@fugood/buttress-server 2.24.2 → 2.24.3

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.
package/README.md CHANGED
@@ -20,6 +20,84 @@ npx bricks-buttress --config ./config.toml
20
20
  npx bricks-buttress
21
21
  ```
22
22
 
23
+ ## Workspace Binding (`bricks buttress`)
24
+
25
+ By default, a buttress-server runs in **public mode**: any client on the LAN can connect, no auth required. To restrict access to a single BRICKS workspace and enable workspace-scoped JWT auth, **bind** the server with the `bricks buttress` CLI commands. Once bound, the server only accepts WebSocket / file-transfer requests carrying a valid access token signed by that workspace's issuer.
26
+
27
+ The `bricks` CLI is the tool that performs the binding and writes the local state file. Install it first — see the [bricks-cli docs](https://docs.bricks.tools/cli) — then `bricks auth login` with the workspace owner's account before running the commands below.
28
+
29
+ ### Bind a server to a workspace
30
+
31
+ ```bash
32
+ # Pair the local machine's buttress-server with the workspace of the current bricks-cli profile
33
+ bricks buttress bind
34
+
35
+ # Override the auto-detected server id, give it a friendly name, or write to a custom state dir
36
+ bricks buttress bind --server-id buttress-mac-studio --name "Studio LLM" --state-dir /etc/buttress
37
+
38
+ # For headless/remote setups: emit state.json to stdout instead of writing to disk
39
+ bricks buttress bind --print > /etc/buttress/state.json
40
+ ```
41
+
42
+ The state file (`~/.bricks-cli/buttress/state.json` by default, or `$BRICKS_BUTTRESS_STATE_DIR`) stores:
43
+
44
+ - `workspace.id` / `workspace.name` — which workspace this server belongs to
45
+ - `workspace.serverId` — the server's stable id (defaults to `buttress-<machineId>`)
46
+ - `workspace.issuerPublicKey` + `workspace.kid` — Ed25519 SPKI used to verify access tokens
47
+
48
+ **Restart `bricks-buttress` after binding** for the change to take effect — the state file is read once at startup.
49
+
50
+ ### Inspect bindings
51
+
52
+ ```bash
53
+ # Show local state.json + the workspace-side bound list
54
+ bricks buttress status
55
+
56
+ # Same, JSON-formatted
57
+ bricks buttress status --json
58
+ ```
59
+
60
+ ### Discover servers on the LAN
61
+
62
+ ```bash
63
+ # UDP scan + HTTP /buttress/info verification (3s timeout by default)
64
+ bricks buttress scan
65
+
66
+ # UDP only (skip the /buttress/info round-trip)
67
+ bricks buttress scan --udp-only
68
+
69
+ # Machine-readable
70
+ bricks buttress scan --json
71
+ ```
72
+
73
+ `scan` lists every buttress-server visible on the LAN, including unbound (public) ones, with their version, auth state (`open` vs `JWT required` + kid), bound workspace, and per-generator hardware caps (`score`, GPU, usable memory). Servers whose workspace matches your current `bricks-cli` profile are highlighted; this is purely a discovery command and does not mint any tokens.
74
+
75
+ ### Unbind
76
+
77
+ ```bash
78
+ # Remove the binding from the workspace and delete the local state.json
79
+ bricks buttress unbind
80
+
81
+ # Keep the local state file (useful if you only want to revoke server-side)
82
+ bricks buttress unbind --keep-local
83
+ ```
84
+
85
+ After unbinding, restart the server to return it to public mode.
86
+
87
+ ### Issue a long-lived access token
88
+
89
+ For headless callers (CI, ctor agents) that already hold a workspace token, mint a long-lived buttress access token instead of relying on a per-launcher session token:
90
+
91
+ ```bash
92
+ # Default 30-day TTL
93
+ bricks buttress issue-token
94
+
95
+ # Custom TTL (seconds), JSON output for scripting
96
+ bricks buttress issue-token --ttl 3600 --json
97
+ ```
98
+
99
+ The token claims `{ k: 'ba', w_id, st: 'ws', sid, jti, exp }` and any buttress-server bound to the same workspace will accept it.
100
+
23
101
  ## Configuration
24
102
 
25
103
  Configuration can be provided via:
@@ -157,6 +235,25 @@ Examples:
157
235
  bricks-buttress --test-caps ggml-stt --test-caps-model-id BricksDisplay/whisper-ggml:ggml-small.bin
158
236
  ```
159
237
 
238
+ ## Compatibility Endpoints (Experimental)
239
+
240
+ The server can expose OpenAI- and Anthropic-compatible HTTP endpoints in addition to the native RPC. Each endpoint is opt-in via the TOML config:
241
+
242
+ ```toml
243
+ [openai_compat]
244
+ enabled = true
245
+ # cors_allowed_origins = "*" # Or a list of origins; defaults to disabled
246
+
247
+ [anthropic_messages]
248
+ enabled = true
249
+ # cors_allowed_origins = ["http://localhost:3000"]
250
+ ```
251
+
252
+ | Endpoint | Config flag |
253
+ | --------------------- | ------------------------------------ |
254
+ | `/oai-compat/v1/*` | `[openai_compat] enabled = true` |
255
+ | `/anthropic-messages` | `[anthropic_messages] enabled = true` |
256
+
160
257
  ## Session State Cache
161
258
 
162
259
  The server supports session state caching for ggml-llm generators, which saves KV cache state to disk after completions. This enables:
@@ -11,15 +11,24 @@
11
11
  # HF_TOKEN = "your_huggingface_token_here"
12
12
  # CUDA_VISIBLE_DEVICES = "0"
13
13
 
14
+ [autodiscover]
15
+ enabled = true
16
+
14
17
  [server]
15
18
  port = 2080
16
19
  log_level = "info"
17
20
  # max_body_size = "100MB" # Supports string (e.g., "100MB", "1GB") or number in bytes
18
21
 
19
22
  [openai_compat]
23
+ enabled = true
20
24
  # cors_allowed_origins = ["http://localhost:3000", "https://example.com"] # Restrict to specific origins
21
25
  # cors_allowed_origins = "*" # Allow all origins (default)
22
26
 
27
+ [anthropic_messages]
28
+ enabled = true
29
+ # cors_allowed_origins = ["http://localhost:3000", "https://example.com"]
30
+ # cors_allowed_origins = "*"
31
+
23
32
  [runtime]
24
33
  cache_dir = "./.buttress-cache"
25
34
  # huggingface_token = "hf_xx"
package/lib/index.d.mts CHANGED
@@ -1,5 +1,6 @@
1
1
 
2
2
  import { AnyElysia, Elysia } from "elysia";
3
+ import crypto from "node:crypto";
3
4
  import { ReadableStream } from "node:stream/web";
4
5
  import { EventEmitter } from "node:events";
5
6
 
@@ -229,9 +230,11 @@ type GeneratorConfig = {
229
230
  type GlobalConfig = {
230
231
  runtime?: RuntimeConfig;
231
232
  openai_compat?: {
233
+ enabled?: boolean;
232
234
  cors_allowed_origins?: string | string[];
233
235
  };
234
236
  anthropic_messages?: {
237
+ enabled?: boolean;
235
238
  cors_allowed_origins?: string | string[];
236
239
  };
237
240
  } & Record<string, any>;
@@ -277,24 +280,40 @@ type Config = {
277
280
  generators: GeneratorConfig[];
278
281
  };
279
282
  type GeneratorInfo = {
280
- type: GeneratorType;
283
+ type: GeneratorType; /** Performance score 0–100 from buttress-hardware-guardrails. */
284
+ score?: number; /** Whether the host has an accelerator (GPU/Metal/etc) for this backend. */
285
+ hasGpu?: boolean; /** Usable memory in bytes for this backend (GPU when present, else CPU). */
286
+ usableBytes?: number;
281
287
  } & Record<string, any>;
282
288
  type ServerInfo = {
283
289
  id: string;
284
290
  name: string;
291
+ version: string;
285
292
  address: string;
286
293
  port: number;
287
294
  url: string;
288
295
  generators: GeneratorInfo[];
289
296
  authentication: {
290
297
  required: boolean;
291
- type: string;
298
+ type: string; /** Issuer key id (when type === 'workspace-jwt'). */
299
+ kid?: string; /** True when buttress is paired with a workspace. */
300
+ bound?: boolean;
301
+ }; /** Workspace identity (only present when paired). */
302
+ workspace?: {
303
+ id: string;
304
+ name?: string;
292
305
  };
293
306
  };
294
307
  //#endregion
295
308
  //#region src/autodiscover/types.d.ts
296
309
  type GetServerInfoFn = () => ServerInfo;
297
310
  //#endregion
311
+ //#region src/autodiscover/udp.d.ts
312
+ interface AnnounceSigner {
313
+ kid: string;
314
+ privateKey: crypto.KeyObject;
315
+ }
316
+ //#endregion
298
317
  //#region src/autodiscover/index.d.ts
299
318
  /**
300
319
  * Autodiscover service that manages discovery transports.
@@ -304,9 +323,10 @@ type GetServerInfoFn = () => ServerInfo;
304
323
  declare class AutodiscoverService {
305
324
  private config;
306
325
  private getServerInfo;
326
+ private signer;
307
327
  private transports;
308
328
  private started;
309
- constructor(config: AutodiscoverConfig, getServerInfo: GetServerInfoFn);
329
+ constructor(config: AutodiscoverConfig, getServerInfo: GetServerInfoFn, signer: AnnounceSigner | null);
310
330
  start(): Promise<void>;
311
331
  stop(): Promise<void>;
312
332
  }
package/lib/index.mjs CHANGED
@@ -1,82 +1,82 @@
1
1
  #!/usr/bin/env node
2
- import{t as e}from"./chunk-C8PTHxhX.mjs";import{node as t}from"@elysiajs/node";import{Elysia as n,file as r,sse as i,t as a}from"elysia";import*as o from"node:stream/web";import{ReadableStream as s}from"node:stream/web";import c,{mkdir as l,open as u,readFile as d,readdir as f,rename as p,stat as m,unlink as h,writeFile as g}from"node:fs/promises";import _ from"node:path";import v from"node:os";import{createHash as y}from"node:crypto";import{gguf as b}from"@huggingface/gguf";import{getBackendDevicesInfo as x,isLibVariantAvailable as S,loadModel as C}from"@fugood/llama.node";import w from"bytes";import{EventEmitter as T}from"node:events";import{initWhisper as E}from"@fugood/whisper.node";import{fileURLToPath as D}from"node:url";import{execFile as ee,execSync as O,spawn as te}from"node:child_process";import k from"node:fs";import A from"@iarna/toml";import{ZodError as ne,z as j}from"zod";import{cors as M}from"@elysiajs/cors";import N from"node-machine-id";import P from"ms";import{Buffer as F}from"node:buffer";import re from"node:dgram";const I=1024**3,L=(e,t,n)=>Math.min(Math.max(e,t),n),ie=e=>e?40:0,ae=(e=0)=>e?L(e/(12*I)*20,0,20):0,R=(e=0)=>e?L(e/(32*I)*10,0,10):0,oe=e=>e?10:0,se=(e=`default`,t=null)=>{let n=String(e).toLowerCase();return n?n.includes(`cuda`)?20:n.includes(`vulkan`)?10:n.includes(`default`)?t===`darwin`||t===`ios`?15:5:0:0},ce=({platform:e,variant:t,hasGpu:n,gpuUsableBytes:r=0,cpuUsableBytes:i=0,ok:a=!0}={})=>{if(!a)return 0;let o=ie(n)+se(t,e)+ae(r),s=R(i),c=oe(a);return Math.min(100,Math.round(o+s+c))},le=({platform:e,variant:t,hasGpu:n,gpuUsableBytes:r=0,cpuUsableBytes:i=0,ok:a=!0}={})=>({gpuPresence:ie(n),variant:se(t,e),gpuMemory:ae(r),cpuMemory:R(i),availability:oe(a)}),ue=[`cuda`,`vulkan`,`snapdragon`,`default`],z=.85,de=.5,fe=e=>!e&&e!==0?[]:Array.isArray(e)?e.filter(e=>e!=null):[e],pe=e=>e&&String(e).trim().toLowerCase()||null,me=({variant:e,preferVariants:t=[],variantPreference:n=[],defaultVariants:r=ue}={})=>{let i=[];e&&i.push(e),i.push(...fe(t)),i.push(...fe(n)),i.push(...r);let a=new Set;for(let e of i){let t=pe(e);t&&a.add(t)}return Array.from(a)},he=(e={})=>{let t=String(e.type||e.deviceType||e.kind||``).toLowerCase();return!!(t.includes(`gpu`)||t.includes(`cuda`)||t.includes(`metal`)||t.includes(`vulkan`)||t.includes(`snapdragon`))},ge=e=>Array.isArray(e)?e.map(e=>({...e})):[],_e=(e,t)=>e===`snapdragon`?t.filter(e=>e.deviceName!==`GPUOpenCL`):t,ve=({platform:e,totalMemoryInBytes:t,variant:n,devices:r,gpuMemoryFraction:i,cpuMemoryFraction:a,ok:o,error:s})=>{let c=ge(_e(n,r)),l=c.some(he),u=c.filter(e=>he(e)&&Number.isFinite(Number(e.maxMemorySize))).reduce((e,t)=>e+t.maxMemorySize,0),d=t,f=l?Math.floor(u*i):0,p=d?Math.floor(d*a):0,m={platform:e,variant:n,hasGpu:l,gpuUsableBytes:f,cpuUsableBytes:p,ok:o};return{platform:e,ok:o,variant:n,hasGpu:l,devices:c,gpuTotalBytes:u,gpuUsableBytes:f,cpuTotalBytes:d,cpuUsableBytes:p,score:ce(m),breakdown:o?le(m):null,error:s,timestamp:new Date().toISOString()}},B=({device:e,modelBytes:t=0,kvCacheBytes:n=0}={})=>{if(!e)return{totalRequiredBytes:t+n,fitsInGpu:!1,fitsInCpu:!1,limiting:`unknown-device`};let r=Math.max(0,Number(t)||0)+Math.max(0,Number(n)||0),i=e.hasGpu&&r>0&&r<=e.gpuUsableBytes,a=r>0&&r<=e.cpuUsableBytes,o=`ok`;return!i&&e.hasGpu&&(o=`gpu-memory`),a||(o=i?`cpu-memory`:`insufficient-memory`),{totalRequiredBytes:r,fitsInGpu:i,fitsInCpu:a,limiting:o}},ye=async({platform:e,variant:t=null,preferVariants:n=[],variantPreference:r=[],gpuMemoryFraction:i=z,cpuMemoryFraction:a=de,includeBreakdown:o=!1,totalMemoryInBytes:s,modelBytes:c=null,kvCacheBytes:l=null,limitedKvCacheBytes:u=null,dependencies:d={},defaultVariants:f=ue}={})=>{let{getBackendDevicesInfo:p,isLibVariantAvailable:m}=d;if(typeof p!=`function`||typeof m!=`function`)throw TypeError(`GGML capability detection requires getBackendDevicesInfo and isLibVariantAvailable functions`);let h=me({variant:t,preferVariants:n,variantPreference:r,defaultVariants:f}),g=[];for(let t of h)try{if(!await m(t))throw Error(`Variant ${t} not available on this platform`);let n=await p(t);g.push(ve({platform:e,totalMemoryInBytes:s,variant:t,devices:n,gpuMemoryFraction:i,cpuMemoryFraction:a,ok:!0}))}catch(n){let r=n instanceof Error?n.message:String(n);g.push(ve({platform:e,totalMemoryInBytes:s,variant:t,devices:[],gpuMemoryFraction:i,cpuMemoryFraction:a,ok:!1,error:r}))}let _=g.filter(e=>e.ok)[0]||null,v={ok:!!_,selected:_?{..._,breakdown:o?_.breakdown:void 0}:null,attempts:g};if(!o&&v.selected&&delete v.selected.breakdown,!v||!c&&!l)return v;let y=e=>{if(!e)return e;let t=B({device:e,modelBytes:c||0,kvCacheBytes:l||0}),n=null;return u!=null&&u!==l&&(n=B({device:e,modelBytes:c||0,kvCacheBytes:u})),{...e,fit:t,...n&&{limitedFit:n}}};return v.selected=y(v.selected),v.attempts=Array.isArray(v.attempts)?v.attempts.map(y):v.attempts,v},be=`ggml-llm`,xe=[`cuda`,`vulkan`,`default`],Se=async({platform:e,variant:t=null,preferVariants:n=[],variantPreference:r=[],gpuMemoryFraction:i=z,cpuMemoryFraction:a=de,includeBreakdown:o=!1,totalMemoryInBytes:s,modelBytes:c=null,processingBytes:l=null,kvCacheBytes:u=null,dependencies:d={}}={})=>ye({platform:e,variant:t,preferVariants:n,variantPreference:r&&r.length>0?r:xe,gpuMemoryFraction:i,cpuMemoryFraction:a,includeBreakdown:o,totalMemoryInBytes:s,modelBytes:c,kvCacheBytes:l??u,dependencies:d,defaultVariants:xe}),Ce=async({platform:e,arch:t=null,unifiedMemoryFraction:n=.85,includeBreakdown:r=!1,totalMemoryInBytes:i,modelBytes:a=null,kvCacheBytes:o=null,limitedKvCacheBytes:s=null}={})=>{let c=[];e!==`darwin`&&c.push(`MLX requires macOS`),t&&t!==`arm64`&&c.push(`MLX requires Apple Silicon (arm64)`);let l=c.length===0,u=l?Math.floor(i*n):0,d={platform:e,variant:`mlx`,hasGpu:l,gpuUsableBytes:u,cpuUsableBytes:u,ok:l},f=ce(d),p=l?le(d):null,m={platform:e,ok:l,variant:`mlx`,hasGpu:l,unifiedMemory:!0,devices:l?[{type:`metal`,deviceName:`Apple Silicon (Unified Memory)`,maxMemorySize:i}]:[],gpuTotalBytes:l?i:0,gpuUsableBytes:u,cpuTotalBytes:i,cpuUsableBytes:u,score:f,breakdown:r?p:void 0,error:l?void 0:c.join(`; `),timestamp:new Date().toISOString()};r||delete m.breakdown;let h={ok:l,selected:l?m:null,attempts:[m],errors:l?[]:c};if(!a&&!o)return h;let g=e=>{if(!e)return e;let t=B({device:e,modelBytes:a||0,kvCacheBytes:o||0}),n=null;return s!=null&&s!==o&&(n=B({device:e,modelBytes:a||0,kvCacheBytes:s})),{...e,fit:t,...n&&{limitedFit:n}}};return h.selected=g(h.selected),h.attempts=h.attempts.map(g),h},we=new Map([[be,ye],[`ggml-stt`,Se],[`mlx-llm`,Ce]]),Te=async({platform:e,totalMemoryInBytes:t,backend:n=be,dependencies:r,...i}={})=>{let a=we.get(n);if(!a)throw Error(`No capability detector registered for backend "${n}"`);return await a({...i,dependencies:r,totalMemoryInBytes:t,platform:e})},Ee={f16:2,f32:4,q8_0:1,q6_k:.75,q5_k:.625,q5_k_m:.625,q5_k_s:.625,q5_1:.625,q5_0:.625,q4_k:.5,q4_k_m:.5,q4_k_s:.5,q4_1:.5,q4_0:.5,iq4_nl:.5},De=e=>Ee[e?String(e).toLowerCase():`f16`]||Ee.f16,Oe=(e,t,n,r,i,a={},{totalLayers:o=null,swaLayers:s=0,swaContext:c=null,swaContextMultiplier:l=1,swaAdditionalTokens:u=0,swaFull:d=!1}={})=>{if(!e||!t||!n||!r||!i)return 0;let f=o!=null&&o!==void 0?Number(o):Number(e),p=Math.max(0,Math.floor(f));if(!p)return 0;let m=De(a.k),h=De(a.v),g=Number(n)*(Number(r)*m+Number(i)*h);if(!g)return 0;let _=Math.max(0,Number(t)||0),v=Math.min(p,Math.max(0,Math.floor(Number(s)||0))),y=Math.max(0,p-v),b=c!=null&&Number.isFinite(Number(c))?Math.max(0,Number(c)):_,x=Math.max(1,Number(l)||1),S=Math.max(0,Number(u)||0),C=b*x+S,w=d?_:Math.min(_,C),T=y*_+v*Math.max(0,Math.floor(w));return Math.round(g*T)},ke=({modelBytes:e=0,audioLengthSeconds:t=30,sampleRate:n=16e3,bytesPerSample:r=4}={})=>{let i=Math.max(0,Number(e)||0),a=Math.max(0,Math.floor(Math.max(0,t)*n*r)),o=1024*1024,s=1024*o,c;c=i<200*o?120*o:i<500*o?140*o:i<2*s?150*o:160*o;let l;l=i<200*o?70*o:i<500*o?135*o:(2*s,220*o);let u;u=i<100*o?20*o:i<200*o?30*o:i<500*o?85*o:i<2*s?215*o:360*o;let d=c+l+u;return{modelBytes:i,audioBufferBytes:a,processingBufferBytes:d,totalBytes:i+d+a}},Ae=e=>e?String(e).trim().toLowerCase():null,je=(e={},t=null)=>{if(!e)return null;let n=Ae(t),r=n?`${n}.attention.sliding_window`:null,i=(r&&e[r]!=null?e[r]:null)??e[`llama.attention.sliding_window`];if(i==null)return null;let a=Number(i);return Number.isFinite(a)?a:null},Me=(e=0,t=0,n=!1)=>{let r=Math.max(0,Math.floor(Number(e)||0)),i=Math.max(0,Math.floor(Number(t)||0));if(!r||i===1)return 0;if(i<=0)return r;let a=Math.max(0,i-1),o=Math.floor(r/i),s=r%i,c=n?Math.max(0,s-1):Math.min(s,a);return o*a+c},Ne=({arch:e,nLayer:t=0})=>({arch:Ae(e),enabled:!1,window:null,pattern:null,denseFirst:!1,type:null,kvLayers:Math.max(0,Math.floor(Number(t)||0)),swaLayers:0}),Pe=new Map([[`llama4`,({nSwa:e})=>e===0?{enabled:!1}:{enabled:!0,window:e&&e>0?e:8192,pattern:4,type:`chunked`}],[`afmoe`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:4,type:`standard`}],[`phi3`,()=>({enabled:!1})],[`gemma2`,({nSwa:e})=>{let t=e&&e>0?e:4096;return t?{enabled:!0,window:t,pattern:2,type:`standard`}:{enabled:!1}}],[`gemma3`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:6,type:`standard`}],[`gemma3n`,({nLayer:e,nSwa:t})=>!t||t<=0?{enabled:!1}:{enabled:!0,window:t,pattern:5,type:`standard`,kvLayers:Math.min(20,e)}],[`gemma-embedding`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:6,type:`symmetric`}],[`cohere2`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:4,type:`standard`}],[`olmo2`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:4,type:`standard`}],[`exaone4`,({nLayer:e,nSwa:t})=>{let n=e>=64,r=null;return t&&t>0?r=t:n&&(r=4096),r?{enabled:!0,window:r,pattern:4,type:`standard`}:{enabled:!1}}],[`gpt-oss`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:2,type:`standard`}],[`gemma4`,({nLayer:e,nSwa:t,metadata:n})=>{if(!t||t<=0)return{enabled:!1};let r=Number(n?.[`gemma4.attention.shared_kv_layers`])||0,i=Math.max(0,e-r),a=n?.[`gemma4.attention.sliding_window_pattern`];return Array.isArray(a)?{enabled:!0,window:t,type:`standard`,swaLayers:a.slice(0,i).filter(e=>Number(e)>0).length,kvLayers:i}:{enabled:!0,window:t,pattern:6,type:`standard`,kvLayers:i}}],[`smallthinker`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:4096,pattern:4,denseFirst:!0,type:`standard`}]]),Fe=({arch:e,metadata:t={},nLayer:n=0}={})=>{let r=Ae(e||t[`general.architecture`]),i=Math.max(0,Math.floor(Number(n)||0)),a=je(t,r),o=r?Pe.get(r):null;if(!o)return Ne({arch:r,nLayer:n});let s=o({nLayer:i,nSwa:a,metadata:t});if(!s||!s.enabled||!s.window||s.window<=0)return Ne({arch:r,nLayer:n});let c=Math.max(0,Math.floor(Number(s.pattern)||0)),l=s.kvLayers!=null&&Number.isFinite(Number(s.kvLayers))?Number(s.kvLayers):i,u=Math.max(0,Math.floor(l)),d=s.swaLayers!=null&&Number.isFinite(Number(s.swaLayers))?Math.max(0,Math.floor(Number(s.swaLayers))):Me(u,c,!!s.denseFirst);return{arch:r,enabled:d>0,window:s.window,pattern:c,denseFirst:!!s.denseFirst,type:s.type||`standard`,kvLayers:u,swaLayers:d}},Ie=new Set([`mamba`,`mamba2`,`rwkv6`,`rwkv6qwen2`,`rwkv7`,`arwkv7`]),Le=new Set([`jamba`,`falcon-h1`,`plamo2`,`granitehybrid`,`lfm2`,`lfm2moe`,`nemotron_h`,`nemotron_h_moe`,`qwen3next`]),Re=e=>e?String(e).trim().toLowerCase():null,ze=e=>{let t=Re(e);return t?Ie.has(t):!1},Be=e=>{let t=Re(e);return t?Le.has(t):!1},Ve=e=>ze(e)?`recurrent`:Be(e)?`hybrid`:`transformer`,He=(e={})=>{let t=e[`general.architecture`],n=(t,n=null)=>{let r=e[t],i=Number(r);return Number.isFinite(i)?i:n},r=(t,n=null)=>{let r=e[t];if(Array.isArray(r))return r;let i=Number(r);return Number.isFinite(i)?i:n},i=t?n(`${t}.context_length`,n(`llama.context_length`)):null,a=t?n(`${t}.block_count`,n(`llama.block_count`)):null,o=t?n(`${t}.embedding_length`,n(`llama.embedding_length`)):null,s=t?n(`${t}.attention.head_count`,n(`llama.attention.head_count`)):null,c=t?r(`${t}.attention.head_count_kv`,r(`llama.attention.head_count_kv`,s)):null,l=null,u=null;if(Array.isArray(c)){let e=c.filter(e=>Number(e)>0);e.length>0?(l=Math.max(...e.map(Number)),u=e.length):(l=0,u=0)}else l=c;let d=t?n(`${t}.attention.key_length`,n(`llama.attention.key_length`)):null,f=t?n(`${t}.attention.value_length`,n(`llama.attention.value_length`)):null,p=e[`general.quantization_version`]||null,m=e[`general.file_type`]||null,h=t?n(`${t}.ssm.conv_kernel`):null,g=t?n(`${t}.ssm.state_size`):null,_=t?n(`${t}.ssm.inner_size`):null,v=t?n(`${t}.ssm.group_count`):null,y=t?n(`${t}.ssm.time_step_rank`):null,b=t?n(`${t}.rwkv.head_size`):null,x=t?n(`${t}.rwkv.token_shift_count`,2):null,S=t?n(`${t}.attention.shared_kv_layers`,0):0,C=u!=null&&a!=null?a-u:null;return{arch:t,nCtxTrain:i,nLayer:a,nEmbd:o,nHead:s,nHeadKv:l,nEmbdHeadK:d,nEmbdHeadV:f,quantVersion:p,fileType:m,attentionLayerCount:u,recurrentLayerCount:C,ssmDConv:h,ssmDState:g,ssmDInner:_,ssmNGroup:v,ssmDtRank:y,rwkvHeadSize:b,rwkvTokenShiftCount:x,sharedKvLayers:S}},Ue=({layerCount:e,headKvCount:t,embdHeadKCount:n,embdHeadVCount:r,cacheTypes:i,swaConfig:a,kvUnified:o=!1,nParallel:s=1,swaFull:c=!1,arch:l=null,attentionLayerCount:u=null})=>{let d=Ve(l);if(d===`recurrent`)return()=>0;let f=d===`hybrid`&&u!=null?Math.max(0,Math.floor(Number(u)||0)):e,p=a?.window&&o?Math.max(1,Number(s)||1):1,m=o?1:Math.max(1,Number(s)||1);return e=>Oe(f,e,t,n,r,i,{totalLayers:f,swaLayers:a?.swaLayers||0,swaContext:a?.window,swaFull:c,swaContextMultiplier:p})*m},We=({nLayer:e,nEmbd:t,recurrentLayerCount:n=null,nSeqMax:r=1,ssmDConv:i=null,ssmDState:a=null,ssmDInner:o=null,ssmNGroup:s=null,ssmDtRank:c=null,rwkvHeadSize:l=null,rwkvTokenShiftCount:u=2,arch:d=null})=>{if(Ve(d)===`transformer`)return 0;let f=n==null?Math.max(0,Math.floor(Number(e)||0)):Math.max(0,Math.floor(Number(n)||0));if(f===0)return 0;let p=Math.max(1,Math.floor(Number(r)||1)),m=0,h=0;if(l!=null&&l>0&&t!=null&&t>0)m=Math.max(1,Number(u)||2)*t,h=t*l;else if(a!=null&&o!=null){let e=Math.max(0,Number(i)||0),t=Math.max(0,Number(a)||0),n=Math.max(0,Number(o)||0),r=Math.max(1,Number(s)||1);Math.max(0,Number(c)||0)>0?(m=e>0?(e-1)*2*r*t:0,h=Math.floor(t*n/2)):(m=e>0?(e-1)*(n+2*r*t):0,h=t*n)}else return 0;let g=(m+h)*p*f*4;return Math.max(0,g)},Ge=({maxCtx:e,availableMemory:t,modelBytes:n,kvBytesForCtx:r})=>{let i=Math.max(1,Math.floor(Number(e)||0));if(!r||t<=n)return i;let a=1,o=i,s=i;for(;a<=o;){let e=Math.floor((a+o)/2);n+r(e)<=t?(s=e,a=e+1):o=e-1}return s},V=new T;V.setMaxListeners(100);const Ke=(e,t,n)=>{e.push({...t,timestamp:t.timestamp||new Date().toISOString()}),e.length>n&&e.shift()};var qe=class{constructor(e=9999){this.maxEntries=e,this.modelLoads=[],this.completions=[],this.transcriptions=[]}addModelLoad(e){Ke(this.modelLoads,e,this.maxEntries),V.emit(`status:modelLoad`,e),V.emit(`status:change`,{type:`modelLoad`,entry:e})}addCompletion(e){Ke(this.completions,e,this.maxEntries),V.emit(`status:completion`,e),V.emit(`status:change`,{type:`completion`,entry:e})}addTranscription(e){Ke(this.transcriptions,e,this.maxEntries),V.emit(`status:transcription`,e),V.emit(`status:change`,{type:`transcription`,entry:e})}getModelLoadHistory(){return[...this.modelLoads].reverse()}getCompletionHistory(){return[...this.completions].reverse()}getTranscriptionHistory(){return[...this.transcriptions].reverse()}clear(){this.modelLoads=[],this.completions=[],this.transcriptions=[]}};const H=new qe,U=new qe;let Je=0;function Ye(e){let t=t=>e(t);return V.on(`status:change`,t),()=>V.off(`status:change`,t)}function Xe(e){return Je+=1,{subscriberId:Je,unsubscribe:Ye(e)}}function Ze(e){let t=[];return{generators:Array.from(e.entries()).filter(([,e])=>e.type===`ggml-llm`).map(([e,n])=>{let{instance:r}=n,i=[];return r.contexts&&(i=Array.from(r.contexts.entries()).map(([n,r])=>{let i={key:n,refCount:r.refCount,hasModel:!!r.context},a=r.context.parallel.getStatus();return i.parallelStatus=a,t.push({generatorId:e,contextKey:n,...a}),i})),{id:e,type:n.type,refCount:n.refCount,repoId:r.info?.model?.repoId||null,quantization:r.info?.model?.quantization||null,variant:r.info?.runtime?.variant||null,nCtx:r.info?.runtime?.n_ctx||null,nParallel:r.info?.runtime?.n_parallel||null,contexts:i}}),parallelStatuses:t,history:{modelLoads:H.getModelLoadHistory().filter(e=>e.variant!==`mlx`),completions:H.getCompletionHistory().filter(e=>e.variant!==`mlx`)}}}function Qe(e){return{generators:Array.from(e.entries()).filter(([,e])=>e.type===`ggml-stt`).map(([e,t])=>{let{instance:n}=t,r=n.getStatus?.()||{},i=r.queueStatus||{processing:!1,queuedCount:0};return{id:e,type:t.type,refCount:t.refCount,repoId:n.info?.model?.repoId||null,quantization:n.info?.model?.quantization||null,modelType:n.info?.model?.modelType||null,variant:n.info?.runtime?.variant||null,hasContext:r.hasContext||!1,contextRefCount:r.contextRefCount||0,queueStatus:i}}),history:{modelLoads:U.getModelLoadHistory(),transcriptions:U.getTranscriptionHistory()}}}function $e(e){return{generators:Array.from(e.entries()).filter(([,e])=>e.type===`mlx-llm`).map(([e,t])=>{let{instance:n}=t,r=n.getStatus?.()||{};return{id:e,type:t.type,refCount:t.refCount,repoId:r.repoId||n.info?.model?.repoId||null,variant:r.variant||`mlx`,contexts:r.contexts||[]}}),history:{modelLoads:H.getModelLoadHistory().filter(e=>e.variant===`mlx`),completions:H.getCompletionHistory().filter(e=>e.variant===`mlx`)}}}function et(e){return{timestamp:new Date().toISOString(),ggmlLlm:Ze(e),ggmlStt:Qe(e),mlxLlm:$e(e)}}const{ReadableStream:tt,WritableStream:nt}=typeof globalThis<`u`&&globalThis.ReadableStream&&globalThis.WritableStream?{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream}:o,rt=(e={},t={})=>(Object.entries(t||{}).forEach(([t,n])=>{n&&typeof n==`object`&&!Array.isArray(n)?((!e[t]||typeof e[t]!=`object`)&&(e[t]={}),rt(e[t],n)):e[t]=n}),e),it=`https://huggingface.co`,at=`https://huggingface.co/api`,W=_.join(v.homedir(),`.buttress`,`models`),ot=[`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`],st=.5,ct={backend:{type:`ggml-llm`,variant:null,variant_preference:[`cuda`,`vulkan`,`snapdragon`,`default`],gpu_memory_fraction:.85,cpu_memory_fraction:st},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:at,base_url:it,enable_mtmd:!1,mmproj_filename:null,mmproj_url:null,mmproj_local_path:null,mmproj_use_gpu:null,mmproj_image_min_tokens:-1,mmproj_image_max_tokens:-1},runtime:{cache_dir:W,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},session_cache:{enabled:!0,max_size_bytes:10*1024*1024*1024,max_entries:1e3},context_release_delay_ms:1e4}},lt=(e,t=[])=>!e&&e!==0?[...t]:Array.isArray(e)?e.filter(e=>e!=null):[e],ut=e=>{if(!e)return null;let t=String(e).toLowerCase();return[`cuda`,`vulkan`,`snapdragon`,`default`].includes(t)?t:null},dt=(e={})=>{let t=structuredClone(ct);if(rt(t,e),t.backend.variant=ut(t.backend.variant),t.backend.variant_preference=Array.from(new Set(lt(t.backend.variant_preference).flatMap(e=>{let t=ut(e);return t?[t]:[]}))),t.backend.variant_preference.length===0&&(t.backend.variant_preference=[`cuda`,`vulkan`,`snapdragon`,`default`]),t.runtime.prefer_variants=Array.from(new Set(lt(t.runtime.prefer_variants).flatMap(e=>{let t=ut(e);return t?