,memoryLimitedCtx:E,recurrentMemoryBytes:M,kvInfo:x,quantization:D}=await v3(j);Y=B,J=L,N=F,H=E,O=M,q=x,_=D}catch(B){}let G=j?.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.gpu_memory_fraction))):void 0,w=j?.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.cpu_memory_fraction))):void 0,z=await _0({...V,platform:process.platform,totalMemoryInBytes:T0.totalmem(),backend:"ggml-llm",includeBreakdown:W,gpuMemoryFraction:G,cpuMemoryFraction:w,dependencies:{getBackendDevicesInfo:$2,isLibVariantAvailable:X2},modelBytes:Y,kvCacheBytes:J,limitedKvCacheBytes:N}),A=z.selected,R=a1(A);A.modelBytes=Y||null,A.kvCacheBytes=J||null,A.memoryLimitedCtx=H||null,A.limitedKvCacheBytes=N||null,A.recurrentMemoryBytes=O||null,A.kvInfo=q||null,A.quantization=_||null;let U=null,Q=null;if($){let B=a1($);Q={...$,score:B};let L="buttress",F="buttress-higher-score";if(!z.ok)L="local",F="buttress-unavailable";else if(!B&&B!==0)L="buttress",F="missing-client-score";else{let{fit:E,limitedFit:M}=Q,x=A?.fit,D=A?.limitedFit,I=E?.fitsInGpu||E?.fitsInCpu||M?.fitsInGpu||M?.fitsInCpu,b=x?.fitsInGpu||x?.fitsInCpu||D?.fitsInGpu||D?.fitsInCpu;if(I&&!b)L="local",F="client-fits-in-memory";else if(b&&!I)L="buttress",F="buttress-fits-in-memory";else if(B>R*Z)L="local",F="client-better";else if(R>B*Z)L="buttress",F="buttress-better";else L="either",F="comparable-scores"}U={buttressScore:R,clientScore:B,threshold:Z,recommendation:L,reason:F}}if(!z.ok&&!U)U={buttressScore:R,clientScore:$?.score??null,threshold:Z,recommendation:"local",reason:"buttress-unavailable"};let K=null;if(j)K={repoId:j.model?.repo_id||null,quantization:j.model?.quantization||null,nCtx:j.model?.n_ctx||null,cacheKType:j.model?.cache_type_k||"f16",cacheVType:j.model?.cache_type_v||"f16"};return{type:"ggml-llm",timestamp:new Date().toISOString(),buttress:z,client:Q,comparison:U,modelConfig:K}}var H3=()=>{if(typeof globalThis<"u"&&globalThis.ReadableStream&&globalThis.WritableStream)return{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream};return G3},O3,Z2,K3,W2=($={},X={})=>{return Object.entries(X||{}).forEach(([Z,W])=>{if(W&&typeof W==="object"&&!Array.isArray(W)){if(!$[Z]||typeof $[Z]!=="object")$[Z]={};W2($[Z],W)}else $[Z]=W}),$},q3=".gguf",j2="https://huggingface.co",Y2="https://huggingface.co/api",n,Q2,t0=0.5,S0,u4=($,X=[])=>{if(!$&&$!==0)return[...X];if(Array.isArray($))return $.filter((Z)=>Z!=null);return[$]},n0=($)=>{if(!$)return null;let X=String($).toLowerCase();if(["cuda","vulkan","snapdragon","default"].includes(X))return X;return null},e0=($={})=>{let X=JSON.parse(JSON.stringify(S0));if(W2(X,$),X.backend.variant=n0(X.backend.variant),X.backend.variant_preference=Array.from(new Set(u4(X.backend.variant_preference).map(n0).filter(Boolean))),X.backend.variant_preference.length===0)X.backend.variant_preference=["cuda","vulkan","snapdragon","default"];if(X.runtime.prefer_variants=Array.from(new Set(u4(X.runtime.prefer_variants).map(n0).filter(Boolean))),X.model.preferred_quantizations=Array.from(new Set(u4(X.model.preferred_quantizations||X.model.quantizations).map((Z)=>Z?String(Z).toLowerCase():null).filter(Boolean))),X.model.quantization){let Z=String(X.model.quantization).toLowerCase();if(!X.model.preferred_quantizations.includes(Z))X.model.preferred_quantizations.unshift(Z)}return X.model.n_parallel=Math.max(1,Number(X.model.n_parallel)||4),X.model.n_batch=Math.max(1,Number(X.model.n_batch)||512),X.model.base_url=X.model.base_url||j2,X.model.api_base=X.model.api_base||Y2,X.runtime.cache_dir=X.runtime.cache_dir?y.resolve(X.runtime.cache_dir):n,X.runtime.session_cache={...S0.runtime.session_cache,...X.runtime.session_cache||{}},X.runtime.context_release_delay_ms=Math.max(0,Number(X.runtime.context_release_delay_ms)||S0.runtime.context_release_delay_ms),X},l1=($)=>{let X=$.toLowerCase();return Q2.find((W)=>X.includes(W))||null},_3=($)=>{let X=[];if($.backend.variant)X.push($.backend.variant);if($.runtime.prefer_variants.length>0)X.push(...$.runtime.prefer_variants);return X.push(...$.backend.variant_preference),X.push("default"),Array.from(new Set(X.map(n0).filter(Boolean)))},H0=async($)=>{await Y3($,{recursive:!0})},L3=($=n)=>y.join($,".metadata-cache"),J2=($,X,Z=n)=>{let W=p4("sha256").update($).digest("hex");return y.join(L3(Z),X,`${W}.json`)},V2=async($,X,Z=n)=>{try{let W=J2($,X,Z),j=await t1(W,"utf-8");return console.log(`[Cache] Hit ${X} cache:`,y.basename(W)),JSON.parse(j,(V,Y)=>{if(typeof Y==="string"&&Y.startsWith("__bigint__"))return BigInt(Y.slice(10));return Y})}catch(W){return null}},g4=async($,X,Z,W=n)=>{try{let j=J2($,X,W);await H0(y.dirname(j)),await e1(j,JSON.stringify(Z,(V,Y)=>{if(typeof Y==="bigint")return`__bigint__${Y.toString()}`;return Y}),"utf-8"),console.log(`[Cache] Wrote ${X} cache:`,y.basename(j))}catch(j){console.warn(`[Cache] Failed to write ${X} cache:`,j.message)}},$4=($=n)=>y.join($,".session-state-cache"),N2=($=n)=>y.join($4($),"cache-map.json"),o0=($=n)=>y.join($4($),"temp"),U2=($=n)=>y.join($4($),"states"),i1=()=>({version:1,entries:{},totalSize:0}),A3=async($=n)=>{try{let X=N2($),Z=await t1(X,"utf-8"),W=JSON.parse(Z);if(!W.entries||typeof W.entries!=="object")return i1();return W}catch{return i1()}},x0=async($,X=n)=>{let Z=N2(X),W=`${Z}.tmp.${Date.now()}`;try{await H0(y.dirname(Z)),await e1(W,JSON.stringify($,null,2),"utf-8"),await f4(W,Z)}catch(j){throw await o(W).catch(()=>{}),j}},z3=($,X)=>{let Z=JSON.stringify({text:$,model:X.modelPath,variant:X.variant,n_gpu_layers:X.n_gpu_layers,n_ctx:X.n_ctx,cacheTypeK:X.cacheTypeK,cacheTypeV:X.cacheTypeV,kvUnified:X.kvUnified,swaFull:X.swaFull,flashAttnType:X.flashAttnType});return p4("sha256").update(Z).digest("hex").slice(0,24)},s1=($,X=n)=>y.join(U2(X),`${$}.bin`),B3=($=n)=>{let X=`${Date.now()}-${Math.random().toString(36).slice(2,10)}`;return y.join(o0($),`${X}.bin`)},m4=($,X)=>$.modelPath===X.modelPath&&$.variant===X.variant&&$.n_gpu_layers===X.n_gpu_layers&&$.n_ctx>=X.n_ctx&&$.cacheTypeK===X.cacheTypeK&&$.cacheTypeV===X.cacheTypeV&&$.kvUnified===X.kvUnified&&$.swaFull===X.swaFull&&$.flashAttnType===X.flashAttnType&&Boolean($.isRecurrent)===Boolean(X.isRecurrent)&&Boolean($.isHybrid)===Boolean(X.isHybrid),R3=($,X)=>{let Z=Math.min($.length,X.length),W=0;while(W<Z&&$[W]===X[W])W+=1;return W},w3=($,X,Z,W=!1)=>{let j=Object.values(Z.entries);console.log(`[SessionCache] Finding match for promptText (${$.length} chars), exactMatch=${W}`),console.log(`[SessionCache] Checking ${j.length} cache entries`);let V=j.filter((J)=>m4(J.metadata,X));if(W){let J=V.find((N)=>N.fullText===$);if(J)return console.log(`[SessionCache] Exact match found: ${J.id} (${J.fullText.length} chars)`),{entry:J,prefixLength:J.fullText.length,exactMatch:!0};return null}let Y=V.reduce((J,N)=>{let H=R3($,N.fullText);if(H>J.prefixLen)return{entry:N,prefixLen:H};if(H===J.prefixLen&&N.fullText.length>(J.entry?.fullText?.length||0))return{entry:N,prefixLen:H};return J},{entry:null,prefixLen:0});if(Y.entry)return console.log(`[SessionCache] Prefix match found: ${Y.entry.id} (${Y.prefixLen}/${Y.entry.fullText.length} chars)`),{entry:Y.entry,prefixLength:Y.prefixLen};return console.log("[SessionCache] No match found"),null},F3=async($,X,Z)=>{let W=Object.values($.entries).sort((J,N)=>new Date(J.lastAccessedAt)-new Date(N.lastAccessedAt)),j=$.totalSize,V=Object.keys($.entries).length,Y=W.filter((J)=>{let N=j>X,H=V>Z;if(!N&&!H)return!1;return j-=(J.stateFileSize||0)+(J.promptStateSize||0),V-=1,!0});return await Promise.all(Y.map(async(J)=>{if(await o(J.stateFilePath).catch(()=>{}),J.promptStatePath)await o(J.promptStatePath).catch(()=>{});delete $.entries[J.id],console.log(`[SessionCache] Evicted entry: ${J.id}`)})),$.totalSize=Math.max(0,j),Y.map((J)=>J.id)},E3=async($,X,Z,W)=>{let j=Object.entries($.entries).filter(([V,Y])=>{if(V===Z)return!1;if(!m4(Y.metadata,W))return!1;return X.startsWith(Y.fullText)&&Y.fullText.length<X.length}).map(([,V])=>V);return await Promise.all(j.map(async(V)=>{if(await o(V.stateFilePath).catch(()=>{}),V.promptStatePath)await o(V.promptStatePath).catch(()=>{});$.totalSize-=(V.stateFileSize||0)+(V.promptStateSize||0),delete $.entries[V.id],console.log(`[SessionCache] Evicted superseded prefix entry: ${V.id} (${V.promptText.length} prompt chars)`)})),j.map((V)=>V.id)},M3=async($=n)=>{let X=o0($);try{let Z=await J3(X),W=Date.now(),j=3600000;await Promise.all(Z.map(async(V)=>{let Y=y.join(X,V),J=await B0(Y).catch(()=>null);if(J&&W-J.mtimeMs>3600000)await o(Y).catch(()=>{}),console.log(`[SessionCache] Cleaned up temp file: ${V}`)}))}catch{}},n1=async($)=>{try{return await B0($),!0}catch{return!1}},x3=($,X)=>{if($==null)return X;if(typeof $==="number")return $;if(typeof $==="string"){let Z=U3.parse($);return Z!=null?Z:X}return X},r1=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,X);if(!Z.ok){let W=await Z.text().catch(()=>"");throw Error(`Failed to fetch ${$}: ${Z.status} ${Z.statusText} ${W}`.trim())}return Z.json()},o1=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,{...X,method:"HEAD"});if(!Z.ok)throw Error(`Failed to fetch headers for ${$}: ${Z.status} ${Z.statusText}`);return Z},G2=async($,X,Z=n)=>{let W=JSON.stringify({url:$,headers:X}),j=await V2(W,"range-metadata",Z);if(j)return j;let V=!/^https?:/i.test($),{metadata:Y}=await V3($,{fetch,additionalFetchHeaders:X,allowLocalFile:V});return await g4(W,"range-metadata",Y,Z),Y},H2=($,X)=>{if($.model.local_path)return y.resolve($.model.local_path);let Z=X.repoId.split("/"),W=y.join($.runtime.cache_dir,...Z,X.revision);return y.join(W,X.filename)},P0=async($,X)=>{try{let Z=await B0($);if(!X)return!0;return Z.size===X}catch(Z){return!1}},a0=async($,X,Z,W,j)=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");await H0(y.dirname(Z));let V=await fetch($,{headers:X});if(!V.ok||!V.body)throw Error(`Failed to download ${$}: ${V.status} ${V.statusText}`);let Y=await Q3(Z,"w"),J=Number(V.headers.get("content-length"))||W||0,N=0,H=0.05;try{await V.body.pipeTo(new K3({async write(O){if(await Y.write(O),N+=O.byteLength,typeof j==="function"&&J>0){let q=Math.min(1,N/J);while(q>=H)j(H),H+=0.05}},async close(){if(await Y.close(),typeof j==="function")j(1)},async abort(O){throw await Y.close().catch(()=>{}),await o(Z).catch(()=>{}),O}}))}catch(O){throw await Y.close().catch(()=>{}),await o(Z).catch(()=>{}),O}if(W){let O=await B0(Z);if(O.size!==W)throw await o(Z).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${W} got ${O.size}`)}},c4=async($)=>{let X=$.model.repo_id||$.model.repository||$.model.model;if(!X)throw Error("`model.repo_id` is required in Buttress backend config");let Z=$.model.revision||"main",W=$.runtime.cache_dir,j=JSON.stringify({repoId:X,revision:Z,filename:$.model.filename,url:$.model.url,quantization:$.model.quantization,preferred_quantizations:$.model.preferred_quantizations}),V=await V2(j,"artifact-info",W);if(V)return V;let Y={...$.runtime.http_headers||{}};if($.runtime.huggingface_token)Y.Authorization=`Bearer ${$.runtime.huggingface_token}`;if($.model.url){let Q=await o1($.model.url,{headers:Y}),K=Number(Q.headers.get("content-length"))||null,B=$.model.filename||$.model.url.split("/").pop(),L={repoId:X,revision:Z,filename:B,url:$.model.url,size:K,headers:Y};return await g4(j,"artifact-info",L,W),L}let{filename:J}=$.model,N=$.model.quantization&&String($.model.quantization).toLowerCase(),H=await r1(`${$.model.api_base}/models/${X}?revision=${Z}&blobs=true`,{headers:Y}),q=(H?.siblings||H?.files||[]).map((Q)=>Q.rfilename||Q.path||Q.filename).filter((Q)=>typeof Q==="string"&&Q.endsWith(q3));if(q.length===0)throw Error(`No GGUF artifacts found in repo ${X}`);let _=$.model.preferred_quantizations.length>0?$.model.preferred_quantizations:Q2,G=()=>{let Q=_.find((K)=>{return q.find((L)=>L.toLowerCase().includes(K))});if(Q)return{filename:q.find((B)=>B.toLowerCase().includes(Q)),quantization:Q};return null};if(!J){let Q=G()||{filename:q[0],quantization:null},{filename:K,quantization:B}=Q;J=K,N=B||l1(J)}else if(!N)N=l1(J);let w=`${$.model.base_url.replace(/\/+$/,"")}/${X}/resolve/${Z}/${J}`,z=/-(\d{5})-of-(\d{5})\.gguf$/,A=J.match(z),R=null;if(A){let[,,Q]=A,K=await r1(`${$.model.api_base}/models/${X}?revision=${Z}&blobs=true`,{headers:Y}),B=K?.siblings||K?.files||[],L=Number(Q);R=0;for(let F=1;F<=L;F+=1){let E=String(F).padStart(5,"0"),M=J.replace(z,`-${E}-of-${Q}.gguf`),x=B.find((I)=>(I.rfilename||I.path||I.filename)===M),D=Number(x?.size);if(Number.isFinite(D)&&D>0)R+=D}}else{let Q=await o1(w,{headers:Y});R=Number(Q.headers.get("content-length"))||null}let U={repoId:X,revision:Z,filename:J,url:w,size:R,quantization:N,headers:Y,isSplit:Boolean(A),splitCount:A?Number(A[2]):0};return await g4(j,"artifact-info",U,W),U},O2=async($,{modelBytes:X=null,kvCacheBytes:Z=null}={})=>{let W=_3($),[j,...V]=W,Y=$.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.gpu_memory_fraction))):S0.backend.gpu_memory_fraction||1,J=$.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):t0,N=await _0({platform:process.platform,totalMemoryInBytes:T0.totalmem(),backend:"ggml-llm",variant:j||null,preferVariants:V,gpuMemoryFraction:Y,cpuMemoryFraction:J,dependencies:{getBackendDevicesInfo:$2,isLibVariantAvailable:X2},modelBytes:X,kvCacheBytes:Z}),H=(_)=>({..._,devices:Array.isArray(_.devices)?_.devices:[],ok:_.ok,hasGpu:Boolean(_.hasGpu),totalMemory:_.gpuTotalBytes||_.totalMemory||0,error:_.ok?null:Error(_.error||`Variant ${_.variant} not available on this platform`)});if(!N.ok||!N.selected){let _=(N.attempts||[]).map((G)=>`${G.variant}: ${G.error||"unknown error"}`).join("; ");throw Error(`Unable to initialize any backend variant (${W.join(", ")}). Errors: ${_}`)}let O=(N.attempts||[]).map(H);return{selected:H(N.selected),attempts:O}},S3=async($)=>{let X=await c4($),Z=await G2(X.url,X.headers,$.runtime.cache_dir),{arch:W,nCtxTrain:j,nLayer:V,nEmbd:Y,nHead:J,nHeadKv:N,nEmbdHeadK:H,nEmbdHeadV:O,quantVersion:q,fileType:_,attentionLayerCount:G,recurrentLayerCount:w,ssmDConv:z,ssmDState:A,ssmDInner:R,ssmNGroup:U,ssmDtRank:Q,rwkvHeadSize:K,rwkvTokenShiftCount:B}=d0(Z),L=Number.isFinite(Number(V))?Number(V):0,F=Number.isFinite(Number(Y))?Number(Y):0,E=Number.isFinite(Number(J))?Number(J):0,M=Number.isFinite(Number(N))?Number(N):E,x=E>0&&F>0?F/E:128,D=H!=null&&Number.isFinite(Number(H))?Number(H):x,I=O!=null&&Number.isFinite(Number(O))?Number(O):x,b=c0({arch:W,metadata:Z,nLayer:L}),u=b&&Number.isFinite(Number(b.kvLayers))?Number(b.kvLayers):L,r=Math.max(0,Math.floor(Number(u)||0)),k={use_mmap:$.model.use_mmap??$.runtime.use_mmap,use_mlock:$.model.use_mlock??$.runtime.use_mlock,n_threads:$.model.n_threads??$.runtime.n_threads,n_ctx:$.model.n_ctx??$.runtime.n_ctx,n_batch:$.model.n_batch??$.runtime.n_batch,n_ubatch:$.model.n_ubatch??$.runtime.n_ubatch,n_cpu_moe:$.model.n_cpu_moe??$.runtime.n_cpu_moe,n_parallel:$.model.n_parallel??$.runtime.n_parallel,cpu_mask:$.model.cpu_mask??$.runtime.cpu_mask,cpu_strict:$.model.cpu_strict??$.runtime.cpu_strict,devices:$.model.devices??$.runtime.devices,n_gpu_layers:$.model.n_gpu_layers??$.runtime.n_gpu_layers,flash_attn_type:$.model.flash_attn_type??$.runtime.flash_attn_type,cache_type_k:$.model.cache_type_k??$.runtime.cache_type_k,cache_type_v:$.model.cache_type_v??$.runtime.cache_type_v,kv_unified:$.model.kv_unified??$.runtime.kv_unified,swa_full:$.model.swa_full??$.runtime.swa_full,ctx_shift:$.model.ctx_shift??$.runtime.ctx_shift},d=k.n_ctx?Number(k.n_ctx):null,v=d||j||4096,S=[],C=[],h=!0;if(d&&j&&d>j){h=!1;let V0=`Requested context length (${d}) exceeds model training context (${j})`;S.push(V0),C.push(V0),v=j}if(d&&!j)S.push("Model metadata missing training context length, using requested value");let p={k:k.cache_type_k,v:k.cache_type_v},f=X.size>0?X.size:0,s=G0({layerCount:r,headKvCount:M,embdHeadKCount:D,embdHeadVCount:I,cacheTypes:p,swaConfig:b,kvUnified:k.kv_unified,nParallel:k.n_parallel,swaFull:k.swa_full,arch:W,attentionLayerCount:G}),W0=l0({nLayer:L,nEmbd:F,recurrentLayerCount:w,nSeqMax:k.n_parallel||4,ssmDConv:z,ssmDState:A,ssmDInner:R,ssmNGroup:U,ssmDtRank:Q,rwkvHeadSize:K,rwkvTokenShiftCount:B,arch:W}),A0=s(v),l=await O2($,{modelBytes:f,kvCacheBytes:A0+W0}),L4=l.selected.totalMemory||0,F0=L4*($.backend.gpu_memory_fraction||1),A4=$.backend.