[t]:[]}))),t.model.preferred_quantizations=Array.from(new Set(lt(t.model.preferred_quantizations||t.model.quantizations).map(e=>e?String(e).toLowerCase():null).filter(Boolean))),t.model.quantization){let e=String(t.model.quantization).toLowerCase();t.model.preferred_quantizations.includes(e)||t.model.preferred_quantizations.unshift(e)}t.model.n_parallel=t.model.n_parallel?Math.max(1,Number(t.model.n_parallel)):void 0,t.model.n_batch=Math.max(1,Number(t.model.n_batch)||512),t.model.base_url=t.model.base_url||it,t.model.api_base=t.model.api_base||at,t.model.enable_mtmd=!!t.model.enable_mtmd;let n=e=>{if(e==null)return-1;let t=Number(e);return Number.isFinite(t)?Math.floor(t):-1};return t.model.mmproj_image_min_tokens=n(t.model.mmproj_image_min_tokens),t.model.mmproj_image_max_tokens=n(t.model.mmproj_image_max_tokens),t.runtime.cache_dir=t.runtime.cache_dir?_.resolve(t.runtime.cache_dir):W,t.runtime.session_cache={...ct.runtime.session_cache,...t.runtime.session_cache||{}},t.runtime.context_release_delay_ms=Math.max(0,Number(t.runtime.context_release_delay_ms)||ct.runtime.context_release_delay_ms),t},ft=e=>{let t=e.toLowerCase();return ot.find(e=>t.includes(e))||null},pt=e=>{let t=[];return e.backend.variant&&t.push(e.backend.variant),e.runtime.prefer_variants.length>0&&t.push(...e.runtime.prefer_variants),t.push(...e.backend.variant_preference),t.push(`default`),Array.from(new Set(t.flatMap(e=>{let t=ut(e);return t?[t]:[]})))},G=async e=>{await l(e,{recursive:!0})},mt=(e=W)=>_.join(e,`.metadata-cache`),ht=(e,t,n=W)=>{let r=y(`sha256`).update(e).digest(`hex`);return _.join(mt(n),t,`${r}.json`)},gt=async(e,t,n=W)=>{try{let r=ht(e,t,n),i=await d(r,`utf-8`);return console.log(`[Cache] Hit ${t} cache:`,_.basename(r)),JSON.parse(i,(e,t)=>typeof t==`string`&&t.startsWith(`__bigint__`)?BigInt(t.slice(10)):t)}catch{return null}},_t=async(e,t,n,r=W)=>{try{let i=ht(e,t,r);await G(_.dirname(i)),await g(i,JSON.stringify(n,(e,t)=>typeof t==`bigint`?`__bigint__${t.toString()}`:t),`utf-8`),console.log(`[Cache] Wrote ${t} cache:`,_.basename(i))}catch(e){console.warn(`[Cache] Failed to write ${t} cache:`,e.message)}},vt=(e=W)=>_.join(e,`.session-state-cache`),yt=(e=W)=>_.join(vt(e),`cache-map.json`),bt=(e=W)=>_.join(vt(e),`temp`),xt=(e=W)=>_.join(vt(e),`states`),St=()=>({version:1,entries:{},totalSize:0}),Ct=async(e=W)=>{try{let t=await d(yt(e),`utf-8`),n=JSON.parse(t);return!n.entries||typeof n.entries!=`object`?St():n}catch{return St()}},wt=async(e,t=W)=>{let n=yt(t),r=`${n}.tmp.${Date.now()}`;try{await G(_.dirname(n)),await g(r,JSON.stringify(e,null,2),`utf-8`),await p(r,n)}catch(e){throw await h(r).catch(()=>{}),e}},Tt=(e,t)=>{let n=JSON.stringify({text:e,model:t.modelPath,variant:t.variant,n_gpu_layers:t.n_gpu_layers,n_ctx:t.n_ctx,cacheTypeK:t.cacheTypeK,cacheTypeV:t.cacheTypeV,kvUnified:t.kvUnified,swaFull:t.swaFull,flashAttnType:t.flashAttnType});return y(`sha256`).update(n).digest(`hex`).slice(0,24)},Et=(e,t=W)=>_.join(xt(t),`${e}.bin`),Dt=(e=W)=>{let t=`${Date.now()}-${Math.random().toString(36).slice(2,10)}`;return _.join(bt(e),`${t}.bin`)},Ot=(e,t)=>e.modelPath===t.modelPath&&e.variant===t.variant&&e.n_gpu_layers===t.n_gpu_layers&&e.n_ctx>=t.n_ctx&&e.cacheTypeK===t.cacheTypeK&&e.cacheTypeV===t.cacheTypeV&&e.kvUnified===t.kvUnified&&e.swaFull===t.swaFull&&e.flashAttnType===t.flashAttnType&&!!e.isRecurrent==!!t.isRecurrent&&!!e.isHybrid==!!t.isHybrid,kt=(e,t)=>{let n=Math.min(e.length,t.length),r=0;for(;r<n&&e[r]===t[r];)r+=1;return r},At=(e,t,n,r=!1)=>{let i=Object.values(n.entries);console.log(`[SessionCache] Finding match for promptText (${e.length} chars), exactMatch=${r}`),console.log(`[SessionCache] Checking ${i.length} cache entries`);let a=i.filter(e=>Ot(e.metadata,t));if(r){let t=a.find(t=>t.fullText===e);return t?(console.log(`[SessionCache] Exact match found: ${t.id} (${t.fullText.length} chars)`),{entry:t,prefixLength:t.fullText.length,exactMatch:!0}):null}let o=a.reduce((t,n)=>{let r=kt(e,n.fullText);return r>t.prefixLen||r===t.prefixLen&&n.fullText.length>(t.entry?.fullText?.length||0)?{entry:n,prefixLen:r}:t},{entry:null,prefixLen:0});return o.entry?(console.log(`[SessionCache] Prefix match found: ${o.entry.id} (${o.prefixLen}/${o.entry.fullText.length} chars)`),{entry:o.entry,prefixLength:o.prefixLen}):(console.log(`[SessionCache] No match found`),null)},jt=async(e,t,n)=>{let r=Object.values(e.entries),i=r.sort((e,t)=>new Date(e.lastAccessedAt)-new Date(t.lastAccessedAt)),a=e.totalSize,o=r.length,s=i.filter(e=>!(a>t)&&!(o>n)?!1:(a-=(e.stateFileSize||0)+(e.promptStateSize||0),--o,!0));return await Promise.all(s.map(async t=>{await h(t.stateFilePath).catch(()=>{}),t.promptStatePath&&await h(t.promptStatePath).catch(()=>{}),delete e.entries[t.id],console.log(`[SessionCache] Evicted entry: ${t.id}`)})),e.totalSize=Math.max(0,a),s.map(e=>e.id)},Mt=async(e,t,n,r)=>{let i=[];for(let[a,o]of Object.entries(e.entries))a!==n&&Ot(o.metadata,r)&&t.startsWith(o.fullText)&&o.fullText.length<t.length&&i.push(o);return await Promise.all(i.map(async t=>{await h(t.stateFilePath).catch(()=>{}),t.promptStatePath&&await h(t.promptStatePath).catch(()=>{}),e.totalSize-=(t.stateFileSize||0)+(t.promptStateSize||0),delete e.entries[t.id],console.log(`[SessionCache] Evicted superseded prefix entry: ${t.id} (${t.promptText.length} prompt chars)`)})),i.map(e=>e.id)},Nt=async(e=W)=>{let t=bt(e);try{let e=await f(t),n=Date.now();await Promise.all(e.map(async e=>{let r=_.join(t,e),i=await m(r).catch(()=>null);i&&n-i.mtimeMs>36e5&&(await h(r).catch(()=>{}),console.log(`[SessionCache] Cleaned up temp file: ${e}`))}))}catch{}},Pt=async e=>{try{return await m(e),!0}catch{return!1}},Ft=(e,t)=>e==null?t:typeof e==`number`?e:typeof e==`string`?w.parse(e)??t:t;var It=class e{constructor(e,t){this.config=e,this.plan=t,this.baseDir=e.runtime.cache_dir,this.enabled=e.runtime.session_cache?.enabled!==!1,this.maxSizeBytes=Ft(e.runtime.session_cache?.max_size_bytes,10*1024*1024*1024),this.maxEntries=e.runtime.session_cache?.max_entries||1e3,this.metadata={variant:t.info?.runtime?.variant||null,n_gpu_layers:t.info?.runtime?.n_gpu_layers||0,n_ctx:t.info?.runtime?.n_ctx||0,modelPath:t.localPath,cacheTypeK:t.info?.runtime?.cache_type_k||`f16`,cacheTypeV:t.info?.runtime?.cache_type_v||`f16`,kvUnified:t.info?.runtime?.kv_unified??null,swaFull:t.info?.runtime?.swa_full??null,flashAttnType:t.info?.runtime?.flash_attn_type||`off`,isRecurrent:!1,isHybrid:!1},this.cacheMap=null,this.initialized=!1}updateModelInfo(e){e&&(this.metadata.isRecurrent=!!e.is_recurrent,this.metadata.isHybrid=!!e.is_hybrid,(this.metadata.isRecurrent||this.metadata.isHybrid)&&console.log(`[SessionCache] Model architecture: recurrent=${this.metadata.isRecurrent}, hybrid=${this.metadata.isHybrid}`))}requiresExactMatch(){return this.metadata.isRecurrent||this.metadata.isHybrid}async persistCacheMap(){try{await wt(this.cacheMap,this.baseDir)}catch(e){console.warn(`[SessionCache] Failed to persist cache map: ${e?.message||e}`)}}static checkTokenPrefixMatch(e,t){if(e.length>t.length)return!1;for(let n=0;n<e.length;n+=1)if(e[n]!==t[n])return!1;return!0}static async tokenizeToArray(e,t){let n=await e.tokenize(t);return Array.from(n?.tokens||[])}async findFormattedMatchForRecurrent(t,n,r){let i=await e.tokenizeToArray(r,n),a=t.map(async t=>{try{let n=await e.tokenizeToArray(r,t.fullText);if(e.checkTokenPrefixMatch(n,i))return{entry:t,usePromptState:!1,tokenCount:n.length};if(t.promptStatePath&&t.promptText){let n=await e.tokenizeToArray(r,t.promptText);if(e.checkTokenPrefixMatch(n,i))return{entry:t,usePromptState:!0,tokenCount:n.length}}return null}catch(e){return console.warn(`[SessionCache] Failed to check entry ${t.id}: ${e.message}`),null}}),o=(await Promise.all(a)).find(e=>e!==null);if(!o)return console.log(`[SessionCache] No token prefix match found for recurrent/hybrid model`),null;let{entry:s,usePromptState:c,tokenCount:l}=o;return console.log(`[SessionCache] Token prefix match: ${s.id} (${l} tokens, usePromptState=${c})`),await Pt(c?s.promptStatePath:s.stateFilePath)?(s.lastAccessedAt=new Date().toISOString(),await this.persistCacheMap(),{entry:s,usePromptState:c}):(await this.removeStaleEntry(s),null)}async initialize(){if(!(!this.enabled||this.initialized))try{await G(vt(this.baseDir)),await G(bt(this.baseDir)),await G(xt(this.baseDir)),this.cacheMap=await Ct(this.baseDir),this.initialized=!0,console.log(`[SessionCache] Initialized with ${Object.keys(this.cacheMap.entries).length} entries`)}catch(e){console.warn(`[SessionCache] Failed to initialize: ${e.message}`),this.enabled=!1}}async removeStaleEntry(e){console.log(`[SessionCache] Removing stale entry: ${e.id}`),e.stateFilePath&&await h(e.stateFilePath).catch(()=>{}),e.promptStatePath&&await h(e.promptStatePath).catch(()=>{}),delete this.cacheMap.entries[e.id],this.cacheMap.totalSize-=(e.stateFileSize||0)+(e.promptStateSize||0),await this.persistCacheMap()}async findMatchingEntry(e,t=null){if(!this.enabled||!this.cacheMap)return null;let n=this.requiresExactMatch();if(n&&t){let n=Object.values(this.cacheMap.entries).filter(e=>Ot(e.metadata,this.metadata)&&e.fullText);return this.findFormattedMatchForRecurrent(n,e,t)}let r=At(e,this.metadata,this.cacheMap,n);if(!r)return null;let{entry:i}=r;return await Pt(i.stateFilePath)?(i.lastAccessedAt=new Date().toISOString(),await this.persistCacheMap(),{entry:i,usePromptState:!1}):(await this.removeStaleEntry(i),null)}async prepareCompletionOptions(e,t,n=null){let r={options:e,cacheEntry:null,promptPrefix:null};if(!this.enabled)return r;let i=await this.findMatchingEntry(t,n);if(!i)return r;let{entry:a,usePromptState:o}=i,s=o?a.promptStatePath:a.stateFilePath,c=o?a.promptText:a.fullText;return console.log(`[SessionCache] Found matching entry: ${a.id} (${c.length} chars, usePromptState=${o})`),{options:{...e,load_state_path:s},cacheEntry:a,promptPrefix:c}}async saveCompletionState(e,t,n,r=0,i=null){if(!this.enabled)return null;let a=e+t,o=Tt(a,this.metadata),s=()=>{n&&h(n).catch(()=>{}),i&&h(i).catch(()=>{})};if(this.cacheMap.entries[o]){console.log(`[SessionCache] Entry already exists for prompt: ${o}, updating position`);let e=this.cacheMap.entries[o];return e.lastAccessedAt=new Date().toISOString(),delete this.cacheMap.entries[o],this.cacheMap.entries[o]=e,await this.persistCacheMap(),s(),e}let c=Et(o,this.baseDir),l=i?Et(`${o}-prompt`,this.baseDir):null;try{await G(_.dirname(c)),await p(n,c);let s=await m(c),u=0;if(i&&l)try{await p(i,l),u=(await m(l)).size,console.log(`[SessionCache] Saved prompt state: ${l}`)}catch(e){console.warn(`[SessionCache] Failed to save prompt state: ${e.message}`)}let d={id:o,promptText:e,completionText:t,fullText:a,promptTokenCount:r,stateFilePath:c,stateFileSize:s.size,promptStatePath:l||null,promptStateSize:u,metadata:{...this.metadata},createdAt:new Date().toISOString(),lastAccessedAt:new Date().toISOString()};return this.cacheMap.entries[o]=d,this.cacheMap.totalSize+=s.size+u,this.requiresExactMatch()||await Mt(this.cacheMap,e,o,this.metadata),await jt(this.cacheMap,this.maxSizeBytes,this.maxEntries),await wt(this.cacheMap,this.baseDir),console.log(`[SessionCache] Saved entry: ${o} (${s.size} bytes, ${a.length} chars)`),d}catch(e){return console.warn(`[SessionCache] Failed to save state: ${e.message}`),s(),null}}async generateTempStatePath(){return await G(bt(this.baseDir)),Dt(this.baseDir)}async cleanup(){await Nt(this.baseDir)}};const Lt=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,t);if(!n.ok){let t=await n.text().catch(()=>``);throw Error(`Failed to fetch ${e}: ${n.status} ${n.statusText} ${t}`.trim())}return n.json()},Rt=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,{...t,method:`HEAD`});if(!n.ok)throw Error(`Failed to fetch headers for ${e}: ${n.status} ${n.statusText}`);return n},zt=async(e,t,n=W)=>{let r=JSON.stringify({url:e,headers:t}),i=await gt(r,`range-metadata`,n);if(i)return i;let a=!/^https?:/i.test(e),{metadata:o}=await b(e,{fetch,additionalFetchHeaders:t,allowLocalFile:a});return await _t(r,`range-metadata`,o,n),o},Bt=(e,t)=>{if(e.model.local_path)return _.resolve(e.model.local_path);let n=t.repoId.split(`/`),r=_.join(e.runtime.cache_dir,...n,t.revision);return _.join(r,t.filename)},K=async(e,t)=>{try{let n=await m(e);return t?n.size===t:!0}catch{return!1}},Vt=async(e,t,n,r,i)=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);await G(_.dirname(n));let a=await fetch(e,{headers:t});if(!a.ok||!a.body)throw Error(`Failed to download ${e}: ${a.status} ${a.statusText}`);let o=await u(n,`w`),s=Number(a.headers.get(`content-length`))||r||0,c=0,l=.05;try{await a.body.pipeTo(new nt({async write(e){if(await o.write(e),c+=e.byteLength,typeof i==`function`&&s>0){let e=Math.min(1,c/s);for(;e>=l;)i(l),l+=.05}},async close(){await o.close(),typeof i==`function`&&i(1)},async abort(e){throw await o.close().catch(()=>{}),await h(n).catch(()=>{}),e}}))}catch(e){throw await o.close().catch(()=>{}),await h(n).catch(()=>{}),e}if(r){let e=await m(n);if(e.size!==r)throw await h(n).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${r} got ${e.size}`)}},Ht=async e=>{let t=e.model.repo_id||e.model.repository||e.model.model;if(!t)throw Error("`model.repo_id` is required in Buttress backend config");let n=e.model.revision||`main`,r=e.runtime.cache_dir,i=JSON.stringify({repoId:t,revision:n,filename:e.model.filename,url:e.model.url,quantization:e.model.quantization,preferred_quantizations:e.model.preferred_quantizations}),a=await gt(i,`artifact-info`,r);if(a)return a;let o={...e.runtime.http_headers||{}};if(e.runtime.huggingface_token&&(o.Authorization=`Bearer ${e.runtime.huggingface_token}`),e.model.url){let a=await Rt(e.model.url,{headers:o}),s=Number(a.headers.get(`content-length`))||null,c={repoId:t,revision:n,filename:e.model.filename||e.model.url.split(`/`).pop(),url:e.model.url,size:s,headers:o};return await _t(i,`artifact-info`,c,r),c}let{filename:s}=e.model,c=e.model.quantization&&String(e.model.quantization).toLowerCase(),l=await Lt(`${e.model.api_base}/models/${t}?revision=${n}&blobs=true`,{headers:o}),u=l?.siblings||l?.files||[],d=[];for(let e of u){let t=e.rfilename||e.path||e.filename;typeof t==`string`&&t.endsWith(`.gguf`)&&d.push(t)}if(d.length===0)throw Error(`No GGUF artifacts found in repo ${t}`);let f=e.model.preferred_quantizations.length>0?e.model.preferred_quantizations:ot,p=d.map(e=>e.toLowerCase()),m=()=>{for(let e of f){let t=p.findIndex(t=>t.includes(e));if(t!==-1)return{filename:d[t],quantization:e}}return null};if(s)c||=ft(s);else{let{filename:e,quantization:t}=m()||{filename:d[0],quantization:null};s=e,c=t||ft(s)}let h=`${e.model.base_url.replace(/\/+$/,``)}/${t}/resolve/${n}/${s}`,g=/-(\d{5})-of-(\d{5})\.gguf$/,_=s.match(g),v=null;if(_){let[,,r]=_,i=await Lt(`${e.model.api_base}/models/${t}?revision=${n}&blobs=true`,{headers:o}),a=i?.siblings||i?.files||[],c=Number(r);v=0;for(let e=1;e<=c;e+=1){let t=String(e).padStart(5,`0`),n=s.replace(g,`-${t}-of-${r}.gguf`),i=a.find(e=>(e.rfilename||e.path||e.filename)===n),o=Number(i?.size);Number.isFinite(o)&&o>0&&(v+=o)}}else{let e=await Rt(h,{headers:o});v=Number(e.headers.get(`content-length`))||null}let y={repoId:t,revision:n,filename:s,url:h,size:v,quantization:c,headers:o,isSplit:!!_,splitCount:_?Number(_[2]):0};return await _t(i,`artifact-info`,y,r),y},Ut=/^mmproj-.*\.gguf$/i,Wt=async(e,t)=>{if(!e.model.enable_mtmd)return null;let n=e.runtime.cache_dir,r={...e.runtime.http_headers||{}};e.runtime.huggingface_token&&(r.Authorization=`Bearer ${e.runtime.huggingface_token}`);let i=t?.repoId||e.model.repo_id,a=t?.revision||e.model.revision||`main`,o=JSON.stringify({kind:`mmproj`,repoId:i,revision:a,mmproj_filename:e.model.mmproj_filename,mmproj_url:e.model.mmproj_url,mmproj_local_path:e.model.mmproj_local_path}),s=await gt(o,`artifact-info`,n);if(s)return s;if(e.model.mmproj_url){let t=await Rt(e.model.mmproj_url,{headers:r}),s=Number(t.headers.get(`content-length`))||null,c={repoId:i,revision:a,filename:e.model.mmproj_filename||e.model.mmproj_url.split(`/`).pop(),url:e.model.mmproj_url,size:s,headers:r};return await _t(o,`artifact-info`,c,n),c}if(e.model.mmproj_local_path){if(!e.model.allow_local_file)throw Error("`model.mmproj_local_path` requires `model.allow_local_file = true`");let t={repoId:i,revision:a,filename:_.basename(e.model.mmproj_local_path),url:null,size:null,headers:r,localPath:_.resolve(e.model.mmproj_local_path)};return await _t(o,`artifact-info`,t,n),t}if(!i)throw Error("Cannot derive mmproj artifact without `model.repo_id`");let c=await Lt(`${e.model.api_base}/models/${i}?revision=${a}&blobs=true`,{headers:r}),l=c?.siblings||c?.files||[],u=l.map(e=>e.rfilename||e.path||e.filename).filter(e=>typeof e==`string`),d=e.model.mmproj_filename;if(d){if(!u.includes(d))throw Error(`mmproj file "${d}" not found in repo ${i}`)}else{let e=u.filter(e=>Ut.test(e));if(e.length===0)return console.warn(`[buttress] enable_mtmd set but no mmproj file found in ${i}; skipping multimodal load`),null;let n=t?.quantization&&String(t.quantization).toLowerCase();d=n&&e.find(e=>e.toLowerCase().includes(n))||e[0]}let f=`${e.model.base_url.replace(/\/+$/,``)}/${i}/resolve/${a}/${d}`,p=l.find(e=>(e.rfilename||e.path||e.filename)===d),m=Number(p?.size);if(!Number.isFinite(m)||m<=0){let e=await Rt(f,{headers:r});m=Number(e.headers.get(`content-length`))||null}let h={repoId:i,revision:a,filename:d,url:f,size:m,headers:r};return await _t(o,`artifact-info`,h,n),h},Gt=(e,t)=>{if(t?.localPath)return t.localPath;if(!t)return null;let n=t.repoId.split(`/`),r=_.join(e.runtime.cache_dir,...n,t.revision);return _.join(r,t.filename)},Kt=async(e,{modelBytes:t=null,kvCacheBytes:n=null}={})=>{let r=pt(e),[i,...a]=r,o=e.backend?.gpu_memory_fraction==null?ct.backend.gpu_memory_fraction||1:Math.min(1,Math.max(0,Number(e.backend.gpu_memory_fraction))),s=e.backend?.cpu_memory_fraction==null?st:Math.min(1,Math.max(0,Number(e.backend.cpu_memory_fraction))),c=await Te({platform:process.platform,totalMemoryInBytes:v.totalmem(),backend:`ggml-llm`,variant:i||null,preferVariants:a,gpuMemoryFraction:o,cpuMemoryFraction:s,dependencies:{getBackendDevicesInfo:x,isLibVariantAvailable:S},modelBytes:t,kvCacheBytes:n}),l=e=>({...e,devices:Array.isArray(e.devices)?e.devices:[],ok:e.ok,hasGpu:!!e.hasGpu,totalMemory:e.gpuTotalBytes||e.totalMemory||0,error:e.ok?null:Error(e.error||`Variant ${e.variant} not available on this platform`)});if(!c.ok||!c.selected){let e=(c.attempts||[]).map(e=>`${e.variant}: ${e.error||`unknown error`}`).join(`; `);throw Error(`Unable to initialize any backend variant (${r.join(`, `)}). Errors: ${e}`)}let u=(c.attempts||[]).map(l);return{selected:l(c.selected),attempts:u}},qt=async e=>{let t=await Ht(e),n=await Wt(e,t),r=await zt(t.url,t.headers,e.runtime.cache_dir),{arch:i,nCtxTrain:a,nLayer:o,nEmbd:s,nHead:c,nHeadKv:l,nEmbdHeadK:u,nEmbdHeadV:d,quantVersion:f,fileType:p,attentionLayerCount:m,recurrentLayerCount:h,ssmDConv:g,ssmDState:_,ssmDInner:y,ssmNGroup:b,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C}=He(r),w=Number.isFinite(Number(o))?Number(o):0,T=Number.isFinite(Number(s))?Number(s):0,E=Number.isFinite(Number(c))?Number(c):0,D=Number.isFinite(Number(l))?Number(l):E,ee=E>0&&T>0?T/E:128,O=u!=null&&Number.isFinite(Number(u))?Number(u):ee,te=d!=null&&Number.isFinite(Number(d))?Number(d):ee,k=Fe({arch:i,metadata:r,nLayer:w}),A=k&&Number.isFinite(Number(k.kvLayers))?Number(k.kvLayers):w,ne=Math.max(0,Math.floor(Number(A)||0)),j={use_mmap:e.model.use_mmap??e.runtime.use_mmap,use_mlock:e.model.use_mlock??e.runtime.use_mlock,no_extra_bufts:e.model.no_extra_bufts??e.runtime.no_extra_bufts,n_threads:e.model.n_threads??e.runtime.n_threads,n_ctx:e.model.n_ctx??e.runtime.n_ctx,n_batch:e.model.n_batch??e.runtime.n_batch,n_ubatch:e.model.n_ubatch??e.runtime.n_ubatch,n_cpu_moe:e.model.n_cpu_moe??e.runtime.n_cpu_moe,n_parallel:(e.model.n_parallel??e.runtime.n_parallel)||4,cpu_mask:e.model.cpu_mask??e.runtime.cpu_mask,cpu_strict:e.model.cpu_strict??e.runtime.cpu_strict,devices:e.model.devices??e.runtime.devices,n_gpu_layers:e.model.n_gpu_layers??e.runtime.n_gpu_layers,flash_attn_type:e.model.flash_attn_type??e.runtime.flash_attn_type,cache_type_k:e.model.cache_type_k??e.runtime.cache_type_k,cache_type_v:e.model.cache_type_v??e.runtime.cache_type_v,kv_unified:e.model.kv_unified??e.runtime.kv_unified,swa_full:e.model.swa_full??e.runtime.swa_full,ctx_shift:e.model.ctx_shift??e.runtime.ctx_shift},M=j.n_ctx?Number(j.n_ctx):null,N=M||a||4096,P=[],F=[],re=!0;if(M&&a&&M>a){re=!1;let e=`Requested context length (${M}) exceeds model training context (${a})`;P.push(e),F.push(e),N=a}M&&!a&&P.push(`Model metadata missing training context length, using requested value`);let I={k:j.cache_type_k,v:j.cache_type_v},L=t.size>0?t.size:0,ie=Ue({layerCount:ne,headKvCount:D,embdHeadKCount:O,embdHeadVCount:te,cacheTypes:I,swaConfig:k,kvUnified:j.kv_unified,nParallel:j.n_parallel,swaFull:j.swa_full,arch:i,attentionLayerCount:m}),ae=We({nLayer:w,nEmbd:T,recurrentLayerCount:h,nSeqMax:j.n_parallel||4,ssmDConv:g,ssmDState:_,ssmDInner:y,ssmNGroup:b,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C,arch:i}),R=await Kt(e,{modelBytes:L,kvCacheBytes:ie(N)+ae}),oe=R.selected.totalMemory||0,se=oe*(e.backend.gpu_memory_fraction||1),ce=e.backend.cpu_memory_fraction==null?st:Math.min(1,Math.max(0,Number(e.backend.cpu_memory_fraction))),le=Math.max(0,v.totalmem()*ce),ue=R.selected.hasGpu?se:le,z=Ge({maxCtx:N,availableMemory:ue,modelBytes:L,kvBytesForCtx:ie});if(!M&&z){let e=a?Math.min(z,a):z,t=Math.max(32,e);t<N&&P.push(`Context length capped to ${t} by memory limits`),N=t}N>z&&(N=z);let de=Math.floor(z);console.log(`[buttress] Memory-limited context length: ${de}`);let fe=ie(N),pe=L+fe+ae,me=w?L/(w+1):L,he=0;R.selected.hasGpu&&me>0&&(he=Math.min(w+1,Math.max(0,Math.floor(se/me)))),console.log(`[buttress] Auto GPU layer capacity (${R.selected.variant}): ${he}/${w+1}`);let ge;ge=j.n_gpu_layers===`auto`||j.n_gpu_layers==null?he:Math.max(0,Math.min(Number(j.n_gpu_layers)||0,w+1));let _e=(()=>{let e=j.flash_attn_type&&String(j.flash_attn_type).toLowerCase();return e===`on`||e===`off`?e:R.selected.hasGpu?`auto`:`off`})(),ve=e.runtime.cache_dir,B=Bt(e,t),ye=await K(B,t.size),be=Gt(e,n),xe=be?await K(be,n?.size):!1,Se=n?{enabled:!0,initialized:!1,filename:n.filename,url:n.url,sizeBytes:n.size,localPath:be,exists:xe,useGpu:e.model.mmproj_use_gpu,imageMinTokens:e.model.mmproj_image_min_tokens,imageMaxTokens:e.model.mmproj_image_max_tokens}:{enabled:!1,requested:!!e.model.enable_mtmd};return{config:e,info:{ok:re,backend:`ggml-llm`,warnings:P,errors:F,model:{repoId:t.repoId,revision:t.revision,filename:t.filename,quantization:t.quantization,url:t.url,sizeBytes:t.size,metadata:{architecture:i,n_ctx_train:a,n_layer:w,n_embd:T,quantization_version:f,file_type:p,kv_layer_count:ne,swa:k?.enabled?{window:k.window,pattern:k.pattern,dense_first:k.denseFirst,type:k.type,layers:k.swaLayers}:null}},runtime:{...j,variant:R.selected.variant,n_ctx:N,requested_ctx:M,n_gpu_layers:ge,auto_gpu_layers:he,flash_attn_type:_e,cache_type_k:I.k,cache_type_v:I.v,estimated_max_n_ctx:de},resources:{modelBytes:L,kvCacheBytes:fe,recurrentMemoryBytes:ae,totalEstimatedBytes:pe,gpuCapacityBytes:oe,gpuUsableBytes:se,cpuUsableBytes:le,fit:R.selected.fit},devices:{selected:R.selected,attempts:R.attempts},download:{cacheDir:ve,localPath:B,exists:ye},multimodal:Se,timestamp:new Date().toISOString()},artifact:t,mmprojArtifact:n,mmprojLocalPath:be,mmprojLocalExists:xe,metadata:{arch:i,nCtxTrain:a,nLayer:w,nEmbd:T},devices:R,cacheTypes:I,localPath:B,localExists:ye}},Jt=(e,t,n=null,r=null)=>{let i,a=Date.now(),o=0;return new tt({async start(s){try{let c=await e.parallel.completion(t,(e,t)=>{t&&(t.token&&(o+=1),s.enqueue({event:`token`,data:{requestId:e,...t}}))}),{requestId:l}=c;i=c.stop;let u=await c.promise;console.log(`[Completion] Result:`,u),s.enqueue({event:`result`,data:{requestId:l,...u}}),s.close();let d=Date.now()-a,f=u.timings||{};H.addCompletion({id:`completion-${l}`,generatorId:n,requestId:l,repoId:r?.repoId||null,quantization:r?.quantization||null,variant:r?.variant||null,cacheTokens:f.cache_n??0,promptTokens:f.prompt_n??0,tokensGenerated:f.predicted_n??o,tokensPerSecond:f.predicted_per_second??0,promptPerSecond:f.prompt_per_second??0,durationMs:d,success:!0,interrupted:u.interrupted||!1,contextFull:u.context_full||u.contextFull||!1})}catch(e){s.enqueue({event:`error`,data:{message:e?.message||String(e)}}),s.error(e),H.addCompletion({id:`completion-${Date.now()}`,generatorId:n,repoId:r?.repoId||null,quantization:r?.quantization||null,variant:r?.variant||null,durationMs:Date.now()-a,tokensGenerated:o,success:!1,error:e?.message||String(e)})}},cancel(){i&&i()}})},Yt=(e,t,n,r,i,a,o=null,s=null,c=null)=>{let l,u=``,d=!1,f=Date.now(),p=0,m=()=>{i&&h(i).catch(()=>{}),c&&h(c).catch(()=>{})};return new tt({async start(h){try{let g=await e.parallel.completion(t,(e,t)=>{t&&(t.token&&(u+=t.token,p+=1),h.enqueue({event:`token`,data:{requestId:e,...t}}))}),{requestId:_}=g;l=g.stop;let v=await g.promise;v.text?u=v.text:v.content&&(u=v.content),d=!v.interrupted&&!v.context_full,console.log(`[Completion] Result:`,v),h.enqueue({event:`result`,data:{requestId:_,...v}}),h.close();let y=Date.now()-f,b=v.timings||{};H.addCompletion({id:`completion-${_}`,generatorId:o,requestId:_,repoId:s?.repoId||null,quantization:s?.quantization||null,variant:s?.variant||null,cacheTokens:b.cache_n??0,promptTokens:b.prompt_n??a??0,tokensGenerated:b.predicted_n??p,tokensPerSecond:b.predicted_per_second??0,promptPerSecond:b.prompt_per_second??0,durationMs:y,success:!0,interrupted:v.interrupted||!1,contextFull:v.context_full||v.contextFull||!1,usedCache:!!t.load_state_path}),d&&n.enabled&&u?n.saveCompletionState(r,u,i,a,c).catch(e=>{console.warn(`[SessionCache] Save failed:`,e.message)}):m()}catch(e){h.enqueue({event:`error`,data:{message:e?.message||String(e)}}),h.error(e),H.addCompletion({id:`completion-${Date.now()}`,generatorId:o,repoId:s?.repoId||null,quantization:s?.quantization||null,variant:s?.variant||null,durationMs:Date.now()-f,tokensGenerated:p,success:!1,error:e?.message||String(e)}),m()}},cancel(){l&&l(),m()}})},Xt=e=>{let t={model:e.plan.localPath,runtime:e.plan.info.runtime};return y(`sha256`).update(JSON.stringify(t)).digest(`hex`).slice(0,24)},Zt=async(e,t,n,r=null)=>{let{config:i,localPath:a,artifact:o}=e;if(e.localExists&&!t.has(a))return e.info.download.exists=!0,typeof n==`function`&&n(.5),a;if(i.model.local_path&&!i.model.allow_local_file)throw Error("Local model path provided but `model.allow_local_file` is not enabled");let s=a;if(r){let t=r.getDownload(s);if(t){console.log(`[ensureModelFile] Waiting for global download: ${o.repoId}`);try{if(await t,await K(a,o.size))return e.localExists=!0,e.info.download.exists=!0,typeof n==`function`&&n(.5),a}catch(e){console.warn(`[ensureModelFile] Global download failed, will retry: ${e.message}`)}}}t.has(s)||t.set(s,(async()=>{if(o.isSplit&&o.splitCount>0){let e=/-(\d{5})-of-(\d{5})\.gguf$/,t=_.dirname(a),r=o.splitCount,s=0;for(let a=1;a<=r;a+=1){let c=String(a).padStart(5,`0`),l=o.filename.replace(e,`-${c}-of-${String(r).padStart(5,`0`)}.gguf`),u=`${i.model.base_url.replace(/\/+$/,``)}/${o.repoId}/resolve/${o.revision}/${l}`,d=_.join(t,l);await K(d)||await Vt(u,o.headers,d,null,e=>{if(e>=0&&Number.isFinite(e)){let t=(s+e)/r,i=Math.round(t*100);console.log(`Downloading model splits: ${Math.min(100,i)}%`),typeof n==`function`&&n(t*.5)}}),s+=1}}else console.log(`Downloading model: 0%`),await Vt(o.url,o.headers,a,o.size,e=>{if(e>=0&&Number.isFinite(e)){let t=Math.round(e*100);console.log(`Downloading model: ${Math.min(100,t)}%`),typeof n==`function`&&n(e*.5)}});e.localExists=!0,e.info.download.exists=!0})());try{await t.get(s)}finally{t.delete(s)}return a},Qt=async(e,t,n,r=null)=>{let{mmprojArtifact:i,mmprojLocalPath:a}=e;if(!i||!a)return null;if(i.localPath){if(!await K(a))throw Error(`mmproj local file not found: ${a}`);return e.mmprojLocalExists=!0,e.info.multimodal.exists=!0,typeof n==`function`&&n(1),a}if(e.mmprojLocalExists&&!t.has(a))return e.info.multimodal.exists=!0,typeof n==`function`&&n(1),a;let o=a;if(r){let t=r.getDownload(o);if(t)try{if(await t,await K(a,i.size))return e.mmprojLocalExists=!0,e.info.multimodal.exists=!0,typeof n==`function`&&n(1),a}catch(e){console.warn(`[ensureMmprojFile] Global download failed, will retry: ${e.message}`)}}t.has(o)||t.set(o,(async()=>{console.log(`Downloading mmproj: 0%`),await Vt(i.url,i.headers,a,i.size,e=>{if(e>=0&&Number.isFinite(e)){let t=Math.round(e*100);console.log(`Downloading mmproj: ${Math.min(100,t)}%`),typeof n==`function`&&n(e)}}),e.mmprojLocalExists=!0,e.info.multimodal.exists=!0})());try{await t.get(o)}finally{t.delete(o)}return a},$t=async(e,t)=>{let n=Xt(e),r=e.contexts.get(n);if(r&&!r.released)return r.releaseTimer&&(clearTimeout(r.releaseTimer),r.releaseTimer=null,console.log(`[Context] Cancelled pending release for context "${n}"`)),r.releaseRequested=!1,r.refCount+=1,console.log(`[Context] Reusing existing context "${n}", refCount=${r.refCount}`),typeof t==`function`&&t(0),r.context||await r.ready,typeof t==`function`&&t(1),r;r?console.log(`[Context] Record exists but released=${r.released}, creating new context`):console.log(`[Context] No existing record for "${n}", creating new context`),r={key:n,refCount:1,ready:null,released:!1},e.contexts.set(n,r),r.ready=(async()=>{let i=Date.now(),a=await Zt(e.plan,e.downloads,t,e.globalDownloadManager);typeof t==`function`&&t(.5);let o={model:a,n_threads:e.plan.info.runtime.n_threads,use_mmap:e.plan.info.runtime.use_mmap,use_mlock:e.plan.info.runtime.use_mlock,no_extra_bufts:e.plan.info.runtime.no_extra_bufts,cpu_mask:e.plan.info.runtime.cpu_mask,cpu_strict:e.plan.info.runtime.cpu_strict,devices:e.plan.info.runtime.devices,n_ctx:e.plan.info.runtime.n_ctx,n_gpu_layers:e.plan.info.runtime.n_gpu_layers,n_parallel:e.plan.info.runtime.n_parallel,n_batch:e.plan.info.runtime.n_batch,n_ubatch:e.plan.info.runtime.n_ubatch,n_cpu_moe:e.plan.info.runtime.n_cpu_moe,flash_attn_type:e.plan.info.runtime.flash_attn_type,ctx_shift:e.plan.info.runtime.ctx_shift,kv_unified:e.plan.info.runtime.kv_unified,swa_full:e.plan.info.runtime.swa_full,lib_variant:e.plan.info.runtime.variant};e.plan.info.runtime.flash_attn_type!