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):t0,h0=Math.max(0,T0.totalmem()*A4),I0=l.selected.hasGpu?F0:h0,U0=i0({maxCtx:v,availableMemory:I0,modelBytes:f,kvBytesForCtx:s});if(!d&&U0){let V0=j?Math.min(U0,j):U0,R4=Math.max(32,V0);if(R4<v)S.push(`Context length capped to ${R4} by memory limits`);v=R4}if(v>U0)v=U0;let C0=Math.floor(U0);console.log(`[buttress] Memory-limited context length: ${C0}`);let A1=s(v),p6=f+A1+W0,z1=L?f/(L+1):f,y0=0;if(l.selected.hasGpu&&z1>0)y0=Math.min(L+1,Math.max(0,Math.floor(F0/z1)));console.log(`[buttress] Auto GPU layer capacity (${l.selected.variant}): ${y0}/${L+1}`);let z4;if(k.n_gpu_layers==="auto"||k.n_gpu_layers==null)z4=y0;else z4=Math.max(0,Math.min(Number(k.n_gpu_layers)||0,L+1));let m6=(()=>{let V0=k.flash_attn_type&&String(k.flash_attn_type).toLowerCase();if(V0==="on"||V0==="off")return V0;if(V0==="auto")return l.selected.hasGpu?"auto":"off";return l.selected.hasGpu?"auto":"off"})(),c6=$.runtime.cache_dir,B4=H2($,X),B1=await P0(B4,X.size),d6={ok:h,backend:"ggml-llm",warnings:S,errors:C,model:{repoId:X.repoId,revision:X.revision,filename:X.filename,quantization:X.quantization,url:X.url,sizeBytes:X.size,metadata:{architecture:W,n_ctx_train:j,n_layer:L,n_embd:F,quantization_version:q,file_type:_,kv_layer_count:r,swa:b?.enabled?{window:b.window,pattern:b.pattern,dense_first:b.denseFirst,type:b.type,layers:b.swaLayers}:null}},runtime:{...k,variant:l.selected.variant,n_ctx:v,requested_ctx:d,n_gpu_layers:z4,auto_gpu_layers:y0,flash_attn_type:m6,cache_type_k:p.k,cache_type_v:p.v,estimated_max_n_ctx:C0},resources:{modelBytes:f,kvCacheBytes:A1,recurrentMemoryBytes:W0,totalEstimatedBytes:p6,gpuCapacityBytes:L4,gpuUsableBytes:F0,cpuUsableBytes:h0,fit:l.selected.fit},devices:{selected:l.selected,attempts:l.attempts},download:{cacheDir:c6,localPath:B4,exists:B1},timestamp:new Date().toISOString()};return{config:$,info:d6,artifact:X,metadata:{arch:W,nCtxTrain:j,nLayer:L,nEmbd:F},devices:l,cacheTypes:p,localPath:B4,localExists:B1}},P3=($,X,Z=null,W=null)=>{let j,V=Date.now(),Y=0;return new Z2({async start(J){try{let N=await $.parallel.completion(X,(G,w)=>{if(!w)return;if(w.token)Y+=1;J.enqueue({event:"token",data:{requestId:G,...w}})}),{requestId:H}=N;j=N.stop;let O=await N.promise;console.log("[Completion] Result:",O),J.enqueue({event:"result",data:{requestId:H,...O}}),J.close();let q=Date.now()-V,_=O.timings||{};$0.addCompletion({id:`completion-${H}`,generatorId:Z,requestId:H,repoId:W?.repoId||null,quantization:W?.quantization||null,variant:W?.variant||null,cacheTokens:_.cache_n??0,promptTokens:_.prompt_n??0,tokensGenerated:_.predicted_n??Y,tokensPerSecond:_.predicted_per_second??0,promptPerSecond:_.prompt_per_second??0,durationMs:q,success:!0,interrupted:O.interrupted||!1,contextFull:O.context_full||O.contextFull||!1})}catch(N){J.enqueue({event:"error",data:{message:N?.message||String(N)}}),J.error(N),$0.addCompletion({id:`completion-${Date.now()}`,generatorId:Z,repoId:W?.repoId||null,quantization:W?.quantization||null,variant:W?.variant||null,durationMs:Date.now()-V,tokensGenerated:Y,success:!1,error:N?.message||String(N)})}},cancel(){if(j)j()}})},T3=($,X,Z,W,j,V,Y=null,J=null,N=null)=>{let H,O="",q=!1,_=Date.now(),G=0,w=()=>{if(j)o(j).catch(()=>{});if(N)o(N).catch(()=>{})};return new Z2({async start(z){try{let A=await $.parallel.completion(X,(B,L)=>{if(!L)return;if(L.token)O+=L.token,G+=1;z.enqueue({event:"token",data:{requestId:B,...L}})}),{requestId:R}=A;H=A.stop;let U=await A.promise;if(U.text)O=U.text;else if(U.content)O=U.content;q=!U.interrupted&&!U.context_full,console.log("[Completion] Result:",U),z.enqueue({event:"result",data:{requestId:R,...U}}),z.close();let Q=Date.now()-_,K=U.timings||{};if($0.addCompletion({id:`completion-${R}`,generatorId:Y,requestId:R,repoId:J?.repoId||null,quantization:J?.quantization||null,variant:J?.variant||null,cacheTokens:K.cache_n??0,promptTokens:K.prompt_n??V??0,tokensGenerated:K.predicted_n??G,tokensPerSecond:K.predicted_per_second??0,promptPerSecond:K.prompt_per_second??0,durationMs:Q,success:!0,interrupted:U.interrupted||!1,contextFull:U.context_full||U.contextFull||!1,usedCache:Boolean(X.load_state_path)}),q&&Z.enabled&&O)Z.saveCompletionState(W,O,j,V,N).catch((B)=>{console.warn("[SessionCache] Save failed:",B.message)});else w()}catch(A){z.enqueue({event:"error",data:{message:A?.message||String(A)}}),z.error(A),$0.addCompletion({id:`completion-${Date.now()}`,generatorId:Y,repoId:J?.repoId||null,quantization:J?.quantization||null,variant:J?.variant||null,durationMs:Date.now()-_,tokensGenerated:G,success:!1,error:A?.message||String(A)}),w()}},cancel(){if(H)H();w()}})},z0=($)=>{let X={model:$.plan.localPath,runtime:$.plan.info.runtime};return p4("sha256").update(JSON.stringify(X)).digest("hex").slice(0,24)},k3=async($,X,Z,W=null)=>{let{config:j,localPath:V,artifact:Y}=$;if($.localExists&&!X.has(V)){if($.info.download.exists=!0,typeof Z==="function")Z(0.5);return V}if(j.model.local_path&&!j.model.allow_local_file)throw Error("Local model path provided but `model.allow_local_file` is not enabled");let J=V;if(W){let N=W.getDownload(J);if(N){console.log(`[ensureModelFile] Waiting for global download: ${Y.repoId}`);try{if(await N,await P0(V,Y.size)){if($.localExists=!0,$.info.download.exists=!0,typeof Z==="function")Z(0.5);return V}}catch(H){console.warn(`[ensureModelFile] Global download failed, will retry: ${H.message}`)}}}if(!X.has(J))X.set(J,(async()=>{if(Y.isSplit&&Y.splitCount>0){let N=/-(\d{5})-of-(\d{5})\.gguf$/,H=y.dirname(V),O=Y.splitCount,q=0;for(let _=1;_<=O;_+=1){let G=String(_).padStart(5,"0"),w=Y.filename.replace(N,`-${G}-of-${String(O).padStart(5,"0")}.gguf`),z=`${j.model.base_url.replace(/\/+$/,"")}/${Y.repoId}/resolve/${Y.revision}/${w}`,A=y.join(H,w);if(!await P0(A))await a0(z,Y.headers,A,null,(U)=>{if(U>=0&&Number.isFinite(U)){let Q=(q+U)/O,K=Math.round(Q*100);if(console.log(`Downloading model splits: ${Math.min(100,K)}%`),typeof Z==="function")Z(Q*0.5)}});q+=1}}else console.log("Downloading model: 0%"),await a0(Y.url,Y.headers,V,Y.size,(N)=>{if(N>=0&&Number.isFinite(N)){let H=Math.round(N*100);if(console.log(`Downloading model: ${Math.min(100,H)}%`),typeof Z==="function")Z(N*0.5)}});$.localExists=!0,$.info.download.exists=!0})());try{await X.get(J)}finally{X.delete(J)}return V},D3=async($,X)=>{let Z=z0($),W=$.contexts.get(Z);if(W&&!W.released){if(W.releaseTimer)clearTimeout(W.releaseTimer),W.releaseTimer=null,console.log(`[Context] Cancelled pending release for context "${Z}"`);if(W.releaseRequested=!1,W.refCount+=1,console.log(`[Context] Reusing existing context "${Z}", refCount=${W.refCount}`),typeof X==="function")X(0);if(!W.context)await W.ready;if(typeof X==="function")X(1);return W}if(W)console.log(`[Context] Record exists but released=${W.released}, creating new context`);else console.log(`[Context] No existing record for "${Z}", creating new context`);W={key:Z,refCount:1,ready:null,released:!1},$.contexts.set(Z,W),W.ready=(async()=>{let j=Date.now(),V=await k3($.plan,$.downloads,X,$.globalDownloadManager);if(typeof X==="function")X(0.5);let Y={model:V,n_threads:$.plan.info.runtime.n_threads,use_mmap:$.plan.info.runtime.use_mmap,use_mlock:$.plan.info.runtime.use_mlock,cpu_mask:$.plan.info.runtime.cpu_mask,cpu_strict:$.plan.info.runtime.cpu_strict,devices:$.plan.info.runtime.devices,n_ctx:$.plan.info.runtime.n_ctx,n_gpu_layers:$.plan.info.runtime.n_gpu_layers,n_parallel:$.plan.info.runtime.n_parallel,n_batch:$.plan.info.runtime.n_batch,n_ubatch:$.plan.info.runtime.n_ubatch,n_cpu_moe:$.plan.info.runtime.n_cpu_moe,flash_attn_type:$.plan.info.runtime.flash_attn_type,ctx_shift:$.plan.info.runtime.ctx_shift,kv_unified:$.plan.info.runtime.kv_unified,swa_full:$.plan.info.runtime.swa_full,lib_variant:$.plan.info.runtime.variant};if($.plan.info.runtime.flash_attn_type!=="off")Y.cache_type_k=$.plan.info.runtime.cache_type_k,Y.cache_type_v=$.plan.info.runtime.cache_type_v;console.log("[Context] Load Options:",Y);let J;try{if(J=await N3(Y,(N)=>{if(typeof X==="function"){if(X(0.5+N*0.25),N%5===0)console.log("[Context] Load Model Progress:",N)}}),$.plan.info.runtime.n_parallel){if(!await J.parallel.enable({n_parallel:$.plan.info.runtime.n_parallel,n_batch:$.plan.info.runtime.n_batch}))throw Error("Failed to enable parallel decoding mode for context")}if(typeof X==="function")X(1);return W.context=J,W.modelInfo=J.getModelInfo(),$0.addModelLoad({id:`${$.id}-${Z}`,generatorId:$.id,contextKey:Z,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,nCtx:$.plan.info.runtime?.n_ctx||null,nGpuLayers:$.plan.info.runtime?.n_gpu_layers||null,durationMs:Date.now()-j,success:!0}),W}catch(N){if($0.addModelLoad({id:`${$.id}-${Z}`,generatorId:$.id,contextKey:Z,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,durationMs:Date.now()-j,success:!1,error:N?.message||String(N)}),J)try{J.release()}catch(H){}throw N}})();try{return await W.ready,W}catch(j){throw $.contexts.delete(Z),j}},r0=async($,X,Z=!1)=>{if(X.released)return!1;if(!Z&&X.refCount>0)return!1;X.released=!0,$.contexts.delete(X.key);try{X.context?.parallel?.disable?.()}catch(W){}return await X.context?.release?.(),!0},b3=async($,X,Z=!1)=>{if(X.releaseRequested=!0,X.releaseTimer)clearTimeout(X.releaseTimer),X.releaseTimer=null;if(Z)X.refCount=0;else if(X.refCount=Math.max(0,X.refCount-1),X.refCount>0)return X.releaseRequested=!1,!1;let W=$.config.runtime.context_release_delay_ms;if(typeof W!=="number"||!Number.isFinite(W))return r0($,X);let j=Math.max(0,Math.floor(W));if(Z||j<=0)return r0($,X);return console.log(`[Context] Scheduling release in ${j}ms for context "${X.key}"`),X.releaseTimer=setTimeout(async()=>{if(X.releaseTimer=null,X.refCount>0){console.log(`[Context] Release cancelled, refCount=${X.refCount} for context "${X.key}"`),X.releaseRequested=!1;return}console.log(`[Context] Releasing context "${X.key}" after ${j}ms delay`),await r0($,X)},j),!0},d4=($)=>{let X=e0($);return X.model.repo_id||X.model.repository||X.model.model||null},a1=($)=>{if(!$)return 0;if(typeof $.score==="number"&&Number.isFinite($.score))return Number($.score);return q0($)};var l4=P(()=>{M0();s0();O3=H3(),{ReadableStream:Z2,WritableStream:K3}=O3,n=y.join(T0.homedir(),".buttress","models"),Q2=["mxfp4","q8_0","q6_k","q6","q5_k_m","q5_k_s","q5_k","q5_1","q5_0","q4_k_m","q4_k_s","q4_k","q4_1","q4_0","q3","q2"],S0={backend:{type:"ggml-llm",variant:null,variant_preference:["cuda","vulkan","snapdragon","default"],gpu_memory_fraction:0.85,cpu_memory_fraction:t0},model:{repo_id:null,revision:"main",filename:null,url:null,quantization:null,preferred_quantizations:[],n_ctx:null,n_gpu_layers:"auto",allow_local_file:!1,local_path:null,api_base:Y2,base_url:j2},runtime:{cache_dir:n,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},session_cache:{enabled:!0,max_size_bytes:10737418240,max_entries:1000},context_release_delay_ms:1e4}}});import j0 from"node:path";import t4 from"node:os";import{stat as B2,mkdir as h3,open as I3,unlink as i4,readFile as C3,writeFile as y3}from"node:fs/promises";import{createHash as u3}from"node:crypto";import{initWhisper as f3}from"@fugood/whisper.node";import{getBackendDevicesInfo as R2,isLibVariantAvailable as w2}from"@fugood/llama.node";import*as g3 from"node:stream/web";class v2{constructor(){this.queue=[],this.processing=!1,this.currentTaskId=null}async enqueue($,X=null){return new Promise((Z,W)=>{this.queue.push({task:$,resolve:Z,reject:W,taskId:X}),this.processNext()})}async processNext(){if(this.processing||this.queue.length===0)return;this.processing=!0;let{task:$,resolve:X,reject:Z,taskId:W}=this.queue.shift();this.currentTaskId=W;try{let j=await $();X(j)}catch(j){Z(j)}finally{this.processing=!1,this.currentTaskId=null,this.processNext()}}getStatus(){return{processing:this.processing,queuedCount:this.queue.length,currentTaskId:this.currentTaskId}}}async function h2($,X,Z={}){let{globalDownloadManager:W=null}=Z,j=j4(X),V=await o3(j),Y={id:$,type:"ggml-stt",config:j,plan:V,info:V.info,contextRecord:null,downloads:new Map,globalDownloadManager:W,queue:new v2,finalized:!1},J=async()=>{if(Y.finalized)return;Y.finalized=!0;let A=Y.contextRecord;if(!A)return;if(A.released)return;if(A.releaseRequested||A.releaseTimer)return;if(A.refCount=Math.max(0,A.refCount-1),A.refCount>0)return;await Z4(Y,A)},N=async(A={})=>{let{onProgress:R}=A;try{let U=await e3(Y,R);return{modelInfo:U.modelInfo&&typeof U.modelInfo==="object"?{...U.modelInfo}:null,runtime:{...Y.plan.info.runtime},download:{...Y.plan.info.download}}}catch(U){throw console.error("[Context] Error initializing context:",U),U}},H=async()=>{if(Y.finalized)return!1;let A=Y.contextRecord;if(!A)return!1;return $9(Y,A)},O=async(A={})=>{let{audioPath:R,audioData:U,options:Q={}}=A,K=Y.contextRecord;if(!K)throw Error("Context not initialized");let B={...Q};if(Y.plan.info.runtime.max_threads&&B.maxThreads==null)B.maxThreads=Y.plan.info.runtime.max_threads;let L=`transcription-${Date.now()}-${Math.random().toString(36).slice(2,8)}`,F=Date.now();return Y.queue.enqueue(async()=>{await K.ready;try{let E;if(U){let M=t3(U),{promise:x}=K.context.transcribeData(M,B);E=await x}else{if(!R)throw Error("audioPath or audioData is required for transcription");let M=j0.resolve(R),{promise:x}=K.context.transcribe(M,B);E=await x}return N0.addTranscription({id:L,generatorId:Y.id,repoId:Y.plan.info.model?.repoId||null,quantization:Y.plan.info.model?.quantization||null,modelType:Y.plan.info.model?.modelType||null,variant:Y.plan.info.runtime?.variant||null,durationMs:Date.now()-F,segmentCount:E?.segments?.length||0,textLength:E?.text?.length||0,success:!0}),E}catch(E){throw N0.addTranscription({id:L,generatorId:Y.id,repoId:Y.plan.info.model?.repoId||null,quantization:Y.plan.info.model?.quantization||null,modelType:Y.plan.info.model?.modelType||null,variant:Y.plan.info.runtime?.variant||null,durationMs:Date.now()-F,success:!1,error:E?.message||String(E)}),E}},L)},q=async(A={})=>O(A),_=async(A={})=>O(A),G=()=>{let A=Y.contextRecord;if(!A)return!1;return!A.released&&(A.releaseRequested||A.releaseTimer||A.refCount>0)},w=()=>{Y.finalized=!1},z=()=>({id:Y.id,type:Y.type,repoId:Y.plan.info.model?.repoId||null,quantization:Y.plan.info.model?.quantization||null,modelType:Y.plan.info.model?.modelType||null,variant:Y.plan.info.runtime?.variant||null,hasContext:Boolean(Y.contextRecord?.context),contextRefCount:Y.contextRecord?.refCount||0,queueStatus:Y.queue.getStatus()});return{id:$,type:"ggml-stt",info:V.info,queue:Y.queue,initContext:N,transcribe:q,transcribeData:_,releaseContext:H,finalize:J,getStatus:z,hasPendingReleases:G,resetFinalized:w}}async function I2($,X,Z={}){let{onProgress:W,onComplete:j,onError:V}=Z;try{let Y=j4($),J=await e4(Y),N=D2(Y,J),{repoId:H}=J;if(await W4(N,J.size)){if(console.log(`[Download] STT model already exists: ${H} at ${N}`),typeof j==="function")j({localPath:N,repoId:H,alreadyExists:!0});return{started:!1,localPath:N,repoId:H,alreadyExists:!0}}let q=X.getDownload(N);if(q)return console.log(`[Download] Already downloading STT model: ${H}`),q.then(()=>{if(typeof j==="function")j({localPath:N,repoId:H,joinedExisting:!0})}).catch((G)=>{if(typeof V==="function")V(G)}),{started:!1,localPath:N,repoId:H,alreadyDownloading:!0};console.log(`[Download] Starting STT model download: ${H}`);let _=(async()=>{try{if(await b2(J.url,J.headers,N,J.size,(G)=>{if(G>=0&&Number.isFinite(G)){if(console.log(`[Download] ${H}: ${Math.round(G*100)}%`),typeof W==="function")W(G)}}),console.log(`[Download] Completed STT model: ${H}`),typeof j==="function")j({localPath:N,repoId:H})}catch(G){if(console.error(`[Download] Failed STT model: ${H}`,G.message),typeof V==="function")V(G);throw G}finally{X.deleteDownload(N)}})();return X.setDownload(N,_),{started:!0,localPath:N,repoId:H}}catch(Y){if(console.error("[Download] Failed to start STT download:",Y.message),typeof V==="function")V(Y);return{started:!1,localPath:null,repoId:null,error:Y.message}}}async function C2($=null,X={}){let{threshold:Z=1.1,includeBreakdown:W=!1,config:j,...V}=X,Y=null,J=null,N=null;if(j)try{let R=j4(j),U=await e4(R);Y=U.size??null,{processingBufferBytes:J}=m0({modelBytes:Y}),N=U.quantization||null}catch(R){}let H=j?.