==`off`&&(o.cache_type_k=e.plan.info.runtime.cache_type_k,o.cache_type_v=e.plan.info.runtime.cache_type_v),console.log(`[Context] Load Options:`,o);let s;try{if(s=await C(o,e=>{typeof t==`function`&&(t(.5+e*.25),e%5==0&&console.log(`[Context] Load Model Progress:`,e))}),e.plan.info.runtime.n_parallel&&!await s.parallel.enable({n_parallel:e.plan.info.runtime.n_parallel,n_batch:e.plan.info.runtime.n_batch}))throw Error(`Failed to enable parallel decoding mode for context`);if(e.plan.mmprojArtifact){let t=await Qt(e.plan,e.downloads,null,e.globalDownloadManager);if(t){let n=e.config.model.mmproj_use_gpu,r={path:t,use_gpu:n==null?(e.plan.info.runtime.n_gpu_layers||0)>0:!!n,image_min_tokens:e.config.model.mmproj_image_min_tokens,image_max_tokens:e.config.model.mmproj_image_max_tokens};console.log(`[Context] initMultimodal:`,r),await s.initMultimodal(r)?e.plan.info.multimodal.initialized=!0:console.warn(`[Context] initMultimodal returned false; multimodal disabled`)}}return typeof t==`function`&&t(1),r.context=s,r.modelInfo=s.getModelInfo(),H.addModelLoad({id:`${e.id}-${n}`,generatorId:e.id,contextKey:n,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,variant:e.plan.info.runtime?.variant||null,nCtx:e.plan.info.runtime?.n_ctx||null,nGpuLayers:e.plan.info.runtime?.n_gpu_layers||null,durationMs:Date.now()-i,success:!0}),r}catch(t){if(H.addModelLoad({id:`${e.id}-${n}`,generatorId:e.id,contextKey:n,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,variant:e.plan.info.runtime?.variant||null,durationMs:Date.now()-i,success:!1,error:t?.message||String(t)}),s)try{s.release()}catch{}throw t}})();try{return await r.ready,r}catch(t){throw e.contexts.delete(n),t}},en=async(e,t,n=!1)=>{if(t.released||!n&&t.refCount>0)return!1;t.released=!0,e.contexts.delete(t.key);try{t.context?.parallel?.disable?.()}catch{}return await t.context?.release?.(),!0},tn=async(e,t,n=!1)=>{if(t.releaseRequested=!0,t.releaseTimer&&=(clearTimeout(t.releaseTimer),null),n)t.refCount=0;else if(t.refCount=Math.max(0,t.refCount-1),t.refCount>0)return t.releaseRequested=!1,!1;let r=e.config.runtime.context_release_delay_ms;if(typeof r!=`number`||!Number.isFinite(r))return en(e,t);let i=Math.max(0,Math.floor(r));return n||i<=0?en(e,t):(console.log(`[Context] Scheduling release in ${i}ms for context "${t.key}"`),t.releaseTimer=setTimeout(async()=>{if(t.releaseTimer=null,t.refCount>0){console.log(`[Context] Release cancelled, refCount=${t.refCount} for context "${t.key}"`),t.releaseRequested=!1;return}console.log(`[Context] Releasing context "${t.key}" after ${i}ms delay`),await en(e,t)},i),!0)};async function nn(e,t,n={}){let{globalDownloadManager:r=null}=n,i=dt(t),a=await qt(i),o=new It(i,a);await o.initialize();let s={id:e,type:`ggml-llm`,config:i,plan:a,info:a.info,contexts:new Map,downloads:new Map,globalDownloadManager:r,sessionCache:o,finalized:!1};return{id:e,type:`ggml-llm`,info:a.info,contexts:s.contexts,initContext:async(e={})=>{let{onProgress:t}=e,n=await $t(s,t);return s.sessionCache.updateModelInfo(n.modelInfo),{modelInfo:n.modelInfo?{...n.modelInfo}:null,runtime:{...s.plan.info.runtime},download:{...s.plan.info.download},multimodal:s.plan.info.multimodal?{...s.plan.info.multimodal}:null}},completion:async(e={})=>{let{options:t={},useCache:n=!0}=e,r=Xt(s),i=s.contexts.get(r);if(!i)throw Error(`Context "${r}" not initialized`);await i.ready;let a=t.prompt||``,o=null,c=null;if(!a&&t.messages){({messages:o}=t),c={chatTemplate:t.chat_template||t.chatTemplate,jinja:t.jinja??!0,tools:t.tools,parallel_tool_calls:t.parallel_tool_calls,tool_choice:t.tool_choice,reasoning_format:t.reasoning_format,enable_thinking:t.enable_thinking,add_generation_prompt:t.add_generation_prompt,now:t.now,chat_template_kwargs:t.chat_template_kwargs,force_pure_content:t.force_pure_content};let e=await i.context.getFormattedChat(o,c.chatTemplate,c);a=e?.prompt||e||``}if(n&&s.sessionCache.enabled&&a){let{options:e}=await s.sessionCache.prepareCompletionOptions(t,a,i.context),n=await s.sessionCache.generateTempStatePath(),r=(await i.context.tokenize(a))?.tokens?.length||0,o={...e,save_state_path:n},c=s.sessionCache.requiresExactMatch(),l=!!o.load_state_path,u=null;c&&!l&&(u=await s.sessionCache.generateTempStatePath(),o.save_prompt_state_path=u);let d={repoId:s.plan.info.model?.repoId||null,quantization:s.plan.info.model?.quantization||null,variant:s.plan.info.runtime?.variant||null};return Yt(i.context,o,s.sessionCache,a,n,r,s.id,d,u)}let l={repoId:s.plan.info.model?.repoId||null,quantization:s.plan.info.model?.quantization||null,variant:s.plan.info.runtime?.variant||null};return Jt(i.context,t,s.id,l)},tokenize:async(e={})=>{let{text:t=``,params:n={}}=e,r=Xt(s),i=s.contexts.get(r);if(!i)throw Error(`Context "${r}" not initialized`);await i.ready;let a=await i.context.tokenize(t,n);if(!a)return{tokens:[]};let o=Array.from(a.tokens??[],Number);return{...a,tokens:o}},detokenize:async(e={})=>{let{tokens:t=[]}=e,n=Xt(s),r=s.contexts.get(n);if(!r)throw Error(`Context "${n}" not initialized`);await r.ready;let i=t.map(e=>Number(e));return r.context.detokenize(i)},applyChatTemplate:async(e={})=>{let{messages:t=[],template:n,params:r}=e,i=Xt(s),a=s.contexts.get(i);if(!a)throw Error(`Context "${i}" not initialized`);return await a.ready,await a.context.getFormattedChat(t,n,r)},releaseContext:async()=>{if(s.finalized)return!1;let e=Xt(s),t=s.contexts.get(e);return t?tn(s,t,!1):!1},finalize:async()=>{if(s.finalized)return;s.finalized=!0;let e=Array.from(s.contexts.values()),t=e.map(e=>e.released||e.releaseRequested||e.releaseTimer||(e.refCount=Math.max(0,e.refCount-1),e.refCount>0)?Promise.resolve(!1):en(s,e));await Promise.allSettled(t),(e.length===0||e.every(e=>e.released))&&await s.sessionCache.cleanup()},getStatus:()=>{let e=[],t=Array.from(s.contexts.entries()).map(([t,n])=>{let r={key:t,refCount:n.refCount,hasModel:!!n.context},i=n.context.parallel.getStatus();return r.parallelStatus=i,e.push({contextKey:t,...i}),r});return{id:s.id,type:s.type,repoId:s.plan.info.model?.repoId||null,quantization:s.plan.info.model?.quantization||null,variant:s.plan.info.runtime?.variant||null,nCtx:s.plan.info.runtime?.n_ctx||null,nParallel:s.plan.info.runtime?.n_parallel||null,contexts:t,parallelStatuses:e}},subscribeParallelStatus:e=>{let t=Array.from(s.contexts.entries()).map(([t,n])=>n.context.parallel.subscribeToStatus(n=>{e({contextKey:t,...n})}));return{remove:()=>{t.forEach(e=>{e?.remove&&e.remove()})}}},hasPendingReleases:()=>Array.from(s.contexts.values()).some(e=>!e.released&&(e.releaseRequested||e.releaseTimer||e.refCount>0)),resetFinalized:()=>{s.finalized=!1}}}const rn=e=>{let t=dt(e);return t.model.repo_id||t.model.repository||t.model.model||null};async function an(e,t,n={}){let{onProgress:r,onComplete:i,onError:a}=n;try{let n=dt(e),o=await Ht(n),s=Bt(n,o),{repoId:c}=o,l=await Wt(n,o).catch(e=>(console.warn(`[Download] Failed to derive mmproj artifact: ${e.message}`),null)),u=Gt(n,l),d=async()=>{if(!l||!u||l.localPath)return;if(await K(u,l.size)){console.log(`[Download] mmproj already exists: ${u}`);return}let e=t.getDownload(u);if(e){await e;return}let n=(async()=>{try{await Vt(l.url,l.headers,u,l.size,e=>{e>=0&&Number.isFinite(e)&&console.log(`[Download] mmproj ${c}: ${Math.round(e*100)}%`)})}finally{t.deleteDownload(u)}})();t.setDownload(u,n),await n};if(await K(s,o.size))return console.log(`[Download] Model already exists: ${c} at ${s}`),await d().catch(e=>{console.error(`[Download] mmproj download failed: ${e.message}`),typeof a==`function`&&a(e)}),typeof i==`function`&&i({localPath:s,repoId:c,alreadyExists:!0}),{started:!1,localPath:s,repoId:c,alreadyExists:!0};let f=t.getDownload(s);if(f)return console.log(`[Download] Already downloading: ${c}`),f.then(()=>{typeof i==`function`&&i({localPath:s,repoId:c,joinedExisting:!0})}).catch(e=>{typeof a==`function`&&a(e)}),{started:!1,localPath:s,repoId:c,alreadyDownloading:!0};console.log(`[Download] Starting download: ${c}`);let p=(async()=>{try{if(o.isSplit&&o.splitCount>0){let e=/-(\d{5})-of-(\d{5})\.gguf$/,t=_.dirname(s),i=o.splitCount,a=0;for(let s=1;s<=i;s+=1){let l=String(s).padStart(5,`0`),u=o.filename.replace(e,`-${l}-of-${String(i).padStart(5,`0`)}.gguf`),d=`${n.model.base_url.replace(/\/+$/,``)}/${o.repoId}/resolve/${o.revision}/${u}`,f=_.join(t,u);await K(f)||await Vt(d,o.headers,f,null,e=>{if(e>=0&&Number.isFinite(e)){let t=(a+e)/i;console.log(`[Download] ${c}: ${Math.round(t*100)}%`),typeof r==`function`&&r(t)}}),a+=1}}else await Vt(o.url,o.headers,s,o.size,e=>{e>=0&&Number.isFinite(e)&&(console.log(`[Download] ${c}: ${Math.round(e*100)}%`),typeof r==`function`&&r(e))});await d(),console.log(`[Download] Completed: ${c}`),typeof i==`function`&&i({localPath:s,repoId:c})}catch(e){throw console.error(`[Download] Failed: ${c}`,e.message),typeof a==`function`&&a(e),e}finally{t.deleteDownload(s)}})();return t.setDownload(s,p),{started:!0,localPath:s,repoId:c}}catch(e){return console.error(`[Download] Failed to start download:`,e.message),typeof a==`function`&&a(e),{started:!1,localPath:null,repoId:null,error:e.message}}}async function on(e){let t=dt(e),n=await Ht(t),r=await zt(n.url,n.headers,t.runtime.cache_dir),{arch:i,nCtxTrain:a,nLayer:o,nEmbd:s,nHead:c,nHeadKv:l,nEmbdHeadK:u,nEmbdHeadV:d,quantVersion:f,fileType:p,attentionLayerCount:m,recurrentLayerCount:h,ssmDConv:g,ssmDState:_,ssmDInner:y,ssmNGroup:b,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C}=He(r),w=Number.isFinite(Number(o))?Number(o):0,T=Number.isFinite(Number(s))?Number(s):0,E=Number.isFinite(Number(c))?Number(c):0,D=Number.isFinite(Number(l))?Number(l):E,ee=E>0&&T>0?T/E:128,O=u!=null&&Number.isFinite(Number(u))?Number(u):ee,te=d!=null&&Number.isFinite(Number(d))?Number(d):ee,k=Fe({arch:i,metadata:r,nLayer:w}),A=k&&Number.isFinite(Number(k.kvLayers))?Number(k.kvLayers):w,ne=Math.max(0,Math.floor(Number(A)||0)),j=(t.model.n_ctx?Number(t.model.n_ctx):null)||a||4096,M={k:t.model.cache_type_k,v:t.model.cache_type_v},N=n.size>0?n.size:0,P=t.model.n_parallel||4,F=Ue({layerCount:ne,headKvCount:D,embdHeadKCount:O,embdHeadVCount:te,cacheTypes:M,swaConfig:k,kvUnified:t.model.kv_unified,nParallel:P,swaFull:t.model.swa_full,arch:i,attentionLayerCount:m}),re=We({nLayer:w,nEmbd:T,recurrentLayerCount:h,nSeqMax:P,ssmDConv:g,ssmDState:_,ssmDInner:y,ssmNGroup:b,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C,arch:i}),I=t.backend?.gpu_memory_fraction==null?ct.backend.gpu_memory_fraction||1:Math.min(1,Math.max(0,Number(t.backend.gpu_memory_fraction))),L=t.backend?.cpu_memory_fraction==null?st:Math.min(1,Math.max(0,Number(t.backend.cpu_memory_fraction))),ie=await Kt(t,{modelBytes:N,kvCacheBytes:F(j)}),ae=(ie.selected.totalMemory||0)*I,R=Math.max(0,v.totalmem()*L),oe=Ge({maxCtx:j,availableMemory:ie.selected.hasGpu?ae:R,modelBytes:N,kvBytesForCtx:F}),se=F(j),ce=F(oe);return{kvInfo:{nCtxTrain:a,nLayer:w,nEmbd:T,nHeadKv:D,nEmbdHeadK:O,nEmbdHeadV:te,nHeadCount:E,nHeadKvCount:D,kvLayerCount:ne,swa:k?.enabled?{window:k.window,pattern:k.pattern,denseFirst:k.denseFirst,type:k.type,layers:k.swaLayers}:null},modelBytes:N,kvCacheBytes:se,limitedKvCacheBytes:ce,memoryLimitedCtx:oe,recurrentMemoryBytes:re,quantization:{name:n.quantization||null,fileType:p,version:f}}}const sn=e=>e?typeof e.score==`number`&&Number.isFinite(e.score)?Number(e.score):ce(e):0;async function cn(e=null,t={}){let{threshold:n=1.1,includeBreakdown:r=!1,config:i,...a}=t,o=null,s=null,c=null,l=null,u=null,d=null,f=null;if(i)try{let{modelBytes:e,kvCacheBytes:t,limitedKvCacheBytes:n,memoryLimitedCtx:r,recurrentMemoryBytes:a,kvInfo:p,quantization:m}=await on(i);o=e,s=t,c=n,l=r,u=a,d=p,f=m}catch{}let p=i?.backend?.gpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.gpu_memory_fraction))),m=i?.backend?.cpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.cpu_memory_fraction))),h=await Te({...a,platform:process.platform,totalMemoryInBytes:v.totalmem(),backend:`ggml-llm`,includeBreakdown:r,gpuMemoryFraction:p,cpuMemoryFraction:m,dependencies:{getBackendDevicesInfo:x,isLibVariantAvailable:S},modelBytes:o,kvCacheBytes:s,limitedKvCacheBytes:c}),g=h.selected,_=sn(g);g.modelBytes=o||null,g.kvCacheBytes=s||null,g.memoryLimitedCtx=l||null,g.limitedKvCacheBytes=c||null,g.recurrentMemoryBytes=u||null,g.kvInfo=d||null,g.quantization=f||null;let y=null,b=null;if(e){let t=sn(e);b={...e,score:t};let r=`buttress`,i=`buttress-higher-score`;if(!h.ok)r=`local`,i=`buttress-unavailable`;else if(!t&&t!==0)r=`buttress`,i=`missing-client-score`;else{let e=b.fit,a=b.limitedFit,o=g?.fit,s=g?.limitedFit,c=e?.fitsInGpu||e?.fitsInCpu||a?.fitsInGpu||a?.fitsInCpu,l=o?.fitsInGpu||o?.fitsInCpu||s?.fitsInGpu||s?.fitsInCpu;c&&!l?(r=`local`,i=`client-fits-in-memory`):l&&!c?(r=`buttress`,i=`buttress-fits-in-memory`):t>_*n?(r=`local`,i=`client-better`):_>t*n?(r=`buttress`,i=`buttress-better`):(r=`either`,i=`comparable-scores`)}y={buttressScore:_,clientScore:t,threshold:n,recommendation:r,reason:i}}!h.ok&&!y&&(y={buttressScore:_,clientScore:e?.score??null,threshold:n,recommendation:`local`,reason:`buttress-unavailable`});let C=null;return i&&(C={repoId:i.model?.repo_id||null,quantization:i.model?.quantization||null,nCtx:i.model?.n_ctx||null,cacheKType:i.model?.cache_type_k||`f16`,cacheVType:i.model?.cache_type_v||`f16`}),{type:`ggml-llm`,timestamp:new Date().toISOString(),buttress:h,client:b,comparison:y,modelConfig:C}}const{WritableStream:ln}=typeof globalThis<`u`&&globalThis.ReadableStream&&globalThis.WritableStream?{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream}:o,un=(e={},t={})=>(Object.entries(t||{}).forEach(([t,n])=>{n&&typeof n==`object`&&!Array.isArray(n)?((!e[t]||typeof e[t]!=`object`)&&(e[t]={}),un(e[t],n)):e[t]=n}),e),dn=`https://huggingface.co`,fn=`https://huggingface.co/api`,pn=_.join(v.homedir(),`.buttress`,`models`),mn=[`cuda`,`vulkan`,`default`],hn=[`q8_0`,`q5_1`,`q5_0`,`q4_1`,`q4_0`],gn=`fp16`,_n=.5,vn=[`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`],yn=e=>{if(!e)return null;let t=e.toLowerCase();return vn.find(e=>t.includes(e))||null},bn={backend:{type:`ggml-stt`,variant:null,variant_preference:mn,gpu_memory_fraction:.85,cpu_memory_fraction:_n},model:{repo_id:`BricksDisplay/whisper-ggml`,revision:`main`,filename:null,url:null,quantization:null,preferred_quantizations:[`q8_0`,gn,`q5_1`],allow_local_file:!1,local_path:null,api_base:fn,base_url:dn,use_gpu:!0,use_flash_attn:`auto`},runtime:{cache_dir:pn,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},max_threads:null,context_release_delay_ms:1e4}},xn=(e,t=[])=>!e&&e!==0?[...t]:Array.isArray(e)?e.filter(e=>e!=null):[e],Sn=e=>{if(!e)return null;let t=String(e).toLowerCase();return[`cuda`,`vulkan`,`default`].includes(t)?t:null},Cn=(e={})=>{let t=structuredClone(bn);if(un(t,e),t.backend.variant=Sn(t.backend.variant),t.backend.variant_preference=Array.from(new Set(xn(t.backend.variant_preference||mn).flatMap(e=>{let t=Sn(e);return t?[t]:[]}))),t.backend.variant_preference.length===0&&(t.backend.variant_preference=[...mn]),t.runtime.prefer_variants=Array.from(new Set(xn(t.runtime.prefer_variants).flatMap(e=>{let t=Sn(e);return t?[t]:[]}))),t.model.preferred_quantizations=Array.from(new Set(xn(t.model.preferred_quantizations||t.model.quantizations).flatMap(e=>{let t=e?String(e).toLowerCase():null;return t?[t]:[]}))),t.model.quantization){let e=String(t.model.quantization).toLowerCase();t.model.preferred_quantizations.includes(e)||t.model.preferred_quantizations.unshift(e)}return t.model.base_url=t.model.base_url||dn,t.model.api_base=t.model.api_base||fn,t.runtime.cache_dir=t.runtime.cache_dir?_.resolve(t.runtime.cache_dir):pn,t.runtime.context_release_delay_ms=Math.max(0,Number(t.runtime.context_release_delay_ms)||bn.runtime.context_release_delay_ms),t},wn=e=>{let t=e.toLowerCase();return hn.find(e=>t.includes(e))||null},Tn=e=>{let t=[];e.backend.variant&&t.push(e.backend.variant),e.runtime.prefer_variants.length>0&&t.push(...e.runtime.prefer_variants),t.push(...e.backend.variant_preference),t.push(`default`);let n=new Set;for(let e of t){let t=Sn(e);t&&n.add(t)}return Array.from(n)},En=async e=>{await l(e,{recursive:!0})},Dn=(e=pn)=>_.join(e,`.metadata-cache`),On=(e,t,n=pn)=>{let r=y(`sha256`).update(e).digest(`hex`);return _.join(Dn(n),t,`${r}.json`)},kn=async(e,t,n=pn)=>{try{let r=await d(On(e,t,n),`utf-8`);return JSON.parse(r)}catch{return null}},An=async(e,t,n,r=pn)=>{try{let i=On(e,t,r);await En(_.dirname(i)),await g(i,JSON.stringify(n),`utf-8`)}catch{}},jn=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,t);if(!n.ok){let t=await n.text().catch(()=>``);throw Error(`Failed to fetch ${e}: ${n.status} ${n.statusText} ${t}`.trim())}return n.json()},Mn=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,{...t,method:`HEAD`});if(!n.ok)throw Error(`Failed to fetch headers for ${e}: ${n.status} ${n.statusText}`);return n},Nn=(e,t)=>{if(e.model.local_path)return _.resolve(e.model.local_path);let n=t.repoId.split(`/`),r=_.join(e.runtime.cache_dir,...n,t.revision);return _.join(r,t.filename)},Pn=async(e,t)=>{try{let n=await m(e);return t?n.size===t:!0}catch{return!1}},Fn=async(e,t,n,r,i)=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);await En(_.dirname(n));let a=await fetch(e,{headers:t});if(!a.ok||!a.body)throw Error(`Failed to download ${e}: ${a.status} ${a.statusText}`);let o=await u(n,`w`),s=Number(a.headers.get(`content-length`))||r||0,c=0,l=.05;try{await a.body.pipeTo(new ln({async write(e){if(await o.write(e),c+=e.byteLength,typeof i==`function`&&s>0){let e=Math.min(1,c/s);for(;e>=l;)i(l),l+=.05}},async close(){await o.close(),typeof i==`function`&&i(1)},async abort(e){throw await o.close().catch(()=>{}),await h(n).catch(()=>{}),e}}))}catch(e){throw await o.close().catch(()=>{}),await h(n).catch(()=>{}),e}if(r){let e=await m(n);if(e.size!==r)throw await h(n).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${r} got ${e.size}`)}},In=async e=>{let t=e.model.repo_id||e.model.repository||e.model.model;if(!t)throw Error("`model.repo_id` is required in Buttress backend config");let n=e.model.revision||`main`,r=e.runtime.cache_dir,i=JSON.stringify({repoId:t,revision:n,filename:e.model.filename,url:e.model.url,quantization:e.model.quantization,preferred_quantizations:e.model.preferred_quantizations}),a=await kn(i,`artifact-info`,r);if(a)return a;let o={...e.runtime.http_headers||{}};if(e.runtime.huggingface_token&&(o.Authorization=`Bearer ${e.runtime.huggingface_token}`),e.model.url){let a=await Mn(e.model.url,{headers:o}),s=Number(a.headers.get(`content-length`))||null,c=e.model.filename||e.model.url.split(`/`).pop(),l={repoId:t,revision:n,filename:c,url:e.model.url,size:s,quantization:wn(c||``),headers:o};return await An(i,`artifact-info`,l,r),l}let{filename:s}=e.model,c=e.model.quantization&&String(e.model.quantization).toLowerCase(),l=await jn(`${e.model.api_base}/models/${t}?revision=${n}&blobs=true`,{headers:o}),u=(l?.siblings||l?.files||[]).map(e=>e.rfilename||e.path||e.filename).filter(e=>typeof e==`string`&&e.endsWith(`.bin`));if(u.length===0)throw Error(`No model artifacts found in repo ${t}`);let d=e.model.preferred_quantizations.length>0?e.model.preferred_quantizations:hn,f=()=>{for(let e of d)if(e===gn){let e=u.find(e=>{let t=e.toLowerCase();return!hn.some(e=>t.includes(e))});if(e)return{filename:e,quantization:null}}else{let t=u.find(t=>t.toLowerCase().includes(e));if(t)return{filename:t,quantization:e}}return null};if(s)c||=wn(s);else{let{filename:e,quantization:t}=f()||{filename:u[0],quantization:null};s=e,c=t||wn(s)}let p=`${e.model.base_url.replace(/\/+$/,``)}/${t}/resolve/${n}/${s}`,m=await Mn(p,{headers:o}),h=Number(m.headers.get(`content-length`))||null,g={repoId:t,revision:n,filename:s,url:p,size:h,quantization:c,headers:o,isSplit:!1,splitCount:0};return await An(i,`artifact-info`,g,r),g},Ln=async(e,{modelBytes:t=null,processingBytes:n=null}={})=>{let r=Tn(e),[i,...a]=r,o=e.backend?.gpu_memory_fraction==null?bn.backend.gpu_memory_fraction||1:Math.min(1,Math.max(0,Number(e.backend.gpu_memory_fraction))),s=e.backend?.cpu_memory_fraction==null?_n:Math.min(1,Math.max(0,Number(e.backend.cpu_memory_fraction))),c=await Te({platform:process.platform,totalMemoryInBytes:v.totalmem(),backend:`ggml-stt`,variant:i||null,preferVariants:a,variantPreference:e.backend.variant_preference,gpuMemoryFraction:o,cpuMemoryFraction:s,dependencies:{getBackendDevicesInfo:x,isLibVariantAvailable:S},modelBytes:t,kvCacheBytes:n}),l=e=>({...e,devices:Array.isArray(e.devices)?e.devices:[],ok:e.ok,hasGpu:!!e.hasGpu,totalMemory:e.gpuTotalBytes||e.totalMemory||0,error:e.ok?null:Error(e.error||`Variant ${e.variant} not available on this platform`)});if(!c.ok||!c.selected){let e=(c.attempts||[]).map(e=>`${e.variant}: ${e.error||`unknown error`}`).join(`; `);throw Error(`Unable to initialize any backend variant (${r.join(`, `)}). Errors: ${e}`)}let u=(c.attempts||[]).map(l);return{selected:l(c.selected),attempts:u}},Rn=async e=>{let t=await In(e),n=ke({modelBytes:t.size>0?t.size:0}),r=await Ln(e,{modelBytes:n.modelBytes,processingBytes:n.processingBufferBytes}),i=r.selected.hasGpu&&(r.selected.fit?.fitsInGpu===void 0?!0:r.selected.fit.fitsInGpu);e.model.use_gpu===!1&&(i=!1);let a=e.model.use_flash_attn&&String(e.model.use_flash_attn).toLowerCase(),o;o=a===`on`||a===`true`?!0:a===`off`||a===`false`?!1:i;let s=e.runtime.cache_dir,c=Nn(e,t),l=await Pn(c,t.size);return{config:e,info:{ok:!0,backend:`ggml-stt`,model:{repoId:t.repoId,revision:t.revision,filename:t.filename,quantization:t.quantization,modelType:yn(t.filename),url:t.url,sizeBytes:t.size},runtime:{variant:r.selected.variant,use_gpu:i,use_flash_attn:o,max_threads:e.runtime.max_threads?Number(e.runtime.max_threads):null},resources:{...n,gpuCapacityBytes:r.selected.gpuTotalBytes,gpuUsableBytes:r.selected.gpuUsableBytes,cpuUsableBytes:r.selected.cpuUsableBytes,fit:r.selected.fit},devices:{selected:r.selected,attempts:r.attempts},download:{cacheDir:s,localPath:c,exists:l},timestamp:new Date().toISOString()},artifact:t,memory:n,devices:r,localPath:c,localExists:l}},zn=async(e,t,n,r=null)=>{let{localPath:i,artifact:a,config:o}=e;if(e.localExists)return typeof n==`function`&&n(1),i;if(r){let t=r.getDownload(i);if(t){console.log(`[ensureModelFile] Waiting for global STT download: ${a.repoId}`);try{if(await t,await Pn(i,a.size))return e.localExists=!0,e.info.download.exists=!0,typeof n==`function`&&n(1),i}catch(e){console.warn(`[ensureModelFile] Global STT download failed, will retry: ${e.message}`)}}}let s=t.get(i);if(s)return await s,typeof n==`function`&&n(1),i;let c=(async()=>{if(o.model.allow_local_file){if(!await Pn(i,a.size))throw Error(`Local model file not found: ${i}`);return i}return await Fn(a.url,a.headers,i,a.size,n),i})();t.set(i,c);try{return await c,i}finally{t.delete(i)}};var Bn=class{constructor(){this.queue=[],this.processing=!1,this.currentTaskId=null}async enqueue(e,t=null){return new Promise((n,r)=>{this.queue.push({task:e,resolve:n,reject:r,taskId:t}),this.processNext()})}async processNext(){if(this.processing||this.queue.length===0)return;this.processing=!0;let{task:e,resolve:t,reject:n,taskId:r}=this.queue.shift();this.currentTaskId=r;try{t(await e())}catch(e){n(e)}finally{this.processing=!1,this.currentTaskId=null,this.processNext()}}getStatus(){return{processing:this.processing,queuedCount:this.queue.length,currentTaskId:this.currentTaskId}}};const Vn=e=>{if(!e)return null;if(e instanceof ArrayBuffer)return e;if(ArrayBuffer.isView(e))return e.buffer;if(typeof e==`string`){let t=e.startsWith(`data:`)?e.split(`,`)[1]||``:e,n=Buffer.from(t,`base64`);return n.buffer.slice(n.byteOffset,n.byteOffset+n.byteLength)}throw Error(`Unsupported audioData format, expected base64 string or ArrayBuffer`)},Hn=async(e,t)=>{if(e.contextRecord&&!e.contextRecord.released)return e.contextRecord.releaseTimer&&(clearTimeout(e.contextRecord.releaseTimer),e.contextRecord.releaseTimer=null,console.log(`[Context] Cancelled pending STT release`)),e.contextRecord.releaseRequested=!1,e.contextRecord.refCount+=1,console.log(`[Context] Reusing existing STT context, refCount=${e.contextRecord.refCount}`),typeof t==`function`&&t(0),e.contextRecord.context||await e.contextRecord.ready,typeof t==`function`&&t(1),e.contextRecord;e.contextRecord?console.log(`[Context] STT record exists but released=${e.contextRecord.released}, creating new context`):console.log(`[Context] No existing STT record, creating new context`);let n={refCount:1,ready:null,released:!1};e.contextRecord=n,n.ready=(async()=>{let r=Date.now();try{typeof t==`function`&&t(0);let i=await zn(e.plan,e.downloads,t,e.globalDownloadManager);typeof t==`function`&&t(.5);let a=await E({filePath:i,useFlashAttn:e.plan.info.runtime.flash_attn_type===`on`,useGpu:e.plan.info.runtime.n_gpu_layers>0,nThreads:e.plan.info.runtime.n_threads},e.plan.info.runtime.variant);typeof t==`function`&&t(1),n.context=a;try{n.modelInfo=a.getModelInfo()}catch{n.modelInfo=null}return U.addModelLoad({id:e.id,generatorId:e.id,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,modelType:e.plan.info.model?.modelType||null,variant:e.plan.info.runtime?.variant||null,useGpu:e.plan.info.runtime?.use_gpu||!1,durationMs:Date.now()-r,success:!0}),n}catch(t){throw U.addModelLoad({id:e.id,generatorId:e.id,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,modelType:e.plan.info.model?.modelType||null,variant:e.plan.info.runtime?.variant||null,durationMs:Date.now()-r,success:!1,error:t?.message||String(t)}),t}})();try{return await n.ready,typeof t==`function`&&t(1),n}catch(t){throw e.contextRecord=null,t}},Un=async(e,t,n=!1)=>t.released||!n&&t.refCount>0?!1:(t.released=!0,e.contextRecord=null,await t.context?.release?.(),!0),Wn=async(e,t,n=!1)=>{if(t.releaseRequested=!0,t.releaseTimer&&=(clearTimeout(t.releaseTimer),null),n)t.refCount=0;else if(t.refCount=Math.max(0,t.refCount-1),t.refCount>0)return t.releaseRequested=!1,!1;let r=e.config.runtime.context_release_delay_ms;if(typeof r!=`number`||!Number.isFinite(r))return Un(e,t);let i=Math.max(0,Math.floor(r));return n||i<=0?Un(e,t):(console.log(`[Context] Scheduling STT release in ${i}ms`),t.releaseTimer=setTimeout(async()=>{if(t.releaseTimer=null,t.refCount>0){console.log(`[Context] STT release cancelled, refCount=${t.refCount}`),t.releaseRequested=!1;return}console.log(`[Context] Releasing STT context after ${i}ms delay`),await Un(e,t)},i),!0)};async function Gn(e,t,n={}){let{globalDownloadManager:r=null}=n,i=Cn(t),a=await Rn(i),o={id:e,type:`ggml-stt`,config:i,plan:a,info:a.info,contextRecord:null,downloads:new Map,globalDownloadManager:r,queue:new Bn,finalized:!1},s=async()=>{if(o.finalized)return;o.finalized=!0;let e=o.contextRecord;e&&(e.released||e.releaseRequested||e.releaseTimer||(e.refCount=Math.max(0,e.refCount-1),!(e.refCount>0)&&await Un(o,e)))},c=async(e={})=>{let{onProgress:t}=e;try{let e=await Hn(o,t);return{modelInfo:e.modelInfo&&typeof e.modelInfo==`object`?{...e.modelInfo}:null,runtime:{...o.plan.info.runtime},download:{...o.plan.info.download}}}catch(e){throw console.error(`[Context] Error initializing context:`,e),e}},l=async()=>{if(o.finalized)return!1;let e=o.contextRecord;return e?Wn(o,e):!1},u=async(e={})=>{let{audioPath:t,audioData:n,options:r={}}=e,i=o.contextRecord;if(!i)throw Error(`Context not initialized`);let a={...r};o.plan.info.runtime.max_threads&&a.maxThreads==null&&(a.maxThreads=o.plan.info.runtime.max_threads);let s=`transcription-${Date.now()}-${Math.random().toString(36).slice(2,8)}`,c=Date.now();return o.queue.enqueue(async()=>{await i.ready;try{let e;if(n){let t=Vn(n),{promise:r}=i.context.transcribeData(t,a);e=await r}else{if(!t)throw Error(`audioPath or audioData is required for transcription`);let n=_.resolve(t),{promise:r}=i.context.transcribe(n,a);e=await r}return U.addTranscription({id:s,generatorId:o.id,repoId:o.plan.info.model?.repoId||null,quantization:o.plan.info.model?.quantization||null,modelType:o.plan.info.model?.modelType||null,variant:o.plan.info.runtime?.variant||null,durationMs:Date.now()-c,segmentCount:e?.segments?.length||0,textLength:e?.text?.length||0,success:!0}),e}catch(e){throw U.addTranscription({id:s,generatorId:o.id,repoId:o.plan.info.model?.repoId||null,quantization:o.plan.info.model?.quantization||null,modelType:o.plan.info.model?.modelType||null,variant:o.plan.info.runtime?.variant||null,durationMs:Date.now()-c,success:!1,error:e?.message||String(e)}),e}},s)};return{id:e,type:`ggml-stt`,info:a.info,queue:o.queue,initContext:c,transcribe:async(e={})=>u(e),transcribeData:async(e={})=>u(e),releaseContext:l,finalize:s,getStatus:()=>({id:o.id,type:o.type,repoId:o.plan.info.model?.repoId||null,quantization:o.plan.info.model?.quantization||null,modelType:o.plan.info.model?.modelType||null,variant:o.plan.info.runtime?.variant||null,hasContext:!!o.contextRecord?.context,contextRefCount:o.contextRecord?.refCount||0,queueStatus:o.queue.getStatus()}),hasPendingReleases:()=>{let e=o.contextRecord;return e?!e.released&&(e.releaseRequested||e.releaseTimer||e.refCount>0):!1},resetFinalized:()=>{o.finalized=!1}}}const Kn=e=>{let t=Cn(e),n=t.model.repo_id||t.model.repository||t.model.model||null;if(!n)return null;let r=yn(t.model.filename);return r?`${n}:${r}`:n};async function qn(e,t,n={}){let{onProgress:r,onComplete:i,onError:a}=n;try{let n=Cn(e),o=await In(n),s=Nn(n,o),{repoId:c}=o;if(await Pn(s,o.size))return console.log(`[Download] STT model already exists: ${c} at ${s}`),typeof i==`function`&&i({localPath:s,repoId:c,alreadyExists:!0}),{started:!1,localPath:s,repoId:c,alreadyExists:!0};let l=t.getDownload(s);if(l)return console.log(`[Download] Already downloading STT model: ${c}`),l.then(()=>{typeof i==`function`&&i({localPath:s,repoId:c,joinedExisting:!0})}).catch(e=>{typeof a==`function`&&a(e)}),{started:!1,localPath:s,repoId:c,alreadyDownloading:!0};console.log(`[Download] Starting STT model download: ${c}`);let u=(async()=>{try{await Fn(o.url,o.headers,s,o.size,e=>{e>=0&&Number.isFinite(e)&&(console.log(`[Download] ${c}: ${Math.round(e*100)}%`),typeof r==`function`&&r(e))}),console.log(`[Download] Completed STT model: ${c}`),typeof i==`function`&&i({localPath:s,repoId:c})}catch(e){throw console.error(`[Download] Failed STT model: ${c}`,e.message),typeof a==`function`&&a(e),e}finally{t.deleteDownload(s)}})();return t.setDownload(s,u),{started:!0,localPath:s,repoId:c}}catch(e){return console.error(`[Download] Failed to start STT download:`,e.message),typeof a==`function`&&a(e),{started:!1,localPath:null,repoId:null,error:e.message}}}const Jn=e=>e?typeof e.score==`number`&&Number.isFinite(e.score)?Number(e.score):ce(e):0;async function Yn(e=null,t={}){let{threshold:n=1.1,includeBreakdown:r=!1,config:i,...a}=t,o=null,s=null,c=null;if(i)try{let e=await In(Cn(i));o=e.size??null,{processingBufferBytes:s}=ke({modelBytes:o}),c=e.quantization||null}catch{}let l=i?.backend?.gpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.gpu_memory_fraction))),u=i?.backend?.cpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.cpu_memory_fraction))),d=await Te({...a,platform:process.platform,totalMemoryInBytes:v.totalmem(),backend:`ggml-stt`,includeBreakdown:r,gpuMemoryFraction:l,cpuMemoryFraction:u,dependencies:{getBackendDevicesInfo:x,isLibVariantAvailable:S},modelBytes:o,kvCacheBytes:s}),f=d.selected,p=Jn(f);f&&(f.modelBytes=o||null,f.processingBytes=s||null,f.quantization=c||null);let m=null,h=null;if(e){let t=Jn(e);h={...e,score:t};let r=`buttress`,i=`buttress-higher-score`;if(!d.ok)r=`local`,i=`buttress-unavailable`;else if(!t&&t!==0)r=`buttress`,i=`missing-client-score`;else if(e.fit&&f?.fit){let a=e.fit.fitsInGpu||e.fit.fitsInCpu,o=f.fit.fitsInGpu||f.fit.fitsInCpu;a&&!o?(r=`local`,i=`client-fits-in-memory`):o&&!a?(r=`buttress`,i=`buttress-fits-in-memory`):t>p*n?(r=`local`,i=`client-better`):p>t*n?(r=`buttress`,i=`buttress-better`):(r=`either`,i=`comparable-scores`)}else t>p*n?(r=`local`,i=`client-better`):p>t*n?(r=`buttress`,i=`buttress-better`):(r=`either`,i=`comparable-scores`);m={buttressScore:p,clientScore:t,threshold:n,recommendation:r,reason:i}}!d.ok&&!m&&(m={buttressScore:p,clientScore:e?.score??null,threshold:n,recommendation:`local`,reason:`buttress-unavailable`});let g=null;return i&&(g={repoId:i.model?.repo_id||null,quantization:i.model?.quantization||null,filename:i.model?.filename||null}),{type:`ggml-stt`,timestamp:new Date().toISOString(),buttress:d,client:h,comparison:m,modelConfig:g}}const{ReadableStream:Xn}=typeof globalThis<`u`&&globalThis.ReadableStream&&globalThis.WritableStream?{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream}:o,Zn=D(import.meta.url),Qn=_.dirname(Zn),$n=_.join(Qn,`mlx-bridge.py`),er=`mlx-vlm==0.4.0`,tr=`mlx-lm==0.31.1`,nr=_.join(v.homedir(),`.buttress`,`models`),rr={backend:{type:`mlx-llm`},model:{repo_id:null,revision:`main`,adapter_path:null,tokenizer_config:null,model_config:null,vlm:`auto`},runtime:{cache_dir:nr,huggingface_token:process.env.HUGGINGFACE_TOKEN||null,mlx_env_dir:null,mlx_lm_package:tr,mlx_vlm_package:er,context_release_delay_ms:1e4,session_cache:{enabled:!0,max_size_bytes:5*1024*1024*1024,max_entries:100}}},ir=(e,t)=>e==null?t:typeof e==`number`?e:typeof e==`string`?w.parse(e)??t:t,ar=(e={},t={})=>(Object.entries(t||{}).forEach(([t,n])=>{n&&typeof n==`object`&&!Array.isArray(n)?((!e[t]||typeof e[t]!=`object`)&&(e[t]={}),ar(e[t],n)):e[t]=n}),e),or=(e={})=>{let t=structuredClone(rr);return ar(t,e),t},sr=async(e,t={})=>{let n=await fetch(e,t);if(!n.ok)throw Error(`HTTP ${n.status}: ${e}`);return n.json()},cr=async e=>{await l(e,{recursive:!0})},lr=(e,t,n)=>{let r=y(`sha256`).update(e).digest(`hex`);return _.join(n,`.metadata-cache`,t,`${r}.json`)},ur=async(e,t,n)=>{try{let r=await d(lr(e,t,n),`utf-8`);return JSON.parse(r)}catch{return null}},dr=async(e,t,n,r)=>{try{let i=lr(e,t,r);await cr(_.dirname(i)),await g(i,JSON.stringify(n),`utf-8`)}catch{}};async function fr(e,{revision:t=`main`,cacheDir:n,token:r}={}){let i=JSON.stringify({repoId:e,revision:t,type:`mlx-model-metadata`});if(n){let e=await ur(i,`mlx-model-metadata`,n);if(e)return e}let a={};r&&(a.Authorization=`Bearer ${r}`);let o=(await sr(`https://huggingface.co/api/models/${e}?revision=${t}&blobs=true`,{headers:a}))?.siblings||[],s=0;for(let e of o){let t=e.rfilename||e.path||e.filename||``;/\.(safetensors|npz)$/.test(t)&&(s+=Number(e.size)||0)}let c=null;try{c=await sr(`https://huggingface.co/${e}/raw/${t}/config.json`,{headers:a})}catch{}let l=c?.text_config||c||{},u=c||{},d=u.model_type||u.architectures?.[0]||null,f=l.hidden_size||l.dim||0,p=l.num_hidden_layers||l.n_layers||0,m=l.num_attention_heads||l.n_heads||0,h=l.num_key_value_heads??m,g=l.vocab_size||0,_=l.max_position_embeddings||0,v=l.intermediate_size||0,y=l.head_dim||l.v_head_dim||(m>0&&f>0&&Number.isInteger(f/m)?f/m:0),b=l.kv_lora_rank||0,x=l.qk_rope_head_dim||0,S=b>0,C=u.quantization||u.quantization_config||null,w=C?.bits||null,T=C?.group_size||null,E=l.dtype||u.torch_dtype||(w?`${w}bit`:null),D={repoId:e,revision:t,modelBytes:s,arch:d,hiddenSize:f,numLayers:p,numHeads:m,numKvHeads:h,headDim:y,vocabSize:g,maxCtx:_,intermediateSize:v,quantBits:w,quantGroupSize:T,dtype:E,isMLA:S,kvLoraRank:b,qkRopeHeadDim:x,fileCount:o.length,config:c};return n&&await dr(i,`mlx-model-metadata`,D,n),D}function pr({numLayers:e,numKvHeads:t,headDim:n,contextLength:r,isMLA:i,kvLoraRank:a,qkRopeHeadDim:o}){return!e||!r?0:i&&a>0?e*(a+(o||0))*r*2:!t||!n?0:e*t*n*r*2*2}const mr=async e=>{try{return await m(e),!0}catch{return!1}},q=(e,t,n={})=>new Promise((r,i)=>{ee(e,t,{timeout:n.timeout||3e5,...n},(t,n,a)=>{if(t){let n=a?.toString().trim()||t.message;i(Error(`${e} failed: ${n}`))}else r({stdout:n?.toString()||``,stderr:a?.toString()||``})})}),hr=new Map;async function gr({envDir:e,mlxLmPackage:t,mlxVlmPackage:n,onProgress:r}){let i=_.resolve(e),a=hr.get(i);if(a){let e=await a;return r?.(1),e}let o=vr({envDir:i,mlxLmPackage:t,mlxVlmPackage:n,onProgress:r});hr.set(i,o);try{return await o}finally{hr.delete(i)}}const _r=[3,10];async function vr({envDir:e,mlxLmPackage:t,mlxVlmPackage:n,onProgress:r}){let i=_.join(e,`bin`,`python3`),a=_.join(e,`bin`,`pip`);if(await mr(i))try{return await q(i,[`-c`,`import mlx_vlm; import torch`],{timeout:1e4}),r?.(1),i}catch{}if(!await mr(i)){r?.(.1);try{let{stdout:e}=await q(`python3`,[`-c`,`import sys; print(f"{sys.version_info.major}.{sys.version_info.minor}")`],{timeout:5e3}),[t,n]=e.trim().split(`.`).map(Number);(t<_r[0]||t===_r[0]&&n<_r[1])&&console.warn(`[mlx-llm] WARNING: System Python is ${t}.${n}, but mlx-vlm requires >= ${_r.join(`.`)}. You may get an older mlx-vlm version with reduced functionality. Consider installing Python >= 3.10 (e.g. via Homebrew).`)}catch{}console.log(`[mlx-llm] Creating venv at ${e}`),await l(e,{recursive:!0}),await q(`python3`,[`-m`,`venv`,e],{timeout:6e4}),r?.(.3)}return console.log(`[mlx-llm] Installing ${n}`),r?.(.4),await q(a,[`install`,t,n,`torch`,`torchvision`],{timeout:6e5,env:{...process.env}}),r?.(.9),await q(i,[`-c`,`import mlx_vlm; import torch; print(mlx_vlm.__version__)`],{timeout:15e3}),r?.(1),console.log(`[mlx-llm] mlx-vlm installed successfully`),i}var yr=class{constructor(){this.process=null,this.pendingRequests=new Map,this.requestCounter=0,this.readyPromise=null,this.buffer=``}spawn(e){return this.process=te(e,[$n],{stdio:[`pipe`,`pipe`,`pipe`],env:{...process.env,PYTHONUNBUFFERED:`1`}}),this.process.stderr.on(`data`,e=>{let t=e.toString().trim();t&&console.log(t)}),this.process.on(`exit`,e=>{console.log(`[mlx-llm] Bridge process exited with code ${e}`);for(let[t,n]of this.pendingRequests)n.reject(Error(`Bridge process exited (code ${e})`)),this.pendingRequests.delete(t);this.process=null}),this.process.stdout.on(`data`,e=>{this.buffer+=e.toString();let t=this.buffer.split(`
2
+ import{t as e}from"./chunk-C8PTHxhX.mjs";import{node as t}from"@elysiajs/node";import{Elysia as n,file as r,sse as i,t as a}from"elysia";import o,{createHash as s}from"node:crypto";import*as c from"node:stream/web";import{ReadableStream as l}from"node:stream/web";import u,{mkdir as d,open as f,readFile as p,readdir as m,rename as h,stat as g,unlink as _,writeFile as v}from"node:fs/promises";import y from"node:path";import b from"node:os";import{gguf as x}from"@huggingface/gguf";import{getBackendDevicesInfo as S,isLibVariantAvailable as C,loadModel as w}from"@fugood/llama.node";import T from"bytes";import{EventEmitter as E}from"node:events";import{initWhisper as D}from"@fugood/whisper.node";import{fileURLToPath as ee}from"node:url";import{execFile as te,execSync as O,spawn as k}from"node:child_process";import A from"node:fs";import j from"@iarna/toml";import{ZodError as M,z as N}from"zod";import{importSPKI as P,jwtVerify as F}from"jose";import{cors as I}from"@elysiajs/cors";import ne from"node-machine-id";import re from"ms";import{Buffer as L}from"node:buffer";import R from"node:dgram";const ie=1024**3,z=(e,t,n)=>Math.min(Math.max(e,t),n),ae=e=>e?40:0,oe=(e=0)=>e?z(e/(12*ie)*20,0,20):0,se=(e=0)=>e?z(e/(32*ie)*10,0,10):0,ce=e=>e?10:0,le=(e=`default`,t=null)=>{let n=String(e).toLowerCase();return n?n.includes(`cuda`)?20:n.includes(`vulkan`)?10:n.includes(`default`)?t===`darwin`||t===`ios`?15:5:0:0},B=({platform:e,variant:t,hasGpu:n,gpuUsableBytes:r=0,cpuUsableBytes:i=0,ok:a=!0}={})=>{if(!a)return 0;let o=ae(n)+le(t,e)+oe(r),s=se(i),c=ce(a);return Math.min(100,Math.round(o+s+c))},ue=({platform:e,variant:t,hasGpu:n,gpuUsableBytes:r=0,cpuUsableBytes:i=0,ok:a=!0}={})=>({gpuPresence:ae(n),variant:le(t,e),gpuMemory:oe(r),cpuMemory:se(i),availability:ce(a)}),de=[`cuda`,`vulkan`,`snapdragon`,`default`],fe=.85,pe=.5,me=e=>!e&&e!==0?[]:Array.isArray(e)?e.filter(e=>e!=null):[e],he=e=>e&&String(e).trim().toLowerCase()||null,ge=({variant:e,preferVariants:t=[],variantPreference:n=[],defaultVariants:r=de}={})=>{let i=[];e&&i.push(e),i.push(...me(t)),i.push(...me(n)),i.push(...r);let a=new Set;for(let e of i){let t=he(e);t&&a.add(t)}return Array.from(a)},_e=(e={})=>{let t=String(e.type||e.deviceType||e.kind||``).toLowerCase();return!!(t.includes(`gpu`)||t.includes(`cuda`)||t.includes(`metal`)||t.includes(`vulkan`)||t.includes(`snapdragon`))},ve=e=>Array.isArray(e)?e.map(e=>({...e})):[],ye=(e,t)=>e===`snapdragon`?t.filter(e=>e.deviceName!==`GPUOpenCL`):t,be=({platform:e,totalMemoryInBytes:t,variant:n,devices:r,gpuMemoryFraction:i,cpuMemoryFraction:a,ok:o,error:s})=>{let c=ve(ye(n,r)),l=c.some(_e),u=c.filter(e=>_e(e)&&Number.isFinite(Number(e.maxMemorySize))).reduce((e,t)=>e+t.maxMemorySize,0),d=t,f=l?Math.floor(u*i):0,p=d?Math.floor(d*a):0,m={platform:e,variant:n,hasGpu:l,gpuUsableBytes:f,cpuUsableBytes:p,ok:o};return{platform:e,ok:o,variant:n,hasGpu:l,devices:c,gpuTotalBytes:u,gpuUsableBytes:f,cpuTotalBytes:d,cpuUsableBytes:p,score:B(m),breakdown:o?ue(m):null,error:s,timestamp:new Date().toISOString()}},xe=({device:e,modelBytes:t=0,kvCacheBytes:n=0}={})=>{if(!e)return{totalRequiredBytes:t+n,fitsInGpu:!1,fitsInCpu:!1,limiting:`unknown-device`};let r=Math.max(0,Number(t)||0)+Math.max(0,Number(n)||0),i=e.hasGpu&&r>0&&r<=e.gpuUsableBytes,a=r>0&&r<=e.cpuUsableBytes,o=`ok`;return!i&&e.hasGpu&&(o=`gpu-memory`),a||(o=i?`cpu-memory`:`insufficient-memory`),{totalRequiredBytes:r,fitsInGpu:i,fitsInCpu:a,limiting:o}},Se=async({platform:e,variant:t=null,preferVariants:n=[],variantPreference:r=[],gpuMemoryFraction:i=fe,cpuMemoryFraction:a=pe,includeBreakdown:o=!1,totalMemoryInBytes:s,modelBytes:c=null,kvCacheBytes:l=null,limitedKvCacheBytes:u=null,dependencies:d={},defaultVariants:f=de}={})=>{let{getBackendDevicesInfo:p,isLibVariantAvailable:m}=d;if(typeof p!=`function`||typeof m!=`function`)throw TypeError(`GGML capability detection requires getBackendDevicesInfo and isLibVariantAvailable functions`);let h=ge({variant:t,preferVariants:n,variantPreference:r,defaultVariants:f}),g=[];for(let t of h)try{if(!await m(t))throw Error(`Variant ${t} not available on this platform`);let n=await p(t);g.push(be({platform:e,totalMemoryInBytes:s,variant:t,devices:n,gpuMemoryFraction:i,cpuMemoryFraction:a,ok:!0}))}catch(n){let r=n instanceof Error?n.message:String(n);g.push(be({platform:e,totalMemoryInBytes:s,variant:t,devices:[],gpuMemoryFraction:i,cpuMemoryFraction:a,ok:!1,error:r}))}let _=g.filter(e=>e.ok)[0]||null,v={ok:!!_,selected:_?{..._,breakdown:o?_.breakdown:void 0}:null,attempts:g};if(!o&&v.selected&&delete v.selected.breakdown,!v||!c&&!l)return v;let y=e=>{if(!e)return e;let t=xe({device:e,modelBytes:c||0,kvCacheBytes:l||0}),n=null;return u!=null&&u!==l&&(n=xe({device:e,modelBytes:c||0,kvCacheBytes:u})),{...e,fit:t,...n&&{limitedFit:n}}};return v.selected=y(v.selected),v.attempts=Array.isArray(v.attempts)?v.attempts.map(y):v.attempts,v},Ce=`ggml-llm`,we=[`cuda`,`vulkan`,`default`],Te=async({platform:e,variant:t=null,preferVariants:n=[],variantPreference:r=[],gpuMemoryFraction:i=fe,cpuMemoryFraction:a=pe,includeBreakdown:o=!1,totalMemoryInBytes:s,modelBytes:c=null,processingBytes:l=null,kvCacheBytes:u=null,dependencies:d={}}={})=>Se({platform:e,variant:t,preferVariants:n,variantPreference:r&&r.length>0?r:we,gpuMemoryFraction:i,cpuMemoryFraction:a,includeBreakdown:o,totalMemoryInBytes:s,modelBytes:c,kvCacheBytes:l??u,dependencies:d,defaultVariants:we}),Ee=async({platform:e,arch:t=null,unifiedMemoryFraction:n=.85,includeBreakdown:r=!1,totalMemoryInBytes:i,modelBytes:a=null,kvCacheBytes:o=null,limitedKvCacheBytes:s=null}={})=>{let c=[];e!==`darwin`&&c.push(`MLX requires macOS`),t&&t!==`arm64`&&c.push(`MLX requires Apple Silicon (arm64)`);let l=c.length===0,u=l?Math.floor(i*n):0,d={platform:e,variant:`mlx`,hasGpu:l,gpuUsableBytes:u,cpuUsableBytes:u,ok:l},f=B(d),p=l?ue(d):null,m={platform:e,ok:l,variant:`mlx`,hasGpu:l,unifiedMemory:!0,devices:l?[{type:`metal`,deviceName:`Apple Silicon (Unified Memory)`,maxMemorySize:i}]:[],gpuTotalBytes:l?i:0,gpuUsableBytes:u,cpuTotalBytes:i,cpuUsableBytes:u,score:f,breakdown:r?p:void 0,error:l?void 0:c.join(`; `),timestamp:new Date().toISOString()};r||delete m.breakdown;let h={ok:l,selected:l?m:null,attempts:[m],errors:l?[]:c};if(!a&&!o)return h;let g=e=>{if(!e)return e;let t=xe({device:e,modelBytes:a||0,kvCacheBytes:o||0}),n=null;return s!=null&&s!==o&&(n=xe({device:e,modelBytes:a||0,kvCacheBytes:s})),{...e,fit:t,...n&&{limitedFit:n}}};return h.selected=g(h.selected),h.attempts=h.attempts.map(g),h},De=new Map([[Ce,Se],[`ggml-stt`,Te],[`mlx-llm`,Ee]]),Oe=async({platform:e,totalMemoryInBytes:t,backend:n=Ce,dependencies:r,...i}={})=>{let a=De.get(n);if(!a)throw Error(`No capability detector registered for backend "${n}"`);return await a({...i,dependencies:r,totalMemoryInBytes:t,platform:e})},ke={f16:2,f32:4,q8_0:1,q6_k:.75,q5_k:.625,q5_k_m:.625,q5_k_s:.625,q5_1:.625,q5_0:.625,q4_k:.5,q4_k_m:.5,q4_k_s:.5,q4_1:.5,q4_0:.5,iq4_nl:.5},Ae=e=>ke[e?String(e).toLowerCase():`f16`]||ke.f16,je=(e,t,n,r,i,a={},{totalLayers:o=null,swaLayers:s=0,swaContext:c=null,swaContextMultiplier:l=1,swaAdditionalTokens:u=0,swaFull:d=!1}={})=>{if(!e||!t||!n||!r||!i)return 0;let f=o!=null&&o!==void 0?Number(o):Number(e),p=Math.max(0,Math.floor(f));if(!p)return 0;let m=Ae(a.k),h=Ae(a.v),g=Number(n)*(Number(r)*m+Number(i)*h);if(!g)return 0;let _=Math.max(0,Number(t)||0),v=Math.min(p,Math.max(0,Math.floor(Number(s)||0))),y=Math.max(0,p-v),b=c!=null&&Number.isFinite(Number(c))?Math.max(0,Number(c)):_,x=Math.max(1,Number(l)||1),S=Math.max(0,Number(u)||0),C=b*x+S,w=d?_:Math.min(_,C),T=y*_+v*Math.max(0,Math.floor(w));return Math.round(g*T)},Me=({modelBytes:e=0,audioLengthSeconds:t=30,sampleRate:n=16e3,bytesPerSample:r=4}={})=>{let i=Math.max(0,Number(e)||0),a=Math.max(0,Math.floor(Math.max(0,t)*n*r)),o=1024*1024,s=1024*o,c;c=i<200*o?120*o:i<500*o?140*o:i<2*s?150*o:160*o;let l;l=i<200*o?70*o:i<500*o?135*o:(2*s,220*o);let u;u=i<100*o?20*o:i<200*o?30*o:i<500*o?85*o:i<2*s?215*o:360*o;let d=c+l+u;return{modelBytes:i,audioBufferBytes:a,processingBufferBytes:d,totalBytes:i+d+a}},Ne=e=>e?String(e).trim().toLowerCase():null,Pe=(e={},t=null)=>{if(!e)return null;let n=Ne(t),r=n?`${n}.attention.sliding_window`:null,i=(r&&e[r]!=null?e[r]:null)??e[`llama.attention.sliding_window`];if(i==null)return null;let a=Number(i);return Number.isFinite(a)?a:null},Fe=(e=0,t=0,n=!1)=>{let r=Math.max(0,Math.floor(Number(e)||0)),i=Math.max(0,Math.floor(Number(t)||0));if(!r||i===1)return 0;if(i<=0)return r;let a=Math.max(0,i-1),o=Math.floor(r/i),s=r%i,c=n?Math.max(0,s-1):Math.min(s,a);return o*a+c},Ie=({arch:e,nLayer:t=0})=>({arch:Ne(e),enabled:!1,window:null,pattern:null,denseFirst:!1,type:null,kvLayers:Math.max(0,Math.floor(Number(t)||0)),swaLayers:0}),Le=new Map([[`llama4`,({nSwa:e})=>e===0?{enabled:!1}:{enabled:!0,window:e&&e>0?e:8192,pattern:4,type:`chunked`}],[`afmoe`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:4,type:`standard`}],[`phi3`,()=>({enabled:!1})],[`gemma2`,({nSwa:e})=>{let t=e&&e>0?e:4096;return t?{enabled:!0,window:t,pattern:2,type:`standard`}:{enabled:!1}}],[`gemma3`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:6,type:`standard`}],[`gemma3n`,({nLayer:e,nSwa:t})=>!t||t<=0?{enabled:!1}:{enabled:!0,window:t,pattern:5,type:`standard`,kvLayers:Math.min(20,e)}],[`gemma-embedding`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:6,type:`symmetric`}],[`cohere2`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:4,type:`standard`}],[`olmo2`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:4,type:`standard`}],[`exaone4`,({nLayer:e,nSwa:t})=>{let n=e>=64,r=null;return t&&t>0?r=t:n&&(r=4096),r?{enabled:!0,window:r,pattern:4,type:`standard`}:{enabled:!1}}],[`gpt-oss`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:e,pattern:2,type:`standard`}],[`gemma4`,({nLayer:e,nSwa:t,metadata:n})=>{if(!t||t<=0)return{enabled:!1};let r=Number(n?.[`gemma4.attention.shared_kv_layers`])||0,i=Math.max(0,e-r),a=n?.[`gemma4.attention.sliding_window_pattern`];return Array.isArray(a)?{enabled:!0,window:t,type:`standard`,swaLayers:a.slice(0,i).filter(e=>Number(e)>0).length,kvLayers:i}:{enabled:!0,window:t,pattern:6,type:`standard`,kvLayers:i}}],[`smallthinker`,({nSwa:e})=>!e||e<=0?{enabled:!1}:{enabled:!0,window:4096,pattern:4,denseFirst:!0,type:`standard`}]]),Re=({arch:e,metadata:t={},nLayer:n=0}={})=>{let r=Ne(e||t[`general.architecture`]),i=Math.max(0,Math.floor(Number(n)||0)),a=Pe(t,r),o=r?Le.get(r):null;if(!o)return Ie({arch:r,nLayer:n});let s=o({nLayer:i,nSwa:a,metadata:t});if(!s||!s.enabled||!s.window||s.window<=0)return Ie({arch:r,nLayer:n});let c=Math.max(0,Math.floor(Number(s.pattern)||0)),l=s.kvLayers!=null&&Number.isFinite(Number(s.kvLayers))?Number(s.kvLayers):i,u=Math.max(0,Math.floor(l)),d=s.swaLayers!=null&&Number.isFinite(Number(s.swaLayers))?Math.max(0,Math.floor(Number(s.swaLayers))):Fe(u,c,!!s.denseFirst);return{arch:r,enabled:d>0,window:s.window,pattern:c,denseFirst:!!s.denseFirst,type:s.type||`standard`,kvLayers:u,swaLayers:d}},ze=new Set([`mamba`,`mamba2`,`rwkv6`,`rwkv6qwen2`,`rwkv7`,`arwkv7`]),Be=new Set([`jamba`,`falcon-h1`,`plamo2`,`granitehybrid`,`lfm2`,`lfm2moe`,`nemotron_h`,`nemotron_h_moe`,`qwen3next`]),Ve=e=>e?String(e).trim().toLowerCase():null,He=e=>{let t=Ve(e);return t?ze.has(t):!1},Ue=e=>{let t=Ve(e);return t?Be.has(t):!1},We=e=>He(e)?`recurrent`:Ue(e)?`hybrid`:`transformer`,Ge=(e={})=>{let t=e[`general.architecture`],n=(t,n=null)=>{let r=e[t],i=Number(r);return Number.isFinite(i)?i:n},r=(t,n=null)=>{let r=e[t];if(Array.isArray(r))return r;let i=Number(r);return Number.isFinite(i)?i:n},i=t?n(`${t}.context_length`,n(`llama.context_length`)):null,a=t?n(`${t}.block_count`,n(`llama.block_count`)):null,o=t?n(`${t}.embedding_length`,n(`llama.embedding_length`)):null,s=t?n(`${t}.attention.head_count`,n(`llama.attention.head_count`)):null,c=t?r(`${t}.attention.head_count_kv`,r(`llama.attention.head_count_kv`,s)):null,l=null,u=null;if(Array.isArray(c)){let e=c.filter(e=>Number(e)>0);e.length>0?(l=Math.max(...e.map(Number)),u=e.length):(l=0,u=0)}else l=c;let d=t?n(`${t}.attention.key_length`,n(`llama.attention.key_length`)):null,f=t?n(`${t}.attention.value_length`,n(`llama.attention.value_length`)):null,p=e[`general.quantization_version`]||null,m=e[`general.file_type`]||null,h=t?n(`${t}.ssm.conv_kernel`):null,g=t?n(`${t}.ssm.state_size`):null,_=t?n(`${t}.ssm.inner_size`):null,v=t?n(`${t}.ssm.group_count`):null,y=t?n(`${t}.ssm.time_step_rank`):null,b=t?n(`${t}.rwkv.head_size`):null,x=t?n(`${t}.rwkv.token_shift_count`,2):null,S=t?n(`${t}.attention.shared_kv_layers`,0):0,C=u!=null&&a!=null?a-u:null;return{arch:t,nCtxTrain:i,nLayer:a,nEmbd:o,nHead:s,nHeadKv:l,nEmbdHeadK:d,nEmbdHeadV:f,quantVersion:p,fileType:m,attentionLayerCount:u,recurrentLayerCount:C,ssmDConv:h,ssmDState:g,ssmDInner:_,ssmNGroup:v,ssmDtRank:y,rwkvHeadSize:b,rwkvTokenShiftCount:x,sharedKvLayers:S}},Ke=({layerCount:e,headKvCount:t,embdHeadKCount:n,embdHeadVCount:r,cacheTypes:i,swaConfig:a,kvUnified:o=!1,nParallel:s=1,swaFull:c=!1,arch:l=null,attentionLayerCount:u=null})=>{let d=We(l);if(d===`recurrent`)return()=>0;let f=d===`hybrid`&&u!=null?Math.max(0,Math.floor(Number(u)||0)):e,p=a?.window&&o?Math.max(1,Number(s)||1):1,m=o?1:Math.max(1,Number(s)||1);return e=>je(f,e,t,n,r,i,{totalLayers:f,swaLayers:a?.swaLayers||0,swaContext:a?.window,swaFull:c,swaContextMultiplier:p})*m},qe=({nLayer:e,nEmbd:t,recurrentLayerCount:n=null,nSeqMax:r=1,ssmDConv:i=null,ssmDState:a=null,ssmDInner:o=null,ssmNGroup:s=null,ssmDtRank:c=null,rwkvHeadSize:l=null,rwkvTokenShiftCount:u=2,arch:d=null})=>{if(We(d)===`transformer`)return 0;let f=n==null?Math.max(0,Math.floor(Number(e)||0)):Math.max(0,Math.floor(Number(n)||0));if(f===0)return 0;let p=Math.max(1,Math.floor(Number(r)||1)),m=0,h=0;if(l!=null&&l>0&&t!=null&&t>0)m=Math.max(1,Number(u)||2)*t,h=t*l;else if(a!=null&&o!=null){let e=Math.max(0,Number(i)||0),t=Math.max(0,Number(a)||0),n=Math.max(0,Number(o)||0),r=Math.max(1,Number(s)||1);Math.max(0,Number(c)||0)>0?(m=e>0?(e-1)*2*r*t:0,h=Math.floor(t*n/2)):(m=e>0?(e-1)*(n+2*r*t):0,h=t*n)}else return 0;let g=(m+h)*p*f*4;return Math.max(0,g)},Je=({maxCtx:e,availableMemory:t,modelBytes:n,kvBytesForCtx:r})=>{let i=Math.max(1,Math.floor(Number(e)||0));if(!r||t<=n)return i;let a=1,o=i,s=i;for(;a<=o;){let e=Math.floor((a+o)/2);n+r(e)<=t?