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.gpu_memory_fraction))):void 0,O=j?.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(j.backend.cpu_memory_fraction))):void 0,q=await _0({...V,platform:process.platform,totalMemoryInBytes:t4.totalmem(),backend:"ggml-stt",includeBreakdown:W,gpuMemoryFraction:H,cpuMemoryFraction:O,dependencies:{getBackendDevicesInfo:R2,isLibVariantAvailable:w2},modelBytes:Y,kvCacheBytes:J}),_=q.selected,G=z2(_);if(_)_.modelBytes=Y||null,_.processingBytes=J||null,_.quantization=N||null;let w=null,z=null;if($){let R=z2($);z={...$,score:R};let U="buttress",Q="buttress-higher-score";if(!q.ok)U="local",Q="buttress-unavailable";else if(!R&&R!==0)U="buttress",Q="missing-client-score";else if($.fit&&_?.fit){let K=$.fit.fitsInGpu||$.fit.fitsInCpu,B=_.fit.fitsInGpu||_.fit.fitsInCpu;if(K&&!B)U="local",Q="client-fits-in-memory";else if(B&&!K)U="buttress",Q="buttress-fits-in-memory";else if(R>G*Z)U="local",Q="client-better";else if(G>R*Z)U="buttress",Q="buttress-better";else U="either",Q="comparable-scores"}else if(R>G*Z)U="local",Q="client-better";else if(G>R*Z)U="buttress",Q="buttress-better";else U="either",Q="comparable-scores";w={buttressScore:G,clientScore:R,threshold:Z,recommendation:U,reason:Q}}if(!q.ok&&!w)w={buttressScore:G,clientScore:$?.score??null,threshold:Z,recommendation:"local",reason:"buttress-unavailable"};let A=null;if(j)A={repoId:j.model?.repo_id||null,quantization:j.model?.quantization||null,filename:j.model?.filename||null};return{type:"ggml-stt",timestamp:new Date().toISOString(),buttress:q,client:z,comparison:w,modelConfig:A}}var p3=()=>{if(typeof globalThis<"u"&&globalThis.ReadableStream&&globalThis.WritableStream)return{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream};return g3},m3,F2=($={},X={})=>{return Object.entries(X||{}).forEach(([Z,W])=>{if(W&&typeof W==="object"&&!Array.isArray(W)){if(!$[Z]||typeof $[Z]!=="object")$[Z]={};F2($[Z],W)}else $[Z]=W}),$},c3=".bin",E2="https://huggingface.co",M2="https://huggingface.co/api",R0,r4,o4,x2="fp16",S2=0.5,d3,P2=($)=>{if(!$)return null;let X=$.toLowerCase();return d3.find((Z)=>X.includes(Z))||null},a4,s4=($,X=[])=>{if(!$&&$!==0)return[...X];if(Array.isArray($))return $.filter((Z)=>Z!=null);return[$]},X4=($)=>{if(!$)return null;let X=String($).toLowerCase();if(["cuda","vulkan","default"].includes(X))return X;return null},j4=($={})=>{let X=JSON.parse(JSON.stringify(a4));if(F2(X,$),X.backend.variant=X4(X.backend.variant),X.backend.variant_preference=Array.from(new Set(s4(X.backend.variant_preference||r4).map(X4).filter(Boolean))),X.backend.variant_preference.length===0)X.backend.variant_preference=[...r4];if(X.runtime.prefer_variants=Array.from(new Set(s4(X.runtime.prefer_variants).map(X4).filter(Boolean))),X.model.preferred_quantizations=Array.from(new Set(s4(X.model.preferred_quantizations||X.model.quantizations).map((Z)=>Z?String(Z).toLowerCase():null).filter(Boolean))),X.model.quantization){let Z=String(X.model.quantization).toLowerCase();if(!X.model.preferred_quantizations.includes(Z))X.model.preferred_quantizations.unshift(Z)}return X.model.base_url=X.model.base_url||E2,X.model.api_base=X.model.api_base||M2,X.runtime.cache_dir=X.runtime.cache_dir?j0.resolve(X.runtime.cache_dir):R0,X.runtime.context_release_delay_ms=Math.max(0,Number(X.runtime.context_release_delay_ms)||a4.runtime.context_release_delay_ms),X},n4=($)=>{let X=$.toLowerCase();return o4.find((W)=>X.includes(W))||null},l3=($)=>{let X=[];if($.backend.variant)X.push($.backend.variant);if($.runtime.prefer_variants.length>0)X.push(...$.runtime.prefer_variants);return X.push(...$.backend.variant_preference),X.push("default"),Array.from(new Set(X.map(X4).filter(Boolean)))},T2=async($)=>{await h3($,{recursive:!0})},i3=($=R0)=>j0.join($,".metadata-cache"),k2=($,X,Z=R0)=>{let W=u3("sha256").update($).digest("hex");return j0.join(i3(Z),X,`${W}.json`)},s3=async($,X,Z=R0)=>{try{let W=k2($,X,Z),j=await C3(W,"utf-8");return JSON.parse(j)}catch(W){return null}},L2=async($,X,Z,W=R0)=>{try{let j=k2($,X,W);await T2(j0.dirname(j)),await y3(j,JSON.stringify(Z),"utf-8")}catch(j){}},n3=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,X);if(!Z.ok){let W=await Z.text().catch(()=>"");throw Error(`Failed to fetch ${$}: ${Z.status} ${Z.statusText} ${W}`.trim())}return Z.json()},A2=async($,X={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let Z=await fetch($,{...X,method:"HEAD"});if(!Z.ok)throw Error(`Failed to fetch headers for ${$}: ${Z.status} ${Z.statusText}`);return Z},D2=($,X)=>{if($.model.local_path)return j0.resolve($.model.local_path);let Z=X.repoId.split("/"),W=j0.join($.runtime.cache_dir,...Z,X.revision);return j0.join(W,X.filename)},W4=async($,X)=>{try{let Z=await B2($);if(!X)return!0;return Z.size===X}catch(Z){return!1}},b2=async($,X,Z,W,j)=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");await T2(j0.dirname(Z));let V=await fetch($,{headers:X});if(!V.ok||!V.body)throw Error(`Failed to download ${$}: ${V.status} ${V.statusText}`);let Y=await I3(Z,"w"),J=Number(V.headers.get("content-length"))||W||0,N=0,H=0.05;try{await V.body.pipeTo(new m3({async write(O){if(await Y.write(O),N+=O.byteLength,typeof j==="function"&&J>0){let q=Math.min(1,N/J);while(q>=H)j(H),H+=0.05}},async close(){if(await Y.close(),typeof j==="function")j(1)},async abort(O){throw await Y.close().catch(()=>{}),await i4(Z).catch(()=>{}),O}}))}catch(O){throw await Y.close().catch(()=>{}),await i4(Z).catch(()=>{}),O}if(W){let O=await B2(Z);if(O.size!==W)throw await i4(Z).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${W} got ${O.size}`)}},e4=async($)=>{let X=$.model.repo_id||$.model.repository||$.model.model;if(!X)throw Error("`model.repo_id` is required in Buttress backend config");let Z=$.model.revision||"main",W=$.runtime.cache_dir,j=JSON.stringify({repoId:X,revision:Z,filename:$.model.filename,url:$.model.url,quantization:$.model.quantization,preferred_quantizations:$.model.preferred_quantizations}),V=await s3(j,"artifact-info",W);if(V)return V;let Y={...$.runtime.http_headers||{}};if($.runtime.huggingface_token)Y.Authorization=`Bearer ${$.runtime.huggingface_token}`;if($.model.url){let U=await A2($.model.url,{headers:Y}),Q=Number(U.headers.get("content-length"))||null,K=$.model.filename||$.model.url.split("/").pop(),B={repoId:X,revision:Z,filename:K,url:$.model.url,size:Q,quantization:n4(K||""),headers:Y};return await L2(j,"artifact-info",B,W),B}let{filename:J}=$.model,N=$.model.quantization&&String($.model.quantization).toLowerCase(),H=await n3(`${$.model.api_base}/models/${X}?revision=${Z}&blobs=true`,{headers:Y}),q=(H?.siblings||H?.files||[]).map((U)=>U.rfilename||U.path||U.filename).filter((U)=>typeof U==="string"&&U.endsWith(c3));if(q.length===0)throw Error(`No model artifacts found in repo ${X}`);let _=$.model.preferred_quantizations.length>0?$.model.preferred_quantizations:o4,G=()=>{for(let U of _)if(U===x2){let Q=q.find((K)=>{let B=K.toLowerCase();return!o4.some((L)=>B.includes(L))});if(Q)return{filename:Q,quantization:null}}else{let Q=q.find((K)=>K.toLowerCase().includes(U));if(Q)return{filename:Q,quantization:U}}return null};if(!J){let U=G()||{filename:q[0],quantization:null},{filename:Q,quantization:K}=U;J=Q,N=K||n4(J)}else if(!N)N=n4(J);let w=`${$.model.base_url.replace(/\/+$/,"")}/${X}/resolve/${Z}/${J}`,z=await A2(w,{headers:Y}),A=Number(z.headers.get("content-length"))||null,R={repoId:X,revision:Z,filename:J,url:w,size:A,quantization:N,headers:Y,isSplit:!1,splitCount:0};return await L2(j,"artifact-info",R,W),R},r3=async($,{modelBytes:X=null,processingBytes:Z=null}={})=>{let W=l3($),[j,...V]=W,Y=$.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.gpu_memory_fraction))):a4.backend.gpu_memory_fraction||1,J=$.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):S2,N=await _0({platform:process.platform,totalMemoryInBytes:t4.totalmem(),backend:"ggml-stt",variant:j||null,preferVariants:V,variantPreference:$.backend.variant_preference,gpuMemoryFraction:Y,cpuMemoryFraction:J,dependencies:{getBackendDevicesInfo:R2,isLibVariantAvailable:w2},modelBytes:X,kvCacheBytes:Z}),H=(_)=>({..._,devices:Array.isArray(_.devices)?_.devices:[],ok:_.ok,hasGpu:Boolean(_.hasGpu),totalMemory:_.gpuTotalBytes||_.totalMemory||0,error:_.ok?null:Error(_.error||`Variant ${_.variant} not available on this platform`)});if(!N.ok||!N.selected){let _=(N.attempts||[]).map((G)=>`${G.variant}: ${G.error||"unknown error"}`).join("; ");throw Error(`Unable to initialize any backend variant (${W.join(", ")}). Errors: ${_}`)}let O=(N.attempts||[]).map(H);return{selected:H(N.selected),attempts:O}},o3=async($)=>{let X=await e4($),Z=m0({modelBytes:X.size>0?X.size:0}),W=await r3($,{modelBytes:Z.modelBytes,processingBytes:Z.processingBufferBytes}),j=W.selected.hasGpu&&(W.selected.fit?.fitsInGpu!==void 0?W.selected.fit.fitsInGpu:!0);if($.model.use_gpu===!1)j=!1;let V=$.model.use_flash_attn&&String($.model.use_flash_attn).toLowerCase(),Y;if(V==="on"||V==="true")Y=!0;else if(V==="off"||V==="false")Y=!1;else Y=j;let J=$.runtime.cache_dir,N=D2($,X),H=await W4(N,X.size),O={ok:!0,backend:"ggml-stt",model:{repoId:X.repoId,revision:X.revision,filename:X.filename,quantization:X.quantization,modelType:P2(X.filename),url:X.url,sizeBytes:X.size},runtime:{variant:W.selected.variant,use_gpu:j,use_flash_attn:Y,max_threads:$.runtime.max_threads?Number($.runtime.max_threads):null},resources:{...Z,gpuCapacityBytes:W.selected.gpuTotalBytes,gpuUsableBytes:W.selected.gpuUsableBytes,cpuUsableBytes:W.selected.cpuUsableBytes,fit:W.selected.fit},devices:{selected:W.selected,attempts:W.attempts},download:{cacheDir:J,localPath:N,exists:H},timestamp:new Date().toISOString()};return{config:$,info:O,artifact:X,memory:Z,devices:W,localPath:N,localExists:H}},a3=async($,X,Z,W=null)=>{let{localPath:j,artifact:V,config:Y}=$;if($.localExists){if(typeof Z==="function")Z(1);return j}if(W){let H=W.getDownload(j);if(H){console.log(`[ensureModelFile] Waiting for global STT download: ${V.repoId}`);try{if(await H,await W4(j,V.size)){if($.localExists=!0,$.info.download.exists=!0,typeof Z==="function")Z(1);return j}}catch(O){console.warn(`[ensureModelFile] Global STT download failed, will retry: ${O.message}`)}}}let J=X.get(j);if(J){if(await J,typeof Z==="function")Z(1);return j}let N=(async()=>{if(Y.model.allow_local_file){if(!await W4(j,V.size))throw Error(`Local model file not found: ${j}`);return j}return await b2(V.url,V.headers,j,V.size,Z),j})();X.set(j,N);try{return await N,j}finally{X.delete(j)}},t3=($)=>{if(!$)return null;if($ instanceof ArrayBuffer)return $;if(ArrayBuffer.isView($))return $.buffer;if(typeof $==="string"){let X=$.startsWith("data:")?$.split(",")[1]||"":$,Z=Buffer.from(X,"base64");return Z.buffer.slice(Z.byteOffset,Z.byteOffset+Z.byteLength)}throw Error("Unsupported audioData format, expected base64 string or ArrayBuffer")},e3=async($,X)=>{if($.contextRecord&&!$.contextRecord.released){if($.contextRecord.releaseTimer)clearTimeout($.contextRecord.releaseTimer),$.contextRecord.releaseTimer=null,console.log("[Context] Cancelled pending STT release");if($.contextRecord.releaseRequested=!1,$.contextRecord.refCount+=1,console.log(`[Context] Reusing existing STT context, refCount=${$.contextRecord.refCount}`),typeof X==="function")X(0);if(!$.contextRecord.context)await $.contextRecord.ready;if(typeof X==="function")X(1);return $.contextRecord}if($.contextRecord)console.log(`[Context] STT record exists but released=${$.contextRecord.released}, creating new context`);else console.log("[Context] No existing STT record, creating new context");let Z={refCount:1,ready:null,released:!1};$.contextRecord=Z,Z.ready=(async()=>{let W=Date.now();try{if(typeof X==="function")X(0);let j=await a3($.plan,$.downloads,X,$.globalDownloadManager);if(typeof X==="function")X(0.5);let V=await f3({filePath:j,useFlashAttn:$.plan.info.runtime.flash_attn_type==="on",useGpu:$.plan.info.runtime.n_gpu_layers>0,nThreads:$.plan.info.runtime.n_threads},$.plan.info.runtime.variant);if(typeof X==="function")X(1);Z.context=V;try{Z.modelInfo=V.getModelInfo()}catch(Y){Z.modelInfo=null}return N0.addModelLoad({id:$.id,generatorId:$.id,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,modelType:$.plan.info.model?.modelType||null,variant:$.plan.info.runtime?.variant||null,useGpu:$.plan.info.runtime?.use_gpu||!1,durationMs:Date.now()-W,success:!0}),Z}catch(j){throw N0.addModelLoad({id:$.id,generatorId:$.id,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,modelType:$.plan.info.model?.modelType||null,variant:$.plan.info.runtime?.variant||null,durationMs:Date.now()-W,success:!1,error:j?.message||String(j)}),j}})();try{if(await Z.ready,typeof X==="function")X(1);return Z}catch(W){throw $.contextRecord=null,W}},Z4=async($,X,Z=!1)=>{if(X.released)return!1;if(!Z&&X.refCount>0)return!1;return X.released=!0,$.contextRecord=null,await X.context?.release?.(),!0},$9=async($,X,Z=!1)=>{if(X.releaseRequested=!0,X.releaseTimer)clearTimeout(X.releaseTimer),X.releaseTimer=null;if(Z)X.refCount=0;else if(X.refCount=Math.max(0,X.refCount-1),X.refCount>0)return X.releaseRequested=!1,!1;let W=$.config.runtime.context_release_delay_ms;if(typeof W!=="number"||!Number.isFinite(W))return Z4($,X);let j=Math.max(0,Math.floor(W));if(Z||j<=0)return Z4($,X);return console.log(`[Context] Scheduling STT release in ${j}ms`),X.releaseTimer=setTimeout(async()=>{if(X.releaseTimer=null,X.refCount>0){console.log(`[Context] STT release cancelled, refCount=${X.refCount}`),X.releaseRequested=!1;return}console.log(`[Context] Releasing STT context after ${j}ms delay`),await Z4($,X)},j),!0},$1=($)=>{let X=j4($),Z=X.model.repo_id||X.model.repository||X.model.model||null;if(!Z)return null;let W=P2(X.model.filename);if(W)return`${Z}:${W}`;return Z},z2=($)=>{if(!$)return 0;if(typeof $.score==="number"&&Number.isFinite($.score))return Number($.score);return q0($)};var X1=P(()=>{M0();s0();({WritableStream:m3}=p3()),R0=j0.join(t4.homedir(),".buttress","models"),r4=["cuda","vulkan","default"],o4=["q8_0","q5_1","q5_0","q4_1","q4_0"],d3=["large-v3-turbo","distil-large-v3","large-v3","large-v2","large-v1","large","distil-medium","medium.en","medium","small.en-tdrz","distil-small.en","small.en","small","base.en","base","tiny.en","tiny"],a4={backend:{type:"ggml-stt",variant:null,variant_preference:r4,gpu_memory_fraction:0.85,cpu_memory_fraction:S2},model:{repo_id:"BricksDisplay/whisper-ggml",revision:"main",filename:null,url:null,quantization:null,preferred_quantizations:["q8_0",x2,"q5_1"],allow_local_file:!1,local_path:null,api_base:M2,base_url:E2,use_gpu:!0,use_flash_attn:"auto"},runtime:{cache_dir:R0,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},max_threads:null,context_release_delay_ms:1e4}}});async function t($,X=null,Z={}){if($==="ggml-llm")return _2(X,Z);if($==="ggml-stt")return C2(X,Z);throw Error(`Unknown backend type: ${$}`)}var Z1=P(()=>{l4();X1()});var g;var y2=P(()=>{g={name:"@fugood/buttress-backend-core",private:!0,type:"module",version:"2.23.0-beta.47",main:"src/index.js",types:"lib/types/index.d.ts",scripts:{build:"tsc --noResolve --noCheck --declaration --emitDeclarationOnly --allowJs --outDir lib/types src/index.js"},dependencies:{"@fugood/buttress-hardware-guardrails":"^2.23.0-beta.47","@fugood/llama.node":"^1.4.15","@fugood/whisper.node":"^1.0.13","@huggingface/gguf":"^0.3.2","@iarna/toml":"^3.0.0",bytes:"^3.1.0"}}});import m from"node:os";import u2 from"node:fs";import f2 from"node:path";import{execSync as Y4}from"node:child_process";import g2 from"@iarna/toml";async function m2({modelIds:$=[],defaultConfig:X=null}={}){let Z=[];if(console.log(`${g.name} v${g.version}`),console.log(`Generating model capabilities comparison...
3
3
  `),Z.push(`${g.name} v${g.version}`),Z.push(`## Model Capabilities Comparison
4
4
  `),!$||$.length===0)console.error("Error: No model IDs provided"),process.exit(1);try{let W=(U={},Q={})=>{let K=Array.isArray(U)?[...U]:{...U};return Object.entries(Q||{}).forEach(([B,L])=>{if(L&&typeof L==="object"&&!Array.isArray(L))K[B]=W(K[B]||{},L);else K[B]=L}),K},j=X||{},{server:V,generators:Y=[],...J}=j,N=(U)=>W(JSON.parse(JSON.stringify(J)),U||{}),H=(U)=>{if(Array.isArray(Y)&&Y.length>0){let Q=Y.filter((K)=>K?.type==="ggml-llm");if(Q.length>0&&U){let K=Q.find((B)=>B.model?.repo_id===U);if(K)return N(K)}}return Object.keys(J).length>0?N({}):null},O=[];for(let U=0;U<$.length;U+=1){let Q=$[U];console.log(`[${U+1}/${$.length}] Analyzing ${Q}...`);let K=H(Q);K={...K||{},model:{...J.runtime,...K?.model||{},repo_id:Q}};let B=await t("ggml-llm",null,{config:K,includeBreakdown:!0});O.push({modelId:Q,capabilities:B,modelInfo:B.buttress?.selected||null,modelConfig:B.modelConfig||null})}let q=(U)=>U?(U/1024/1024/1024).toFixed(2):"N/A",_=(U)=>U?"✅":"\uD83D\uDEAB";Z.push("| Model ID | Quantization | Size (GB) | Context Size | KV Cache Size (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |"),Z.push("|----------|--------------|-----------|--------------|--------------------|-----------------------------|-----------------|-----------------|"),O.forEach(({modelId:U,modelInfo:Q,modelConfig:K})=>{let B=Q?.quantization?.name?.toUpperCase()||"N/A",L=q(Q?.modelBytes),F=K?.nCtx||Q?.kvInfo?.nCtxTrain||"N/A",E=G0(Q),M=Number(F),x=Q?.kvCacheBytes||(E&&Number.isFinite(M)&&M>0?E(M):E&&E(Q?.kvInfo?.nCtxTrain||0))||null,D=q(x),I=q(Q?.modelBytes&&x?Q.modelBytes+x:Q?.fit?.totalRequiredBytes),b=_(Q?.fit?.fitsInGpu),u=_(Q?.fit?.fitsInCpu);Z.push(`| ${U} | ${B} | ${L} | ${F} | ${D} | ${I} | ${b} | ${u} |`);let r=Q?.memoryLimitedCtx!=null||Q?.limitedFit!=null,k=!Q?.fit?.fitsInGpu||!Q?.fit?.fitsInCpu;if(r&&k){let d=Q?.memoryLimitedCtx||F,v=Number(d),S=Q?.limitedKvCacheBytes||E&&Number.isFinite(v)&&v>0&&E(v)||null,C=q(S),h=q(Q?.modelBytes&&S?Q.modelBytes+S:Q?.limitedFit?.totalRequiredBytes),p=_(Q?.limitedFit?.fitsInGpu),f=_(Q?.limitedFit?.fitsInCpu);if(d!==F||C!==D||h!==I)Z.push(`| ↳ Limited | - | ${L} | ${d} | ${C} | ${h} | ${p} | ${f} |`)}}),Z.push(`
5
5
  ---`),Z.push(`
@@ -36,18 +36,18 @@ STT model capabilities table saved to: ${R}`),process.exit(0)}catch(W){console.e
36
36
  --- GPU Details ---`),z.devices.filter((Q)=>Q.type==="gpu").forEach((Q)=>{console.log(`GPU Backend: ${Q.backend}`),console.log(`GPU Name: ${Q.deviceName}`),console.log(`GPU Total Memory: ${(Q.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let K=z.gpuTotalBytes>0?(z.gpuUsableBytes/z.gpuTotalBytes*100).toFixed(0):0;if(console.log(`GPU Usable Memory: ${(z.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(z.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Q.metadata){if(Q.metadata.hasBFloat16)console.log("Supports BFloat16: Yes");if(Q.metadata.hasUnifiedMemory)console.log("Unified Memory: Yes")}});else console.log("GPU: Not available");if(console.log(`
37
37
  Backend Variant: ${z.variant}`),console.log(`Performance Score: ${z.score}`),z.fit)console.log(`
38
38
  --- Model Fit Analysis ---`),console.log(`Fits in GPU: ${z.fit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU: ${z.fit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor: ${z.fit.limiting}`)}console.log(`
39
- === Full Capabilities JSON ===`),console.log(JSON.stringify(_,null,2)),process.exit(0)}catch(Z){console.error("Failed to get capabilities:",Z.message),process.exit(1)}}var p2=($)=>{if(!$)return{repoId:null,filename:null};let[X,Z]=$.split(":");return{repoId:X,filename:Z||null}};var i2=P(()=>{M0();Z1();y2()});var J4={};i6(J4,{testGgmlSttCapabilities:()=>l2,testGgmlLlmCapabilities:()=>c2,status:()=>V9,startModelDownload:()=>Q4,startGenerator:()=>W9,showSttModelsTable:()=>d2,showModelsTable:()=>m2,globalDownloadManager:()=>W1,ggmlStt:()=>Q9,ggmlLlm:()=>Y9,getModelIdentifier:()=>J9,getCapabilities:()=>t,generatorRegistry:()=>X0,finalizeGenerator:()=>j9});async function W9($,X){let W={"ggml-llm":{create:K2,getId:d4},"ggml-stt":{create:h2,getId:$1}}[$];if(!W)throw Error(`Unsupported backend type: ${$}`);let j=W.getId(X);if(!j)throw Error("Buttress generator config missing repo identifier");let V=`${$}:${j}`,Y=X0.get(V);if(Y)return Y.refCount+=1,Y.instance.resetFinalized?.(),{id:Y.id,info:Y.instance.info};let J=await W.create(V,X,{globalDownloadManager:W1}),N={id:V,type:J.type,instance:J,refCount:1};return X0.set(V,N),{id:V,info:J.info}}async function j9($){let X=X0.get($);if(!X)return!1;if(X.refCount-=1,X.refCount<=0){if(await X.instance.finalize(),!(X.instance.hasPendingReleases?.()??!1))X0.delete($)}return!0}function J9($,X){if($==="ggml-llm")return d4(X);if($==="ggml-stt")return $1(X);return null}async function Q4($,X,Z={}){let j={"ggml-llm":q2,"ggml-stt":I2}[$];if(!j)return{started:!1,localPath:null,repoId:null,error:`Unknown backend type: ${$}`};return j(X,W1,Z)}var X0,W1,Z9=($)=>{let X=X0.get($);if(!X)throw Error(`Unknown generator id "${$}"`);return X},O0=($,X)=>{let Z=Z9($);if(Z.type!==X)throw Error(`Generator "${$}" does not support ${X} backend`);return Z.instance},Y9,Q9,V9;var V4=P(()=>{l4();X1();Z1();s0();i2();X0=new Map,W1={downloads:new Map,getDownload($){return this.downloads.get($)||null},setDownload($,X){this.downloads.set($,X)},deleteDownload($){this.downloads.delete($)},isDownloading($){return this.downloads.has($)},getActiveDownloads(){return Array.from(this.downloads.entries()).map(([$,X])=>({localPath:$,promise:X}))}};Y9={async initContext($,X){return O0($,"ggml-llm").initContext(X)},async completion($,X){return O0($,"ggml-llm").completion(X)},async tokenize($,X){return O0($,"ggml-llm").tokenize(X)},async detokenize($,X){return O0($,"ggml-llm").detokenize(X)},async applyChatTemplate($,X){return O0($,"ggml-llm").applyChatTemplate(X)},async releaseContext($,X){let Z=X0.get($);if(!Z)return{released:!0,alreadyReleased:!0};if(Z.type!=="ggml-llm")throw Error(`Generator "${$}" does not support ggml-llm backend`);return Z.instance.releaseContext(X)}},Q9={async initContext($,X){return O0($,"ggml-stt").initContext(X)},async transcribe($,X){return O0($,"ggml-stt").transcribe(X)},async transcribeData($,X){return O0($,"ggml-stt").transcribeData(X)},async releaseContext($,X){let Z=X0.get($);if(!Z)return{released:!0,alreadyReleased:!0};if(Z.type!=="ggml-stt")throw Error(`Generator "${$}" does not support ggml-stt backend`);return Z.instance.releaseContext(X)}};V9={getFullStatus:()=>d1(X0),getGgmlLlmStatus:()=>C4(X0),getGgmlSttStatus:()=>y4(X0),subscribeToStatus:I4,subscribeToStatusWithId:c1,llmStatusTracker:$0,sttStatusTracker:N0,statusEmitter:e}});import{node as N9}from"@elysiajs/node";import{Elysia as U9}from"elysia";var G9,K0=($)=>new U9({adapter:G9?N9():void 0,...$});var k0=P(()=>{G9=typeof process<"u"&&process.versions&&process.versions.node});import{t as Y0}from"elysia";var H9,O9=({store:{serverInfo:$}})=>({id:$.id,name:$.name,version:$.version,generators:$.generators,authentication:$.authentication}),j1=($)=>{let X=K0(),Z=$.autodiscover.http?.path??"/buttress/info";return X.get(Z,O9,{response:H9}),X};var s2=P(()=>{k0();H9=Y0.Object({id:Y0.String(),name:Y0.String(),version:Y0.String(),generators:Y0.Array(Y0.Object({type:Y0.String()})),authentication:Y0.Object({required:Y0.Boolean(),type:Y0.Literal("device-group")})})});import{t as Q0,file as K9}from"elysia";import{writeFile as n2}from"node:fs/promises";import Y1 from"node:path";var q9,Q1;var r2=P(()=>{k0();q9=typeof process<"u"&&process.versions!=null&&process.versions.node!=null,Q1=K0().post("/buttress/upload",async({body:{file:$},store:{config:X}})=>{let Z=`${Date.now()}-${$.name.replace(/[^\dA-Za-z]/g,"_")}`,W=Y1.join(X.server.temp_file_dir,Z);try{if(q9)await n2(W,await $.stream());else await n2(W,await $.arrayBuffer());return{ok:!0,filename:Z}}catch(j){return{ok:!1,error:String(j)}}},{body:Q0.Object({file:Q0.File()}),response:Q0.Object({ok:Q0.Boolean(),filename:Q0.Optional(Q0.String()),error:Q0.Optional(Q0.String())})}).get("/buttress/download/:filename",async({params:{filename:$},store:{config:X},status:Z})=>{let W=Y1.join(X.server.temp_file_dir,$);if(Y1.relative(X.server.temp_file_dir,W).includes(".."))return Z(400),"Invalid file path";return K9(W)},{params:Q0.Object({filename:Q0.String()})})});import J1 from"node:path";import t2 from"node:fs/promises";import{fileURLToPath as _9}from"node:url";var o2,L9=async()=>{let $=[J1.join(o2,"..","public","status.html"),J1.join(o2,"..","..","public","status.html")];return(await Promise.all($.map((Z)=>t2.access(Z).then(()=>Z,()=>null)))).find((Z)=>Z!==null)??null},A9=($)=>{let{status:X}=$;if(X?.getFullStatus)return X.getFullStatus();return{timestamp:new Date().toISOString(),ggmlLlm:{generators:[],history:{}},ggmlStt:{generators:[],history:{}}}},a2=async()=>{let $=await L9();if(!$)return console.error("[Status] Failed to find status.html in candidate paths"),new Response("Status page not found",{status:404,headers:{"Content-Type":"text/plain"}});try{let X=await t2.readFile($,"utf-8");return new Response(X,{headers:{"Content-Type":"text/html; charset=utf-8"}})}catch(X){return console.error("[Status] Failed to serve status page:",X),new Response("Status page not found",{status:404,headers:{"Content-Type":"text/plain"}})}},V1;var e2=P(()=>{k0();o2=J1.dirname(_9(import.meta.url)),V1=K0().get("/status",a2).get("/status/",a2).get("/buttress/status",({store:{backend:$}})=>A9($))});import{t as T,sse as N4}from"elysia";import{cors as z9}from"@elysiajs/cors";async function R9($,X,Z){let j=(X.generators||[]).filter((z)=>z.type==="ggml-llm");if(j.length===0)throw Error('No ggml-llm generator configured. Add a [[generators]] with type = "ggml-llm" to your config.');let V=j[0],Y=Z||V.model?.repo_id;if(Z){let z=j.find((A)=>A.model?.repo_id===Z);if(z)V=z}else Y=V.model?.repo_id;let J=Y,N=$6.get(J);if(N?.initialized)return N;let{generators:H,server:O,...q}=X.global||{},_={...q,...V,model:{...V.model,repo_id:Y}};console.log(`[OpenAI] Creating generator for ${J}`);let{id:G}=await $.startGenerator("ggml-llm",_),w={id:G,config:_,repoId:Y,initialized:!1};return $6.set(J,w),await $.ggmlLlm.initContext(G,{}),w.initialized=!0,console.log(`[OpenAI] Generator ready: ${J}`),w}async function w9($,X,Z,W){let j=$.getReader(),V="",Y=null,J=null,N="stop",H={prompt_tokens:0,completion_tokens:0,total_tokens:0};try{let q=!1;while(!q){let _=await j.read();if({done:q}=_,q)break;let{event:G,data:w}=_.value;if(G==="token"){if(w.content!=null)V+=w.content;else if(w.token!=null)V+=w.token}else if(G==="result"){if(w.text)V=w.text;else if(w.content)V=w.content;if(w.reasoning_content)Y=w.reasoning_content;if(w.tool_calls?.length>0)J=w.tool_calls.map((z,A)=>({id:z.id||`call_${X}_${A}`,type:"function",function:{name:z.function?.name||"",arguments:z.function?.arguments||""}})),N="tool_calls";else N=w.interrupted?"length":"stop";H={prompt_tokens:w.prompt_tokens||w.promptTokens||0,completion_tokens:w.tokens_predicted||w.tokensPredicted||0,total_tokens:(w.prompt_tokens||w.promptTokens||0)+(w.tokens_predicted||w.tokensPredicted||0)}}else if(G==="error")throw Error(w.message)}}finally{j.cancel().catch(()=>{})}let O={role:"assistant",content:V||null};if(Y)O.reasoning_content=Y;if(J)O.tool_calls=J;return{id:X,object:"chat.completion",created:Z,model:W,choices:[{index:0,message:O,finish_reason:N}],usage:H}}function U4({global:$}){let X=K0({prefix:"/oai-compat"});return X.use(z9({origin:$?.openai_compat?.cors_allowed_origins??!1,methods:["GET","POST","OPTIONS"],allowedHeaders:["Content-Type","Authorization"],maxAge:86400,preflight:!0})),X.get("/v1/models",({store:Z})=>{let{config:W}=Z,Y=(W.generators||[]).filter((J)=>J.type==="ggml-llm").map((J)=>{return{id:J.model?.repo_id||"ggml-llm",object:"model",created:Math.floor(Date.now()/1000),owned_by:"local"}});if(Y.length===0)Y.push({id:"ggml-llm",object:"model",created:Math.floor(Date.now()/1000),owned_by:"local"});return{object:"list",data:Y}}),X.post("/v1/chat/completions",async function*({body:W,set:j,store:V}){let{config:Y,backend:J}=V,{messages:N=[],stream:H=!1,model:O,tools:q,temperature:_,stop:G,top_p:w,max_tokens:z,presence_penalty:A,frequency_penalty:R,tool_choice:U,stream_options:Q}=W;if(!N||N.length===0)return j.status=400,{error:{message:"messages is required and must not be empty",type:"invalid_request_error"}};try{let K=await R9(J,Y,O),B=B9(),L=Math.floor(Date.now()/1000),F=K.repoId||"ggml-llm",E={reasoning_format:"auto",messages:N,jinja:!0,add_generation_prompt:!0};if(_!=null)E.temperature=_;if(w!=null)E.top_p=w;if(z!=null)E.n_predict=z;if(G!=null)E.stop=Array.isArray(G)?G:[G];if(A!=null)E.presence_penalty=A;if(R!=null)E.frequency_penalty=R;if(q!=null)E.tools=q;if(U!=null)E.tool_choice=U;let M=await J.ggmlLlm.completion(K.id,{options:E});if(!H)return await w9(M,B,L,F);let x=Q?.include_usage===!0,D=M.getReader(),I="",b="",u=new Map,r=new Map;try{let k=!1;while(!k){let d=await D.read();if({done:k}=d,k)break;let{event:v,data:S}=d.value;if(v==="token"){let C={};if(S.content!=null){let h=S.content;if(h.length>I.length)C.content=h.slice(I.length),I=h}if(S.reasoning_content!=null){let h=S.reasoning_content;if(h.length>b.length)C.reasoning_content=h.slice(b.length),b=h}if(S.tool_calls?.length>0){let h=[];if(S.tool_calls.forEach((p,f)=>{let s={index:f};if(!r.has(f))r.set(f,p.id||`call_${B}_${f}`),s.id=r.get(f),s.type="function";let W0=p.function?.arguments||"",A0=u.get(f)||"",l={};if(!u.has(f)&&p.function?.name)l.name=p.function.name;if(W0.length>A0.length)l.arguments=W0.slice(A0.length),u.set(f,W0);if(Object.keys(l).length>0)s.function=l,h.push(s);else if(s.id)s.function={name:p.function?.name||"",arguments:""},h.push(s)}),h.length>0)C.tool_calls=h}if(Object.keys(C).length>0)yield N4({data:JSON.stringify({id:B,object:"chat.completion.chunk",created:L,model:F,choices:[{index:0,delta:C,finish_reason:null}]})})}else if(v==="result"){let C="stop";if(S.tool_calls?.length>0||r.size>0)C="tool_calls";else if(S.interrupted)C="length";let h={id:B,object:"chat.completion.chunk",created:L,model:F,choices:[{index:0,delta:{},finish_reason:C}]};if(x)h.usage={prompt_tokens:S.prompt_tokens||S.promptTokens||0,completion_tokens:S.tokens_predicted||S.tokensPredicted||0,total_tokens:(S.prompt_tokens||S.promptTokens||0)+(S.tokens_predicted||S.tokensPredicted||0)};yield N4({data:JSON.stringify(h)})}else if(v==="error")yield N4({data:JSON.stringify({error:{message:S.message,type:"server_error"}})})}yield N4({data:"[DONE]"})}finally{D.cancel().catch(()=>{})}}catch(K){return console.error("[OpenAI] Chat completion error:",K),j.status=500,{error:{message:K.message||"Internal server error",type:"server_error"}}}},{body:T.Object({model:T.Optional(T.String()),messages:T.Array(T.Any()),stream:T.Optional(T.Boolean()),temperature:T.Optional(T.Number()),top_p:T.Optional(T.Number()),max_tokens:T.Optional(T.Number()),stop:T.Optional(T.Union([T.String(),T.Array(T.String())])),presence_penalty:T.Optional(T.Number()),frequency_penalty:T.Optional(T.Number()),tools:T.Optional(T.Array(T.Any())),tool_choice:T.Optional(T.Any()),stream_options:T.Optional(T.Object({include_usage:T.Optional(T.Boolean())}))})}),X}var B9=()=>`chatcmpl-${Date.now()}-${Math.random().toString(36).slice(2,9)}`,$6;var X6=P(()=>{k0();$6=new Map});var Z6=P(()=>{s2();r2();e2();X6()});import E9 from"node:os";import M9 from"node:path";import x9 from"node-machine-id";import S9 from"bytes";import P9 from"ms";var D0=($={},X={})=>{let Z=Array.isArray($)?[...$]:{...$};return Object.entries(X||{}).forEach(([W,j])=>{if(j&&typeof j==="object"&&!Array.isArray(j))Z[W]=D0(Z[W]||{},j);else Z[W]=j}),Z},w0=($)=>{if(!$)return null;if(typeof $==="object")return JSON.parse(JSON.stringify($));return null},N1=($,X)=>{let Z=w0($)||{},W=w0(X)||{};return D0(Z,W)},W6=($,X)=>D0(JSON.parse(JSON.stringify($.global)),X||{}),U1=($,X,Z,W)=>{if($.generators.length>0){let j=$.generators.filter((V)=>V?.type===Z);if(j.length>0){if(W){let V=j.find((Y)=>X.getModelIdentifier(Z,Y)===W);if(V)return W6($,V)}}}if(Object.keys($.global).length>0)return W6($,{});return null},j6,T9=($)=>{if(!$)return null;if($===!0)return{...j6};return D0(j6,$)},V6=($,X)=>{if(!