(s=e,a=e+1):o=e-1}return s},V=new E;V.setMaxListeners(100);const Ye=(e,t,n)=>{e.push({...t,timestamp:t.timestamp||new Date().toISOString()}),e.length>n&&e.shift()};var Xe=class{constructor(e=9999){this.maxEntries=e,this.modelLoads=[],this.completions=[],this.transcriptions=[]}addModelLoad(e){Ye(this.modelLoads,e,this.maxEntries),V.emit(`status:modelLoad`,e),V.emit(`status:change`,{type:`modelLoad`,entry:e})}addCompletion(e){Ye(this.completions,e,this.maxEntries),V.emit(`status:completion`,e),V.emit(`status:change`,{type:`completion`,entry:e})}addTranscription(e){Ye(this.transcriptions,e,this.maxEntries),V.emit(`status:transcription`,e),V.emit(`status:change`,{type:`transcription`,entry:e})}getModelLoadHistory(){return[...this.modelLoads].reverse()}getCompletionHistory(){return[...this.completions].reverse()}getTranscriptionHistory(){return[...this.transcriptions].reverse()}clear(){this.modelLoads=[],this.completions=[],this.transcriptions=[]}};const H=new Xe,U=new Xe;let Ze=0;function Qe(e){let t=t=>e(t);return V.on(`status:change`,t),()=>V.off(`status:change`,t)}function $e(e){return Ze+=1,{subscriberId:Ze,unsubscribe:Qe(e)}}function et(e){let t=[];return{generators:Array.from(e.entries()).filter(([,e])=>e.type===`ggml-llm`).map(([e,n])=>{let{instance:r}=n,i=[];return r.contexts&&(i=Array.from(r.contexts.entries()).map(([n,r])=>{let i={key:n,refCount:r.refCount,hasModel:!!r.context},a=r.context.parallel.getStatus();return i.parallelStatus=a,t.push({generatorId:e,contextKey:n,...a}),i})),{id:e,type:n.type,refCount:n.refCount,repoId:r.info?.model?.repoId||null,quantization:r.info?.model?.quantization||null,variant:r.info?.runtime?.variant||null,nCtx:r.info?.runtime?.n_ctx||null,nParallel:r.info?.runtime?.n_parallel||null,contexts:i}}),parallelStatuses:t,history:{modelLoads:H.getModelLoadHistory().filter(e=>e.variant!==`mlx`),completions:H.getCompletionHistory().filter(e=>e.variant!==`mlx`)}}}function tt(e){return{generators:Array.from(e.entries()).filter(([,e])=>e.type===`ggml-stt`).map(([e,t])=>{let{instance:n}=t,r=n.getStatus?.()||{},i=r.queueStatus||{processing:!1,queuedCount:0};return{id:e,type:t.type,refCount:t.refCount,repoId:n.info?.model?.repoId||null,quantization:n.info?.model?.quantization||null,modelType:n.info?.model?.modelType||null,variant:n.info?.runtime?.variant||null,hasContext:r.hasContext||!1,contextRefCount:r.contextRefCount||0,queueStatus:i}}),history:{modelLoads:U.getModelLoadHistory(),transcriptions:U.getTranscriptionHistory()}}}function nt(e){return{generators:Array.from(e.entries()).filter(([,e])=>e.type===`mlx-llm`).map(([e,t])=>{let{instance:n}=t,r=n.getStatus?.()||{};return{id:e,type:t.type,refCount:t.refCount,repoId:r.repoId||n.info?.model?.repoId||null,variant:r.variant||`mlx`,contexts:r.contexts||[]}}),history:{modelLoads:H.getModelLoadHistory().filter(e=>e.variant===`mlx`),completions:H.getCompletionHistory().filter(e=>e.variant===`mlx`)}}}function rt(e){return{timestamp:new Date().toISOString(),ggmlLlm:et(e),ggmlStt:tt(e),mlxLlm:nt(e)}}const{ReadableStream:it,WritableStream:at}=typeof globalThis<`u`&&globalThis.ReadableStream&&globalThis.WritableStream?{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream}:c,ot=(e={},t={})=>(Object.entries(t||{}).forEach(([t,n])=>{n&&typeof n==`object`&&!Array.isArray(n)?((!e[t]||typeof e[t]!=`object`)&&(e[t]={}),ot(e[t],n)):e[t]=n}),e),st=`https://huggingface.co`,ct=`https://huggingface.co/api`,W=y.join(b.homedir(),`.buttress`,`models`),lt=[`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`],ut=.5,dt={backend:{type:`ggml-llm`,variant:null,variant_preference:[`cuda`,`vulkan`,`snapdragon`,`default`],gpu_memory_fraction:.85,cpu_memory_fraction:ut},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:ct,base_url:st,enable_mtmd:!1,mmproj_filename:null,mmproj_url:null,mmproj_local_path:null,mmproj_use_gpu:null,mmproj_image_min_tokens:-1,mmproj_image_max_tokens:-1},runtime:{cache_dir:W,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},session_cache:{enabled:!0,max_size_bytes:10*1024*1024*1024,max_entries:1e3},context_release_delay_ms:1e4}},ft=(e,t=[])=>!e&&e!==0?[...t]:Array.isArray(e)?e.filter(e=>e!=null):[e],pt=e=>{if(!e)return null;let t=String(e).toLowerCase();return[`cuda`,`vulkan`,`snapdragon`,`default`].includes(t)?t:null},mt=(e={})=>{let t=structuredClone(dt);if(ot(t,e),t.backend.variant=pt(t.backend.variant),t.backend.variant_preference=Array.from(new Set(ft(t.backend.variant_preference).flatMap(e=>{let t=pt(e);return t?[t]:[]}))),t.backend.variant_preference.length===0&&(t.backend.variant_preference=[`cuda`,`vulkan`,`snapdragon`,`default`]),t.runtime.prefer_variants=Array.from(new Set(ft(t.runtime.prefer_variants).flatMap(e=>{let t=pt(e);return t?[t]:[]}))),t.model.preferred_quantizations=Array.from(new Set(ft(t.model.preferred_quantizations||t.model.quantizations).map(e=>e?String(e).toLowerCase():null).filter(Boolean))),t.model.quantization){let e=String(t.model.quantization).toLowerCase();t.model.preferred_quantizations.includes(e)||t.model.preferred_quantizations.unshift(e)}t.model.n_parallel=t.model.n_parallel?Math.max(1,Number(t.model.n_parallel)):void 0,t.model.n_batch=Math.max(1,Number(t.model.n_batch)||512),t.model.base_url=t.model.base_url||st,t.model.api_base=t.model.api_base||ct,t.model.enable_mtmd=!!t.model.enable_mtmd;let n=e=>{if(e==null)return-1;let t=Number(e);return Number.isFinite(t)?Math.floor(t):-1};return t.model.mmproj_image_min_tokens=n(t.model.mmproj_image_min_tokens),t.model.mmproj_image_max_tokens=n(t.model.mmproj_image_max_tokens),t.runtime.cache_dir=t.runtime.cache_dir?y.resolve(t.runtime.cache_dir):W,t.runtime.session_cache={...dt.runtime.session_cache,...t.runtime.session_cache||{}},t.runtime.context_release_delay_ms=Math.max(0,Number(t.runtime.context_release_delay_ms)||dt.runtime.context_release_delay_ms),t},ht=e=>{let t=e.toLowerCase();return lt.find(e=>t.includes(e))||null},gt=e=>{let t=[];return e.backend.variant&&t.push(e.backend.variant),e.runtime.prefer_variants.length>0&&t.push(...e.runtime.prefer_variants),t.push(...e.backend.variant_preference),t.push(`default`),Array.from(new Set(t.flatMap(e=>{let t=pt(e);return t?[t]:[]})))},G=async e=>{await d(e,{recursive:!0})},_t=(e=W)=>y.join(e,`.metadata-cache`),vt=(e,t,n=W)=>{let r=s(`sha256`).update(e).digest(`hex`);return y.join(_t(n),t,`${r}.json`)},yt=async(e,t,n=W)=>{try{let r=vt(e,t,n),i=await p(r,`utf-8`);return console.log(`[Cache] Hit ${t} cache:`,y.basename(r)),JSON.parse(i,(e,t)=>typeof t==`string`&&t.startsWith(`__bigint__`)?BigInt(t.slice(10)):t)}catch{return null}},bt=async(e,t,n,r=W)=>{try{let i=vt(e,t,r);await G(y.dirname(i)),await v(i,JSON.stringify(n,(e,t)=>typeof t==`bigint`?`__bigint__${t.toString()}`:t),`utf-8`),console.log(`[Cache] Wrote ${t} cache:`,y.basename(i))}catch(e){console.warn(`[Cache] Failed to write ${t} cache:`,e.message)}},xt=(e=W)=>y.join(e,`.session-state-cache`),St=(e=W)=>y.join(xt(e),`cache-map.json`),Ct=(e=W)=>y.join(xt(e),`temp`),wt=(e=W)=>y.join(xt(e),`states`),Tt=()=>({version:1,entries:{},totalSize:0}),Et=async(e=W)=>{try{let t=await p(St(e),`utf-8`),n=JSON.parse(t);return!n.entries||typeof n.entries!=`object`?Tt():n}catch{return Tt()}},Dt=async(e,t=W)=>{let n=St(t),r=`${n}.tmp.${Date.now()}`;try{await G(y.dirname(n)),await v(r,JSON.stringify(e,null,2),`utf-8`),await h(r,n)}catch(e){throw await _(r).catch(()=>{}),e}},Ot=(e,t)=>{let n=JSON.stringify({text:e,model:t.modelPath,variant:t.variant,n_gpu_layers:t.n_gpu_layers,n_ctx:t.n_ctx,cacheTypeK:t.cacheTypeK,cacheTypeV:t.cacheTypeV,kvUnified:t.kvUnified,swaFull:t.swaFull,flashAttnType:t.flashAttnType});return s(`sha256`).update(n).digest(`hex`).slice(0,24)},kt=(e,t=W)=>y.join(wt(t),`${e}.bin`),At=(e=W)=>{let t=`${Date.now()}-${Math.random().toString(36).slice(2,10)}`;return y.join(Ct(e),`${t}.bin`)},jt=(e,t)=>e.modelPath===t.modelPath&&e.variant===t.variant&&e.n_gpu_layers===t.n_gpu_layers&&e.n_ctx>=t.n_ctx&&e.cacheTypeK===t.cacheTypeK&&e.cacheTypeV===t.cacheTypeV&&e.kvUnified===t.kvUnified&&e.swaFull===t.swaFull&&e.flashAttnType===t.flashAttnType&&!!e.isRecurrent==!!t.isRecurrent&&!!e.isHybrid==!!t.isHybrid,Mt=(e,t)=>{let n=Math.min(e.length,t.length),r=0;for(;r<n&&e[r]===t[r];)r+=1;return r},Nt=(e,t,n,r=!1)=>{let i=Object.values(n.entries);console.log(`[SessionCache] Finding match for promptText (${e.length} chars), exactMatch=${r}`),console.log(`[SessionCache] Checking ${i.length} cache entries`);let a=i.filter(e=>jt(e.metadata,t));if(r){let t=a.find(t=>t.fullText===e);return t?(console.log(`[SessionCache] Exact match found: ${t.id} (${t.fullText.length} chars)`),{entry:t,prefixLength:t.fullText.length,exactMatch:!0}):null}let o=a.reduce((t,n)=>{let r=Mt(e,n.fullText);return r>t.prefixLen||r===t.prefixLen&&n.fullText.length>(t.entry?.fullText?.length||0)?{entry:n,prefixLen:r}:t},{entry:null,prefixLen:0});return o.entry?(console.log(`[SessionCache] Prefix match found: ${o.entry.id} (${o.prefixLen}/${o.entry.fullText.length} chars)`),{entry:o.entry,prefixLength:o.prefixLen}):(console.log(`[SessionCache] No match found`),null)},Pt=async(e,t,n)=>{let r=Object.values(e.entries),i=r.sort((e,t)=>new Date(e.lastAccessedAt)-new Date(t.lastAccessedAt)),a=e.totalSize,o=r.length,s=i.filter(e=>!(a>t)&&!(o>n)?!1:(a-=(e.stateFileSize||0)+(e.promptStateSize||0),--o,!0));return await Promise.all(s.map(async t=>{await _(t.stateFilePath).catch(()=>{}),t.promptStatePath&&await _(t.promptStatePath).catch(()=>{}),delete e.entries[t.id],console.log(`[SessionCache] Evicted entry: ${t.id}`)})),e.totalSize=Math.max(0,a),s.map(e=>e.id)},Ft=async(e,t,n,r)=>{let i=[];for(let[a,o]of Object.entries(e.entries))a!==n&&jt(o.metadata,r)&&t.startsWith(o.fullText)&&o.fullText.length<t.length&&i.push(o);return await Promise.all(i.map(async t=>{await _(t.stateFilePath).catch(()=>{}),t.promptStatePath&&await _(t.promptStatePath).catch(()=>{}),e.totalSize-=(t.stateFileSize||0)+(t.promptStateSize||0),delete e.entries[t.id],console.log(`[SessionCache] Evicted superseded prefix entry: ${t.id} (${t.promptText.length} prompt chars)`)})),i.map(e=>e.id)},It=async(e=W)=>{let t=Ct(e);try{let e=await m(t),n=Date.now();await Promise.all(e.map(async e=>{let r=y.join(t,e),i=await g(r).catch(()=>null);i&&n-i.mtimeMs>36e5&&(await _(r).catch(()=>{}),console.log(`[SessionCache] Cleaned up temp file: ${e}`))}))}catch{}},Lt=async e=>{try{return await g(e),!0}catch{return!1}},Rt=(e,t)=>e==null?t:typeof e==`number`?e:typeof e==`string`?T.parse(e)??t:t;var zt=class e{constructor(e,t){this.config=e,this.plan=t,this.baseDir=e.runtime.cache_dir,this.enabled=e.runtime.session_cache?.enabled!==!1,this.maxSizeBytes=Rt(e.runtime.session_cache?.max_size_bytes,10*1024*1024*1024),this.maxEntries=e.runtime.session_cache?.max_entries||1e3,this.metadata={variant:t.info?.runtime?.variant||null,n_gpu_layers:t.info?.runtime?.n_gpu_layers||0,n_ctx:t.info?.runtime?.n_ctx||0,modelPath:t.localPath,cacheTypeK:t.info?.runtime?.cache_type_k||`f16`,cacheTypeV:t.info?.runtime?.cache_type_v||`f16`,kvUnified:t.info?.runtime?.kv_unified??null,swaFull:t.info?.runtime?.swa_full??null,flashAttnType:t.info?.runtime?.flash_attn_type||`off`,isRecurrent:!1,isHybrid:!1},this.cacheMap=null,this.initialized=!1}updateModelInfo(e){e&&(this.metadata.isRecurrent=!!e.is_recurrent,this.metadata.isHybrid=!!e.is_hybrid,(this.metadata.isRecurrent||this.metadata.isHybrid)&&console.log(`[SessionCache] Model architecture: recurrent=${this.metadata.isRecurrent}, hybrid=${this.metadata.isHybrid}`))}requiresExactMatch(){return this.metadata.isRecurrent||this.metadata.isHybrid}async persistCacheMap(){try{await Dt(this.cacheMap,this.baseDir)}catch(e){console.warn(`[SessionCache] Failed to persist cache map: ${e?.message||e}`)}}static checkTokenPrefixMatch(e,t){if(e.length>t.length)return!1;for(let n=0;n<e.length;n+=1)if(e[n]!==t[n])return!1;return!0}static async tokenizeToArray(e,t){let n=await e.tokenize(t);return Array.from(n?.tokens||[])}async findFormattedMatchForRecurrent(t,n,r){let i=await e.tokenizeToArray(r,n),a=t.map(async t=>{try{let n=await e.tokenizeToArray(r,t.fullText);if(e.checkTokenPrefixMatch(n,i))return{entry:t,usePromptState:!1,tokenCount:n.length};if(t.promptStatePath&&t.promptText){let n=await e.tokenizeToArray(r,t.promptText);if(e.checkTokenPrefixMatch(n,i))return{entry:t,usePromptState:!0,tokenCount:n.length}}return null}catch(e){return console.warn(`[SessionCache] Failed to check entry ${t.id}: ${e.message}`),null}}),o=(await Promise.all(a)).find(e=>e!==null);if(!o)return console.log(`[SessionCache] No token prefix match found for recurrent/hybrid model`),null;let{entry:s,usePromptState:c,tokenCount:l}=o;return console.log(`[SessionCache] Token prefix match: ${s.id} (${l} tokens, usePromptState=${c})`),await Lt(c?s.promptStatePath:s.stateFilePath)?(s.lastAccessedAt=new Date().toISOString(),await this.persistCacheMap(),{entry:s,usePromptState:c}):(await this.removeStaleEntry(s),null)}async initialize(){if(!(!this.enabled||this.initialized))try{await G(xt(this.baseDir)),await G(Ct(this.baseDir)),await G(wt(this.baseDir)),this.cacheMap=await Et(this.baseDir),this.initialized=!0,console.log(`[SessionCache] Initialized with ${Object.keys(this.cacheMap.entries).length} entries`)}catch(e){console.warn(`[SessionCache] Failed to initialize: ${e.message}`),this.enabled=!1}}async removeStaleEntry(e){console.log(`[SessionCache] Removing stale entry: ${e.id}`),e.stateFilePath&&await _(e.stateFilePath).catch(()=>{}),e.promptStatePath&&await _(e.promptStatePath).catch(()=>{}),delete this.cacheMap.entries[e.id],this.cacheMap.totalSize-=(e.stateFileSize||0)+(e.promptStateSize||0),await this.persistCacheMap()}async findMatchingEntry(e,t=null){if(!this.enabled||!this.cacheMap)return null;let n=this.requiresExactMatch();if(n&&t){let n=Object.values(this.cacheMap.entries).filter(e=>jt(e.metadata,this.metadata)&&e.fullText);return this.findFormattedMatchForRecurrent(n,e,t)}let r=Nt(e,this.metadata,this.cacheMap,n);if(!r)return null;let{entry:i}=r;return await Lt(i.stateFilePath)?(i.lastAccessedAt=new Date().toISOString(),await this.persistCacheMap(),{entry:i,usePromptState:!1}):(await this.removeStaleEntry(i),null)}async prepareCompletionOptions(e,t,n=null){let r={options:e,cacheEntry:null,promptPrefix:null};if(!this.enabled)return r;let i=await this.findMatchingEntry(t,n);if(!i)return r;let{entry:a,usePromptState:o}=i,s=o?a.promptStatePath:a.stateFilePath,c=o?a.promptText:a.fullText;return console.log(`[SessionCache] Found matching entry: ${a.id} (${c.length} chars, usePromptState=${o})`),{options:{...e,load_state_path:s},cacheEntry:a,promptPrefix:c}}async saveCompletionState(e,t,n,r=0,i=null){if(!this.enabled)return null;let a=e+t,o=Ot(a,this.metadata),s=()=>{n&&_(n).catch(()=>{}),i&&_(i).catch(()=>{})};if(this.cacheMap.entries[o]){console.log(`[SessionCache] Entry already exists for prompt: ${o}, updating position`);let e=this.cacheMap.entries[o];return e.lastAccessedAt=new Date().toISOString(),delete this.cacheMap.entries[o],this.cacheMap.entries[o]=e,await this.persistCacheMap(),s(),e}let c=kt(o,this.baseDir),l=i?kt(`${o}-prompt`,this.baseDir):null;try{await G(y.dirname(c)),await h(n,c);let s=await g(c),u=0;if(i&&l)try{await h(i,l),u=(await g(l)).size,console.log(`[SessionCache] Saved prompt state: ${l}`)}catch(e){console.warn(`[SessionCache] Failed to save prompt state: ${e.message}`)}let d={id:o,promptText:e,completionText:t,fullText:a,promptTokenCount:r,stateFilePath:c,stateFileSize:s.size,promptStatePath:l||null,promptStateSize:u,metadata:{...this.metadata},createdAt:new Date().toISOString(),lastAccessedAt:new Date().toISOString()};return this.cacheMap.entries[o]=d,this.cacheMap.totalSize+=s.size+u,this.requiresExactMatch()||await Ft(this.cacheMap,e,o,this.metadata),await Pt(this.cacheMap,this.maxSizeBytes,this.maxEntries),await Dt(this.cacheMap,this.baseDir),console.log(`[SessionCache] Saved entry: ${o} (${s.size} bytes, ${a.length} chars)`),d}catch(e){return console.warn(`[SessionCache] Failed to save state: ${e.message}`),s(),null}}async generateTempStatePath(){return await G(Ct(this.baseDir)),At(this.baseDir)}async cleanup(){await It(this.baseDir)}};const Bt=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,t);if(!n.ok){let t=await n.text().catch(()=>``);throw Error(`Failed to fetch ${e}: ${n.status} ${n.statusText} ${t}`.trim())}return n.json()},Vt=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,{...t,method:`HEAD`});if(!n.ok)throw Error(`Failed to fetch headers for ${e}: ${n.status} ${n.statusText}`);return n},Ht=async(e,t,n=W)=>{let r=JSON.stringify({url:e,headers:t}),i=await yt(r,`range-metadata`,n);if(i)return i;let a=!/^https?:/i.test(e),{metadata:o}=await x(e,{fetch,additionalFetchHeaders:t,allowLocalFile:a});return await bt(r,`range-metadata`,o,n),o},Ut=(e,t)=>{if(e.model.local_path)return y.resolve(e.model.local_path);let n=t.repoId.split(`/`),r=y.join(e.runtime.cache_dir,...n,t.revision);return y.join(r,t.filename)},K=async(e,t)=>{try{let n=await g(e);return t?n.size===t:!0}catch{return!1}},Wt=async(e,t,n,r,i)=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);await G(y.dirname(n));let a=await fetch(e,{headers:t});if(!a.ok||!a.body)throw Error(`Failed to download ${e}: ${a.status} ${a.statusText}`);let o=await f(n,`w`),s=Number(a.headers.get(`content-length`))||r||0,c=0,l=.05;try{await a.body.pipeTo(new at({async write(e){if(await o.write(e),c+=e.byteLength,typeof i==`function`&&s>0){let e=Math.min(1,c/s);for(;e>=l;)i(l),l+=.05}},async close(){await o.close(),typeof i==`function`&&i(1)},async abort(e){throw await o.close().catch(()=>{}),await _(n).catch(()=>{}),e}}))}catch(e){throw await o.close().catch(()=>{}),await _(n).catch(()=>{}),e}if(r){let e=await g(n);if(e.size!==r)throw await _(n).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${r} got ${e.size}`)}},Gt=async e=>{let t=e.model.repo_id||e.model.repository||e.model.model;if(!t)throw Error("`model.repo_id` is required in Buttress backend config");let n=e.model.revision||`main`,r=e.runtime.cache_dir,i=JSON.stringify({repoId:t,revision:n,filename:e.model.filename,url:e.model.url,quantization:e.model.quantization,preferred_quantizations:e.model.preferred_quantizations}),a=await yt(i,`artifact-info`,r);if(a)return a;let o={...e.runtime.http_headers||{}};if(e.runtime.huggingface_token&&(o.Authorization=`Bearer ${e.runtime.huggingface_token}`),e.model.url){let a=await Vt(e.model.url,{headers:o}),s=Number(a.headers.get(`content-length`))||null,c={repoId:t,revision:n,filename:e.model.filename||e.model.url.split(`/`).pop(),url:e.model.url,size:s,headers:o};return await bt(i,`artifact-info`,c,r),c}let{filename:s}=e.model,c=e.model.quantization&&String(e.model.quantization).toLowerCase(),l=await Bt(`${e.model.api_base}/models/${t}?revision=${n}&blobs=true`,{headers:o}),u=l?.siblings||l?.files||[],d=[];for(let e of u){let t=e.rfilename||e.path||e.filename;typeof t==`string`&&t.endsWith(`.gguf`)&&d.push(t)}if(d.length===0)throw Error(`No GGUF artifacts found in repo ${t}`);let f=e.model.preferred_quantizations.length>0?e.model.preferred_quantizations:lt,p=d.map(e=>e.toLowerCase()),m=()=>{for(let e of f){let t=p.findIndex(t=>t.includes(e));if(t!==-1)return{filename:d[t],quantization:e}}return null};if(s)c||=ht(s);else{let{filename:e,quantization:t}=m()||{filename:d[0],quantization:null};s=e,c=t||ht(s)}let h=`${e.model.base_url.replace(/\/+$/,``)}/${t}/resolve/${n}/${s}`,g=/-(\d{5})-of-(\d{5})\.gguf$/,_=s.match(g),v=null;if(_){let[,,r]=_,i=await Bt(`${e.model.api_base}/models/${t}?revision=${n}&blobs=true`,{headers:o}),a=i?.siblings||i?.files||[],c=Number(r);v=0;for(let e=1;e<=c;e+=1){let t=String(e).padStart(5,`0`),n=s.replace(g,`-${t}-of-${r}.gguf`),i=a.find(e=>(e.rfilename||e.path||e.filename)===n),o=Number(i?.size);Number.isFinite(o)&&o>0&&(v+=o)}}else{let e=await Vt(h,{headers:o});v=Number(e.headers.get(`content-length`))||null}let y={repoId:t,revision:n,filename:s,url:h,size:v,quantization:c,headers:o,isSplit:!!_,splitCount:_?Number(_[2]):0};return await bt(i,`artifact-info`,y,r),y},Kt=/^mmproj-.*\.gguf$/i,qt=async(e,t)=>{if(!e.model.enable_mtmd)return null;let n=e.runtime.cache_dir,r={...e.runtime.http_headers||{}};e.runtime.huggingface_token&&(r.Authorization=`Bearer ${e.runtime.huggingface_token}`);let i=t?.repoId||e.model.repo_id,a=t?.revision||e.model.revision||`main`,o=JSON.stringify({kind:`mmproj`,repoId:i,revision:a,mmproj_filename:e.model.mmproj_filename,mmproj_url:e.model.mmproj_url,mmproj_local_path:e.model.mmproj_local_path}),s=await yt(o,`artifact-info`,n);if(s)return s;if(e.model.mmproj_url){let t=await Vt(e.model.mmproj_url,{headers:r}),s=Number(t.headers.get(`content-length`))||null,c={repoId:i,revision:a,filename:e.model.mmproj_filename||e.model.mmproj_url.split(`/`).pop(),url:e.model.mmproj_url,size:s,headers:r};return await bt(o,`artifact-info`,c,n),c}if(e.model.mmproj_local_path){if(!e.model.allow_local_file)throw Error("`model.mmproj_local_path` requires `model.allow_local_file = true`");let t={repoId:i,revision:a,filename:y.basename(e.model.mmproj_local_path),url:null,size:null,headers:r,localPath:y.resolve(e.model.mmproj_local_path)};return await bt(o,`artifact-info`,t,n),t}if(!i)throw Error("Cannot derive mmproj artifact without `model.repo_id`");let c=await Bt(`${e.model.api_base}/models/${i}?revision=${a}&blobs=true`,{headers:r}),l=c?.siblings||c?.files||[],u=l.map(e=>e.rfilename||e.path||e.filename).filter(e=>typeof e==`string`),d=e.model.mmproj_filename;if(d){if(!u.includes(d))throw Error(`mmproj file "${d}" not found in repo ${i}`)}else{let e=u.filter(e=>Kt.test(e));if(e.length===0)return console.warn(`[buttress] enable_mtmd set but no mmproj file found in ${i}; skipping multimodal load`),null;let n=t?.quantization&&String(t.quantization).toLowerCase();d=n&&e.find(e=>e.toLowerCase().includes(n))||e[0]}let f=`${e.model.base_url.replace(/\/+$/,``)}/${i}/resolve/${a}/${d}`,p=l.find(e=>(e.rfilename||e.path||e.filename)===d),m=Number(p?.size);if(!Number.isFinite(m)||m<=0){let e=await Vt(f,{headers:r});m=Number(e.headers.get(`content-length`))||null}let h={repoId:i,revision:a,filename:d,url:f,size:m,headers:r};return await bt(o,`artifact-info`,h,n),h},Jt=(e,t)=>{if(t?.localPath)return t.localPath;if(!t)return null;let n=t.repoId.split(`/`),r=y.join(e.runtime.cache_dir,...n,t.revision);return y.join(r,t.filename)},Yt=async(e,{modelBytes:t=null,kvCacheBytes:n=null}={})=>{let r=gt(e),[i,...a]=r,o=e.backend?.gpu_memory_fraction==null?dt.backend.gpu_memory_fraction||1:Math.min(1,Math.max(0,Number(e.backend.gpu_memory_fraction))),s=e.backend?.cpu_memory_fraction==null?ut:Math.min(1,Math.max(0,Number(e.backend.cpu_memory_fraction))),c=await Oe({platform:process.platform,totalMemoryInBytes:b.totalmem(),backend:`ggml-llm`,variant:i||null,preferVariants:a,gpuMemoryFraction:o,cpuMemoryFraction:s,dependencies:{getBackendDevicesInfo:S,isLibVariantAvailable:C},modelBytes:t,kvCacheBytes:n}),l=e=>({...e,devices:Array.isArray(e.devices)?e.devices:[],ok:e.ok,hasGpu:!!e.hasGpu,totalMemory:e.gpuTotalBytes||e.totalMemory||0,error:e.ok?null:Error(e.error||`Variant ${e.variant} not available on this platform`)});if(!c.ok||!c.selected){let e=(c.attempts||[]).map(e=>`${e.variant}: ${e.error||`unknown error`}`).join(`; `);throw Error(`Unable to initialize any backend variant (${r.join(`, `)}). Errors: ${e}`)}let u=(c.attempts||[]).map(l);return{selected:l(c.selected),attempts:u}},Xt=async e=>{let t=await Gt(e),n=await qt(e,t),r=await Ht(t.url,t.headers,e.runtime.cache_dir),{arch:i,nCtxTrain:a,nLayer:o,nEmbd:s,nHead:c,nHeadKv:l,nEmbdHeadK:u,nEmbdHeadV:d,quantVersion:f,fileType:p,attentionLayerCount:m,recurrentLayerCount:h,ssmDConv:g,ssmDState:_,ssmDInner:v,ssmNGroup:y,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C}=Ge(r),w=Number.isFinite(Number(o))?Number(o):0,T=Number.isFinite(Number(s))?Number(s):0,E=Number.isFinite(Number(c))?Number(c):0,D=Number.isFinite(Number(l))?Number(l):E,ee=E>0&&T>0?T/E:128,te=u!=null&&Number.isFinite(Number(u))?Number(u):ee,O=d!=null&&Number.isFinite(Number(d))?Number(d):ee,k=Re({arch:i,metadata:r,nLayer:w}),A=k&&Number.isFinite(Number(k.kvLayers))?Number(k.kvLayers):w,j=Math.max(0,Math.floor(Number(A)||0)),M={use_mmap:e.model.use_mmap??e.runtime.use_mmap,use_mlock:e.model.use_mlock??e.runtime.use_mlock,no_extra_bufts:e.model.no_extra_bufts??e.runtime.no_extra_bufts,n_threads:e.model.n_threads??e.runtime.n_threads,n_ctx:e.model.n_ctx??e.runtime.n_ctx,n_batch:e.model.n_batch??e.runtime.n_batch,n_ubatch:e.model.n_ubatch??e.runtime.n_ubatch,n_cpu_moe:e.model.n_cpu_moe??e.runtime.n_cpu_moe,n_parallel:(e.model.n_parallel??e.runtime.n_parallel)||4,cpu_mask:e.model.cpu_mask??e.runtime.cpu_mask,cpu_strict:e.model.cpu_strict??e.runtime.cpu_strict,devices:e.model.devices??e.runtime.devices,n_gpu_layers:e.model.n_gpu_layers??e.runtime.n_gpu_layers,flash_attn_type:e.model.flash_attn_type??e.runtime.flash_attn_type,cache_type_k:e.model.cache_type_k??e.runtime.cache_type_k,cache_type_v:e.model.cache_type_v??e.runtime.cache_type_v,kv_unified:e.model.kv_unified??e.runtime.kv_unified,swa_full:e.model.swa_full??e.runtime.swa_full,ctx_shift:e.model.ctx_shift??e.runtime.ctx_shift},N=M.n_ctx?Number(M.n_ctx):null,P=N||a||4096,F=[],I=[],ne=!0;if(N&&a&&N>a){ne=!1;let e=`Requested context length (${N}) exceeds model training context (${a})`;F.push(e),I.push(e),P=a}N&&!a&&F.push(`Model metadata missing training context length, using requested value`);let re={k:M.cache_type_k,v:M.cache_type_v},L=t.size>0?t.size:0,R=Ke({layerCount:j,headKvCount:D,embdHeadKCount:te,embdHeadVCount:O,cacheTypes:re,swaConfig:k,kvUnified:M.kv_unified,nParallel:M.n_parallel,swaFull:M.swa_full,arch:i,attentionLayerCount:m}),ie=qe({nLayer:w,nEmbd:T,recurrentLayerCount:h,nSeqMax:M.n_parallel||4,ssmDConv:g,ssmDState:_,ssmDInner:v,ssmNGroup:y,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C,arch:i}),z=await Yt(e,{modelBytes:L,kvCacheBytes:R(P)+ie}),ae=z.selected.totalMemory||0,oe=ae*(e.backend.gpu_memory_fraction||1),se=e.backend.cpu_memory_fraction==null?ut:Math.min(1,Math.max(0,Number(e.backend.cpu_memory_fraction))),ce=Math.max(0,b.totalmem()*se),le=z.selected.hasGpu?oe:ce,B=Je({maxCtx:P,availableMemory:le,modelBytes:L,kvBytesForCtx:R});if(!N&&B){let e=a?Math.min(B,a):B,t=Math.max(32,e);t<P&&F.push(`Context length capped to ${t} by memory limits`),P=t}P>B&&(P=B);let ue=Math.floor(B);console.log(`[buttress] Memory-limited context length: ${ue}`);let de=R(P),fe=L+de+ie,pe=w?L/(w+1):L,me=0;z.selected.hasGpu&&pe>0&&(me=Math.min(w+1,Math.max(0,Math.floor(oe/pe)))),console.log(`[buttress] Auto GPU layer capacity (${z.selected.variant}): ${me}/${w+1}`);let he;he=M.n_gpu_layers===`auto`||M.n_gpu_layers==null?me:Math.max(0,Math.min(Number(M.n_gpu_layers)||0,w+1));let ge=(()=>{let e=M.flash_attn_type&&String(M.flash_attn_type).toLowerCase();return e===`on`||e===`off`?e:z.selected.hasGpu?`auto`:`off`})(),_e=e.runtime.cache_dir,ve=Ut(e,t),ye=await K(ve,t.size),be=Jt(e,n),xe=be?await K(be,n?.size):!1,Se=n?{enabled:!0,initialized:!1,filename:n.filename,url:n.url,sizeBytes:n.size,localPath:be,exists:xe,useGpu:e.model.mmproj_use_gpu,imageMinTokens:e.model.mmproj_image_min_tokens,imageMaxTokens:e.model.mmproj_image_max_tokens}:{enabled:!1,requested:!!e.model.enable_mtmd};return{config:e,info:{ok:ne,backend:`ggml-llm`,warnings:F,errors:I,model:{repoId:t.repoId,revision:t.revision,filename:t.filename,quantization:t.quantization,url:t.url,sizeBytes:t.size,metadata:{architecture:i,n_ctx_train:a,n_layer:w,n_embd:T,quantization_version:f,file_type:p,kv_layer_count:j,swa:k?.enabled?