$.generators||$.generators.length===0)return X.map((W)=>({type:W}));let Z=new Set;if($.generators.forEach((W)=>{if(W.type)Z.add(W.type)}),Z.size===0)return X.map((W)=>({type:W}));return Array.from(Z).map((W)=>({type:W}))},Y6=($,X,Z)=>{if($===void 0)return Z;if(typeof $==="number")return $;return X($)??Z},k9=2080,Q6=60000,J6=52428800,G1=($)=>{let X=x9.machineIdSync(),Z={server:{id:`buttress-${X}`,name:`Buttress Server (${X.slice(-8)})`,port:k9,temp_file_dir:M9.join(E9.tmpdir(),".buttress"),session_timeout:Q6,max_body_size:J6},autodiscover:!1},W=D0(Z,w0($)||{}),j=Array.isArray(W.generators)?W.generators:[],{server:V,generators:Y,autodiscover:J,...N}=W;return{autodiscover:T9(J),server:{id:V.id,name:V.name,port:V.port,log_level:V.log_level,temp_file_dir:V.temp_file_dir,max_body_size:Y6(V.max_body_size,S9.parse,J6),session_timeout:Y6(V.session_timeout,P9,Q6)},global:N,generators:j}};var b0=P(()=>{j6={udp:{port:8089,announcements:{enabled:!0,interval:5000},requests:{enabled:!0,responseDelay:100}},http:{enabled:!0,path:"/buttress/info",cors:!0}}});import{z as Z0}from"zod";var N6,U6;var G6=P(()=>{b0();N6={getCapabilities:Z0.tuple([Z0.object({type:Z0.string().optional().default("ggml-llm"),config:Z0.any().optional(),currentClientCapabilities:Z0.any().optional(),options:Z0.any().optional()}).nullable().optional()]),startGenerator:Z0.tuple([Z0.string(),Z0.any().optional()]),finalizeGenerator:Z0.tuple([Z0.string()])},U6={async getCapabilities({backend:$,config:X},Z=null){console.log("[Server] Get Capabilities:",Z);let W=Z||{type:"ggml-llm"},{type:j="ggml-llm",config:V,currentClientCapabilities:Y=null,options:J={}}=W,N=w0(V),H=$.getModelIdentifier(j,N),O=U1(X,$,j,H),q=N1(O,V);if(Object.keys(q).length===0)throw Error("Buttress server missing generator configuration");if(q.backend=q.backend||{},!q.backend.type)q.backend.type=j;return $.getCapabilities(j,Y,{...J,config:q})},async startGenerator({backend:$,config:X,session:Z},W,j){console.log("[Server] Start Generator:",W,j);let V=w0(j),Y=$.getModelIdentifier(W,V),J=U1(X,$,W,Y),N=N1(J,j);if(Object.keys(N).length===0)throw Error("Buttress server missing generator configuration");if(N.backend=N.backend||{},!N.backend.type)N.backend.type=W;let H=await $.startGenerator(W,N);return Z.generators.add(H.id),H},async finalizeGenerator({backend:$,session:X},Z){return console.log("[Server] Finalize Generator:",Z),X.generators.delete(Z),$.finalizeGenerator(Z)}}});import{z as i}from"zod";import{ReadableStream as D9}from"node:stream/web";var H6,O6;var K6=P(()=>{H6={initContext:i.tuple([i.string(),i.any().optional()]),completion:i.tuple([i.string(),i.any().optional()]),tokenize:i.tuple([i.string(),i.any()]),detokenize:i.tuple([i.string(),i.any()]),applyChatTemplate:i.tuple([i.string(),i.any()]),releaseContext:i.tuple([i.string()])},O6={initContext({backend:$,session:X},Z,W){return new D9({async start(j){try{let V=await $.ggmlLlm.initContext(Z,{...W,onProgress:(N)=>{j.enqueue({event:"progress",data:{progress:N}})}});X.initializedContexts.add(Z),await new Promise((N)=>setTimeout(N,1000));let{download:Y,...J}=V||{};j.enqueue({event:"result",data:{result:J}}),j.close()}catch(V){j.error(V)}}})},completion({backend:$},X,Z){return console.log("[Server] Completion:",{id:X,property:Z}),$.ggmlLlm.completion(X,Z)},async tokenize({backend:$},X,Z){return console.log("[Server] Tokenize:",{id:X,property:Z}),$.ggmlLlm.tokenize(X,Z)},async detokenize({backend:$},X,Z){return console.log("[Server] Detokenize:",{id:X,property:Z}),$.ggmlLlm.detokenize(X,Z)},async applyChatTemplate({backend:$},X,Z){return console.log("[Server] Apply Chat Template:",{id:X,property:Z}),$.ggmlLlm.applyChatTemplate(X,Z)},async releaseContext({backend:$,session:X},Z,W){if(console.log("[Server] Release Context:",{id:Z,force:W}),!X.initializedContexts.has(Z))return console.log("[Server] Release Context skipped - not initialized by this session:",{id:Z}),{released:!1,skipped:!0};return X.initializedContexts.delete(Z),$.ggmlLlm.releaseContext(Z,{force:W})}}});import{z as a}from"zod";import{ReadableStream as b9}from"node:stream/web";import v9 from"node:path";var q6,_6;var L6=P(()=>{q6={initContext:a.tuple([a.string(),a.any().optional()]),transcribe:a.tuple([a.string(),a.string(),a.any().optional()]),transcribeData:a.tuple([a.string(),a.union([a.instanceof(Buffer),a.instanceof(Uint8Array)]),a.any().optional()]),releaseContext:a.tuple([a.string()])},_6={initContext({backend:$,session:X},Z,W){return new b9({async start(j){try{let V=await $.ggmlStt.initContext(Z,{...W,onProgress:(N)=>{j.enqueue({event:"progress",data:{progress:N}})}});X.initializedContexts.add(Z),await new Promise((N)=>setTimeout(N,1000));let{download:Y,...J}=V||{};j.enqueue({event:"result",data:{result:J}}),j.close()}catch(V){j.error(V)}}})},async transcribe({backend:$,config:{server:X}},Z,W,j){return console.log("[Server] Transcribe:",{id:Z,audioPath:W,options:j}),$.ggmlStt.transcribe(Z,{audioPath:v9.join(X.temp_file_dir,W),options:j})},async transcribeData({backend:$},X,Z,W){return console.log("[Server] Transcribe Data:",{id:X,audioDataLength:Z?.length||0,options:W}),$.ggmlStt.transcribeData(X,{audioData:Z,options:W})},async releaseContext({backend:$,session:X},Z,W){if(console.log("[Server] Release STT Context:",{id:Z,force:W}),!X.initializedContexts.has(Z))return console.log("[Server] Release STT Context skipped - not initialized by this session:",{id:Z}),{released:!1,skipped:!0};return X.initializedContexts.delete(Z),$.ggmlStt.releaseContext(Z,{force:W})}}});var h9,A6,z6;var B6=P(()=>{G6();K6();L6();h9={common:U6,ggmlLlm:O6,ggmlStt:_6},A6={common:N6,ggmlLlm:H6,ggmlStt:q6},z6=h9});import{Buffer as G4}from"node:buffer";var R6=($)=>{try{return JSON.parse($,(X,Z)=>{if(!Z)return Z;if(Z?.type==="Buffer"&&Z?.data)return G4.from(Z.data,"base64");if(Z?.type==="Uint8Array"&&Z?.data){let W=G4.from(Z.data,"base64");return W.buffer.slice(W.byteOffset,W.byteOffset+W.byteLength)}if(Z?.type==="Error"&&Z?.name&&Z?.message)return Error(Z.name,Z.message);return Z})}catch{return $}},H4=($)=>{try{return JSON.stringify($,(X,Z)=>{if(Z instanceof Error)return{type:"Error",name:Z.name,message:Z.message};if(Z instanceof G4)return{type:"Buffer",data:Z.toString("base64")};if(Z instanceof Uint8Array)return{type:"Uint8Array",data:G4.from(Z).toString("base64")};return Z})}catch{return $}};var w6=()=>{};var H1="1.0",F6=8089;import I9 from"node:dgram";class O1{name="udp";socket=null;announcementTimer=null;config;getServerInfo;port;constructor($,X){this.config=$,this.getServerInfo=X,this.port=$.port??F6}async start(){if(this.socket=I9.createSocket({type:"udp4",reuseAddr:!0}),this.socket.on("message",($,X)=>{this.handleMessage($,X)}),this.socket.on("error",($)=>{console.error("[Autodiscover UDP] Socket error:",$.message)}),await new Promise(($,X)=>{this.socket.bind(this.port,()=>{this.socket.setBroadcast(!0),console.log(`[Autodiscover UDP] Listening on port ${this.port}`),$()}),this.socket.once("error",X)}),this.config.announcements.enabled){let $=this.config.announcements.interval??5000;this.announcementTimer=setInterval(()=>{this.sendAnnouncement()},$),this.sendAnnouncement()}}async stop(){if(this.announcementTimer)clearInterval(this.announcementTimer),this.announcementTimer=null;if(this.socket)await new Promise(($)=>{this.socket.close(()=>$())}),this.socket=null}handleMessage($,X){try{let Z=JSON.parse($.toString());if(Z.t==="QUERY"&&this.config.requests.enabled){let W=Z.d,j=this.config.requests.responseDelay??0,V=j>0?Math.random()*j:0;setTimeout(()=>{this.sendResponse(W.id,X)},V)}}catch{}}sendAnnouncement(){if(!this.socket)return;let $=this.getServerInfo(),X={t:"ANNOUNCE",v:H1,d:{info:$}},Z=Buffer.from(JSON.stringify(X));this.socket.send(Z,0,Z.length,this.port,"255.255.255.255",(W)=>{if(W)console.error("[Autodiscover UDP] Announcement error:",W.message)})}sendResponse($,X){if(!this.socket)return;let Z=this.getServerInfo(),W={t:"RESPONSE",v:H1,d:{request_id:$,info:Z}},j=Buffer.from(JSON.stringify(W));this.socket.send(j,0,j.length,X.port,X.address,(V)=>{if(V)console.error("[Autodiscover UDP] Response error:",V.message)})}}var E6=()=>{};class K1{config;getServerInfo;transports=[];started=!1;constructor($,X){this.config=$;this.getServerInfo=X;if($.udp?.announcements?.enabled||$.udp?.requests?.enabled)this.transports.push(new O1($.udp,X))}async start(){if(this.started)return;(await Promise.allSettled(this.transports.map((X)=>X.start()))).forEach((X,Z)=>{if(X.status==="rejected")console.error(`[Autodiscover] Failed to start ${this.transports[Z].name}:`,X.reason)}),this.started=!0}async stop(){if(!this.started)return;await Promise.allSettled(this.transports.map(($)=>$.stop())),this.started=!1}}var M6=P(()=>{E6()});import C9 from"node:os";var O4=()=>{let $=C9.networkInterfaces();return Object.values($).flat().find((Z)=>Z?.family==="IPv4"&&!Z?.internal)?.address||null};var q1=()=>{};import c from"node:os";import x6 from"node:fs";import S6 from"node:path";import{execSync as K4}from"node:child_process";import P6 from"@iarna/toml";async function _1({modelIds:$=[],defaultConfig:X=null}={}){let Z=[];if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.45"}`),console.log(`Generating model capabilities comparison...
40
- `),Z.push(`${"@fugood/buttress-server"} v${"2.23.0-beta.45"}`),Z.push(`## Model Capabilities Comparison
39
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(_,null,2)),process.exit(0)}catch(Z){console.error("Failed to get capabilities:",Z.message),process.exit(1)}}var p2=($)=>{if(!$)return{repoId:null,filename:null};let[X,Z]=$.split(":");return{repoId:X,filename:Z||null}};var i2=P(()=>{M0();Z1();y2()});var J4={};i6(J4,{testGgmlSttCapabilities:()=>l2,testGgmlLlmCapabilities:()=>c2,status:()=>V9,startModelDownload:()=>Q4,startGenerator:()=>W9,showSttModelsTable:()=>d2,showModelsTable:()=>m2,globalDownloadManager:()=>W1,ggmlStt:()=>Q9,ggmlLlm:()=>Y9,getModelIdentifier:()=>J9,getCapabilities:()=>t,generatorRegistry:()=>X0,finalizeGenerator:()=>j9});async function W9($,X){let W={"ggml-llm":{create:K2,getId:d4},"ggml-stt":{create:h2,getId:$1}}[$];if(!W)throw Error(`Unsupported backend type: ${$}`);let j=W.getId(X);if(!j)throw Error("Buttress generator config missing repo identifier");let V=`${$}:${j}`,Y=X0.get(V);if(Y)return Y.refCount+=1,Y.instance.resetFinalized?.(),{id:Y.id,info:Y.instance.info};let J=await W.create(V,X,{globalDownloadManager:W1}),N={id:V,type:J.type,instance:J,refCount:1};return X0.set(V,N),{id:V,info:J.info}}async function j9($){let X=X0.get($);if(!X)return!1;if(X.refCount-=1,X.refCount<=0){if(await X.instance.finalize(),!(X.instance.hasPendingReleases?.()??!1))X0.delete($)}return!0}function J9($,X){if($==="ggml-llm")return d4(X);if($==="ggml-stt")return $1(X);return null}async function Q4($,X,Z={}){let j={"ggml-llm":q2,"ggml-stt":I2}[$];if(!j)return{started:!1,localPath:null,repoId:null,error:`Unknown backend type: ${$}`};return j(X,W1,Z)}var X0,W1,Z9=($)=>{let X=X0.get($);if(!X)throw Error(`Unknown generator id "${$}"`);return X},O0=($,X)=>{let Z=Z9($);if(Z.type!==X)throw Error(`Generator "${$}" does not support ${X} backend`);return Z.instance},Y9,Q9,V9;var V4=P(()=>{l4();X1();Z1();s0();i2();X0=new Map,W1={downloads:new Map,getDownload($){return this.downloads.get($)||null},setDownload($,X){this.downloads.set($,X)},deleteDownload($){this.downloads.delete($)},isDownloading($){return this.downloads.has($)},getActiveDownloads(){return Array.from(this.downloads.entries()).map(([$,X])=>({localPath:$,promise:X}))}};Y9={async initContext($,X){return O0($,"ggml-llm").initContext(X)},async completion($,X){return O0($,"ggml-llm").completion(X)},async tokenize($,X){return O0($,"ggml-llm").tokenize(X)},async detokenize($,X){return O0($,"ggml-llm").detokenize(X)},async applyChatTemplate($,X){return O0($,"ggml-llm").applyChatTemplate(X)},async releaseContext($,X){let Z=X0.get($);if(!Z)return{released:!0,alreadyReleased:!0};if(Z.type!=="ggml-llm")throw Error(`Generator "${$}" does not support ggml-llm backend`);return Z.instance.releaseContext(X)}},Q9={async initContext($,X){return O0($,"ggml-stt").initContext(X)},async transcribe($,X){return O0($,"ggml-stt").transcribe(X)},async transcribeData($,X){return O0($,"ggml-stt").transcribeData(X)},async releaseContext($,X){let Z=X0.get($);if(!Z)return{released:!0,alreadyReleased:!0};if(Z.type!=="ggml-stt")throw Error(`Generator "${$}" does not support ggml-stt backend`);return Z.instance.releaseContext(X)}};V9={getFullStatus:()=>d1(X0),getGgmlLlmStatus:()=>C4(X0),getGgmlSttStatus:()=>y4(X0),subscribeToStatus:I4,subscribeToStatusWithId:c1,llmStatusTracker:$0,sttStatusTracker:N0,statusEmitter:e}});import{node as N9}from"@elysiajs/node";import{Elysia as U9}from"elysia";var G9,K0=($)=>new U9({adapter:G9?N9():void 0,...$});var k0=P(()=>{G9=typeof process<"u"&&process.versions&&process.versions.node});import{t as Y0}from"elysia";var H9,O9=({store:{serverInfo:$}})=>({id:$.id,name:$.name,version:$.version,generators:$.generators,authentication:$.authentication}),j1=($)=>{let X=K0(),Z=$.autodiscover.http?.path??"/buttress/info";return X.get(Z,O9,{response:H9}),X};var s2=P(()=>{k0();H9=Y0.Object({id:Y0.String(),name:Y0.String(),version:Y0.String(),generators:Y0.Array(Y0.Object({type:Y0.String()})),authentication:Y0.Object({required:Y0.Boolean(),type:Y0.Literal("device-group")})})});import{t as Q0,file as K9}from"elysia";import{writeFile as n2}from"node:fs/promises";import Y1 from"node:path";var q9,Q1;var r2=P(()=>{k0();q9=typeof process<"u"&&process.versions!=null&&process.versions.node!=null,Q1=K0().post("/buttress/upload",async({body:{file:$},store:{config:X}})=>{let Z=`${Date.now()}-${$.name.replace(/[^\dA-Za-z]/g,"_")}`,W=Y1.join(X.server.temp_file_dir,Z);try{if(q9)await n2(W,await $.stream());else await n2(W,await $.arrayBuffer());return{ok:!0,filename:Z}}catch(j){return{ok:!1,error:String(j)}}},{body:Q0.Object({file:Q0.File()}),response:Q0.Object({ok:Q0.Boolean(),filename:Q0.Optional(Q0.String()),error:Q0.Optional(Q0.String())})}).get("/buttress/download/:filename",async({params:{filename:$},store:{config:X},status:Z})=>{let W=Y1.join(X.server.temp_file_dir,$);if(Y1.relative(X.server.temp_file_dir,W).includes(".."))return Z(400),"Invalid file path";return K9(W)},{params:Q0.Object({filename:Q0.String()})})});import J1 from"node:path";import t2 from"node:fs/promises";import{fileURLToPath as _9}from"node:url";var o2,L9=async()=>{let $=[J1.join(o2,"..","public","status.html"),J1.join(o2,"..","..","public","status.html")];return(await Promise.all($.map((Z)=>t2.access(Z).then(()=>Z,()=>null)))).find((Z)=>Z!==null)??null},A9=($)=>{let{status:X}=$;if(X?.getFullStatus)return X.getFullStatus();return{timestamp:new Date().toISOString(),ggmlLlm:{generators:[],history:{}},ggmlStt:{generators:[],history:{}}}},a2=async()=>{let $=await L9();if(!$)return console.error("[Status] Failed to find status.html in candidate paths"),new Response("Status page not found",{status:404,headers:{"Content-Type":"text/plain"}});try{let X=await t2.readFile($,"utf-8");return new Response(X,{headers:{"Content-Type":"text/html; charset=utf-8"}})}catch(X){return console.error("[Status] Failed to serve status page:",X),new Response("Status page not found",{status:404,headers:{"Content-Type":"text/plain"}})}},V1;var e2=P(()=>{k0();o2=J1.dirname(_9(import.meta.url)),V1=K0().get("/status",a2).get("/status/",a2).get("/buttress/status",({store:{backend:$}})=>A9($))});import{t as T,sse as N4}from"elysia";import{cors as z9}from"@elysiajs/cors";async function R9($,X,Z){let j=(X.generators||[]).filter((z)=>z.type==="ggml-llm");if(j.length===0)throw Error('No ggml-llm generator configured. Add a [[generators]] with type = "ggml-llm" to your config.');let V=j[0],Y=Z||V.model?.repo_id;if(Z){let z=j.find((A)=>A.model?.repo_id===Z);if(z)V=z}else Y=V.model?.repo_id;let J=Y,N=$6.get(J);if(N?.initialized)return N;let{generators:H,server:O,...q}=X.global||{},_={...q,...V,model:{...V.model,repo_id:Y}};console.log(`[OpenAI] Creating generator for ${J}`);let{id:G}=await $.startGenerator("ggml-llm",_),w={id:G,config:_,repoId:Y,initialized:!1};return $6.set(J,w),await $.ggmlLlm.initContext(G,{}),w.initialized=!0,console.log(`[OpenAI] Generator ready: ${J}`),w}async function w9($,X,Z,W){let j=$.getReader(),V="",Y=null,J=null,N="stop",H={prompt_tokens:0,completion_tokens:0,total_tokens:0};try{let q=!1;while(!q){let _=await j.read();if({done:q}=_,q)break;let{event:G,data:w}=_.value;if(G==="token"){if(w.content!=null)V+=w.content;else if(w.token!=null)V+=w.token}else if(G==="result"){if(w.text)V=w.text;else if(w.content)V=w.content;if(w.reasoning_content)Y=w.reasoning_content;if(w.tool_calls?.length>0)J=w.tool_calls.map((z,A)=>({id:z.id||`call_${X}_${A}`,type:"function",function:{name:z.function?.name||"",arguments:z.function?.arguments||""}})),N="tool_calls";else N=w.interrupted?"length":"stop";H={prompt_tokens:w.prompt_tokens||w.promptTokens||0,completion_tokens:w.tokens_predicted||w.tokensPredicted||0,total_tokens:(w.prompt_tokens||w.promptTokens||0)+(w.tokens_predicted||w.tokensPredicted||0)}}else if(G==="error")throw Error(w.message)}}finally{j.cancel().catch(()=>{})}let O={role:"assistant",content:V||null};if(Y)O.reasoning_content=Y;if(J)O.tool_calls=J;return{id:X,object:"chat.completion",created:Z,model:W,choices:[{index:0,message:O,finish_reason:N}],usage:H}}function U4({global:$}){let X=K0({prefix:"/oai-compat"});return X.use(z9({origin:$?.openai_compat?.cors_allowed_origins??!1,methods:["GET","POST","OPTIONS"],allowedHeaders:["Content-Type","Authorization"],maxAge:86400,preflight:!0})),X.get("/v1/models",({store:Z})=>{let{config:W}=Z,Y=(W.generators||[]).filter((J)=>J.type==="ggml-llm").map((J)=>{return{id:J.model?.repo_id||"ggml-llm",object:"model",created:Math.floor(Date.now()/1000),owned_by:"local"}});if(Y.length===0)Y.push({id:"ggml-llm",object:"model",created:Math.floor(Date.now()/1000),owned_by:"local"});return{object:"list",data:Y}}),X.post("/v1/chat/completions",async function*({body:W,set:j,store:V}){let{config:Y,backend:J}=V,{messages:N=[],stream:H=!1,model:O,tools:q,temperature:_,stop:G,top_p:w,max_tokens:z,presence_penalty:A,frequency_penalty:R,tool_choice:U,stream_options:Q}=W;if(!N||N.length===0)return j.status=400,{error:{message:"messages is required and must not be empty",type:"invalid_request_error"}};try{let K=await R9(J,Y,O),B=B9(),L=Math.floor(Date.now()/1000),F=K.repoId||"ggml-llm",E={reasoning_format:"auto",messages:N,jinja:!0,add_generation_prompt:!0};if(_!=null)E.temperature=_;if(w!=null)E.top_p=w;if(z!=null)E.n_predict=z;if(G!=null)E.stop=Array.isArray(G)?G:[G];if(A!=null)E.presence_penalty=A;if(R!=null)E.frequency_penalty=R;if(q!=null)E.tools=q;if(U!=null)E.tool_choice=U;let M=await J.ggmlLlm.completion(K.id,{options:E});if(!H)return await w9(M,B,L,F);let x=Q?.include_usage===!0,D=M.getReader(),I="",b="",u=new Map,r=new Map;try{let k=!1;while(!k){let d=await D.read();if({done:k}=d,k)break;let{event:v,data:S}=d.value;if(v==="token"){let C={};if(S.content!=null){let h=S.content;if(h.length>I.length)C.content=h.slice(I.length),I=h}if(S.reasoning_content!=null){let h=S.reasoning_content;if(h.length>b.length)C.reasoning_content=h.slice(b.length),b=h}if(S.tool_calls?.length>0){let h=[];if(S.tool_calls.forEach((p,f)=>{let s={index:f};if(!r.has(f))r.set(f,p.id||`call_${B}_${f}`),s.id=r.get(f),s.type="function";let W0=p.function?.arguments||"",A0=u.get(f)||"",l={};if(!u.has(f)&&p.function?.name)l.name=p.function.name;if(W0.length>A0.length)l.arguments=W0.slice(A0.length),u.set(f,W0);if(Object.keys(l).length>0)s.function=l,h.push(s);else if(s.id)s.function={name:p.function?.name||"",arguments:""},h.push(s)}),h.length>0)C.tool_calls=h}if(Object.keys(C).length>0)yield N4({data:JSON.stringify({id:B,object:"chat.completion.chunk",created:L,model:F,choices:[{index:0,delta:C,finish_reason:null}]})})}else if(v==="result"){let C="stop";if(S.tool_calls?.length>0||r.size>0)C="tool_calls";else if(S.interrupted)C="length";let h={id:B,object:"chat.completion.chunk",created:L,model:F,choices:[{index:0,delta:{},finish_reason:C}]};if(x)h.usage={prompt_tokens:S.prompt_tokens||S.promptTokens||0,completion_tokens:S.tokens_predicted||S.tokensPredicted||0,total_tokens:(S.prompt_tokens||S.promptTokens||0)+(S.tokens_predicted||S.tokensPredicted||0)};yield N4({data:JSON.stringify(h)})}else if(v==="error")yield N4({data:JSON.stringify({error:{message:S.message,type:"server_error"}})})}yield N4({data:"[DONE]"})}finally{D.cancel().catch(()=>{})}}catch(K){return console.error("[OpenAI] Chat completion error:",K),j.status=500,{error:{message:K.message||"Internal server error",type:"server_error"}}}},{body:T.Object({model:T.Optional(T.String()),messages:T.Array(T.Any()),stream:T.Optional(T.Boolean()),temperature:T.Optional(T.Number()),top_p:T.Optional(T.Number()),max_tokens:T.Optional(T.Number()),stop:T.Optional(T.Union([T.String(),T.Array(T.String())])),presence_penalty:T.Optional(T.Number()),frequency_penalty:T.Optional(T.Number()),tools:T.Optional(T.Array(T.Any())),tool_choice:T.Optional(T.Any()),stream_options:T.Optional(T.Object({include_usage:T.Optional(T.Boolean())}))})}),X}var B9=()=>`chatcmpl-${Date.now()}-${Math.random().toString(36).slice(2,9)}`,$6;var X6=P(()=>{k0();$6=new Map});var Z6=P(()=>{s2();r2();e2();X6()});import E9 from"node:os";import M9 from"node:path";import x9 from"node-machine-id";import S9 from"bytes";import P9 from"ms";var D0=($={},X={})=>{let Z=Array.isArray($)?[...$]:{...$};return Object.entries(X||{}).forEach(([W,j])=>{if(j&&typeof j==="object"&&!Array.isArray(j))Z[W]=D0(Z[W]||{},j);else Z[W]=j}),Z},w0=($)=>{if(!$)return null;if(typeof $==="object")return JSON.parse(JSON.stringify($));return null},N1=($,X)=>{let Z=w0($)||{},W=w0(X)||{};return D0(Z,W)},W6=($,X)=>D0(JSON.parse(JSON.stringify($.global)),X||{}),U1=($,X,Z,W)=>{if($.generators.length>0){let j=$.generators.filter((V)=>V?.type===Z);if(j.length>0){if(W){let V=j.find((Y)=>X.getModelIdentifier(Z,Y)===W);if(V)return W6($,V)}}}if(Object.keys($.global).length>0)return W6($,{});return null},j6,T9=($)=>{if(!$)return null;if($===!0)return{...j6};return D0(j6,$)},V6=($,X)=>{if(!$.generators||$.generators.length===0)return X.map((W)=>({type:W}));let Z=new Set;if($.generators.forEach((W)=>{if(W.type)Z.add(W.type)}),Z.size===0)return X.map((W)=>({type:W}));return Array.from(Z).map((W)=>({type:W}))},Y6=($,X,Z)=>{if($===void 0)return Z;if(typeof $==="number")return $;return X($)??Z},k9=2080,Q6=60000,J6=52428800,G1=($)=>{let X=x9.machineIdSync(),Z={server:{id:`buttress-${X}`,name:`Buttress Server (${X.slice(-8)})`,port:k9,temp_file_dir:M9.join(E9.tmpdir(),".buttress"),session_timeout:Q6,max_body_size:J6},autodiscover:!1},W=D0(Z,w0($)||{}),j=Array.isArray(W.generators)?W.generators:[],{server:V,generators:Y,autodiscover:J,...N}=W;return{autodiscover:T9(J),server:{id:V.id,name:V.name,port:V.port,log_level:V.log_level,temp_file_dir:V.temp_file_dir,max_body_size:Y6(V.max_body_size,S9.parse,J6),session_timeout:Y6(V.session_timeout,P9,Q6)},global:N,generators:j}};var b0=P(()=>{j6={udp:{port:8089,announcements:{enabled:!0,interval:5000},requests:{enabled:!0,responseDelay:100}},http:{enabled:!0,path:"/buttress/info",cors:!0}}});import{z as Z0}from"zod";var N6,U6;var G6=P(()=>{b0();N6={getCapabilities:Z0.tuple([Z0.object({type:Z0.string().optional().default("ggml-llm"),config:Z0.any().optional(),currentClientCapabilities:Z0.any().optional(),options:Z0.any().optional()}).nullable().optional()]),startGenerator:Z0.tuple([Z0.string(),Z0.any().optional()]),finalizeGenerator:Z0.tuple([Z0.string()])},U6={async getCapabilities({backend:$,config:X},Z=null){console.log("[Server] Get Capabilities:",Z);let W=Z||{type:"ggml-llm"},{type:j="ggml-llm",config:V,currentClientCapabilities:Y=null,options:J={}}=W,N=w0(V),H=$.getModelIdentifier(j,N),O=U1(X,$,j,H),q=N1(O,V);if(Object.keys(q).length===0)throw Error("Buttress server missing generator configuration");if(q.backend=q.backend||{},!q.backend.type)q.backend.type=j;return $.getCapabilities(j,Y,{...J,config:q})},async startGenerator({backend:$,config:X,session:Z},W,j){console.log("[Server] Start Generator:",W,j);let V=w0(j),Y=$.getModelIdentifier(W,V),J=U1(X,$,W,Y),N=N1(J,j);if(Object.keys(N).length===0)throw Error("Buttress server missing generator configuration");if(N.backend=N.backend||{},!N.backend.type)N.backend.type=W;let H=await $.startGenerator(W,N);return Z.generators.add(H.id),H},async finalizeGenerator({backend:$,session:X},Z){return console.log("[Server] Finalize Generator:",Z),X.generators.delete(Z),$.finalizeGenerator(Z)}}});import{z as i}from"zod";import{ReadableStream as D9}from"node:stream/web";var H6,O6;var K6=P(()=>{H6={initContext:i.tuple([i.string(),i.any().optional()]),completion:i.tuple([i.string(),i.any().optional()]),tokenize:i.tuple([i.string(),i.any()]),detokenize:i.tuple([i.string(),i.any()]),applyChatTemplate:i.tuple([i.string(),i.any()]),releaseContext:i.tuple([i.string()])},O6={initContext({backend:$,session:X},Z,W){return new D9({async start(j){try{let V=await $.ggmlLlm.initContext(Z,{...W,onProgress:(N)=>{j.enqueue({event:"progress",data:{progress:N}})}});X.initializedContexts.add(Z),await new Promise((N)=>setTimeout(N,1000));let{download:Y,...J}=V||{};j.enqueue({event:"result",data:{result:J}}),j.close()}catch(V){j.error(V)}}})},completion({backend:$},X,Z){return console.log("[Server] Completion:",{id:X,property:Z}),$.ggmlLlm.completion(X,Z)},async tokenize({backend:$},X,Z){return console.log("[Server] Tokenize:",{id:X,property:Z}),$.ggmlLlm.tokenize(X,Z)},async detokenize({backend:$},X,Z){return console.log("[Server] Detokenize:",{id:X,property:Z}),$.ggmlLlm.detokenize(X,Z)},async applyChatTemplate({backend:$},X,Z){return console.log("[Server] Apply Chat Template:",{id:X,property:Z}),$.ggmlLlm.applyChatTemplate(X,Z)},async releaseContext({backend:$,session:X},Z,W){if(console.log("[Server] Release Context:",{id:Z,force:W}),!X.initializedContexts.has(Z))return console.log("[Server] Release Context skipped - not initialized by this session:",{id:Z}),{released:!1,skipped:!0};return X.initializedContexts.delete(Z),$.ggmlLlm.releaseContext(Z,{force:W})}}});import{z as a}from"zod";import{ReadableStream as b9}from"node:stream/web";import v9 from"node:path";var q6,_6;var L6=P(()=>{q6={initContext:a.tuple([a.string(),a.any().optional()]),transcribe:a.tuple([a.string(),a.string(),a.any().optional()]),transcribeData:a.tuple([a.string(),a.union([a.instanceof(Buffer),a.instanceof(Uint8Array)]),a.any().optional()]),releaseContext:a.tuple([a.string()])},_6={initContext({backend:$,session:X},Z,W){return new b9({async start(j){try{let V=await $.ggmlStt.initContext(Z,{...W,onProgress:(N)=>{j.enqueue({event:"progress",data:{progress:N}})}});X.initializedContexts.add(Z),await new Promise((N)=>setTimeout(N,1000));let{download:Y,...J}=V||{};j.enqueue({event:"result",data:{result:J}}),j.close()}catch(V){j.error(V)}}})},async transcribe({backend:$,config:{server:X}},Z,W,j){return console.log("[Server] Transcribe:",{id:Z,audioPath:W,options:j}),$.ggmlStt.transcribe(Z,{audioPath:v9.join(X.temp_file_dir,W),options:j})},async transcribeData({backend:$},X,Z,W){return console.log("[Server] Transcribe Data:",{id:X,audioDataLength:Z?.length||0,options:W}),$.ggmlStt.transcribeData(X,{audioData:Z,options:W})},async releaseContext({backend:$,session:X},Z,W){if(console.log("[Server] Release STT Context:",{id:Z,force:W}),!X.initializedContexts.has(Z))return console.log("[Server] Release STT Context skipped - not initialized by this session:",{id:Z}),{released:!1,skipped:!0};return X.initializedContexts.delete(Z),$.ggmlStt.releaseContext(Z,{force:W})}}});var h9,A6,z6;var B6=P(()=>{G6();K6();L6();h9={common:U6,ggmlLlm:O6,ggmlStt:_6},A6={common:N6,ggmlLlm:H6,ggmlStt:q6},z6=h9});import{Buffer as G4}from"node:buffer";var R6=($)=>{try{return JSON.parse($,(X,Z)=>{if(!Z)return Z;if(Z?.type==="Buffer"&&Z?.data)return G4.from(Z.data,"base64");if(Z?.type==="Uint8Array"&&Z?.data){let W=G4.from(Z.data,"base64");return W.buffer.slice(W.byteOffset,W.byteOffset+W.byteLength)}if(Z?.type==="Error"&&Z?.name&&Z?.message)return Error(Z.name,Z.message);return Z})}catch{return $}},H4=($)=>{try{return JSON.stringify($,(X,Z)=>{if(Z instanceof Error)return{type:"Error",name:Z.name,message:Z.message};if(Z instanceof G4)return{type:"Buffer",data:Z.toString("base64")};if(Z instanceof Uint8Array)return{type:"Uint8Array",data:G4.from(Z).toString("base64")};return Z})}catch{return $}};var w6=()=>{};var H1="1.0",F6=8089;import I9 from"node:dgram";class O1{name="udp";socket=null;announcementTimer=null;config;getServerInfo;port;constructor($,X){this.config=$,this.getServerInfo=X,this.port=$.port??F6}async start(){if(this.socket=I9.createSocket({type:"udp4",reuseAddr:!0}),this.socket.on("message",($,X)=>{this.handleMessage($,X)}),this.socket.on("error",($)=>{console.error("[Autodiscover UDP] Socket error:",$.message)}),await new Promise(($,X)=>{this.socket.bind(this.port,()=>{this.socket.setBroadcast(!0),console.log(`[Autodiscover UDP] Listening on port ${this.port}`),$()}),this.socket.once("error",X)}),this.config.announcements.enabled){let $=this.config.announcements.interval??5000;this.announcementTimer=setInterval(()=>{this.sendAnnouncement()},$),this.sendAnnouncement()}}async stop(){if(this.announcementTimer)clearInterval(this.announcementTimer),this.announcementTimer=null;if(this.socket)await new Promise(($)=>{this.socket.close(()=>$())}),this.socket=null}handleMessage($,X){try{let Z=JSON.parse($.toString());if(Z.t==="QUERY"&&this.config.requests.enabled){let W=Z.d,j=this.config.requests.responseDelay??0,V=j>0?Math.random()*j:0;setTimeout(()=>{this.sendResponse(W.id,X)},V)}}catch{}}sendAnnouncement(){if(!this.socket)return;let $=this.getServerInfo(),X={t:"ANNOUNCE",v:H1,d:{info:$}},Z=Buffer.from(JSON.stringify(X));this.socket.send(Z,0,Z.length,this.port,"255.255.255.255",(W)=>{if(W)console.error("[Autodiscover UDP] Announcement error:",W.message)})}sendResponse($,X){if(!this.socket)return;let Z=this.getServerInfo(),W={t:"RESPONSE",v:H1,d:{request_id:$,info:Z}},j=Buffer.from(JSON.stringify(W));this.socket.send(j,0,j.length,X.port,X.address,(V)=>{if(V)console.error("[Autodiscover UDP] Response error:",V.message)})}}var E6=()=>{};class K1{config;getServerInfo;transports=[];started=!1;constructor($,X){this.config=$;this.getServerInfo=X;if($.udp?.announcements?.enabled||$.udp?.requests?.enabled)this.transports.push(new O1($.udp,X))}async start(){if(this.started)return;(await Promise.allSettled(this.transports.map((X)=>X.start()))).forEach((X,Z)=>{if(X.status==="rejected")console.error(`[Autodiscover] Failed to start ${this.transports[Z].name}:`,X.reason)}),this.started=!0}async stop(){if(!this.started)return;await Promise.allSettled(this.transports.map(($)=>$.stop())),this.started=!1}}var M6=P(()=>{E6()});import C9 from"node:os";var O4=()=>{let $=C9.networkInterfaces();return Object.values($).flat().find((Z)=>Z?.family==="IPv4"&&!Z?.internal)?.address||null};var q1=()=>{};import c from"node:os";import x6 from"node:fs";import S6 from"node:path";import{execSync as K4}from"node:child_process";import P6 from"@iarna/toml";async function _1({modelIds:$=[],defaultConfig:X=null}={}){let Z=[];if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.47"}`),console.log(`Generating model capabilities comparison...
40
+ `),Z.push(`${"@fugood/buttress-server"} v${"2.23.0-beta.47"}`),Z.push(`## Model Capabilities Comparison
41
41
  `),!$||$.length===0)console.error("Error: No model IDs provided"),process.exit(1);try{let W=(U={},Q={})=>{let K=Array.isArray(U)?[...U]:{...U};return Object.entries(Q||{}).forEach(([B,L])=>{if(L&&typeof L==="object"&&!Array.isArray(L))K[B]=W(K[B]||{},L);else K[B]=L}),K},j=X||{},{server:V,generators:Y=[],...J}=j,N=(U)=>W(JSON.parse(JSON.stringify(J)),U||{}),H=(U)=>{if(Array.isArray(Y)&&Y.length>0){let Q=Y.filter((K)=>K?.type==="ggml-llm");if(Q.length>0&&U){let K=Q.find((B)=>B.model?.repo_id===U);if(K)return N(K)}}return Object.keys(J).length>0?N({}):null},O=[];for(let U=0;U<$.length;U+=1){let Q=$[U];console.log(`[${U+1}/${$.length}] Analyzing ${Q}...`);let K=H(Q);K={...K||{},model:{...J.runtime,...K?.model||{},repo_id:Q}};let B=await t("ggml-llm",null,{config:K,includeBreakdown:!0});O.push({modelId:Q,capabilities:B,modelInfo:B.buttress?.selected||null,modelConfig:B.modelConfig||null})}let q=(U)=>U?(U/1024/1024/1024).toFixed(2):"N/A",_=(U)=>U?"✅":"\uD83D\uDEAB";Z.push("| Model ID | Size (GB) | Context Size | KV Cache Size (GB) | Recurrent Mem (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |"),Z.push("|----------|-----------|--------------|--------------------|--------------------|----------------------------|-----------------|-----------------|"),O.forEach(({modelId:U,modelInfo:Q,modelConfig:K})=>{let B=q(Q?.modelBytes),L=K?.nCtx||Q?.kvInfo?.nCtxTrain||"N/A",F=G0(Q),E=Number(L),M=Q?.kvCacheBytes||(F&&Number.isFinite(E)&&E>0?F(E):F&&F(Q?.kvInfo?.nCtxTrain||0))||null,x=q(M),D=Q?.recurrentMemoryBytes||0,I=D>0?q(D):"-",b=q(Q?.modelBytes&&(M!=null||D>0)?Q.modelBytes+(M||0)+D:Q?.fit?.totalRequiredBytes),u=_(Q?.fit?.fitsInGpu),r=_(Q?.fit?.fitsInCpu);Z.push(`| ${U} | ${B} | ${L} | ${x} | ${I} | ${b} | ${u} | ${r} |`);let k=Q?.memoryLimitedCtx!=null||Q?.limitedFit!=null,d=!Q?.fit?.fitsInGpu||!Q?.fit?.fitsInCpu;if(k&&d){let v=Q?.memoryLimitedCtx||L,S=Number(v),C=Q?.limitedKvCacheBytes||F&&Number.isFinite(S)&&S>0&&F(S)||null,h=q(C),p=q(Q?.modelBytes&&(C!=null||D>0)?Q.modelBytes+(C||0)+D:Q?.limitedFit?.totalRequiredBytes),f=_(Q?.limitedFit?.fitsInGpu),s=_(Q?.limitedFit?.fitsInCpu);if(v!==L||h!==x||p!==b)Z.push(`| ↳ Limited | ${B} | ${v} | ${h} | ${I} | ${p} | ${f} | ${s} |`)}}),Z.push(`
42
42
  ---`),Z.push(`
43
43
  ### System Information`);let G=null;if(process.platform!=="win32")try{G=K4("uname -a",{encoding:"utf8"}).trim()}catch{}if(G)Z.push(`- **System:** ${G}`);else Z.push(`- **Hostname:** ${c.hostname()}`),Z.push(`- **OS:** ${c.type()} ${c.release()}`);if(Z.push(`- **Platform:** ${process.platform}`),Z.push(`- **CPU Cores:** ${c.cpus().length}`),Z.push(`- **Total System Memory:** ${(c.totalmem()/1024/1024/1024).toFixed(2)} GB`),O.length>0){let Q=O[0].capabilities.buttress?.selected;if(Q){let K=Q.cpuTotalBytes>0?(Q.cpuUsableBytes/Q.cpuTotalBytes*100).toFixed(0):0;if(Z.push(`- **Usable CPU Memory:** ${(Q.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(Q.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Q.hasGpu){let B=Q.devices.filter((L)=>L.type==="gpu");if(B.length>0){let L=B[0];Z.push(`- **GPU Backend:** ${L.backend}`),Z.push(`- **GPU Name:** ${L.deviceName}`),Z.push(`- **GPU Total Memory:** ${(L.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let F=Q.gpuTotalBytes>0?(Q.gpuUsableBytes/Q.gpuTotalBytes*100).toFixed(0):0;Z.push(`- **GPU Usable Memory:** ${(Q.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${F}% of ${(Q.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else Z.push("- **GPU:** Not available")}}Z.push(`
44
44
  ### Command Used`);let w=process.argv.slice(2).join(" ");if(Z.push(`\`\`\`bash
45
45
  ${process.argv[0]} ${process.argv[1]} ${w}
46
46
  \`\`\``),Z.push(`
47
- ### Package Information`),Z.push(`- **Name:** ${"@fugood/buttress-server"}`),Z.push(`- **Version:** ${"2.23.0-beta.45"}`),X&&Object.keys(X).length>0){Z.push(`
47
+ ### Package Information`),Z.push(`- **Name:** ${"@fugood/buttress-server"}`),Z.push(`- **Version:** ${"2.23.0-beta.47"}`),X&&Object.keys(X).length>0){Z.push(`
48
48
  ### Configuration`),Z.push("<details>"),Z.push("<summary>Click to expand TOML configuration</summary>"),Z.push("\n```toml");try{let U=P6.stringify(X);Z.push(U)}catch(U){Z.push("# Error serializing config"),Z.push(JSON.stringify(X,null,2))}Z.push("```"),Z.push("</details>")}let A=`ggml-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,"-").split("T")[0]}.md`,R=S6.join(process.cwd(),A);x6.writeFileSync(R,Z.join(`
49
49
  `),"utf8"),console.log(`
50
- Model capabilities table saved to: ${R}`),process.exit(0)}catch(W){console.error("Failed to generate model table:",W.message),process.exit(1)}}async function k6({modelId:$=null,defaultConfig:X=null}={}){if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.45"}`),console.log("Testing capabilities for backend: ggml-llm"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let Z=X||{},{server:W,generators:j=[],...V}=Z,Y=(G={},w={})=>{let z=Array.isArray(G)?[...G]:{...G};return Object.entries(w||{}).forEach(([A,R])=>{if(R&&typeof R==="object"&&!Array.isArray(R))z[A]=Y(z[A]||{},R);else z[A]=R}),z},J=(G)=>Y(JSON.parse(JSON.stringify(V)),G||{}),H=((G)=>{if(Array.isArray(j)&&j.length>0){let w=j.filter((z)=>z?.type==="ggml-llm");if(w.length>0){if(G){let z=w.find((A)=>A.model?.repo_id===G);if(z)return J(z)}}}if(Object.keys(V).length>0)return J({});return null})($);if($)H={...H||{},model:{...H?.model||{},repo_id:$}};let O=await t("ggml-llm",null,{config:H,includeBreakdown:!0}),q=O.buttress?.selected||null,_=O.modelConfig||null;if($||_?.repoId){console.log(`
50
+ Model capabilities table saved to: ${R}`),process.exit(0)}catch(W){console.error("Failed to generate model table:",W.message),process.exit(1)}}async function k6({modelId:$=null,defaultConfig:X=null}={}){if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.47"}`),console.log("Testing capabilities for backend: ggml-llm"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let Z=X||{},{server:W,generators:j=[],...V}=Z,Y=(G={},w={})=>{let z=Array.isArray(G)?[...G]:{...G};return Object.entries(w||{}).forEach(([A,R])=>{if(R&&typeof R==="object"&&!Array.isArray(R))z[A]=Y(z[A]||{},R);else z[A]=R}),z},J=(G)=>Y(JSON.parse(JSON.stringify(V)),G||{}),H=((G)=>{if(Array.isArray(j)&&j.length>0){let w=j.filter((z)=>z?.type==="ggml-llm");if(w.length>0){if(G){let z=w.find((A)=>A.model?.repo_id===G);if(z)return J(z)}}}if(Object.keys(V).length>0)return J({});return null})($);if($)H={...H||{},model:{...H?.model||{},repo_id:$}};let O=await t("ggml-llm",null,{config:H,includeBreakdown:!0}),q=O.buttress?.selected||null,_=O.modelConfig||null;if($||_?.repoId){console.log(`
51
51
  === Model Information ===`);let G=$||_?.repoId;if(console.log(`Repository ID: ${G}`),_?.quantization)console.log(`Quantization: ${_.quantization}`);if(_?.nCtx)console.log(`Context Length: ${_.nCtx}`);if(q?.quantization){let{fileType:A}=q.quantization;if(A!=null)console.log(`Model File Type (GGUF): ${A}`)}let w=_?.cache_type_k||"f16",z=_?.cache_type_v||"f16";if(console.log(`KV Cache Type: K=${w}, V=${z}`),q?.modelBytes&&q?.kvCacheBytes!=null){if(console.log(`Model Size: ${(q.modelBytes/1024/1024/1024).toFixed(2)} GB`),q.kvInfo)console.log(`KV Cache Size: ${(q.kvCacheBytes/1024/1024/1024).toFixed(2)} GB (KV info: ${JSON.stringify(q.kvInfo)})`);else console.log(`KV Cache Size: ${(q.kvCacheBytes/1024/1024/1024).toFixed(2)} GB`);let A=q.recurrentMemoryBytes||0;if(A>0)console.log(`Recurrent Memory: ${(A/1024/1024/1024).toFixed(2)} GB`);let R=q.modelBytes+q.kvCacheBytes+A;if(console.log(`Total Required Memory: ${(R/1024/1024/1024).toFixed(2)} GB`),q.memoryLimitedCtx!=null){let U=q.memoryLimitedCtx,Q=q.kvInfo?.nCtxTrain;if(Q)console.log(`
52
52
  Memory-Limited Context: ${U} (Train: ${Q})`);else console.log(`
53
53
  Memory-Limited Context: ${U}`);if(q.limitedKvCacheBytes!=null)console.log(`Limited KV Cache Size: ${(q.limitedKvCacheBytes/1024/1024/1024).toFixed(2)} GB`)}}else if(O.buttress?.selected?.fit){let{totalRequiredBytes:A}=O.buttress.selected.fit;console.log(`Total Required Memory: ${(A/1024/1024/1024).toFixed(2)} GB`)}}if(O.buttress?.selected){let{selected:G}=O.buttress;console.log(`
@@ -56,25 +56,25 @@ Memory-Limited Context: ${U}`);if(q.limitedKvCacheBytes!=null)console.log(`Limit
56
56
  Backend Variant: ${G.variant}`),console.log(`Performance Score: ${G.score}`),G.fit){if(console.log(`
57
57
  --- Model Fit Analysis ---`),console.log(`Fits in GPU: ${G.fit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU: ${G.fit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor: ${G.fit.limiting}`),G.limitedFit)console.log(`
58
58
  --- Memory-Limited Fit Analysis ---`),console.log(`Limited Total Required: ${(G.limitedFit.totalRequiredBytes/1024/1024/1024).toFixed(2)} GB`),console.log(`Fits in GPU (Limited): ${G.limitedFit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU (Limited): ${G.limitedFit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor (Limited): ${G.limitedFit.limiting}`)}}console.log(`
59
- === Full Capabilities JSON ===`),console.log(JSON.stringify(O,null,2)),process.exit(0)}catch(Z){console.error("Failed to get capabilities:",Z.message),process.exit(1)}}async function L1({modelIds:$=[],defaultConfig:X=null}={}){let Z=[];if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.45"}`),console.log(`Generating STT model capabilities comparison...