{window:k.window,pattern:k.pattern,dense_first:k.denseFirst,type:k.type,layers:k.swaLayers}:null}},runtime:{...M,variant:z.selected.variant,n_ctx:P,requested_ctx:N,n_gpu_layers:he,auto_gpu_layers:me,flash_attn_type:ge,cache_type_k:re.k,cache_type_v:re.v,estimated_max_n_ctx:ue},resources:{modelBytes:L,kvCacheBytes:de,recurrentMemoryBytes:ie,totalEstimatedBytes:fe,gpuCapacityBytes:ae,gpuUsableBytes:oe,cpuUsableBytes:ce,fit:z.selected.fit},devices:{selected:z.selected,attempts:z.attempts},download:{cacheDir:_e,localPath:ve,exists:ye},multimodal:Se,timestamp:new Date().toISOString()},artifact:t,mmprojArtifact:n,mmprojLocalPath:be,mmprojLocalExists:xe,metadata:{arch:i,nCtxTrain:a,nLayer:w,nEmbd:T},devices:z,cacheTypes:re,localPath:ve,localExists:ye}},Zt=(e,t,n=null,r=null)=>{let i,a=Date.now(),o=0;return new it({async start(s){try{let c=await e.parallel.completion(t,(e,t)=>{t&&(t.token&&(o+=1),s.enqueue({event:`token`,data:{requestId:e,...t}}))}),{requestId:l}=c;i=c.stop;let u=await c.promise;console.log(`[Completion] Result:`,u),s.enqueue({event:`result`,data:{requestId:l,...u}}),s.close();let d=Date.now()-a,f=u.timings||{};H.addCompletion({id:`completion-${l}`,generatorId:n,requestId:l,repoId:r?.repoId||null,quantization:r?.quantization||null,variant:r?.variant||null,cacheTokens:f.cache_n??0,promptTokens:f.prompt_n??0,tokensGenerated:f.predicted_n??o,tokensPerSecond:f.predicted_per_second??0,promptPerSecond:f.prompt_per_second??0,durationMs:d,success:!0,interrupted:u.interrupted||!1,contextFull:u.context_full||u.contextFull||!1})}catch(e){s.enqueue({event:`error`,data:{message:e?.message||String(e)}}),s.error(e),H.addCompletion({id:`completion-${Date.now()}`,generatorId:n,repoId:r?.repoId||null,quantization:r?.quantization||null,variant:r?.variant||null,durationMs:Date.now()-a,tokensGenerated:o,success:!1,error:e?.message||String(e)})}},cancel(){i&&i()}})},Qt=(e,t,n,r,i,a,o=null,s=null,c=null)=>{let l,u=``,d=!1,f=Date.now(),p=0,m=()=>{i&&_(i).catch(()=>{}),c&&_(c).catch(()=>{})};return new it({async start(h){try{let g=await e.parallel.completion(t,(e,t)=>{t&&(t.token&&(u+=t.token,p+=1),h.enqueue({event:`token`,data:{requestId:e,...t}}))}),{requestId:_}=g;l=g.stop;let v=await g.promise;v.text?u=v.text:v.content&&(u=v.content),d=!v.interrupted&&!v.context_full,console.log(`[Completion] Result:`,v),h.enqueue({event:`result`,data:{requestId:_,...v}}),h.close();let y=Date.now()-f,b=v.timings||{};H.addCompletion({id:`completion-${_}`,generatorId:o,requestId:_,repoId:s?.repoId||null,quantization:s?.quantization||null,variant:s?.variant||null,cacheTokens:b.cache_n??0,promptTokens:b.prompt_n??a??0,tokensGenerated:b.predicted_n??p,tokensPerSecond:b.predicted_per_second??0,promptPerSecond:b.prompt_per_second??0,durationMs:y,success:!0,interrupted:v.interrupted||!1,contextFull:v.context_full||v.contextFull||!1,usedCache:!!t.load_state_path}),d&&n.enabled&&u?n.saveCompletionState(r,u,i,a,c).catch(e=>{console.warn(`[SessionCache] Save failed:`,e.message)}):m()}catch(e){h.enqueue({event:`error`,data:{message:e?.message||String(e)}}),h.error(e),H.addCompletion({id:`completion-${Date.now()}`,generatorId:o,repoId:s?.repoId||null,quantization:s?.quantization||null,variant:s?.variant||null,durationMs:Date.now()-f,tokensGenerated:p,success:!1,error:e?.message||String(e)}),m()}},cancel(){l&&l(),m()}})},$t=e=>{let t={model:e.plan.localPath,runtime:e.plan.info.runtime};return s(`sha256`).update(JSON.stringify(t)).digest(`hex`).slice(0,24)},en=async(e,t,n,r=null)=>{let{config:i,localPath:a,artifact:o}=e;if(e.localExists&&!t.has(a))return e.info.download.exists=!0,typeof n==`function`&&n(.5),a;if(i.model.local_path&&!i.model.allow_local_file)throw Error("Local model path provided but `model.allow_local_file` is not enabled");let s=a;if(r){let t=r.getDownload(s);if(t){console.log(`[ensureModelFile] Waiting for global download: ${o.repoId}`);try{if(await t,await K(a,o.size))return e.localExists=!0,e.info.download.exists=!0,typeof n==`function`&&n(.5),a}catch(e){console.warn(`[ensureModelFile] Global download failed, will retry: ${e.message}`)}}}t.has(s)||t.set(s,(async()=>{if(o.isSplit&&o.splitCount>0){let e=/-(\d{5})-of-(\d{5})\.gguf$/,t=y.dirname(a),r=o.splitCount,s=0;for(let a=1;a<=r;a+=1){let c=String(a).padStart(5,`0`),l=o.filename.replace(e,`-${c}-of-${String(r).padStart(5,`0`)}.gguf`),u=`${i.model.base_url.replace(/\/+$/,``)}/${o.repoId}/resolve/${o.revision}/${l}`,d=y.join(t,l);await K(d)||await Wt(u,o.headers,d,null,e=>{if(e>=0&&Number.isFinite(e)){let t=(s+e)/r,i=Math.round(t*100);console.log(`Downloading model splits: ${Math.min(100,i)}%`),typeof n==`function`&&n(t*.5)}}),s+=1}}else console.log(`Downloading model: 0%`),await Wt(o.url,o.headers,a,o.size,e=>{if(e>=0&&Number.isFinite(e)){let t=Math.round(e*100);console.log(`Downloading model: ${Math.min(100,t)}%`),typeof n==`function`&&n(e*.5)}});e.localExists=!0,e.info.download.exists=!0})());try{await t.get(s)}finally{t.delete(s)}return a},tn=async(e,t,n,r=null)=>{let{mmprojArtifact:i,mmprojLocalPath:a}=e;if(!i||!a)return null;if(i.localPath){if(!await K(a))throw Error(`mmproj local file not found: ${a}`);return e.mmprojLocalExists=!0,e.info.multimodal.exists=!0,typeof n==`function`&&n(1),a}if(e.mmprojLocalExists&&!t.has(a))return e.info.multimodal.exists=!0,typeof n==`function`&&n(1),a;let o=a;if(r){let t=r.getDownload(o);if(t)try{if(await t,await K(a,i.size))return e.mmprojLocalExists=!0,e.info.multimodal.exists=!0,typeof n==`function`&&n(1),a}catch(e){console.warn(`[ensureMmprojFile] Global download failed, will retry: ${e.message}`)}}t.has(o)||t.set(o,(async()=>{console.log(`Downloading mmproj: 0%`),await Wt(i.url,i.headers,a,i.size,e=>{if(e>=0&&Number.isFinite(e)){let t=Math.round(e*100);console.log(`Downloading mmproj: ${Math.min(100,t)}%`),typeof n==`function`&&n(e)}}),e.mmprojLocalExists=!0,e.info.multimodal.exists=!0})());try{await t.get(o)}finally{t.delete(o)}return a},nn=async(e,t)=>{let n=$t(e),r=e.contexts.get(n);if(r&&!r.released)return r.releaseTimer&&(clearTimeout(r.releaseTimer),r.releaseTimer=null,console.log(`[Context] Cancelled pending release for context "${n}"`)),r.releaseRequested=!1,r.refCount+=1,console.log(`[Context] Reusing existing context "${n}", refCount=${r.refCount}`),typeof t==`function`&&t(0),r.context||await r.ready,typeof t==`function`&&t(1),r;r?console.log(`[Context] Record exists but released=${r.released}, creating new context`):console.log(`[Context] No existing record for "${n}", creating new context`),r={key:n,refCount:1,ready:null,released:!1},e.contexts.set(n,r),r.ready=(async()=>{let i=Date.now(),a=await en(e.plan,e.downloads,t,e.globalDownloadManager);typeof t==`function`&&t(.5);let o={model:a,n_threads:e.plan.info.runtime.n_threads,use_mmap:e.plan.info.runtime.use_mmap,use_mlock:e.plan.info.runtime.use_mlock,no_extra_bufts:e.plan.info.runtime.no_extra_bufts,cpu_mask:e.plan.info.runtime.cpu_mask,cpu_strict:e.plan.info.runtime.cpu_strict,devices:e.plan.info.runtime.devices,n_ctx:e.plan.info.runtime.n_ctx,n_gpu_layers:e.plan.info.runtime.n_gpu_layers,n_parallel:e.plan.info.runtime.n_parallel,n_batch:e.plan.info.runtime.n_batch,n_ubatch:e.plan.info.runtime.n_ubatch,n_cpu_moe:e.plan.info.runtime.n_cpu_moe,flash_attn_type:e.plan.info.runtime.flash_attn_type,ctx_shift:e.plan.info.runtime.ctx_shift,kv_unified:e.plan.info.runtime.kv_unified,swa_full:e.plan.info.runtime.swa_full,lib_variant:e.plan.info.runtime.variant};e.plan.info.runtime.flash_attn_type!==`off`&&(o.cache_type_k=e.plan.info.runtime.cache_type_k,o.cache_type_v=e.plan.info.runtime.cache_type_v),console.log(`[Context] Load Options:`,o);let s;try{if(s=await w(o,e=>{typeof t==`function`&&(t(.5+e*.25),e%5==0&&console.log(`[Context] Load Model Progress:`,e))}),e.plan.info.runtime.n_parallel&&!await s.parallel.enable({n_parallel:e.plan.info.runtime.n_parallel,n_batch:e.plan.info.runtime.n_batch}))throw Error(`Failed to enable parallel decoding mode for context`);if(e.plan.mmprojArtifact){let t=await tn(e.plan,e.downloads,null,e.globalDownloadManager);if(t){let n=e.config.model.mmproj_use_gpu,r={path:t,use_gpu:n==null?(e.plan.info.runtime.n_gpu_layers||0)>0:!!n,image_min_tokens:e.config.model.mmproj_image_min_tokens,image_max_tokens:e.config.model.mmproj_image_max_tokens};console.log(`[Context] initMultimodal:`,r),await s.initMultimodal(r)?e.plan.info.multimodal.initialized=!0:console.warn(`[Context] initMultimodal returned false; multimodal disabled`)}}return typeof t==`function`&&t(1),r.context=s,r.modelInfo=s.getModelInfo(),H.addModelLoad({id:`${e.id}-${n}`,generatorId:e.id,contextKey:n,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,variant:e.plan.info.runtime?.variant||null,nCtx:e.plan.info.runtime?.n_ctx||null,nGpuLayers:e.plan.info.runtime?.n_gpu_layers||null,durationMs:Date.now()-i,success:!0}),r}catch(t){if(H.addModelLoad({id:`${e.id}-${n}`,generatorId:e.id,contextKey:n,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,variant:e.plan.info.runtime?.variant||null,durationMs:Date.now()-i,success:!1,error:t?.message||String(t)}),s)try{s.release()}catch{}throw t}})();try{return await r.ready,r}catch(t){throw e.contexts.delete(n),t}},rn=async(e,t,n=!1)=>{if(t.released||!n&&t.refCount>0)return!1;t.released=!0,e.contexts.delete(t.key);try{t.context?.parallel?.disable?.()}catch{}return await t.context?.release?.(),!0},an=async(e,t,n=!1)=>{if(t.releaseRequested=!0,t.releaseTimer&&=(clearTimeout(t.releaseTimer),null),n)t.refCount=0;else if(t.refCount=Math.max(0,t.refCount-1),t.refCount>0)return t.releaseRequested=!1,!1;let r=e.config.runtime.context_release_delay_ms;if(typeof r!=`number`||!Number.isFinite(r))return rn(e,t);let i=Math.max(0,Math.floor(r));return n||i<=0?rn(e,t):(console.log(`[Context] Scheduling release in ${i}ms for context "${t.key}"`),t.releaseTimer=setTimeout(async()=>{if(t.releaseTimer=null,t.refCount>0){console.log(`[Context] Release cancelled, refCount=${t.refCount} for context "${t.key}"`),t.releaseRequested=!1;return}console.log(`[Context] Releasing context "${t.key}" after ${i}ms delay`),await rn(e,t)},i),!0)};async function on(e,t,n={}){let{globalDownloadManager:r=null}=n,i=mt(t),a=await Xt(i),o=new zt(i,a);await o.initialize();let s={id:e,type:`ggml-llm`,config:i,plan:a,info:a.info,contexts:new Map,downloads:new Map,globalDownloadManager:r,sessionCache:o,finalized:!1};return{id:e,type:`ggml-llm`,info:a.info,contexts:s.contexts,initContext:async(e={})=>{let{onProgress:t}=e,n=await nn(s,t);return s.sessionCache.updateModelInfo(n.modelInfo),{modelInfo:n.modelInfo?{...n.modelInfo}:null,runtime:{...s.plan.info.runtime},download:{...s.plan.info.download},multimodal:s.plan.info.multimodal?{...s.plan.info.multimodal}:null}},completion:async(e={})=>{let{options:t={},useCache:n=!0}=e,r=$t(s),i=s.contexts.get(r);if(!i)throw Error(`Context "${r}" not initialized`);await i.ready;let a=t.prompt||``,o=null,c=null;if(!a&&t.messages){({messages:o}=t),c={chatTemplate:t.chat_template||t.chatTemplate,jinja:t.jinja??!0,tools:t.tools,parallel_tool_calls:t.parallel_tool_calls,tool_choice:t.tool_choice,reasoning_format:t.reasoning_format,enable_thinking:t.enable_thinking,add_generation_prompt:t.add_generation_prompt,now:t.now,chat_template_kwargs:t.chat_template_kwargs,force_pure_content:t.force_pure_content};let e=await i.context.getFormattedChat(o,c.chatTemplate,c);a=e?.prompt||e||``}if(n&&s.sessionCache.enabled&&a){let{options:e}=await s.sessionCache.prepareCompletionOptions(t,a,i.context),n=await s.sessionCache.generateTempStatePath(),r=(await i.context.tokenize(a))?.tokens?.length||0,o={...e,save_state_path:n},c=s.sessionCache.requiresExactMatch(),l=!!o.load_state_path,u=null;c&&!l&&(u=await s.sessionCache.generateTempStatePath(),o.save_prompt_state_path=u);let d={repoId:s.plan.info.model?.repoId||null,quantization:s.plan.info.model?.quantization||null,variant:s.plan.info.runtime?.variant||null};return Qt(i.context,o,s.sessionCache,a,n,r,s.id,d,u)}let l={repoId:s.plan.info.model?.repoId||null,quantization:s.plan.info.model?.quantization||null,variant:s.plan.info.runtime?.variant||null};return Zt(i.context,t,s.id,l)},tokenize:async(e={})=>{let{text:t=``,params:n={}}=e,r=$t(s),i=s.contexts.get(r);if(!i)throw Error(`Context "${r}" not initialized`);await i.ready;let a=await i.context.tokenize(t,n);if(!a)return{tokens:[]};let o=Array.from(a.tokens??[],Number);return{...a,tokens:o}},detokenize:async(e={})=>{let{tokens:t=[]}=e,n=$t(s),r=s.contexts.get(n);if(!r)throw Error(`Context "${n}" not initialized`);await r.ready;let i=t.map(e=>Number(e));return r.context.detokenize(i)},applyChatTemplate:async(e={})=>{let{messages:t=[],template:n,params:r}=e,i=$t(s),a=s.contexts.get(i);if(!a)throw Error(`Context "${i}" not initialized`);return await a.ready,await a.context.getFormattedChat(t,n,r)},releaseContext:async()=>{if(s.finalized)return!1;let e=$t(s),t=s.contexts.get(e);return t?an(s,t,!1):!1},finalize:async()=>{if(s.finalized)return;s.finalized=!0;let e=Array.from(s.contexts.values()),t=e.map(e=>e.released||e.releaseRequested||e.releaseTimer||(e.refCount=Math.max(0,e.refCount-1),e.refCount>0)?Promise.resolve(!1):rn(s,e));await Promise.allSettled(t),(e.length===0||e.every(e=>e.released))&&await s.sessionCache.cleanup()},getStatus:()=>{let e=[],t=Array.from(s.contexts.entries()).map(([t,n])=>{let r={key:t,refCount:n.refCount,hasModel:!!n.context},i=n.context.parallel.getStatus();return r.parallelStatus=i,e.push({contextKey:t,...i}),r});return{id:s.id,type:s.type,repoId:s.plan.info.model?.repoId||null,quantization:s.plan.info.model?.quantization||null,variant:s.plan.info.runtime?.variant||null,nCtx:s.plan.info.runtime?.n_ctx||null,nParallel:s.plan.info.runtime?.n_parallel||null,contexts:t,parallelStatuses:e}},subscribeParallelStatus:e=>{let t=Array.from(s.contexts.entries()).map(([t,n])=>n.context.parallel.subscribeToStatus(n=>{e({contextKey:t,...n})}));return{remove:()=>{t.forEach(e=>{e?.remove&&e.remove()})}}},hasPendingReleases:()=>Array.from(s.contexts.values()).some(e=>!e.released&&(e.releaseRequested||e.releaseTimer||e.refCount>0)),resetFinalized:()=>{s.finalized=!1}}}const sn=e=>{let t=mt(e);return t.model.repo_id||t.model.repository||t.model.model||null};async function cn(e,t,n={}){let{onProgress:r,onComplete:i,onError:a}=n;try{let n=mt(e),o=await Gt(n),s=Ut(n,o),{repoId:c}=o,l=await qt(n,o).catch(e=>(console.warn(`[Download] Failed to derive mmproj artifact: ${e.message}`),null)),u=Jt(n,l),d=async()=>{if(!l||!u||l.localPath)return;if(await K(u,l.size)){console.log(`[Download] mmproj already exists: ${u}`);return}let e=t.getDownload(u);if(e){await e;return}let n=(async()=>{try{await Wt(l.url,l.headers,u,l.size,e=>{e>=0&&Number.isFinite(e)&&console.log(`[Download] mmproj ${c}: ${Math.round(e*100)}%`)})}finally{t.deleteDownload(u)}})();t.setDownload(u,n),await n};if(await K(s,o.size))return console.log(`[Download] Model already exists: ${c} at ${s}`),await d().catch(e=>{console.error(`[Download] mmproj download failed: ${e.message}`),typeof a==`function`&&a(e)}),typeof i==`function`&&i({localPath:s,repoId:c,alreadyExists:!0}),{started:!1,localPath:s,repoId:c,alreadyExists:!0};let f=t.getDownload(s);if(f)return console.log(`[Download] Already downloading: ${c}`),f.then(()=>{typeof i==`function`&&i({localPath:s,repoId:c,joinedExisting:!0})}).catch(e=>{typeof a==`function`&&a(e)}),{started:!1,localPath:s,repoId:c,alreadyDownloading:!0};console.log(`[Download] Starting download: ${c}`);let p=(async()=>{try{if(o.isSplit&&o.splitCount>0){let e=/-(\d{5})-of-(\d{5})\.gguf$/,t=y.dirname(s),i=o.splitCount,a=0;for(let s=1;s<=i;s+=1){let l=String(s).padStart(5,`0`),u=o.filename.replace(e,`-${l}-of-${String(i).padStart(5,`0`)}.gguf`),d=`${n.model.base_url.replace(/\/+$/,``)}/${o.repoId}/resolve/${o.revision}/${u}`,f=y.join(t,u);await K(f)||await Wt(d,o.headers,f,null,e=>{if(e>=0&&Number.isFinite(e)){let t=(a+e)/i;console.log(`[Download] ${c}: ${Math.round(t*100)}%`),typeof r==`function`&&r(t)}}),a+=1}}else await Wt(o.url,o.headers,s,o.size,e=>{e>=0&&Number.isFinite(e)&&(console.log(`[Download] ${c}: ${Math.round(e*100)}%`),typeof r==`function`&&r(e))});await d(),console.log(`[Download] Completed: ${c}`),typeof i==`function`&&i({localPath:s,repoId:c})}catch(e){throw console.error(`[Download] Failed: ${c}`,e.message),typeof a==`function`&&a(e),e}finally{t.deleteDownload(s)}})();return t.setDownload(s,p),{started:!0,localPath:s,repoId:c}}catch(e){return console.error(`[Download] Failed to start download:`,e.message),typeof a==`function`&&a(e),{started:!1,localPath:null,repoId:null,error:e.message}}}async function ln(e){let t=mt(e),n=await Gt(t),r=await Ht(n.url,n.headers,t.runtime.cache_dir),{arch:i,nCtxTrain:a,nLayer:o,nEmbd:s,nHead:c,nHeadKv:l,nEmbdHeadK:u,nEmbdHeadV:d,quantVersion:f,fileType:p,attentionLayerCount:m,recurrentLayerCount:h,ssmDConv:g,ssmDState:_,ssmDInner:v,ssmNGroup:y,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C}=Ge(r),w=Number.isFinite(Number(o))?Number(o):0,T=Number.isFinite(Number(s))?Number(s):0,E=Number.isFinite(Number(c))?Number(c):0,D=Number.isFinite(Number(l))?Number(l):E,ee=E>0&&T>0?T/E:128,te=u!=null&&Number.isFinite(Number(u))?Number(u):ee,O=d!=null&&Number.isFinite(Number(d))?Number(d):ee,k=Re({arch:i,metadata:r,nLayer:w}),A=k&&Number.isFinite(Number(k.kvLayers))?Number(k.kvLayers):w,j=Math.max(0,Math.floor(Number(A)||0)),M=(t.model.n_ctx?Number(t.model.n_ctx):null)||a||4096,N={k:t.model.cache_type_k,v:t.model.cache_type_v},P=n.size>0?n.size:0,F=t.model.n_parallel||4,I=Ke({layerCount:j,headKvCount:D,embdHeadKCount:te,embdHeadVCount:O,cacheTypes:N,swaConfig:k,kvUnified:t.model.kv_unified,nParallel:F,swaFull:t.model.swa_full,arch:i,attentionLayerCount:m}),ne=qe({nLayer:w,nEmbd:T,recurrentLayerCount:h,nSeqMax:F,ssmDConv:g,ssmDState:_,ssmDInner:v,ssmNGroup:y,ssmDtRank:x,rwkvHeadSize:S,rwkvTokenShiftCount:C,arch:i}),re=t.backend?.gpu_memory_fraction==null?dt.backend.gpu_memory_fraction||1:Math.min(1,Math.max(0,Number(t.backend.gpu_memory_fraction))),L=t.backend?.cpu_memory_fraction==null?ut:Math.min(1,Math.max(0,Number(t.backend.cpu_memory_fraction))),R=await Yt(t,{modelBytes:P,kvCacheBytes:I(M)}),ie=(R.selected.totalMemory||0)*re,z=Math.max(0,b.totalmem()*L),ae=Je({maxCtx:M,availableMemory:R.selected.hasGpu?ie:z,modelBytes:P,kvBytesForCtx:I}),oe=I(M),se=I(ae);return{kvInfo:{nCtxTrain:a,nLayer:w,nEmbd:T,nHeadKv:D,nEmbdHeadK:te,nEmbdHeadV:O,nHeadCount:E,nHeadKvCount:D,kvLayerCount:j,swa:k?.enabled?{window:k.window,pattern:k.pattern,denseFirst:k.denseFirst,type:k.type,layers:k.swaLayers}:null},modelBytes:P,kvCacheBytes:oe,limitedKvCacheBytes:se,memoryLimitedCtx:ae,recurrentMemoryBytes:ne,quantization:{name:n.quantization||null,fileType:p,version:f}}}const un=e=>e?typeof e.score==`number`&&Number.isFinite(e.score)?Number(e.score):B(e):0;async function dn(e=null,t={}){let{threshold:n=1.1,includeBreakdown:r=!1,config:i,...a}=t,o=null,s=null,c=null,l=null,u=null,d=null,f=null;if(i)try{let{modelBytes:e,kvCacheBytes:t,limitedKvCacheBytes:n,memoryLimitedCtx:r,recurrentMemoryBytes:a,kvInfo:p,quantization:m}=await ln(i);o=e,s=t,c=n,l=r,u=a,d=p,f=m}catch{}let p=i?.backend?.gpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.gpu_memory_fraction))),m=i?.backend?.cpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.cpu_memory_fraction))),h=await Oe({...a,platform:process.platform,totalMemoryInBytes:b.totalmem(),backend:`ggml-llm`,includeBreakdown:r,gpuMemoryFraction:p,cpuMemoryFraction:m,dependencies:{getBackendDevicesInfo:S,isLibVariantAvailable:C},modelBytes:o,kvCacheBytes:s,limitedKvCacheBytes:c}),g=h.selected,_=un(g);g.modelBytes=o||null,g.kvCacheBytes=s||null,g.memoryLimitedCtx=l||null,g.limitedKvCacheBytes=c||null,g.recurrentMemoryBytes=u||null,g.kvInfo=d||null,g.quantization=f||null;let v=null,y=null;if(e){let t=un(e);y={...e,score:t};let r=`buttress`,i=`buttress-higher-score`;if(!h.ok)r=`local`,i=`buttress-unavailable`;else if(!t&&t!==0)r=`buttress`,i=`missing-client-score`;else{let e=y.fit,a=y.limitedFit,o=g?.fit,s=g?.limitedFit,c=e?.fitsInGpu||e?.fitsInCpu||a?.fitsInGpu||a?.fitsInCpu,l=o?.fitsInGpu||o?.fitsInCpu||s?.fitsInGpu||s?.fitsInCpu;c&&!l?(r=`local`,i=`client-fits-in-memory`):l&&!c?(r=`buttress`,i=`buttress-fits-in-memory`):t>_*n?(r=`local`,i=`client-better`):_>t*n?(r=`buttress`,i=`buttress-better`):(r=`either`,i=`comparable-scores`)}v={buttressScore:_,clientScore:t,threshold:n,recommendation:r,reason:i}}!h.ok&&!v&&(v={buttressScore:_,clientScore:e?.score??null,threshold:n,recommendation:`local`,reason:`buttress-unavailable`});let x=null;return i&&(x={repoId:i.model?.repo_id||null,quantization:i.model?.quantization||null,nCtx:i.model?.n_ctx||null,cacheKType:i.model?.cache_type_k||`f16`,cacheVType:i.model?.cache_type_v||`f16`}),{type:`ggml-llm`,timestamp:new Date().toISOString(),buttress:h,client:y,comparison:v,modelConfig:x}}const{WritableStream:fn}=typeof globalThis<`u`&&globalThis.ReadableStream&&globalThis.WritableStream?{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream}:c,pn=(e={},t={})=>(Object.entries(t||{}).forEach(([t,n])=>{n&&typeof n==`object`&&!Array.isArray(n)?((!e[t]||typeof e[t]!=`object`)&&(e[t]={}),pn(e[t],n)):e[t]=n}),e),mn=`https://huggingface.co`,hn=`https://huggingface.co/api`,gn=y.join(b.homedir(),`.buttress`,`models`),_n=[`cuda`,`vulkan`,`default`],vn=[`q8_0`,`q5_1`,`q5_0`,`q4_1`,`q4_0`],yn=`fp16`,bn=.5,xn=[`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`],Sn=e=>{if(!e)return null;let t=e.toLowerCase();return xn.find(e=>t.includes(e))||null},Cn={backend:{type:`ggml-stt`,variant:null,variant_preference:_n,gpu_memory_fraction:.85,cpu_memory_fraction:bn},model:{repo_id:`BricksDisplay/whisper-ggml`,revision:`main`,filename:null,url:null,quantization:null,preferred_quantizations:[`q8_0`,yn,`q5_1`],allow_local_file:!1,local_path:null,api_base:hn,base_url:mn,use_gpu:!0,use_flash_attn:`auto`},runtime:{cache_dir:gn,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},max_threads:null,context_release_delay_ms:1e4}},wn=(e,t=[])=>!e&&e!==0?[...t]:Array.isArray(e)?e.filter(e=>e!=null):[e],Tn=e=>{if(!e)return null;let t=String(e).toLowerCase();return[`cuda`,`vulkan`,`default`].includes(t)?t:null},En=(e={})=>{let t=structuredClone(Cn);if(pn(t,e),t.backend.variant=Tn(t.backend.variant),t.backend.variant_preference=Array.from(new Set(wn(t.backend.variant_preference||_n).flatMap(e=>{let t=Tn(e);return t?[t]:[]}))),t.backend.variant_preference.length===0&&(t.backend.variant_preference=[..._n]),t.runtime.prefer_variants=Array.from(new Set(wn(t.runtime.prefer_variants).flatMap(e=>{let t=Tn(e);return t?[t]:[]}))),t.model.preferred_quantizations=Array.from(new Set(wn(t.model.preferred_quantizations||t.model.quantizations).flatMap(e=>{let t=e?String(e).toLowerCase():null;return t?[t]:[]}))),t.model.quantization){let e=String(t.model.quantization).toLowerCase();t.model.preferred_quantizations.includes(e)||t.model.preferred_quantizations.unshift(e)}return t.model.base_url=t.model.base_url||mn,t.model.api_base=t.model.api_base||hn,t.runtime.cache_dir=t.runtime.cache_dir?y.resolve(t.runtime.cache_dir):gn,t.runtime.context_release_delay_ms=Math.max(0,Number(t.runtime.context_release_delay_ms)||Cn.runtime.context_release_delay_ms),t},Dn=e=>{let t=e.toLowerCase();return vn.find(e=>t.includes(e))||null},On=e=>{let t=[];e.backend.variant&&t.push(e.backend.variant),e.runtime.prefer_variants.length>0&&t.push(...e.runtime.prefer_variants),t.push(...e.backend.variant_preference),t.push(`default`);let n=new Set;for(let e of t){let t=Tn(e);t&&n.add(t)}return Array.from(n)},kn=async e=>{await d(e,{recursive:!0})},An=(e=gn)=>y.join(e,`.metadata-cache`),jn=(e,t,n=gn)=>{let r=s(`sha256`).update(e).digest(`hex`);return y.join(An(n),t,`${r}.json`)},Mn=async(e,t,n=gn)=>{try{let r=await p(jn(e,t,n),`utf-8`);return JSON.parse(r)}catch{return null}},Nn=async(e,t,n,r=gn)=>{try{let i=jn(e,t,r);await kn(y.dirname(i)),await v(i,JSON.stringify(n),`utf-8`)}catch{}},Pn=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,t);if(!n.ok){let t=await n.text().catch(()=>``);throw Error(`Failed to fetch ${e}: ${n.status} ${n.statusText} ${t}`.trim())}return n.json()},Fn=async(e,t={})=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);let n=await fetch(e,{...t,method:`HEAD`});if(!n.ok)throw Error(`Failed to fetch headers for ${e}: ${n.status} ${n.statusText}`);return n},In=(e,t)=>{if(e.model.local_path)return y.resolve(e.model.local_path);let n=t.repoId.split(`/`),r=y.join(e.runtime.cache_dir,...n,t.revision);return y.join(r,t.filename)},Ln=async(e,t)=>{try{let n=await g(e);return t?n.size===t:!0}catch{return!1}},Rn=async(e,t,n,r,i)=>{if(typeof fetch!=`function`)throw Error(`Global fetch is not available in this runtime`);await kn(y.dirname(n));let a=await fetch(e,{headers:t});if(!a.ok||!a.body)throw Error(`Failed to download ${e}: ${a.status} ${a.statusText}`);let o=await f(n,`w`),s=Number(a.headers.get(`content-length`))||r||0,c=0,l=.05;try{await a.body.pipeTo(new fn({async write(e){if(await o.write(e),c+=e.byteLength,typeof i==`function`&&s>0){let e=Math.min(1,c/s);for(;e>=l;)i(l),l+=.05}},async close(){await o.close(),typeof i==`function`&&i(1)},async abort(e){throw await o.close().catch(()=>{}),await _(n).catch(()=>{}),e}}))}catch(e){throw await o.close().catch(()=>{}),await _(n).catch(()=>{}),e}if(r){let e=await g(n);if(e.size!==r)throw await _(n).