60
- `),Z.push(`${"@fugood/buttress-server"} v${"2.23.0-beta.45"}`),Z.push(`## STT Model Capabilities Comparison
59
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(O,null,2)),process.exit(0)}catch(Z){console.error("Failed to get capabilities:",Z.message),process.exit(1)}}async function L1({modelIds:$=[],defaultConfig:X=null}={}){let Z=[];if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.47"}`),console.log(`Generating STT model capabilities comparison...
60
+ `),Z.push(`${"@fugood/buttress-server"} v${"2.23.0-beta.47"}`),Z.push(`## STT Model Capabilities Comparison
61
61
  `),!$||$.length===0)console.error("Error: No model IDs provided"),process.exit(1);try{let W=(U={},Q={})=>{let K=Array.isArray(U)?[...U]:{...U};return Object.entries(Q||{}).forEach(([B,L])=>{if(L&&typeof L==="object"&&!Array.isArray(L))K[B]=W(K[B]||{},L);else K[B]=L}),K},j=X||{},{server:V,generators:Y=[],...J}=j,N=(U)=>W(JSON.parse(JSON.stringify(J)),U||{}),H=(U)=>{if(Array.isArray(Y)&&Y.length>0){let Q=Y.filter((K)=>K?.type==="ggml-stt");if(Q.length>0&&U){let K=Q.find((B)=>B.model?.repo_id===U);if(K)return N(K)}}return Object.keys(J).length>0?N({}):null},O=[];for(let U=0;U<$.length;U+=1){let Q=$[U],{repoId:K,filename:B}=T6(Q);console.log(`[${U+1}/${$.length}] Analyzing ${Q}...`);let L=H(K);L={...L||{},model:{...L?.model||{},repo_id:K,...B&&{filename:B}}};let F=await t("ggml-stt",null,{config:L,includeBreakdown:!0});O.push({modelId:Q,repoId:K,filename:B,capabilities:F,modelInfo:F.buttress?.selected||null,modelConfig:F.modelConfig||null})}let q=(U)=>U?(U/1024/1024).toFixed(1):"N/A",_=(U)=>U?"✅":"\uD83D\uDEAB";Z.push("| Model | Size (MB) | Processing Buffer (MB) | Total Required (MB) | Fits GPU | Fits CPU |"),Z.push("|-------|-----------|------------------------|---------------------|----------|----------|"),O.forEach(({modelId:U,modelInfo:Q})=>{let K=q(Q?.modelBytes),B=q(Q?.processingBytes||Q?.kvCacheBytes),L=q(Q?.fit?.totalRequiredBytes),F=_(Q?.fit?.fitsInGpu),E=_(Q?.fit?.fitsInCpu);Z.push(`| ${U} | ${K} | ${B} | ${L} | ${F} | ${E} |`)}),Z.push(`
62
62
  ---`),Z.push(`
63
63
  ### System Information`);let G=null;if(process.platform!=="win32")try{G=K4("uname -a",{encoding:"utf8"}).trim()}catch{}if(G)Z.push(`- **System:** ${G}`);else Z.push(`- **Hostname:** ${c.hostname()}`),Z.push(`- **OS:** ${c.type()} ${c.release()}`);if(Z.push(`- **Platform:** ${process.platform}`),Z.push(`- **CPU Cores:** ${c.cpus().length}`),Z.push(`- **Total System Memory:** ${(c.totalmem()/1024/1024/1024).toFixed(2)} GB`),O.length>0){let Q=O[0].capabilities.buttress?.selected;if(Q){let K=Q.cpuTotalBytes>0?(Q.cpuUsableBytes/Q.cpuTotalBytes*100).toFixed(0):0;if(Z.push(`- **Usable CPU Memory:** ${(Q.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(Q.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Q.hasGpu){let B=Q.devices.filter((L)=>L.type==="gpu");if(B.length>0){let L=B[0];Z.push(`- **GPU Backend:** ${L.backend}`),Z.push(`- **GPU Name:** ${L.deviceName}`),Z.push(`- **GPU Total Memory:** ${(L.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let F=Q.gpuTotalBytes>0?(Q.gpuUsableBytes/Q.gpuTotalBytes*100).toFixed(0):0;Z.push(`- **GPU Usable Memory:** ${(Q.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${F}% of ${(Q.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else Z.push("- **GPU:** Not available")}}Z.push(`
64
64
  ### Command Used`);let w=process.argv.slice(2).join(" ");if(Z.push(`\`\`\`bash
65
65
  ${process.argv[0]} ${process.argv[1]} ${w}
66
66
  \`\`\``),Z.push(`
67
- ### Package Information`),Z.push(`- **Name:** ${"@fugood/buttress-server"}`),Z.push(`- **Version:** ${"2.23.0-beta.45"}`),X&&Object.keys(X).length>0){Z.push(`
67
+ ### Package Information`),Z.push(`- **Name:** ${"@fugood/buttress-server"}`),Z.push(`- **Version:** ${"2.23.0-beta.47"}`),X&&Object.keys(X).length>0){Z.push(`
68
68
  ### Configuration`),Z.push("<details>"),Z.push("<summary>Click to expand TOML configuration</summary>"),Z.push("\n```toml");try{let U=P6.stringify(X);Z.push(U)}catch(U){Z.push("# Error serializing config"),Z.push(JSON.stringify(X,null,2))}Z.push("```"),Z.push("</details>")}let A=`ggml-stt-model-capabilities-${new Date().toISOString().replace(/[.:]/g,"-").split("T")[0]}.md`,R=S6.join(process.cwd(),A);x6.writeFileSync(R,Z.join(`
69
69
  `),"utf8"),console.log(`
70
- STT model capabilities table saved to: ${R}`),process.exit(0)}catch(W){console.error("Failed to generate STT model table:",W.message),process.exit(1)}}async function D6({modelId:$=null,defaultConfig:X=null}={}){if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.45"}`),console.log("Testing capabilities for backend: ggml-stt"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let{repoId:Z,filename:W}=T6($),j=X||{},{server:V,generators:Y=[],...J}=j,N=(z={},A={})=>{let R=Array.isArray(z)?[...z]:{...z};return Object.entries(A||{}).forEach(([U,Q])=>{if(Q&&typeof Q==="object"&&!Array.isArray(Q))R[U]=N(R[U]||{},Q);else R[U]=Q}),R},H=(z)=>N(JSON.parse(JSON.stringify(J)),z||{}),q=((z)=>{if(Array.isArray(Y)&&Y.length>0){let A=Y.filter((R)=>R?.type==="ggml-stt");if(A.length>0){if(z){let R=A.find((U)=>U.model?.repo_id===z);if(R)return H(R)}}}if(Object.keys(J).length>0)return H({});return null})(Z);if(Z)q={...q||{},model:{...J.runtime,...q?.model||{},repo_id:Z,...W&&{filename:W}}};let _=await t("ggml-stt",null,{config:q,includeBreakdown:!0}),G=_.buttress?.selected||null,w=_.modelConfig||null;if(Z||w?.repoId){console.log(`
70
+ STT model capabilities table saved to: ${R}`),process.exit(0)}catch(W){console.error("Failed to generate STT model table:",W.message),process.exit(1)}}async function D6({modelId:$=null,defaultConfig:X=null}={}){if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.47"}`),console.log("Testing capabilities for backend: ggml-stt"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let{repoId:Z,filename:W}=T6($),j=X||{},{server:V,generators:Y=[],...J}=j,N=(z={},A={})=>{let R=Array.isArray(z)?[...z]:{...z};return Object.entries(A||{}).forEach(([U,Q])=>{if(Q&&typeof Q==="object"&&!Array.isArray(Q))R[U]=N(R[U]||{},Q);else R[U]=Q}),R},H=(z)=>N(JSON.parse(JSON.stringify(J)),z||{}),q=((z)=>{if(Array.isArray(Y)&&Y.length>0){let A=Y.filter((R)=>R?.type==="ggml-stt");if(A.length>0){if(z){let R=A.find((U)=>U.model?.repo_id===z);if(R)return H(R)}}}if(Object.keys(J).length>0)return H({});return null})(Z);if(Z)q={...q||{},model:{...J.runtime,...q?.model||{},repo_id:Z,...W&&{filename:W}}};let _=await t("ggml-stt",null,{config:q,includeBreakdown:!0}),G=_.buttress?.selected||null,w=_.modelConfig||null;if(Z||w?.repoId){console.log(`
71
71
  === Model Information ===`);let z=Z||w?.repoId;if(console.log(`Repository ID: ${z}`),W)console.log(`Filename: ${W}`);if(G?.modelBytes)console.log(`Model Size: ${(G.modelBytes/1024/1024).toFixed(1)} MB`);let A=G?.processingBytes||G?.kvCacheBytes;if(A)console.log(`Processing Buffer: ${(A/1024/1024).toFixed(1)} MB`);if(G?.modelBytes&&A)console.log(`Total Required Memory: ${((G.modelBytes+A)/1024/1024).toFixed(1)} MB`);else if(_.buttress?.selected?.fit){let{totalRequiredBytes:R}=_.buttress.selected.fit;console.log(`Total Required Memory: ${(R/1024/1024).toFixed(1)} MB`)}}if(_.buttress?.selected){let{selected:z}=_.buttress;console.log(`
72
72
  === Hardware Information ===`);let A=null;if(process.platform!=="win32")try{A=K4("uname -a",{encoding:"utf8"}).trim()}catch{}if(A)console.log(`System: ${A}`);else console.log(`Hostname: ${c.hostname()}`),console.log(`OS: ${c.type()} ${c.release()}`);console.log(`Platform: ${z.platform}`),console.log(`CPU Cores: ${c.cpus().length}`),console.log(`Total System Memory: ${(c.totalmem()/1024/1024/1024).toFixed(2)} GB`);let R=z.cpuTotalBytes>0?(z.cpuUsableBytes/z.cpuTotalBytes*100).toFixed(0):0;if(console.log(`Usable CPU Memory: ${(z.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${R}% of ${(z.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),z.hasGpu)console.log(`
73
73
  --- GPU Details ---`),z.devices.filter((Q)=>Q.type==="gpu").forEach((Q)=>{console.log(`GPU Backend: ${Q.backend}`),console.log(`GPU Name: ${Q.deviceName}`),console.log(`GPU Total Memory: ${(Q.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let K=z.gpuTotalBytes>0?(z.gpuUsableBytes/z.gpuTotalBytes*100).toFixed(0):0;if(console.log(`GPU Usable Memory: ${(z.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(z.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Q.metadata){if(Q.metadata.hasBFloat16)console.log("Supports BFloat16: Yes");if(Q.metadata.hasUnifiedMemory)console.log("Unified Memory: Yes")}});else console.log("GPU: Not available");if(console.log(`
74
74
  Backend Variant: ${z.variant}`),console.log(`Performance Score: ${z.score}`),z.fit)console.log(`
75
75
  --- Model Fit Analysis ---`),console.log(`Fits in GPU: ${z.fit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU: ${z.fit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor: ${z.fit.limiting}`)}console.log(`
76
- === Full Capabilities JSON ===`),console.log(JSON.stringify(_,null,2)),process.exit(0)}catch(Z){console.error("Failed to get capabilities:",Z.message),process.exit(1)}}var T6=($)=>{if(!$)return{repoId:null,filename:null};let[X,Z]=$.split(":");return{repoId:X,filename:Z||null}};var b6=P(()=>{V4();M0()});var d9={};import y9 from"node:fs";import u9 from"node:path";import f9 from"@iarna/toml";async function g9($){if(!$?.generators||!Array.isArray($.generators))return;let X=$.generators.filter((O)=>{if(!O.model?.download)return!1;let{type:q}=O;if(!q||q!=="ggml-llm"&&q!=="ggml-stt")return console.warn(`[Download] Skipping unknown generator type: ${q}`),!1;return!0});if(X.length===0)return;let{server:Z,generators:W,...j}=$,V=X.map((O)=>{let{type:q}=O,_=O.model?.repo_id;console.log(`[Download] Starting pre-download for ${q}: ${_}`);let G={...j,backend:O.backend||{},model:O.model||{},runtime:{...j.runtime,...O.runtime||{}}};return Q4(q,G,{onProgress:()=>{},onComplete:({repoId:w,alreadyExists:z})=>{if(z)console.log(`[Download] Pre-download complete (already exists): ${w}`);else console.log(`[Download] Pre-download complete: ${w}`)},onError:(w)=>{console.error(`[Download] Pre-download failed for ${_}:`,w.message)}})}),Y=await Promise.all(V),J=Y.filter((O)=>O.started).length,N=Y.filter((O)=>O.alreadyExists).length,H=Y.filter((O)=>O.alreadyDownloading).length;console.log(`[Download] Pre-download summary: ${J} started, ${N} already exist, ${H} already downloading`)}var v6,h6,I6,q4,J0=null,p9,m9,C6,_4,c9;var y6=P(async()=>{b6();b0();q1();await g6();if(process.argv.includes("--version")||process.argv.includes("-v"))console.log("2.23.0-beta.45"),process.exit(0);if(process.argv.includes("--help")||process.argv.includes("-h"))console.log(`
77
- bricks-buttress v${"2.23.0-beta.45"}
76
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(_,null,2)),process.exit(0)}catch(Z){console.error("Failed to get capabilities:",Z.message),process.exit(1)}}var T6=($)=>{if(!$)return{repoId:null,filename:null};let[X,Z]=$.split(":");return{repoId:X,filename:Z||null}};var b6=P(()=>{V4();M0()});var d9={};import y9 from"node:fs";import u9 from"node:path";import f9 from"@iarna/toml";async function g9($){if(!$?.generators||!Array.isArray($.generators))return;let X=$.generators.filter((O)=>{if(!O.model?.download)return!1;let{type:q}=O;if(!q||q!=="ggml-llm"&&q!=="ggml-stt")return console.warn(`[Download] Skipping unknown generator type: ${q}`),!1;return!0});if(X.length===0)return;let{server:Z,generators:W,...j}=$,V=X.map((O)=>{let{type:q}=O,_=O.model?.repo_id;console.log(`[Download] Starting pre-download for ${q}: ${_}`);let G={...j,backend:O.backend||{},model:O.model||{},runtime:{...j.runtime,...O.runtime||{}}};return Q4(q,G,{onProgress:()=>{},onComplete:({repoId:w,alreadyExists:z})=>{if(z)console.log(`[Download] Pre-download complete (already exists): ${w}`);else console.log(`[Download] Pre-download complete: ${w}`)},onError:(w)=>{console.error(`[Download] Pre-download failed for ${_}:`,w.message)}})}),Y=await Promise.all(V),J=Y.filter((O)=>O.started).length,N=Y.filter((O)=>O.alreadyExists).length,H=Y.filter((O)=>O.alreadyDownloading).length;console.log(`[Download] Pre-download summary: ${J} started, ${N} already exist, ${H} already downloading`)}var v6,h6,I6,q4,J0=null,p9,m9,C6,_4,c9;var y6=P(async()=>{b6();b0();q1();await g6();if(process.argv.includes("--version")||process.argv.includes("-v"))console.log("2.23.0-beta.47"),process.exit(0);if(process.argv.includes("--help")||process.argv.includes("-h"))console.log(`
77
+ bricks-buttress v${"2.23.0-beta.47"}
78
78
 
79
79
  Buttress server for remote inference with GGML backends.