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${r} got ${e.size}`)}},zn=async e=>{let t=e.model.repo_id||e.model.repository||e.model.model;if(!t)throw Error("`model.repo_id` is required in Buttress backend config");let n=e.model.revision||`main`,r=e.runtime.cache_dir,i=JSON.stringify({repoId:t,revision:n,filename:e.model.filename,url:e.model.url,quantization:e.model.quantization,preferred_quantizations:e.model.preferred_quantizations}),a=await Mn(i,`artifact-info`,r);if(a)return a;let o={...e.runtime.http_headers||{}};if(e.runtime.huggingface_token&&(o.Authorization=`Bearer ${e.runtime.huggingface_token}`),e.model.url){let a=await Fn(e.model.url,{headers:o}),s=Number(a.headers.get(`content-length`))||null,c=e.model.filename||e.model.url.split(`/`).pop(),l={repoId:t,revision:n,filename:c,url:e.model.url,size:s,quantization:Dn(c||``),headers:o};return await Nn(i,`artifact-info`,l,r),l}let{filename:s}=e.model,c=e.model.quantization&&String(e.model.quantization).toLowerCase(),l=await Pn(`${e.model.api_base}/models/${t}?revision=${n}&blobs=true`,{headers:o}),u=(l?.siblings||l?.files||[]).map(e=>e.rfilename||e.path||e.filename).filter(e=>typeof e==`string`&&e.endsWith(`.bin`));if(u.length===0)throw Error(`No model artifacts found in repo ${t}`);let d=e.model.preferred_quantizations.length>0?e.model.preferred_quantizations:vn,f=()=>{for(let e of d)if(e===yn){let e=u.find(e=>{let t=e.toLowerCase();return!vn.some(e=>t.includes(e))});if(e)return{filename:e,quantization:null}}else{let t=u.find(t=>t.toLowerCase().includes(e));if(t)return{filename:t,quantization:e}}return null};if(s)c||=Dn(s);else{let{filename:e,quantization:t}=f()||{filename:u[0],quantization:null};s=e,c=t||Dn(s)}let p=`${e.model.base_url.replace(/\/+$/,``)}/${t}/resolve/${n}/${s}`,m=await Fn(p,{headers:o}),h=Number(m.headers.get(`content-length`))||null,g={repoId:t,revision:n,filename:s,url:p,size:h,quantization:c,headers:o,isSplit:!1,splitCount:0};return await Nn(i,`artifact-info`,g,r),g},Bn=async(e,{modelBytes:t=null,processingBytes:n=null}={})=>{let r=On(e),[i,...a]=r,o=e.backend?.gpu_memory_fraction==null?Cn.backend.gpu_memory_fraction||1:Math.min(1,Math.max(0,Number(e.backend.gpu_memory_fraction))),s=e.backend?.cpu_memory_fraction==null?bn:Math.min(1,Math.max(0,Number(e.backend.cpu_memory_fraction))),c=await Oe({platform:process.platform,totalMemoryInBytes:b.totalmem(),backend:`ggml-stt`,variant:i||null,preferVariants:a,variantPreference:e.backend.variant_preference,gpuMemoryFraction:o,cpuMemoryFraction:s,dependencies:{getBackendDevicesInfo:S,isLibVariantAvailable:C},modelBytes:t,kvCacheBytes:n}),l=e=>({...e,devices:Array.isArray(e.devices)?e.devices:[],ok:e.ok,hasGpu:!!e.hasGpu,totalMemory:e.gpuTotalBytes||e.totalMemory||0,error:e.ok?null:Error(e.error||`Variant ${e.variant} not available on this platform`)});if(!c.ok||!c.selected){let e=(c.attempts||[]).map(e=>`${e.variant}: ${e.error||`unknown error`}`).join(`; `);throw Error(`Unable to initialize any backend variant (${r.join(`, `)}). Errors: ${e}`)}let u=(c.attempts||[]).map(l);return{selected:l(c.selected),attempts:u}},Vn=async e=>{let t=await zn(e),n=Me({modelBytes:t.size>0?t.size:0}),r=await Bn(e,{modelBytes:n.modelBytes,processingBytes:n.processingBufferBytes}),i=r.selected.hasGpu&&(r.selected.fit?.fitsInGpu===void 0?!0:r.selected.fit.fitsInGpu);e.model.use_gpu===!1&&(i=!1);let a=e.model.use_flash_attn&&String(e.model.use_flash_attn).toLowerCase(),o;o=a===`on`||a===`true`?!0:a===`off`||a===`false`?!1:i;let s=e.runtime.cache_dir,c=In(e,t),l=await Ln(c,t.size);return{config:e,info:{ok:!0,backend:`ggml-stt`,model:{repoId:t.repoId,revision:t.revision,filename:t.filename,quantization:t.quantization,modelType:Sn(t.filename),url:t.url,sizeBytes:t.size},runtime:{variant:r.selected.variant,use_gpu:i,use_flash_attn:o,max_threads:e.runtime.max_threads?Number(e.runtime.max_threads):null},resources:{...n,gpuCapacityBytes:r.selected.gpuTotalBytes,gpuUsableBytes:r.selected.gpuUsableBytes,cpuUsableBytes:r.selected.cpuUsableBytes,fit:r.selected.fit},devices:{selected:r.selected,attempts:r.attempts},download:{cacheDir:s,localPath:c,exists:l},timestamp:new Date().toISOString()},artifact:t,memory:n,devices:r,localPath:c,localExists:l}},Hn=async(e,t,n,r=null)=>{let{localPath:i,artifact:a,config:o}=e;if(e.localExists)return typeof n==`function`&&n(1),i;if(r){let t=r.getDownload(i);if(t){console.log(`[ensureModelFile] Waiting for global STT download: ${a.repoId}`);try{if(await t,await Ln(i,a.size))return e.localExists=!0,e.info.download.exists=!0,typeof n==`function`&&n(1),i}catch(e){console.warn(`[ensureModelFile] Global STT download failed, will retry: ${e.message}`)}}}let s=t.get(i);if(s)return await s,typeof n==`function`&&n(1),i;let c=(async()=>{if(o.model.allow_local_file){if(!await Ln(i,a.size))throw Error(`Local model file not found: ${i}`);return i}return await Rn(a.url,a.headers,i,a.size,n),i})();t.set(i,c);try{return await c,i}finally{t.delete(i)}};var Un=class{constructor(){this.queue=[],this.processing=!1,this.currentTaskId=null}async enqueue(e,t=null){return new Promise((n,r)=>{this.queue.push({task:e,resolve:n,reject:r,taskId:t}),this.processNext()})}async processNext(){if(this.processing||this.queue.length===0)return;this.processing=!0;let{task:e,resolve:t,reject:n,taskId:r}=this.queue.shift();this.currentTaskId=r;try{t(await e())}catch(e){n(e)}finally{this.processing=!1,this.currentTaskId=null,this.processNext()}}getStatus(){return{processing:this.processing,queuedCount:this.queue.length,currentTaskId:this.currentTaskId}}};const Wn=e=>{if(!e)return null;if(e instanceof ArrayBuffer)return e;if(ArrayBuffer.isView(e))return e.buffer;if(typeof e==`string`){let t=e.startsWith(`data:`)?e.split(`,`)[1]||``:e,n=Buffer.from(t,`base64`);return n.buffer.slice(n.byteOffset,n.byteOffset+n.byteLength)}throw Error(`Unsupported audioData format, expected base64 string or ArrayBuffer`)},Gn=async(e,t)=>{if(e.contextRecord&&!e.contextRecord.released)return e.contextRecord.releaseTimer&&(clearTimeout(e.contextRecord.releaseTimer),e.contextRecord.releaseTimer=null,console.log(`[Context] Cancelled pending STT release`)),e.contextRecord.releaseRequested=!1,e.contextRecord.refCount+=1,console.log(`[Context] Reusing existing STT context, refCount=${e.contextRecord.refCount}`),typeof t==`function`&&t(0),e.contextRecord.context||await e.contextRecord.ready,typeof t==`function`&&t(1),e.contextRecord;e.contextRecord?console.log(`[Context] STT record exists but released=${e.contextRecord.released}, creating new context`):console.log(`[Context] No existing STT record, creating new context`);let n={refCount:1,ready:null,released:!1};e.contextRecord=n,n.ready=(async()=>{let r=Date.now();try{typeof t==`function`&&t(0);let i=await Hn(e.plan,e.downloads,t,e.globalDownloadManager);typeof t==`function`&&t(.5);let a=await D({filePath:i,useFlashAttn:e.plan.info.runtime.flash_attn_type===`on`,useGpu:e.plan.info.runtime.n_gpu_layers>0,nThreads:e.plan.info.runtime.n_threads},e.plan.info.runtime.variant);typeof t==`function`&&t(1),n.context=a;try{n.modelInfo=a.getModelInfo()}catch{n.modelInfo=null}return U.addModelLoad({id:e.id,generatorId:e.id,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,modelType:e.plan.info.model?.modelType||null,variant:e.plan.info.runtime?.variant||null,useGpu:e.plan.info.runtime?.use_gpu||!1,durationMs:Date.now()-r,success:!0}),n}catch(t){throw U.addModelLoad({id:e.id,generatorId:e.id,repoId:e.plan.info.model?.repoId||null,quantization:e.plan.info.model?.quantization||null,modelType:e.plan.info.model?.modelType||null,variant:e.plan.info.runtime?.variant||null,durationMs:Date.now()-r,success:!1,error:t?.message||String(t)}),t}})();try{return await n.ready,typeof t==`function`&&t(1),n}catch(t){throw e.contextRecord=null,t}},Kn=async(e,t,n=!1)=>t.released||!n&&t.refCount>0?!1:(t.released=!0,e.contextRecord=null,await t.context?.release?.(),!0),qn=async(e,t,n=!1)=>{if(t.releaseRequested=!0,t.releaseTimer&&=(clearTimeout(t.releaseTimer),null),n)t.refCount=0;else if(t.refCount=Math.max(0,t.refCount-1),t.refCount>0)return t.releaseRequested=!1,!1;let r=e.config.runtime.context_release_delay_ms;if(typeof r!=`number`||!Number.isFinite(r))return Kn(e,t);let i=Math.max(0,Math.floor(r));return n||i<=0?Kn(e,t):(console.log(`[Context] Scheduling STT release in ${i}ms`),t.releaseTimer=setTimeout(async()=>{if(t.releaseTimer=null,t.refCount>0){console.log(`[Context] STT release cancelled, refCount=${t.refCount}`),t.releaseRequested=!1;return}console.log(`[Context] Releasing STT context after ${i}ms delay`),await Kn(e,t)},i),!0)};async function Jn(e,t,n={}){let{globalDownloadManager:r=null}=n,i=En(t),a=await Vn(i),o={id:e,type:`ggml-stt`,config:i,plan:a,info:a.info,contextRecord:null,downloads:new Map,globalDownloadManager:r,queue:new Un,finalized:!1},s=async()=>{if(o.finalized)return;o.finalized=!0;let e=o.contextRecord;e&&(e.released||e.releaseRequested||e.releaseTimer||(e.refCount=Math.max(0,e.refCount-1),!(e.refCount>0)&&await Kn(o,e)))},c=async(e={})=>{let{onProgress:t}=e;try{let e=await Gn(o,t);return{modelInfo:e.modelInfo&&typeof e.modelInfo==`object`?{...e.modelInfo}:null,runtime:{...o.plan.info.runtime},download:{...o.plan.info.download}}}catch(e){throw console.error(`[Context] Error initializing context:`,e),e}},l=async()=>{if(o.finalized)return!1;let e=o.contextRecord;return e?qn(o,e):!1},u=async(e={})=>{let{audioPath:t,audioData:n,options:r={}}=e,i=o.contextRecord;if(!i)throw Error(`Context not initialized`);let a={...r};o.plan.info.runtime.max_threads&&a.maxThreads==null&&(a.maxThreads=o.plan.info.runtime.max_threads);let s=`transcription-${Date.now()}-${Math.random().toString(36).slice(2,8)}`,c=Date.now();return o.queue.enqueue(async()=>{await i.ready;try{let e;if(n){let t=Wn(n),{promise:r}=i.context.transcribeData(t,a);e=await r}else{if(!t)throw Error(`audioPath or audioData is required for transcription`);let n=y.resolve(t),{promise:r}=i.context.transcribe(n,a);e=await r}return U.addTranscription({id:s,generatorId:o.id,repoId:o.plan.info.model?.repoId||null,quantization:o.plan.info.model?.quantization||null,modelType:o.plan.info.model?.modelType||null,variant:o.plan.info.runtime?.variant||null,durationMs:Date.now()-c,segmentCount:e?.segments?.length||0,textLength:e?.text?.length||0,success:!0}),e}catch(e){throw U.addTranscription({id:s,generatorId:o.id,repoId:o.plan.info.model?.repoId||null,quantization:o.plan.info.model?.quantization||null,modelType:o.plan.info.model?.modelType||null,variant:o.plan.info.runtime?.variant||null,durationMs:Date.now()-c,success:!1,error:e?.message||String(e)}),e}},s)};return{id:e,type:`ggml-stt`,info:a.info,queue:o.queue,initContext:c,transcribe:async(e={})=>u(e),transcribeData:async(e={})=>u(e),releaseContext:l,finalize:s,getStatus:()=>({id:o.id,type:o.type,repoId:o.plan.info.model?.repoId||null,quantization:o.plan.info.model?.quantization||null,modelType:o.plan.info.model?.modelType||null,variant:o.plan.info.runtime?.variant||null,hasContext:!!o.contextRecord?.context,contextRefCount:o.contextRecord?.refCount||0,queueStatus:o.queue.getStatus()}),hasPendingReleases:()=>{let e=o.contextRecord;return e?!e.released&&(e.releaseRequested||e.releaseTimer||e.refCount>0):!1},resetFinalized:()=>{o.finalized=!1}}}const Yn=e=>{let t=En(e),n=t.model.repo_id||t.model.repository||t.model.model||null;if(!n)return null;let r=Sn(t.model.filename);return r?`${n}:${r}`:n};async function Xn(e,t,n={}){let{onProgress:r,onComplete:i,onError:a}=n;try{let n=En(e),o=await zn(n),s=In(n,o),{repoId:c}=o;if(await Ln(s,o.size))return console.log(`[Download] STT model already exists: ${c} at ${s}`),typeof i==`function`&&i({localPath:s,repoId:c,alreadyExists:!0}),{started:!1,localPath:s,repoId:c,alreadyExists:!0};let l=t.getDownload(s);if(l)return console.log(`[Download] Already downloading STT model: ${c}`),l.then(()=>{typeof i==`function`&&i({localPath:s,repoId:c,joinedExisting:!0})}).catch(e=>{typeof a==`function`&&a(e)}),{started:!1,localPath:s,repoId:c,alreadyDownloading:!0};console.log(`[Download] Starting STT model download: ${c}`);let u=(async()=>{try{await Rn(o.url,o.headers,s,o.size,e=>{e>=0&&Number.isFinite(e)&&(console.log(`[Download] ${c}: ${Math.round(e*100)}%`),typeof r==`function`&&r(e))}),console.log(`[Download] Completed STT model: ${c}`),typeof i==`function`&&i({localPath:s,repoId:c})}catch(e){throw console.error(`[Download] Failed STT model: ${c}`,e.message),typeof a==`function`&&a(e),e}finally{t.deleteDownload(s)}})();return t.setDownload(s,u),{started:!0,localPath:s,repoId:c}}catch(e){return console.error(`[Download] Failed to start STT download:`,e.message),typeof a==`function`&&a(e),{started:!1,localPath:null,repoId:null,error:e.message}}}const Zn=e=>e?typeof e.score==`number`&&Number.isFinite(e.score)?Number(e.score):B(e):0;async function Qn(e=null,t={}){let{threshold:n=1.1,includeBreakdown:r=!1,config:i,...a}=t,o=null,s=null,c=null;if(i)try{let e=await zn(En(i));o=e.size??null,{processingBufferBytes:s}=Me({modelBytes:o}),c=e.quantization||null}catch{}let l=i?.backend?.gpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.gpu_memory_fraction))),u=i?.backend?.cpu_memory_fraction==null?void 0:Math.min(1,Math.max(0,Number(i.backend.cpu_memory_fraction))),d=await Oe({...a,platform:process.platform,totalMemoryInBytes:b.totalmem(),backend:`ggml-stt`,includeBreakdown:r,gpuMemoryFraction:l,cpuMemoryFraction:u,dependencies:{getBackendDevicesInfo:S,isLibVariantAvailable:C},modelBytes:o,kvCacheBytes:s}),f=d.selected,p=Zn(f);f&&(f.modelBytes=o||null,f.processingBytes=s||null,f.quantization=c||null);let m=null,h=null;if(e){let t=Zn(e);h={...e,score:t};let r=`buttress`,i=`buttress-higher-score`;if(!d.ok)r=`local`,i=`buttress-unavailable`;else if(!t&&t!==0)r=`buttress`,i=`missing-client-score`;else if(e.fit&&f?.fit){let a=e.fit.fitsInGpu||e.fit.fitsInCpu,o=f.fit.fitsInGpu||f.fit.fitsInCpu;a&&!o?(r=`local`,i=`client-fits-in-memory`):o&&!a?(r=`buttress`,i=`buttress-fits-in-memory`):t>p*n?(r=`local`,i=`client-better`):p>t*n?(r=`buttress`,i=`buttress-better`):(r=`either`,i=`comparable-scores`)}else t>p*n?(r=`local`,i=`client-better`):p>t*n?(r=`buttress`,i=`buttress-better`):(r=`either`,i=`comparable-scores`);m={buttressScore:p,clientScore:t,threshold:n,recommendation:r,reason:i}}!d.ok&&!m&&(m={buttressScore:p,clientScore:e?.score??null,threshold:n,recommendation:`local`,reason:`buttress-unavailable`});let g=null;return i&&(g={repoId:i.model?.repo_id||null,quantization:i.model?.quantization||null,filename:i.model?.filename||null}),{type:`ggml-stt`,timestamp:new Date().toISOString(),buttress:d,client:h,comparison:m,modelConfig:g}}const{ReadableStream:$n}=typeof globalThis<`u`&&globalThis.ReadableStream&&globalThis.WritableStream?{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream}:c,er=ee(import.meta.url),tr=y.dirname(er),nr=y.join(tr,`mlx-bridge.py`),rr=`mlx-vlm==0.4.0`,ir=`mlx-lm==0.31.1`,ar=y.join(b.homedir(),`.buttress`,`models`),or={backend:{type:`mlx-llm`},model:{repo_id:null,revision:`main`,adapter_path:null,tokenizer_config:null,model_config:null,vlm:`auto`},runtime:{cache_dir:ar,huggingface_token:process.env.HUGGINGFACE_TOKEN||null,mlx_env_dir:null,mlx_lm_package:ir,mlx_vlm_package:rr,context_release_delay_ms:1e4,session_cache:{enabled:!0,max_size_bytes:5*1024*1024*1024,max_entries:100}}},sr=(e,t)=>e==null?t:typeof e==`number`?e:typeof e==`string`?T.parse(e)??t:t,cr=(e={},t={})=>(Object.entries(t||{}).forEach(([t,n])=>{n&&typeof n==`object`&&!Array.isArray(n)?((!e[t]||typeof e[t]!=`object`)&&(e[t]={}),cr(e[t],n)):e[t]=n}),e),lr=(e={})=>{let t=structuredClone(or);return cr(t,e),t},ur=async(e,t={})=>{let n=await fetch(e,t);if(!n.ok)throw Error(`HTTP ${n.status}: ${e}`);return n.json()},dr=async e=>{await d(e,{recursive:!0})},fr=(e,t,n)=>{let r=s(`sha256`).update(e).digest(`hex`);return y.join(n,`.metadata-cache`,t,`${r}.json`)},pr=async(e,t,n)=>{try{let r=await p(fr(e,t,n),`utf-8`);return JSON.parse(r)}catch{return null}},mr=async(e,t,n,r)=>{try{let i=fr(e,t,r);await dr(y.dirname(i)),await v(i,JSON.stringify(n),`utf-8`)}catch{}};async function hr(e,{revision:t=`main`,cacheDir:n,token:r}={}){let i=JSON.stringify({repoId:e,revision:t,type:`mlx-model-metadata`});if(n){let e=await pr(i,`mlx-model-metadata`,n);if(e)return e}let a={};r&&(a.Authorization=`Bearer ${r}`);let o=(await ur(`https://huggingface.co/api/models/${e}?revision=${t}&blobs=true`,{headers:a}))?.siblings||[],s=0;for(let e of o){let t=e.rfilename||e.path||e.filename||``;/\.(safetensors|npz)$/.test(t)&&(s+=Number(e.size)||0)}let c=null;try{c=await ur(`https://huggingface.co/${e}/raw/${t}/config.json`,{headers:a})}catch{}let l=c?.text_config||c||{},u=c||{},d=u.model_type||u.architectures?.[0]||null,f=l.hidden_size||l.dim||0,p=l.num_hidden_layers||l.n_layers||0,m=l.num_attention_heads||l.n_heads||0,h=l.num_key_value_heads??m,g=l.vocab_size||0,_=l.max_position_embeddings||0,v=l.intermediate_size||0,y=l.head_dim||l.v_head_dim||(m>0&&f>0&&Number.isInteger(f/m)?f/m:0),b=l.kv_lora_rank||0,x=l.qk_rope_head_dim||0,S=b>0,C=u.quantization||u.quantization_config||null,w=C?.bits||null,T=C?.group_size||null,E=l.dtype||u.torch_dtype||(w?`${w}bit`:null),D={repoId:e,revision:t,modelBytes:s,arch:d,hiddenSize:f,numLayers:p,numHeads:m,numKvHeads:h,headDim:y,vocabSize:g,maxCtx:_,intermediateSize:v,quantBits:w,quantGroupSize:T,dtype:E,isMLA:S,kvLoraRank:b,qkRopeHeadDim:x,fileCount:o.length,config:c};return n&&await mr(i,`mlx-model-metadata`,D,n),D}function gr({numLayers:e,numKvHeads:t,headDim:n,contextLength:r,isMLA:i,kvLoraRank:a,qkRopeHeadDim:o}){return!e||!r?0:i&&a>0?e*(a+(o||0))*r*2:!t||!n?0:e*t*n*r*2*2}const _r=async e=>{try{return await g(e),!0}catch{return!1}},q=(e,t,n={})=>new Promise((r,i)=>{te(e,t,{timeout:n.timeout||3e5,...n},(t,n,a)=>{if(t){let n=a?.toString().trim()||t.message;i(Error(`${e} failed: ${n}`))}else r({stdout:n?.toString()||``,stderr:a?.toString()||``})})}),vr=new Map;async function yr({envDir:e,mlxLmPackage:t,mlxVlmPackage:n,onProgress:r}){let i=y.resolve(e),a=vr.get(i);if(a){let e=await a;return r?.(1),e}let o=xr({envDir:i,mlxLmPackage:t,mlxVlmPackage:n,onProgress:r});vr.set(i,o);try{return await o}finally{vr.delete(i)}}const br=[3,10];async function xr({envDir:e,mlxLmPackage:t,mlxVlmPackage:n,onProgress:r}){let i=y.join(e,`bin`,`python3`),a=y.join(e,`bin`,`pip`);if(await _r(i))try{return await q(i,[`-c`,`import mlx_vlm; import torch`],{timeout:1e4}),r?.(1),i}catch{}if(!await _r(i)){r?.(.1);try{let{stdout:e}=await q(`python3`,[`-c`,`import sys; print(f"{sys.version_info.major}.{sys.version_info.minor}")`],{timeout:5e3}),[t,n]=e.trim().split(`.`).map(Number);(t<br[0]||t===br[0]&&n<br[1])&&console.warn(`[mlx-llm] WARNING: System Python is ${t}.${n}, but mlx-vlm requires >= ${br.join(`.`)}. You may get an older mlx-vlm version with reduced functionality. Consider installing Python >= 3.10 (e.g. via Homebrew).`)}catch{}console.log(`[mlx-llm] Creating venv at ${e}`),await d(e,{recursive:!0}),await q(`python3`,[`-m`,`venv`,e],{timeout:6e4}),r?.(.3)}return console.log(`[mlx-llm] Installing ${n}`),r?.(.4),await q(a,[`install`,t,n,`torch`,`torchvision`],{timeout:6e5,env:{...process.env}}),r?.(.9),await q(i,[`-c`,`import mlx_vlm; import torch; print(mlx_vlm.__version__)`],{timeout:15e3}),r?.(1),console.log(`[mlx-llm] mlx-vlm installed successfully`),i}var Sr=class{constructor(){this.process=null,this.pendingRequests=new Map,this.requestCounter=0,this.readyPromise=null,this.buffer=``}spawn(e){return this.process=k(e,[nr],{stdio:[`pipe`,`pipe`,`pipe`],env:{...process.env,PYTHONUNBUFFERED:`1`}}),this.process.stderr.on(`data`,e=>{let t=e.toString().trim();t&&console.log(t)}),this.process.on(`exit`,e=>{console.log(`[mlx-llm] Bridge process exited with code ${e}`);for(let[t,n]of this.pendingRequests)n.reject(Error(`Bridge process exited (code ${e})`)),this.pendingRequests.delete(t);this.process=null}),this.process.stdout.on(`data`,e=>{this.buffer+=e.toString();let t=this.buffer.split(`
3
3
  `);this.buffer=t.pop();for(let e of t)if(e.trim())try{this.handleMessage(JSON.parse(e))}catch(t){console.error(`[mlx-llm] Failed to parse bridge message:`,e,t)}}),this.readyPromise=new Promise((e,t)=>{this.pendingRequests.set(`__init__`,{resolve:()=>e(),reject:t}),setTimeout(()=>t(Error(`Bridge startup timeout`)),3e4)}),this.readyPromise}handleMessage(e){let t=this.pendingRequests.get(e.id);t&&(e.error?(t.reject(Error(e.error.message)),this.pendingRequests.delete(e.id)):e.event?e.event===`result`?(t.resolve(e.data),this.pendingRequests.delete(e.id)):t.onEvent?.(e.event,e.data):e.result!==void 0&&(t.resolve(e.result),this.pendingRequests.delete(e.id)))}async call(e,t={}){if(!this.process)throw Error(`Bridge not running`);let n=String(++this.requestCounter);return new Promise((r,i)=>{this.pendingRequests.set(n,{resolve:r,reject:i}),this.write({id:n,method:e,params:t})})}stream(e,t,n){if(!this.process)throw Error(`Bridge not running`);let r=String(++this.requestCounter);return{id:r,promise:new Promise((i,a)=>{this.pendingRequests.set(r,{resolve:i,reject:a,onEvent:n}),this.write({id:r,method:e,params:t})})}}cancel(e){this.process&&this.write({id:`cancel-${e}`,method:`cancel`,params:{request_id:e}})}write(e){this.process?.stdin?.write(JSON.stringify(e)+`
4
- `)}kill(){this.process&&=(this.process.kill(),null),this.pendingRequests.clear()}get alive(){return this.process!=null&&!this.process.killed}};function br(){let e=[];return process.platform!==`darwin`&&e.push(`MLX requires macOS (Apple Silicon)`),v.arch()!==`arm64`&&e.push(`MLX requires Apple Silicon (arm64)`),e}function xr(e){let t=br();return{config:e,info:{ok:t.length===0,backend:`mlx-llm`,warnings:[],errors:[...t],model:{repoId:e.model.repo_id,revision:e.model.revision},runtime:{variant:`mlx`},resources:{},devices:{selected:{variant:`mlx`,hasGpu:!0}},download:{cacheDir:e.runtime.cache_dir,localPath:null,exists:!1},timestamp:new Date().toISOString()}}}const Sr=e=>{if(!e)return[];let t=[];for(let n of e)if(Array.isArray(n.content))for(let e of n.content)e.type===`image_url`&&e.image_url?.url&&t.push(e.image_url.url);return t};var Cr=class{constructor(){this.queue=[],this.processing=!1,this.currentTaskId=null}async enqueue(e,t=null){return new Promise((n,r)=>{this.queue.push({task:e,resolve:n,reject:r,taskId:t}),this.processNext()})}async processNext(){if(this.processing||this.queue.length===0)return;this.processing=!0;let{task:e,resolve:t,reject:n,taskId:r}=this.queue.shift();this.currentTaskId=r;try{t(await e())}catch(e){n(e)}finally{this.processing=!1,this.currentTaskId=null,this.processNext()}}getStatus(){return{processing:this.processing,queuedCount:this.queue.length,currentTaskId:this.currentTaskId}}};const wr=`</think>`;function Tr(e,t){if(!t)return{reasoningContent:``,content:e};let n=e.indexOf(wr);if(n!==-1)return{reasoningContent:e.slice(0,n).replace(/^\n+/,``),content:e.slice(n+8).replace(/^\n+/,``)};let r=e.length;for(let t=1;t<=8&&t<=e.length;t++)if(wr.startsWith(e.slice(-t))){r=e.length-t;break}return{reasoningContent:e.slice(0,r).replace(/^\n+/,``),content:``}}function Er(e,t,n,r,{enableThinking:i=!1}={}){let a=null,o=Date.now(),s=0,c=``;return new Xn({start(l){let{id:u,promise:d}=e.stream(`generate`,t,(e,t)=>{if(e===`token`){s+=1,c+=t.token||``;let e=Tr(c,i);l.enqueue({event:`token`,data:{requestId:u,token:t.token,token_id:t.token_id,text:c,content:e.content,reasoning_content:e.reasoningContent}})}});a=u,d.then(e=>{let t={prompt_n:e.prompt_tokens??0,prompt_per_second:e.prompt_tps??0,predicted_n:e.generation_tokens??s,predicted_per_second:e.generation_tps??0},a=Tr(c,i);l.enqueue({event:`result`,data:{requestId:u,text:c,content:a.content,reasoning_content:a.reasoningContent,timings:t,prompt_tokens:t.prompt_n,tokens_predicted:t.predicted_n,interrupted:e.interrupted||!1,peak_memory:e.peak_memory}}),l.close(),H.addCompletion({id:`completion-${u}`,generatorId:n,requestId:u,repoId:r?.repoId||null,quantization:r?.quantization||null,variant:`mlx`,promptTokens:t.prompt_n,tokensGenerated:t.predicted_n,tokensPerSecond:t.predicted_per_second,promptPerSecond:t.prompt_per_second,durationMs:Date.now()-o,success:!0,interrupted:e.interrupted||!1})}).catch(e=>{l.enqueue({event:`error`,data:{message:e?.message||String(e)}}),l.error(e),H.addCompletion({id:`completion-${Date.now()}`,generatorId:n,repoId:r?.repoId||null,variant:`mlx`,durationMs:Date.now()-o,tokensGenerated:s,success:!1,error:e?.message||String(e)})})},cancel(){a&&e.cancel(a)}})}async function Dr(e,t,n={}){let r=or(t),i=xr(r);i.info.ok||console.error(`[mlx-llm] Platform check failed:`,i.info.errors);let a={id:e,type:`mlx-llm`,config:r,plan:i,info:i.info,contexts:new Map,bridge:null,queue:new Cr,finalized:!1},o=`mlx:${r.model.repo_id}`,s=async(e={})=>{let{onProgress:t}=e,n=a.contexts.get(o);if(n&&!n.released)return n.refCount+=1,n.releaseTimer&&=(clearTimeout(n.releaseTimer),null),await n.ready,{modelInfo:n.modelInfo,runtime:{...a.info.runtime},download:{...a.info.download}};let i={key:o,refCount:1,ready:null,released:!1,releaseRequested:!1,releaseTimer:null,modelInfo:null};a.contexts.set(o,i);let s=Date.now();i.ready=(async()=>{let e=r.runtime.cache_dir||nr,n=await gr({envDir:r.runtime.mlx_env_dir||_.join(e,`mlx-env`),mlxLmPackage:r.runtime.mlx_lm_package||tr,mlxVlmPackage:r.runtime.mlx_vlm_package||er,onProgress:t?e=>t(e*.3):void 0});(!a.bridge||!a.bridge.alive)&&(a.bridge=new yr,await a.bridge.spawn(n)),t?.(.4);let o={model:r.model.repo_id};r.model.revision&&(o.revision=r.model.revision),r.model.adapter_path&&(o.adapter_path=r.model.adapter_path),r.model.vlm!=null&&(o.vlm=r.model.vlm),r.runtime.huggingface_token&&(process.env.HF_TOKEN=r.runtime.huggingface_token),await a.bridge.call(`load`,o),t?.(.9);let c=await a.bridge.call(`get_info`);i.modelInfo={model:c.model,peak_memory:c.peak_memory,active_memory:c.active_memory};let l=r.runtime.session_cache;l?.enabled!==!1&&await a.bridge.call(`configure_cache`,{enabled:!0,cache_dir:_.join(e,`mlx-session-cache`),max_entries:l?.max_entries||100,max_size_bytes:ir(l?.max_size_bytes,5*1024*1024*1024)}),a.info.download.exists=!0,t?.(1),H.addModelLoad({id:`load-${Date.now()}`,generatorId:a.id,repoId:r.model.repo_id,variant:`mlx`,durationMs:Date.now()-s,success:!0})})();try{await i.ready}catch(e){throw i.released=!0,a.contexts.delete(o),H.addModelLoad({id:`load-${Date.now()}`,generatorId:a.id,repoId:r.model.repo_id,variant:`mlx`,durationMs:Date.now()-s,success:!1,error:e?.message||String(e)}),e}return{modelInfo:i.modelInfo,runtime:{...a.info.runtime},download:{...a.info.download}}},c=async e=>{if(e.released)return!1;e.released=!0;try{a.bridge?.alive&&await a.bridge.call(`release`)}catch(e){console.error(`[mlx-llm] Error releasing context:`,e.message)}return a.contexts.delete(e.key),!0};return{id:e,type:`mlx-llm`,info:i.info,contexts:a.contexts,queue:a.queue,initContext:s,completion:async(e={})=>{let{options:t={}}=e,n=a.contexts.get(o);if(!n)throw Error(`Context "${o}" not initialized`);await n.ready;let r=Sr(t.messages),i=t.prompt||``;if(!i&&t.messages){let e={messages:t.messages,add_generation_prompt:t.add_generation_prompt??!0,tools:t.tools,...t.chat_template_kwargs};t.enable_thinking!=null&&(e.enable_thinking=t.enable_thinking),i=(await a.bridge.call(`apply_chat_template`,e)).text}let s={prompt:i,max_tokens:t.n_predict??t.max_tokens??256};r.length>0&&(s.image=r),t.temperature!=null&&(s.temperature=t.temperature),t.top_p!=null&&(s.top_p=t.top_p),t.top_k!=null&&(s.top_k=t.top_k),t.min_p!=null&&(s.min_p=t.min_p),t.seed!=null&&(s.seed=t.seed),t.repetition_penalty!=null&&(s.repetition_penalty=t.repetition_penalty),t.stop&&(s.stop=t.stop);let c={repoId:a.info.model?.repoId||null},l=`completion-${Date.now()}-${Math.random().toString(36).slice(2,8)}`;return new Xn({start(e){a.queue.enqueue(async()=>{let n=Er(a.bridge,s,a.id,c,{enableThinking:!!t.enable_thinking}).getReader();try{for(;;){let{value:t,done:r}=await n.read();if(r)break;e.enqueue(t)}e.close()}catch(t){throw e.error(t),t}},l).catch(t=>{try{e.error(t)}catch{}})},cancel(){a.bridge?.alive}})},tokenize:async(e={})=>{let{text:t=``}=e,n=a.contexts.get(o);if(!n)throw Error(`Context "${o}" not initialized`);return await n.ready,a.bridge.call(`tokenize`,{text:t})},detokenize:async(e={})=>{let{tokens:t=[]}=e,n=a.contexts.get(o);if(!n)throw Error(`Context "${o}" not initialized`);return await n.ready,(await a.bridge.call(`detokenize`,{tokens:t})).text},applyChatTemplate:async(e={})=>{let{messages:t=[],params:n={}}=e,r=a.contexts.get(o);if(!r)throw Error(`Context "${o}" not initialized`);await r.ready;let i={messages:t,add_generation_prompt:n.add_generation_prompt??!0,tools:n.tools,...n.chat_template_kwargs};return(await a.bridge.call(`apply_chat_template`,i)).text},releaseContext:async()=>{if(a.finalized)return!1;let e=a.contexts.get(o);if(!e||(e.releaseRequested=!0,e.refCount=Math.max(0,e.refCount-1),e.refCount>0))return!1;let t=r.runtime.context_release_delay_ms??1e4;return t>0?new Promise(n=>{e.releaseTimer=setTimeout(async()=>{e.releaseTimer=null,e.refCount<=0&&!e.released?n(await c(e)):n(!1)},t)}):c(e)},finalize:async()=>{if(a.finalized)return;a.finalized=!0;let e=Array.from(a.contexts.values());for(let t of e)t.released||(t.refCount=0,await c(t));a.bridge?.kill(),a.bridge=null},getStatus:()=>({id:a.id,type:a.type,repoId:a.info.model?.repoId||null,variant:`mlx`,contexts:Array.from(a.contexts.entries()).map(([e,t])=>({key:e,refCount:t.refCount,hasModel:!t.released})),queueStatus:a.queue.getStatus()}),hasPendingReleases:()=>Array.from(a.contexts.values()).some(e=>!e.released&&(e.releaseRequested||e.releaseTimer||e.refCount>0)),resetFinalized:()=>{a.finalized=!1}}}const Or=e=>or(e).model.repo_id||null;async function kr(e=null,t={}){let{includeBreakdown:n=!1,config:r}=t,i=br(),a=i.length===0,o=!1,s=!1;if(a){try{await q(`python3`,[`--version`],{timeout:5e3}),o=!0}catch{}if(o)try{await q(`python3`,[`-c`,`import mlx`],{timeout:1e4}),s=!0}catch{}}let c=await Ce({platform:process.platform,arch:v.arch(),totalMemoryInBytes:v.totalmem(),includeBreakdown:n}),l=null,u=null,d=null,f=null,p=null,m=null,h=null;if(r){let e=or(r),t=e.model.repo_id;if(t)try{l=await fr(t,{revision:e.model.revision,cacheDir:e.runtime.cache_dir||nr,token:e.runtime.huggingface_token}),u=l.modelBytes||0,f=l.maxCtx||4096;let n={numLayers:l.numLayers,numKvHeads:l.numKvHeads,headDim:l.headDim,isMLA:l.isMLA,kvLoraRank:l.kvLoraRank,qkRopeHeadDim:l.qkRopeHeadDim};d=pr({...n,contextLength:f});let r=c.ok?c.selected.gpuUsableBytes:0;if(r>0&&u>0&&l.numLayers){let e=r-u;if(e>0){let t;t=l.isMLA&&l.kvLoraRank>0?l.numLayers*(l.kvLoraRank+(l.qkRopeHeadDim||0))*2:l.numKvHeads&&l.headDim?l.numLayers*l.numKvHeads*l.headDim*2*2:0,t>0&&(m=Math.floor(e/t),m=Math.min(m,f))}else m=0;m!=null&&m<f&&(p=pr({...n,contextLength:m}))}h={repoId:t,revision:e.model.revision,nCtx:f,architecture:l.arch,quantBits:l.quantBits,quantGroupSize:l.quantGroupSize}}catch{}}let g=c.ok?{...c.selected,modelBytes:u,kvCacheBytes:d,memoryLimitedCtx:m,limitedKvCacheBytes:p,kvInfo:l?{nCtxTrain:l.maxCtx||null,nLayer:l.numLayers,nEmbd:l.hiddenSize,nHeadKv:l.numKvHeads,headDim:l.headDim}:null,quantization:l?{bits:l.quantBits,groupSize:l.quantGroupSize,dtype:l.dtype}:null}:null;if(c.ok&&u!=null&&u>0){let e=u+(d||0),t=c.selected.gpuUsableBytes,n=e<=t;if(g.fit={totalRequiredBytes:e,fitsInGpu:n,fitsInCpu:e<=t,limiting:n?`none`:`insufficient-memory`},p!=null&&p!==d){let e=u+p;g.limitedFit={totalRequiredBytes:e,fitsInGpu:e<=t,fitsInCpu:e<=t,limiting:e<=t?`none`:`insufficient-memory`}}}return{type:`mlx-llm`,available:a,platform:{ok:a,os:process.platform,arch:v.arch(),errors:i},python:{available:o},mlx:{systemAvailable:s,venvSupported:o},buttress:c.ok?{ok:c.ok,selected:g,attempts:c.attempts}:{ok:!1,selected:null,attempts:c.attempts||[],errors:c.errors},modelConfig:h,timestamp:new Date().toISOString()}}async function Ar(e,t,n={}){let{onProgress:r,onComplete:i,onError:a}=n,o=or(e),s=o.model.repo_id;if(!s)return{started:!1,localPath:null,repoId:null,error:`Missing repo_id`};let c=br();if(c.length>0)return{started:!1,localPath:null,repoId:s,error:c.join(`; `)};let l=`mlx:${s}`;if(t?.isDownloading(l))return{started:!1,localPath:null,repoId:s,alreadyDownloading:!0};let u=(async()=>{try{let e=o.runtime.cache_dir||nr,t=await gr({envDir:o.runtime.mlx_env_dir||_.join(e,`mlx-env`),mlxLmPackage:o.runtime.mlx_lm_package||tr,mlxVlmPackage:o.runtime.mlx_vlm_package||er,onProgress:r?e=>r(e*.3):void 0});r?.(.3);let n=`