80
80
 
@@ -108,4 +108,4 @@ Examples:
108
108
  bricks-buttress --test-caps ggml-llm --test-models-default
109
109
  bricks-buttress --test-caps ggml-stt --test-caps-model-id BricksDisplay/whisper-ggml:ggml-small.bin
110
110
  `),process.exit(0);v6=process.argv.findIndex(($)=>$==="--port"||$==="-p"),h6=v6>=0?Number(process.argv[v6+1]):void 0,I6=process.argv.findIndex(($)=>$==="--config"||$==="-c"),q4=I6>=0?process.argv[I6+1]:null;if(q4){let $;if(q4.includes(`
111
- `))$=q4;else{let X=u9.resolve(q4);try{$=y9.readFileSync(X,"utf8")}catch(Z){console.error(`Failed to read Buttress config at ${X}:`,Z),process.exit(1)}}try{let X=f9.parse($);if(X.env&&typeof X.env==="object")Object.entries(X.env).forEach(([Z,W])=>{if(process.env[Z]===void 0)process.env[Z]=String(W)}),delete X.env;J0=X}catch(X){console.error("Failed to parse TOML config:",X),process.exit(1)}}p9=["ggml-org/gpt-oss-20b-GGUF","ggml-org/gpt-oss-120b-GGUF","unsloth/Nemotron-3-Nano-30B-A3B-GGUF","unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF","bartowski/Mistral-Nemo-Instruct-2407-GGUF","mistralai/Magistral-Small-2509-GGUF","mistralai/Ministral-3-14B-Reasoning-2512-GGUF","bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF","bartowski/mistralai_Devstral-2-123B-Instruct-2512-GGUF","ggml-org/gemma-3-12b-it-qat-GGUF","ggml-org/gemma-3-27b-it-qat-GGUF","unsloth/phi-4-GGUF"],m9=["BricksDisplay/whisper-ggml:ggml-small.bin","BricksDisplay/whisper-ggml:ggml-small-q8_0.bin","BricksDisplay/whisper-ggml:ggml-medium.bin","BricksDisplay/whisper-ggml:ggml-medium-q8_0.bin","BricksDisplay/whisper-ggml:ggml-large-v3-turbo.bin","BricksDisplay/whisper-ggml:ggml-large-v3-turbo-q8_0.bin","BricksDisplay/whisper-ggml:ggml-large-v3.bin"],C6=process.argv.findIndex(($)=>$==="--test-caps");if(C6>=0){let $=process.argv[C6+1]||"ggml-llm";if($!=="ggml-llm"&&$!=="ggml-stt")console.error("Only ggml-llm and ggml-stt backends are supported for testing capabilities"),process.exit(1);let X=process.argv.findIndex((W)=>W==="--test-models"),Z=process.argv.includes("--test-models-default");if($==="ggml-stt")if(X>=0){let W=process.argv[X+1];if(!W)console.error("Error: --test-models requires a comma-separated list of model IDs"),process.exit(1);let j=W.split(",").map((V)=>V.trim());await L1({modelIds:j,defaultConfig:J0})}else if(Z)await L1({modelIds:m9,defaultConfig:J0});else{let W=process.argv.findIndex((V)=>V==="--test-caps-model-id"),j=W>=0?process.argv[W+1]:null;await D6({modelId:j,defaultConfig:J0})}else if(X>=0){let W=process.argv[X+1];if(!W)console.error("Error: --test-models requires a comma-separated list of model IDs"),process.exit(1);let j=W.split(",").map((V)=>V.trim());await _1({modelIds:j,defaultConfig:J0})}else if(Z)await _1({modelIds:p9,defaultConfig:J0});else{let W=process.argv.findIndex((V)=>V==="--test-caps-model-id"),j=W>=0?process.argv[W+1]:null;await k6({modelId:j,defaultConfig:J0})}}_4=G1(J0);if(h6)_4.server.port=h6;if(!_4.server.port)_4.server.port=Number(process.env.BUTTRESS_PORT)||2080;c9=process.env.ENABLE_OPENAI_COMPAT_ENDPOINT==="1";f6({config:_4,enableOpenAICompat:c9}).then(async({port:$,openaiEnabled:X,autoDiscover:Z})=>{let W=O4();if(console.log(`Buttress server listening on port ${$}`),console.log("--------------------------------"),await u6(),console.log(),console.log("Current supported Generators:"),console.log("- LLM (GGML)"),console.log("- STT (GGML)"),console.log(),console.log("Please configure `Buttress (Remote Inference)` in the Generator to connect to this server."),console.log(),console.log(`- Use http://${W}:${$} to connect to this server via LAN.`),console.log(`- Visit http://${W}:${$}/status to see status via LAN.`),console.log(),X)console.log("OpenAI-compatible API [EXPERIMENTAL]:"),console.log(`- Base URL: http://${W}:${$}/oai-compat/v1`),console.log(`- Chat completions: POST http://${W}:${$}/oai-compat/v1/chat/completions`),console.log(`- Models: GET http://${W}:${$}/oai-compat/v1/models`),console.log();else console.log("OpenAI-compatible API [EXPERIMENTAL]: disabled"),console.log(" Set ENABLE_OPENAI_COMPAT_ENDPOINT=1 to enable"),console.log();if(Z)console.log("Auto-discover enabled"),console.log();if(J0)await g9(J0)}).catch(($)=>{console.error("Failed to start Buttress server:",$),process.exitCode=1})});import{node as l9}from"@elysiajs/node";import{Elysia as i9,t as v0}from"elysia";import{ReadableStream as s9}from"node:stream/web";import n9 from"node:fs/promises";import{ZodError as r9}from"zod";var o9=async()=>{let $=`https://registry.npmjs.org/${"@fugood/buttress-server"}/latest`;try{let X=new AbortController,Z=setTimeout(()=>X.abort(),3000),W=await fetch($,{headers:{Accept:"application/json"},signal:X.signal});if(clearTimeout(Z),!W.ok)return null;return(await W.json()).version||null}catch{return null}},a9=($,X)=>{if(!X)return!1;let Z=$.split(/[.-]/),W=X.split(/[.-]/);for(let j=0;j<Math.max(Z.length,W.length);j+=1){let V=parseInt(Z[j])||0,Y=parseInt(W[j])||0;if(Y>V)return!0;if(Y<V)return!1}return!1},t9=($)=>{console.log(""),console.log("\x1B[33m╭─────────────────────────────────────────────────╮\x1B[0m"),console.log("\x1B[33m│\x1B[0m Update available! \x1B[2m%s\x1B[0m → \x1B[32m%s\x1B[0m","2.23.0-beta.45".padEnd(12),$.padEnd(12),"\x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m \x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m Run to upgrade: \x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m \x1B[36mnpm install -g %s\x1B[0m \x1B[33m│\x1B[0m","@fugood/buttress-server".padEnd(27)),console.log("\x1B[33m╰─────────────────────────────────────────────────╯\x1B[0m"),console.log("")},u6=async()=>{try{let $=await o9();if($&&a9("2.23.0-beta.45",$))t9($)}catch($){}},e9,$$=async({backend:$,router:X,config:Z,enableOpenAICompat:W})=>{try{await n9.mkdir(Z.server.temp_file_dir,{recursive:!0})}catch{}let j=O4()||"0.0.0.0",V={id:Z.server.id,name:Z.server.name,version:"2.23.0-beta.45",address:j,port:Z.server.port,url:`http://${j}:${Z.server.port}`,generators:V6(Z,Z.generators.map((N)=>N.type)),authentication:{required:!0,type:"device-group"}},Y=new i9({serve:{maxRequestBodySize:Z.server.max_body_size},websocket:{idleTimeout:Math.ceil(Z.server.session_timeout/1000)},adapter:e9?l9():void 0}).state({sessions:new Map,backend:$||J4,config:Z,serverInfo:V});if(X)Y.use(X);if(Z.autodiscover?.http?.enabled)Y.use(j1(Z));if(Y.use(Q1),Y.use(V1),W)Y.use(U4(Z));let J={INVALID_REQUEST:-32600,INVALID_PARAMS:-32602,METHOD_NOT_FOUND:-32601,INTERNAL_ERROR:-32603};return Y.ws("/buttress/rpc",{parse:(N,H)=>{if(typeof H==="string")try{return JSON.parse(H)}catch{return N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INVALID_REQUEST,message:"Invalid request"},id:null})),null}return H},body:v0.Object({jsonrpc:v0.String(),method:v0.String(),params:v0.String(),id:v0.String()}),open(N){let H=N.id??N.raw?.id??N.remoteAddress;if(console.log(`[Request] New connection: ${H}`),!N.data.store.sessions.has(H))N.data.store.sessions.set(H,{streamReaders:new Map,generators:new Set,initializedContexts:new Set,timeout:null});else{let O=N.data.store.sessions.get(H);clearTimeout(O.timeout),O.timeout=null}},async message(N,{id:H,method:O,params:q}){let _=N.id??N.raw?.id??N.remoteAddress;console.log(`[Request] Received request from ${_}: ${O}`);let G=N.data.store.sessions.get(_),[w,z]=O.split("."),A=z6[w]?.[z];if(!A){N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.METHOD_NOT_FOUND,message:"Method not found"},id:H}));return}try{if(O==="cancel"){if(G.streamReaders.has(H))G.streamReaders.get(H)?.cancel(),G.streamReaders.delete(H);return}if(O==="ping"){N.send(JSON.stringify({jsonrpc:"2.0",result:"pong",id:H}));return}let R=R6(q),U=A6[w]?.[z],Q=U?U.parse(R):R,K={...N.data.store,peerId:_,session:G},B=await A(K,...Q);if(B instanceof s9){let L=B.getReader();G.streamReaders.set(H,L),N.send(JSON.stringify({jsonrpc:"2.0",result:{type:"stream"},id:H}));try{while(!0){let{value:F,done:E}=await L.read();if(E)break;let{event:M,data:x}=F;N.send(JSON.stringify({jsonrpc:"2.0",method:`notification/${M}`,params:H4(x),id:H}))}N.send(JSON.stringify({jsonrpc:"2.0",method:"notification/_end",id:H}))}catch(F){console.error(F),N.send(JSON.stringify({jsonrpc:"2.0",method:"notification/_error",params:H4(F),id:H}))}G.streamReaders.delete(H)}else N.send(JSON.stringify({jsonrpc:"2.0",result:H4(B),id:H}))}catch(R){if(R instanceof r9){N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INVALID_PARAMS,message:"Invalid params",data:R.issues},id:H}));return}console.error(R),N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INTERNAL_ERROR,message:String(R)},id:H}))}},async close(N){let H=N.id??N.raw?.id??N.remoteAddress;console.log(`[Request] Connection closed: ${H}`);let{backend:O,sessions:q}=N.data.store,_=q.get(H);if(!_)return;_.streamReaders.forEach((G)=>G.cancel()),_.streamReaders.clear(),_.timeout=setTimeout(()=>{q.delete(H),console.log(`[Request] Session timed out: ${H}`);let{generators:G}=_;G.forEach((w)=>{O.finalizeGenerator(w)})},Z.server.session_timeout)}}),{app:Y,config:Z}},f6=async({backend:$,router:X,config:Z,enableOpenAICompat:W=!1})=>{let{app:j,config:V}=await $$({backend:$,router:X,config:Z,enableOpenAICompat:W}),{server:{port:Y}}=V,J=[new Promise((H)=>j.listen(Y,H))],N=null;if(V.autodiscover)N=new K1(V.autodiscover,()=>j.store.serverInfo),J.push(N.start());return await Promise.all(J),{app:j,port:Y,openaiEnabled:W,autoDiscover:N}};var g6=P(async()=>{V4();Z6();B6();w6();b0();M6();q1();V4();b0();e9=typeof process<"u"&&process.versions&&process.versions.node;if(R1.main==R1.module)await y6().then(() => d9)});await g6();export{f6 as startServer,Q4 as startModelDownload,G1 as processConfig,t9 as logUpdateMessage,$$ as createServer,a9 as compareVersions,o9 as checkForUpdates,u6 as checkAndNotifyUpdates};
111
+ `))$=q4;else{let X=u9.resolve(q4);try{$=y9.readFileSync(X,"utf8")}catch(Z){console.error(`Failed to read Buttress config at ${X}:`,Z),process.exit(1)}}try{let X=f9.parse($);if(X.env&&typeof X.env==="object")Object.entries(X.env).forEach(([Z,W])=>{if(process.env[Z]===void 0)process.env[Z]=String(W)}),delete X.env;J0=X}catch(X){console.error("Failed to parse TOML config:",X),process.exit(1)}}p9=["ggml-org/gpt-oss-20b-GGUF","ggml-org/gpt-oss-120b-GGUF","unsloth/Nemotron-3-Nano-30B-A3B-GGUF","unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF","bartowski/Mistral-Nemo-Instruct-2407-GGUF","mistralai/Magistral-Small-2509-GGUF","mistralai/Ministral-3-14B-Reasoning-2512-GGUF","bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF","bartowski/mistralai_Devstral-2-123B-Instruct-2512-GGUF","ggml-org/gemma-3-12b-it-qat-GGUF","ggml-org/gemma-3-27b-it-qat-GGUF","unsloth/phi-4-GGUF"],m9=["BricksDisplay/whisper-ggml:ggml-small.bin","BricksDisplay/whisper-ggml:ggml-small-q8_0.bin","BricksDisplay/whisper-ggml:ggml-medium.bin","BricksDisplay/whisper-ggml:ggml-medium-q8_0.bin","BricksDisplay/whisper-ggml:ggml-large-v3-turbo.bin","BricksDisplay/whisper-ggml:ggml-large-v3-turbo-q8_0.bin","BricksDisplay/whisper-ggml:ggml-large-v3.bin"],C6=process.argv.findIndex(($)=>$==="--test-caps");if(C6>=0){let $=process.argv[C6+1]||"ggml-llm";if($!=="ggml-llm"&&$!=="ggml-stt")console.error("Only ggml-llm and ggml-stt backends are supported for testing capabilities"),process.exit(1);let X=process.argv.findIndex((W)=>W==="--test-models"),Z=process.argv.includes("--test-models-default");if($==="ggml-stt")if(X>=0){let W=process.argv[X+1];if(!W)console.error("Error: --test-models requires a comma-separated list of model IDs"),process.exit(1);let j=W.split(",").map((V)=>V.trim());await L1({modelIds:j,defaultConfig:J0})}else if(Z)await L1({modelIds:m9,defaultConfig:J0});else{let W=process.argv.findIndex((V)=>V==="--test-caps-model-id"),j=W>=0?process.argv[W+1]:null;await D6({modelId:j,defaultConfig:J0})}else if(X>=0){let W=process.argv[X+1];if(!W)console.error("Error: --test-models requires a comma-separated list of model IDs"),process.exit(1);let j=W.split(",").map((V)=>V.trim());await _1({modelIds:j,defaultConfig:J0})}else if(Z)await _1({modelIds:p9,defaultConfig:J0});else{let W=process.argv.findIndex((V)=>V==="--test-caps-model-id"),j=W>=0?process.argv[W+1]:null;await k6({modelId:j,defaultConfig:J0})}}_4=G1(J0);if(h6)_4.server.port=h6;if(!_4.server.port)_4.server.port=Number(process.env.BUTTRESS_PORT)||2080;c9=process.env.ENABLE_OPENAI_COMPAT_ENDPOINT==="1";f6({config:_4,enableOpenAICompat:c9}).then(async({port:$,openaiEnabled:X,autoDiscover:Z})=>{let W=O4();if(console.log(`Buttress server listening on port ${$}`),console.log("--------------------------------"),await u6(),console.log(),console.log("Current supported Generators:"),console.log("- LLM (GGML)"),console.log("- STT (GGML)"),console.log(),console.log("Please configure `Buttress (Remote Inference)` in the Generator to connect to this server."),console.log(),console.log(`- Use http://${W}:${$} to connect to this server via LAN.`),console.log(`- Visit http://${W}:${$}/status to see status via LAN.`),console.log(),X)console.log("OpenAI-compatible API [EXPERIMENTAL]:"),console.log(`- Base URL: http://${W}:${$}/oai-compat/v1`),console.log(`- Chat completions: POST http://${W}:${$}/oai-compat/v1/chat/completions`),console.log(`- Models: GET http://${W}:${$}/oai-compat/v1/models`),console.log();else console.log("OpenAI-compatible API [EXPERIMENTAL]: disabled"),console.log(" Set ENABLE_OPENAI_COMPAT_ENDPOINT=1 to enable"),console.log();if(Z)console.log("Auto-discover enabled"),console.log();if(J0)await g9(J0)}).catch(($)=>{console.error("Failed to start Buttress server:",$),process.exitCode=1})});import{node as l9}from"@elysiajs/node";import{Elysia as i9,t as v0}from"elysia";import{ReadableStream as s9}from"node:stream/web";import n9 from"node:fs/promises";import{ZodError as r9}from"zod";var o9=async()=>{let $=`https://registry.npmjs.org/${"@fugood/buttress-server"}/latest`;try{let X=new AbortController,Z=setTimeout(()=>X.abort(),3000),W=await fetch($,{headers:{Accept:"application/json"},signal:X.signal});if(clearTimeout(Z),!W.ok)return null;return(await W.json()).version||null}catch{return null}},a9=($,X)=>{if(!X)return!1;let Z=$.split(/[.-]/),W=X.split(/[.-]/);for(let j=0;j<Math.max(Z.length,W.length);j+=1){let V=parseInt(Z[j])||0,Y=parseInt(W[j])||0;if(Y>V)return!0;if(Y<V)return!1}return!1},t9=($)=>{console.log(""),console.log("\x1B[33m╭─────────────────────────────────────────────────╮\x1B[0m"),console.log("\x1B[33m│\x1B[0m Update available! \x1B[2m%s\x1B[0m → \x1B[32m%s\x1B[0m","2.23.0-beta.47".padEnd(12),$.padEnd(12),"\x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m \x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m Run to upgrade: \x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m \x1B[36mnpm install -g %s\x1B[0m \x1B[33m│\x1B[0m","@fugood/buttress-server".padEnd(27)),console.log("\x1B[33m╰─────────────────────────────────────────────────╯\x1B[0m"),console.log("")},u6=async()=>{try{let $=await o9();if($&&a9("2.23.0-beta.47",$))t9($)}catch($){}},e9,$$=async({backend:$,router:X,config:Z,enableOpenAICompat:W})=>{try{await n9.mkdir(Z.server.temp_file_dir,{recursive:!0})}catch{}let j=O4()||"0.0.0.0",V={id:Z.server.id,name:Z.server.name,version:"2.23.0-beta.47",address:j,port:Z.server.port,url:`http://${j}:${Z.server.port}`,generators:V6(Z,Z.generators.map((N)=>N.type)),authentication:{required:!0,type:"device-group"}},Y=new i9({serve:{maxRequestBodySize:Z.server.max_body_size},websocket:{idleTimeout:Math.ceil(Z.server.session_timeout/1000)},adapter:e9?l9():void 0}).state({sessions:new Map,backend:$||J4,config:Z,serverInfo:V});if(X)Y.use(X);if(Z.autodiscover?.http?.enabled)Y.use(j1(Z));if(Y.use(Q1),Y.use(V1),W)Y.use(U4(Z));let J={INVALID_REQUEST:-32600,INVALID_PARAMS:-32602,METHOD_NOT_FOUND:-32601,INTERNAL_ERROR:-32603};return Y.ws("/buttress/rpc",{parse:(N,H)=>{if(typeof H==="string")try{return JSON.parse(H)}catch{return N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INVALID_REQUEST,message:"Invalid request"},id:null})),null}return H},body:v0.Object({jsonrpc:v0.String(),method:v0.String(),params:v0.String(),id:v0.String()}),open(N){let H=N.id??N.raw?.id??N.remoteAddress;if(console.log(`[Request] New connection: ${H}`),!N.data.store.sessions.has(H))N.data.store.sessions.set(H,{streamReaders:new Map,generators:new Set,initializedContexts:new Set,timeout:null});else{let O=N.data.store.sessions.get(H);clearTimeout(O.timeout),O.timeout=null}},async message(N,{id:H,method:O,params:q}){let _=N.id??N.raw?.id??N.remoteAddress;console.log(`[Request] Received request from ${_}: ${O}`);let G=N.data.store.sessions.get(_),[w,z]=O.split("."),A=z6[w]?.[z];if(!A){N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.METHOD_NOT_FOUND,message:"Method not found"},id:H}));return}try{if(O==="cancel"){if(G.streamReaders.has(H))G.streamReaders.get(H)?.cancel(),G.streamReaders.delete(H);return}if(O==="ping"){N.send(JSON.stringify({jsonrpc:"2.0",result:"pong",id:H}));return}let R=R6(q),U=A6[w]?.[z],Q=U?U.parse(R):R,K={...N.data.store,peerId:_,session:G},B=await A(K,...Q);if(B instanceof s9){let L=B.getReader();G.streamReaders.set(H,L),N.send(JSON.stringify({jsonrpc:"2.0",result:{type:"stream"},id:H}));try{while(!0){let{value:F,done:E}=await L.read();if(E)break;let{event:M,data:x}=F;N.send(JSON.stringify({jsonrpc:"2.0",method:`notification/${M}`,params:H4(x),id:H}))}N.send(JSON.stringify({jsonrpc:"2.0",method:"notification/_end",id:H}))}catch(F){console.error(F),N.send(JSON.stringify({jsonrpc:"2.0",method:"notification/_error",params:H4(F),id:H}))}G.streamReaders.delete(H)}else N.send(JSON.stringify({jsonrpc:"2.0",result:H4(B),id:H}))}catch(R){if(R instanceof r9){N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INVALID_PARAMS,message:"Invalid params",data:R.issues},id:H}));return}console.error(R),N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INTERNAL_ERROR,message:String(R)},id:H}))}},async close(N){let H=N.id??N.raw?.id??N.remoteAddress;console.log(`[Request] Connection closed: ${H}`);let{backend:O,sessions:q}=N.data.store,_=q.get(H);if(!_)return;_.streamReaders.forEach((G)=>G.cancel()),_.streamReaders.clear(),_.timeout=setTimeout(()=>{q.delete(H),console.log(`[Request] Session timed out: ${H}`);let{generators:G}=_;G.forEach((w)=>{O.finalizeGenerator(w)})},Z.server.session_timeout)}}),{app:Y,config:Z}},f6=async({backend:$,router:X,config:Z,enableOpenAICompat:W=!1})=>{let{app:j,config:V}=await $$({backend:$,router:X,config:Z,enableOpenAICompat:W}),{server:{port:Y}}=V,J=[new Promise((H)=>j.listen(Y,H))],N=null;if(V.autodiscover)N=new K1(V.autodiscover,()=>j.store.serverInfo),J.push(N.start());return await Promise.all(J),{app:j,port:Y,openaiEnabled:W,autoDiscover:N}};var g6=P(async()=>{V4();Z6();B6();w6();b0();M6();q1();V4();b0();e9=typeof process<"u"&&process.versions&&process.versions.node;if(R1.main==R1.module)await y6().then(() => d9)});await g6();export{f6 as startServer,Q4 as startModelDownload,G1 as processConfig,t9 as logUpdateMessage,$$ as createServer,a9 as compareVersions,o9 as checkForUpdates,u6 as checkAndNotifyUpdates};
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@fugood/buttress-server",
3
- "version": "2.23.0-beta.45",
3
+ "version": "2.23.0-beta.47",
4
4
  "main": "dist/index.mjs",
5
5
  "type": "module",
6
6
  "bin": {
@@ -28,7 +28,7 @@
28
28
  "dependencies": {
29
29
  "@elysiajs/cors": "^1.1.1",
30
30
  "@elysiajs/node": "^1.4.2",
31
- "@fugood/llama.node": "^1.4.14",
31
+ "@fugood/llama.node": "^1.4.15",
32
32
  "@fugood/whisper.node": "^1.0.13",
33
33
  "@huggingface/gguf": "^0.3.2",
34
34
  "@iarna/toml": "^3.0.0",
@@ -43,5 +43,5 @@
43
43
  "oxc-transform": "^0.105.0",
44
44
  "typescript": "^5.9.3"
45
45
  },
46
- "gitHead": "fa67a0c333ca6e372e7b7b01035d8e871b216b9f"
46
+ "gitHead": "cc06e63df3f83592832a838bbe72e0107b6e92b6"
47
47
  }