4
+ `)}kill(){this.process&&=(this.process.kill(),null),this.pendingRequests.clear()}get alive(){return this.process!=null&&!this.process.killed}};function Cr(){let e=[];return process.platform!==`darwin`&&e.push(`MLX requires macOS (Apple Silicon)`),b.arch()!==`arm64`&&e.push(`MLX requires Apple Silicon (arm64)`),e}function wr(e){let t=Cr();return{config:e,info:{ok:t.length===0,backend:`mlx-llm`,warnings:[],errors:[...t],model:{repoId:e.model.repo_id,revision:e.model.revision},runtime:{variant:`mlx`},resources:{},devices:{selected:{variant:`mlx`,hasGpu:!0}},download:{cacheDir:e.runtime.cache_dir,localPath:null,exists:!1},timestamp:new Date().toISOString()}}}const Tr=e=>{if(!e)return[];let t=[];for(let n of e)if(Array.isArray(n.content))for(let e of n.content)e.type===`image_url`&&e.image_url?.url&&t.push(e.image_url.url);return t};var Er=class{constructor(){this.queue=[],this.processing=!1,this.currentTaskId=null}async enqueue(e,t=null){return new Promise((n,r)=>{this.queue.push({task:e,resolve:n,reject:r,taskId:t}),this.processNext()})}async processNext(){if(this.processing||this.queue.length===0)return;this.processing=!0;let{task:e,resolve:t,reject:n,taskId:r}=this.queue.shift();this.currentTaskId=r;try{t(await e())}catch(e){n(e)}finally{this.processing=!1,this.currentTaskId=null,this.processNext()}}getStatus(){return{processing:this.processing,queuedCount:this.queue.length,currentTaskId:this.currentTaskId}}};const Dr=`</think>`;function Or(e,t){if(!t)return{reasoningContent:``,content:e};let n=e.indexOf(Dr);if(n!==-1)return{reasoningContent:e.slice(0,n).replace(/^\n+/,``),content:e.slice(n+8).replace(/^\n+/,``)};let r=e.length;for(let t=1;t<=8&&t<=e.length;t++)if(Dr.startsWith(e.slice(-t))){r=e.length-t;break}return{reasoningContent:e.slice(0,r).replace(/^\n+/,``),content:``}}function kr(e,t,n,r,{enableThinking:i=!1}={}){let a=null,o=Date.now(),s=0,c=``;return new $n({start(l){let{id:u,promise:d}=e.stream(`generate`,t,(e,t)=>{if(e===`token`){s+=1,c+=t.token||``;let e=Or(c,i);l.enqueue({event:`token`,data:{requestId:u,token:t.token,token_id:t.token_id,text:c,content:e.content,reasoning_content:e.reasoningContent}})}});a=u,d.then(e=>{let t={prompt_n:e.prompt_tokens??0,prompt_per_second:e.prompt_tps??0,predicted_n:e.generation_tokens??s,predicted_per_second:e.generation_tps??0},a=Or(c,i);l.enqueue({event:`result`,data:{requestId:u,text:c,content:a.content,reasoning_content:a.reasoningContent,timings:t,prompt_tokens:t.prompt_n,tokens_predicted:t.predicted_n,interrupted:e.interrupted||!1,peak_memory:e.peak_memory}}),l.close(),H.addCompletion({id:`completion-${u}`,generatorId:n,requestId:u,repoId:r?.repoId||null,quantization:r?.quantization||null,variant:`mlx`,promptTokens:t.prompt_n,tokensGenerated:t.predicted_n,tokensPerSecond:t.predicted_per_second,promptPerSecond:t.prompt_per_second,durationMs:Date.now()-o,success:!0,interrupted:e.interrupted||!1})}).catch(e=>{l.enqueue({event:`error`,data:{message:e?.message||String(e)}}),l.error(e),H.addCompletion({id:`completion-${Date.now()}`,generatorId:n,repoId:r?.repoId||null,variant:`mlx`,durationMs:Date.now()-o,tokensGenerated:s,success:!1,error:e?.message||String(e)})})},cancel(){a&&e.cancel(a)}})}async function Ar(e,t,n={}){let r=lr(t),i=wr(r);i.info.ok||console.error(`[mlx-llm] Platform check failed:`,i.info.errors);let a={id:e,type:`mlx-llm`,config:r,plan:i,info:i.info,contexts:new Map,bridge:null,queue:new Er,finalized:!1},o=`mlx:${r.model.repo_id}`,s=async(e={})=>{let{onProgress:t}=e,n=a.contexts.get(o);if(n&&!n.released)return n.refCount+=1,n.releaseTimer&&=(clearTimeout(n.releaseTimer),null),await n.ready,{modelInfo:n.modelInfo,runtime:{...a.info.runtime},download:{...a.info.download}};let i={key:o,refCount:1,ready:null,released:!1,releaseRequested:!1,releaseTimer:null,modelInfo:null};a.contexts.set(o,i);let s=Date.now();i.ready=(async()=>{let e=r.runtime.cache_dir||ar,n=await yr({envDir:r.runtime.mlx_env_dir||y.join(e,`mlx-env`),mlxLmPackage:r.runtime.mlx_lm_package||ir,mlxVlmPackage:r.runtime.mlx_vlm_package||rr,onProgress:t?e=>t(e*.3):void 0});(!a.bridge||!a.bridge.alive)&&(a.bridge=new Sr,await a.bridge.spawn(n)),t?.(.4);let o={model:r.model.repo_id};r.model.revision&&(o.revision=r.model.revision),r.model.adapter_path&&(o.adapter_path=r.model.adapter_path),r.model.vlm!=null&&(o.vlm=r.model.vlm),r.runtime.huggingface_token&&(process.env.HF_TOKEN=r.runtime.huggingface_token),await a.bridge.call(`load`,o),t?.(.9);let c=await a.bridge.call(`get_info`);i.modelInfo={model:c.model,peak_memory:c.peak_memory,active_memory:c.active_memory};let l=r.runtime.session_cache;l?.enabled!==!1&&await a.bridge.call(`configure_cache`,{enabled:!0,cache_dir:y.join(e,`mlx-session-cache`),max_entries:l?.max_entries||100,max_size_bytes:sr(l?.max_size_bytes,5*1024*1024*1024)}),a.info.download.exists=!0,t?.(1),H.addModelLoad({id:`load-${Date.now()}`,generatorId:a.id,repoId:r.model.repo_id,variant:`mlx`,durationMs:Date.now()-s,success:!0})})();try{await i.ready}catch(e){throw i.released=!0,a.contexts.delete(o),H.addModelLoad({id:`load-${Date.now()}`,generatorId:a.id,repoId:r.model.repo_id,variant:`mlx`,durationMs:Date.now()-s,success:!1,error:e?.message||String(e)}),e}return{modelInfo:i.modelInfo,runtime:{...a.info.runtime},download:{...a.info.download}}},c=async e=>{if(e.released)return!1;e.released=!0;try{a.bridge?.alive&&await a.bridge.call(`release`)}catch(e){console.error(`[mlx-llm] Error releasing context:`,e.message)}return a.contexts.delete(e.key),!0};return{id:e,type:`mlx-llm`,info:i.info,contexts:a.contexts,queue:a.queue,initContext:s,completion:async(e={})=>{let{options:t={}}=e,n=a.contexts.get(o);if(!n)throw Error(`Context "${o}" not initialized`);await n.ready;let r=Tr(t.messages),i=t.prompt||``;if(!i&&t.messages){let e={messages:t.messages,add_generation_prompt:t.add_generation_prompt??!0,tools:t.tools,...t.chat_template_kwargs};t.enable_thinking!=null&&(e.enable_thinking=t.enable_thinking),i=(await a.bridge.call(`apply_chat_template`,e)).text}let s={prompt:i,max_tokens:t.n_predict??t.max_tokens??256};r.length>0&&(s.image=r),t.temperature!=null&&(s.temperature=t.temperature),t.top_p!=null&&(s.top_p=t.top_p),t.top_k!=null&&(s.top_k=t.top_k),t.min_p!=null&&(s.min_p=t.min_p),t.seed!=null&&(s.seed=t.seed),t.repetition_penalty!=null&&(s.repetition_penalty=t.repetition_penalty),t.stop&&(s.stop=t.stop);let c={repoId:a.info.model?.repoId||null},l=`completion-${Date.now()}-${Math.random().toString(36).slice(2,8)}`;return new $n({start(e){a.queue.enqueue(async()=>{let n=kr(a.bridge,s,a.id,c,{enableThinking:!!t.enable_thinking}).getReader();try{for(;;){let{value:t,done:r}=await n.read();if(r)break;e.enqueue(t)}e.close()}catch(t){throw e.error(t),t}},l).catch(t=>{try{e.error(t)}catch{}})},cancel(){a.bridge?.alive}})},tokenize:async(e={})=>{let{text:t=``}=e,n=a.contexts.get(o);if(!n)throw Error(`Context "${o}" not initialized`);return await n.ready,a.bridge.call(`tokenize`,{text:t})},detokenize:async(e={})=>{let{tokens:t=[]}=e,n=a.contexts.get(o);if(!n)throw Error(`Context "${o}" not initialized`);return await n.ready,(await a.bridge.call(`detokenize`,{tokens:t})).text},applyChatTemplate:async(e={})=>{let{messages:t=[],params:n={}}=e,r=a.contexts.get(o);if(!r)throw Error(`Context "${o}" not initialized`);await r.ready;let i={messages:t,add_generation_prompt:n.add_generation_prompt??!0,tools:n.tools,...n.chat_template_kwargs};return(await a.bridge.call(`apply_chat_template`,i)).text},releaseContext:async()=>{if(a.finalized)return!1;let e=a.contexts.get(o);if(!e||(e.releaseRequested=!0,e.refCount=Math.max(0,e.refCount-1),e.refCount>0))return!1;let t=r.runtime.context_release_delay_ms??1e4;return t>0?new Promise(n=>{e.releaseTimer=setTimeout(async()=>{e.releaseTimer=null,e.refCount<=0&&!e.released?n(await c(e)):n(!1)},t)}):c(e)},finalize:async()=>{if(a.finalized)return;a.finalized=!0;let e=Array.from(a.contexts.values());for(let t of e)t.released||(t.refCount=0,await c(t));a.bridge?.kill(),a.bridge=null},getStatus:()=>({id:a.id,type:a.type,repoId:a.info.model?.repoId||null,variant:`mlx`,contexts:Array.from(a.contexts.entries()).map(([e,t])=>({key:e,refCount:t.refCount,hasModel:!t.released})),queueStatus:a.queue.getStatus()}),hasPendingReleases:()=>Array.from(a.contexts.values()).some(e=>!e.released&&(e.releaseRequested||e.releaseTimer||e.refCount>0)),resetFinalized:()=>{a.finalized=!1}}}const jr=e=>lr(e).model.repo_id||null;async function Mr(e=null,t={}){let{includeBreakdown:n=!1,config:r}=t,i=Cr(),a=i.length===0,o=!1,s=!1;if(a){try{await q(`python3`,[`--version`],{timeout:5e3}),o=!0}catch{}if(o)try{await q(`python3`,[`-c`,`import mlx`],{timeout:1e4}),s=!0}catch{}}let c=await Ee({platform:process.platform,arch:b.arch(),totalMemoryInBytes:b.totalmem(),includeBreakdown:n}),l=null,u=null,d=null,f=null,p=null,m=null,h=null;if(r){let e=lr(r),t=e.model.repo_id;if(t)try{l=await hr(t,{revision:e.model.revision,cacheDir:e.runtime.cache_dir||ar,token:e.runtime.huggingface_token}),u=l.modelBytes||0,f=l.maxCtx||4096;let n={numLayers:l.numLayers,numKvHeads:l.numKvHeads,headDim:l.headDim,isMLA:l.isMLA,kvLoraRank:l.kvLoraRank,qkRopeHeadDim:l.qkRopeHeadDim};d=gr({...n,contextLength:f});let r=c.ok?c.selected.gpuUsableBytes:0;if(r>0&&u>0&&l.numLayers){let e=r-u;if(e>0){let t;t=l.isMLA&&l.kvLoraRank>0?l.numLayers*(l.kvLoraRank+(l.qkRopeHeadDim||0))*2:l.numKvHeads&&l.headDim?l.numLayers*l.numKvHeads*l.headDim*2*2:0,t>0&&(m=Math.floor(e/t),m=Math.min(m,f))}else m=0;m!=null&&m<f&&(p=gr({...n,contextLength:m}))}h={repoId:t,revision:e.model.revision,nCtx:f,architecture:l.arch,quantBits:l.quantBits,quantGroupSize:l.quantGroupSize}}catch{}}let g=c.ok?{...c.selected,modelBytes:u,kvCacheBytes:d,memoryLimitedCtx:m,limitedKvCacheBytes:p,kvInfo:l?{nCtxTrain:l.maxCtx||null,nLayer:l.numLayers,nEmbd:l.hiddenSize,nHeadKv:l.numKvHeads,headDim:l.headDim}:null,quantization:l?{bits:l.quantBits,groupSize:l.quantGroupSize,dtype:l.dtype}:null}:null;if(c.ok&&u!=null&&u>0){let e=u+(d||0),t=c.selected.gpuUsableBytes,n=e<=t;if(g.fit={totalRequiredBytes:e,fitsInGpu:n,fitsInCpu:e<=t,limiting:n?`none`:`insufficient-memory`},p!=null&&p!==d){let e=u+p;g.limitedFit={totalRequiredBytes:e,fitsInGpu:e<=t,fitsInCpu:e<=t,limiting:e<=t?`none`:`insufficient-memory`}}}return{type:`mlx-llm`,available:a,platform:{ok:a,os:process.platform,arch:b.arch(),errors:i},python:{available:o},mlx:{systemAvailable:s,venvSupported:o},buttress:c.ok?{ok:c.ok,selected:g,attempts:c.attempts}:{ok:!1,selected:null,attempts:c.attempts||[],errors:c.errors},modelConfig:h,timestamp:new Date().toISOString()}}async function Nr(e,t,n={}){let{onProgress:r,onComplete:i,onError:a}=n,o=lr(e),s=o.model.repo_id;if(!s)return{started:!1,localPath:null,repoId:null,error:`Missing repo_id`};let c=Cr();if(c.length>0)return{started:!1,localPath:null,repoId:s,error:c.join(`; `)};let l=`mlx:${s}`;if(t?.isDownloading(l))return{started:!1,localPath:null,repoId:s,alreadyDownloading:!0};let u=(async()=>{try{let e=o.runtime.cache_dir||ar,t=await yr({envDir:o.runtime.mlx_env_dir||y.join(e,`mlx-env`),mlxLmPackage:o.runtime.mlx_lm_package||ir,mlxVlmPackage:o.runtime.mlx_vlm_package||rr,onProgress:r?e=>r(e*.3):void 0});r?.(.3);let n=`
5
5
  from huggingface_hub import snapshot_download
6
6
  path = snapshot_download("${s}", revision="${o.model.revision||`main`}")
7
7
  print(path)
8
8
  `.trim(),a={...process.env};o.runtime.huggingface_token&&(a.HF_TOKEN=o.runtime.huggingface_token);let c=await q(t,[`-c`,n],{timeout:6e5,env:a});r?.(1);let l=c.stdout.trim().split(`
9
- `).pop();i?.({localPath:l,repoId:s,alreadyExists:!1})}catch(e){throw a?.(e),e}finally{t?.deleteDownload(l)}})();return t?.setDownload(l,u),{started:!0,localPath:null,repoId:s}}async function J(e,t=null,n={}){if(e===`ggml-llm`)return cn(t,n);if(e===`ggml-stt`)return Yn(t,n);if(e===`mlx-llm`)return kr(t,n);throw Error(`Unknown backend type: ${e}`)}var Y=`@fugood/buttress-backend-core`,X=`2.24.1`,jr={name:Y,private:!0,type:`module`,version:X,main:`src/index.js`,types:`lib/types/index.d.ts`,scripts:{build:`tsc --noCheck --declaration --emitDeclarationOnly --allowJs --outDir lib/types src/index.js`},dependencies:{"@fugood/buttress-hardware-guardrails":`^2.24.0`,"@fugood/llama.node":`^1.7.0-rc.11`,"@fugood/whisper.node":`^1.0.18`,"@huggingface/gguf":`^0.3.2`,"@iarna/toml":`^3.0.0`,bytes:`^3.1.0`}};const Mr=e=>{if(!e)return{repoId:null,filename:null};let[t,n]=e.split(`:`);return{repoId:t,filename:n||null}};async function Nr({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${Y} v${X}`),console.log(`Generating model capabilities comparison...
9
+ `).pop();i?.({localPath:l,repoId:s,alreadyExists:!1})}catch(e){throw a?.(e),e}finally{t?.deleteDownload(l)}})();return t?.setDownload(l,u),{started:!0,localPath:null,repoId:s}}async function J(e,t=null,n={}){if(e===`ggml-llm`)return dn(t,n);if(e===`ggml-stt`)return Qn(t,n);if(e===`mlx-llm`)return Mr(t,n);throw Error(`Unknown backend type: ${e}`)}var Y=`@fugood/buttress-backend-core`,X=`2.24.1`,Pr={name:Y,private:!0,type:`module`,version:X,main:`src/index.js`,types:`lib/types/index.d.ts`,scripts:{build:`tsc --noCheck --declaration --emitDeclarationOnly --allowJs --outDir lib/types src/index.js`},dependencies:{"@fugood/buttress-hardware-guardrails":`^2.24.0`,"@fugood/llama.node":`^1.7.0`,"@fugood/whisper.node":`^1.0.18`,"@huggingface/gguf":`^0.3.2`,"@iarna/toml":`^3.0.0`,bytes:`^3.1.0`}};const Fr=e=>{if(!e)return{repoId:null,filename:null};let[t,n]=e.split(`:`);return{repoId:t,filename:n||null}};async function Ir({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${Y} v${X}`),console.log(`Generating model capabilities comparison...
10
10
  `),n.push(`${Y} v${X}`),n.push(`## Model Capabilities Comparison
11
- `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t];console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let r=c(n);r={...r||{},model:{...o.runtime,...r?.model||{},repo_id:n}};let i=await J(`ggml-llm`,null,{config:r,includeBreakdown:!0});l.push({modelId:n,capabilities:i,modelInfo:i.buttress?.selected||null,modelConfig:i.modelConfig||null})}let u=e=>e?(e/1024/1024/1024).toFixed(2):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model ID | Quantization | Size (GB) | Context Size | KV Cache Size (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |`),n.push(`|----------|--------------|-----------|--------------|--------------------|-----------------------------|-----------------|-----------------|`),l.forEach(({modelId:e,modelInfo:t,modelConfig:r})=>{let i=t?.quantization?.name?.toUpperCase()||`N/A`,a=u(t?.modelBytes),o=r?.nCtx||t?.kvInfo?.nCtxTrain||`N/A`,s=Ue(t),c=Number(o),l=t?.kvCacheBytes||(s&&Number.isFinite(c)&&c>0?s(c):s&&s(t?.kvInfo?.nCtxTrain||0))||null,f=u(l),p=u(t?.modelBytes&&l?t.modelBytes+l:t?.fit?.totalRequiredBytes),m=d(t?.fit?.fitsInGpu),h=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${i} | ${a} | ${o} | ${f} | ${p} | ${m} | ${h} |`);let g=t?.memoryLimitedCtx!=null||t?.limitedFit!=null,_=!t?.fit?.fitsInGpu||!t?.fit?.fitsInCpu;if(g&&_){let e=t?.memoryLimitedCtx||o,r=Number(e),i=t?.limitedKvCacheBytes||s&&Number.isFinite(r)&&r>0&&s(r)||null,c=u(i),l=u(t?.modelBytes&&i?t.modelBytes+i:t?.limitedFit?.totalRequiredBytes),m=d(t?.limitedFit?.fitsInGpu),h=d(t?.limitedFit?.fitsInCpu);(e!==o||c!==f||l!==p)&&n.push(`| ↳ Limited | - | ${a} | ${e} | ${c} | ${l} | ${m} | ${h} |`)}}),n.push(`
11
+ `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t];console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let r=c(n);r={...r||{},model:{...o.runtime,...r?.model||{},repo_id:n}};let i=await J(`ggml-llm`,null,{config:r,includeBreakdown:!0});l.push({modelId:n,capabilities:i,modelInfo:i.buttress?.selected||null,modelConfig:i.modelConfig||null})}let u=e=>e?(e/1024/1024/1024).toFixed(2):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model ID | Quantization | Size (GB) | Context Size | KV Cache Size (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |`),n.push(`|----------|--------------|-----------|--------------|--------------------|-----------------------------|-----------------|-----------------|`),l.forEach(({modelId:e,modelInfo:t,modelConfig:r})=>{let i=t?.quantization?.name?.toUpperCase()||`N/A`,a=u(t?.modelBytes),o=r?.nCtx||t?.kvInfo?.nCtxTrain||`N/A`,s=Ke(t),c=Number(o),l=t?.kvCacheBytes||(s&&Number.isFinite(c)&&c>0?s(c):s&&s(t?.kvInfo?.nCtxTrain||0))||null,f=u(l),p=u(t?.modelBytes&&l?t.modelBytes+l:t?.fit?.totalRequiredBytes),m=d(t?.fit?.fitsInGpu),h=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${i} | ${a} | ${o} | ${f} | ${p} | ${m} | ${h} |`);let g=t?.memoryLimitedCtx!=null||t?.limitedFit!=null,_=!t?.fit?.fitsInGpu||!t?.fit?.fitsInCpu;if(g&&_){let e=t?.memoryLimitedCtx||o,r=Number(e),i=t?.limitedKvCacheBytes||s&&Number.isFinite(r)&&r>0&&s(r)||null,c=u(i),l=u(t?.modelBytes&&i?t.modelBytes+i:t?.limitedFit?.totalRequiredBytes),m=d(t?.limitedFit?.fitsInGpu),h=d(t?.limitedFit?.fitsInCpu);(e!==o||c!==f||l!==p)&&n.push(`| ↳ Limited | - | ${a} | ${e} | ${c} | ${l} | ${m} | ${h} |`)}}),n.push(`
12
12
  ---`),n.push(`
13
- ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${v.hostname()}`),n.push(`- **OS:** ${v.type()} ${v.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${v.cpus().length}`),n.push(`- **Total System Memory:** ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
13
+ ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${b.hostname()}`),n.push(`- **OS:** ${b.type()} ${b.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${b.cpus().length}`),n.push(`- **Total System Memory:** ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
14
14
  ### Command Used`);let p=process.argv.slice(2).join(` `);if(n.push(`\`\`\`bash\n${process.argv[0]} ${process.argv[1]} ${p}\n\`\`\``),n.push(`
15
- ### Package Information`),n.push(`- **Name:** ${Y}`),n.push(`- **Version:** ${X}`),jr.description&&n.push(`- **Description:** ${jr.description}`),t&&Object.keys(t).length>0){n.push(`
16
- ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=A.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=_.join(process.cwd(),m);k.writeFileSync(h,n.join(`
17
- `),`utf8`),console.log(`\nModel capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate model table:`,e.message),process.exit(1)}}async function Pr({modelId:e=null,defaultConfig:t=null}={}){console.log(`${Y} v${X}`),console.log(`Testing capabilities for backend: ggml-llm`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{server:n,generators:r=[],...i}=t||{},a=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=a(n[e]||{},t):n[e]=t}),n},o=e=>a(structuredClone(i),e||{}),s=(e=>{if(Array.isArray(r)&&r.length>0){let t=r.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return o(n)}}return Object.keys(i).length>0?o({}):null})(e);e&&(s={...s||{},model:{...s?.model||{},repo_id:e}});let c=await J(`ggml-llm`,null,{config:s,includeBreakdown:!0}),l=c.buttress?.selected||null,u=c.modelConfig||null;if(e||u?.repoId){console.log(`
15
+ ### Package Information`),n.push(`- **Name:** ${Y}`),n.push(`- **Version:** ${X}`),Pr.description&&n.push(`- **Description:** ${Pr.description}`),t&&Object.keys(t).length>0){n.push(`
16
+ ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=j.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=y.join(process.cwd(),m);A.writeFileSync(h,n.join(`
17
+ `),`utf8`),console.log(`\nModel capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate model table:`,e.message),process.exit(1)}}async function Lr({modelId:e=null,defaultConfig:t=null}={}){console.log(`${Y} v${X}`),console.log(`Testing capabilities for backend: ggml-llm`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{server:n,generators:r=[],...i}=t||{},a=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=a(n[e]||{},t):n[e]=t}),n},o=e=>a(structuredClone(i),e||{}),s=(e=>{if(Array.isArray(r)&&r.length>0){let t=r.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return o(n)}}return Object.keys(i).length>0?o({}):null})(e);e&&(s={...s||{},model:{...s?.model||{},repo_id:e}});let c=await J(`ggml-llm`,null,{config:s,includeBreakdown:!0}),l=c.buttress?.selected||null,u=c.modelConfig||null;if(e||u?.repoId){console.log(`
18
18
  === Model Information ===`);let t=e||u?.repoId;console.log(`Repository ID: ${t}`),u?.quantization&&console.log(`Quantization: ${u.quantization}`),u?.nCtx&&console.log(`Context Length: ${u.nCtx}`),l?.quantization?.name&&console.log(`Model Quantization: ${l.quantization.name.toUpperCase()}`);let n=u?.cache_type_k||`f16`,r=u?.cache_type_v||`f16`;if(console.log(`KV Cache Type: K=${n}, V=${r}`),l?.modelBytes&&l?.kvCacheBytes){if(console.log(`Model Size: ${(l.modelBytes/1024/1024/1024).toFixed(2)} GB`),l.kvInfo?console.log(`KV Cache Size: ${(l.kvCacheBytes/1024/1024/1024).toFixed(2)} GB (KV info: ${JSON.stringify(l.kvInfo)})`):console.log(`KV Cache Size: ${(l.kvCacheBytes/1024/1024/1024).toFixed(2)} GB`),console.log(`Total Required Memory: ${((l.modelBytes+l.kvCacheBytes)/1024/1024/1024).toFixed(2)} GB`),l.memoryLimitedCtx!=null){let e=l.memoryLimitedCtx,t=l.kvInfo?.nCtxTrain;t?console.log(`\nMemory-Limited Context: ${e} (Train: ${t})`):console.log(`\nMemory-Limited Context: ${e}`),l.limitedKvCacheBytes!=null&&console.log(`Limited KV Cache Size: ${(l.limitedKvCacheBytes/1024/1024/1024).toFixed(2)} GB`)}}else if(c.buttress?.selected?.fit){let{totalRequiredBytes:e}=c.buttress.selected.fit;console.log(`Total Required Memory: ${(e/1024/1024/1024).toFixed(2)} GB`)}}if(c.buttress?.selected){let{selected:e}=c.buttress;console.log(`
19
- === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${v.hostname()}`),console.log(`OS: ${v.type()} ${v.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${v.cpus().length}`),console.log(`Total System Memory: ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
19
+ === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${b.hostname()}`),console.log(`OS: ${b.type()} ${b.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${b.cpus().length}`),console.log(`Total System Memory: ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
20
20
  --- GPU Details ---`),e.devices.filter(e=>e.type===`gpu`).forEach(t=>{console.log(`GPU Backend: ${t.backend}`),console.log(`GPU Name: ${t.deviceName}`),console.log(`GPU Total Memory: ${(t.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let n=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;console.log(`GPU Usable Memory: ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),t.metadata&&(t.metadata.hasBFloat16&&console.log(`Supports BFloat16: Yes`),t.metadata.hasUnifiedMemory&&console.log(`Unified Memory: Yes`))})):console.log(`GPU: Not available`),console.log(`\nBackend Variant: ${e.variant}`),console.log(`Performance Score: ${e.score}`),e.fit&&(console.log(`
21
21
  --- Model Fit Analysis ---`),console.log(`Fits in GPU: ${e.fit.fitsInGpu?`Yes`:`No`}`),console.log(`Fits in CPU: ${e.fit.fitsInCpu?`Yes`:`No`}`),console.log(`Limiting Factor: ${e.fit.limiting}`),e.limitedFit&&(console.log(`
22
22
  --- Memory-Limited Fit Analysis ---`),console.log(`Limited Total Required: ${(e.limitedFit.totalRequiredBytes/1024/1024/1024).toFixed(2)} GB`),console.log(`Fits in GPU (Limited): ${e.limitedFit.fitsInGpu?`Yes`:`No`}`),console.log(`Fits in CPU (Limited): ${e.limitedFit.fitsInCpu?`Yes`:`No`}`),console.log(`Limiting Factor (Limited): ${e.limitedFit.limiting}`)))}console.log(`
23
- === Full Capabilities JSON ===`),console.log(JSON.stringify(c,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}async function Fr({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${Y} v${X}`),console.log(`Generating STT model capabilities comparison...
23
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(c,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}async function Rr({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${Y} v${X}`),console.log(`Generating STT model capabilities comparison...
24
24
  `),n.push(`${Y} v${X}`),n.push(`## STT Model Capabilities Comparison
25
- `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t],{repoId:r,filename:i}=Mr(n);console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let a=c(r);a={...a||{},model:{...a?.model||{},repo_id:r,...i&&{filename:i}}};let o=await J(`ggml-stt`,null,{config:a,includeBreakdown:!0});l.push({modelId:n,repoId:r,filename:i,capabilities:o,modelInfo:o.buttress?.selected||null,modelConfig:o.modelConfig||null})}let u=e=>e?(e/1024/1024).toFixed(1):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model | Size (MB) | Processing Buffer (MB) | Total Required (MB) | Fits GPU | Fits CPU |`),n.push(`|-------|-----------|------------------------|---------------------|----------|----------|`),l.forEach(({modelId:e,modelInfo:t})=>{let r=u(t?.modelBytes),i=u(t?.processingBytes||t?.kvCacheBytes),a=u(t?.fit?.totalRequiredBytes),o=d(t?.fit?.fitsInGpu),s=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${r} | ${i} | ${a} | ${o} | ${s} |`)}),n.push(`
25
+ `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t],{repoId:r,filename:i}=Fr(n);console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let a=c(r);a={...a||{},model:{...a?.model||{},repo_id:r,...i&&{filename:i}}};let o=await J(`ggml-stt`,null,{config:a,includeBreakdown:!0});l.push({modelId:n,repoId:r,filename:i,capabilities:o,modelInfo:o.buttress?.selected||null,modelConfig:o.modelConfig||null})}let u=e=>e?(e/1024/1024).toFixed(1):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model | Size (MB) | Processing Buffer (MB) | Total Required (MB) | Fits GPU | Fits CPU |`),n.push(`|-------|-----------|------------------------|---------------------|----------|----------|`),l.forEach(({modelId:e,modelInfo:t})=>{let r=u(t?.modelBytes),i=u(t?.processingBytes||t?.kvCacheBytes),a=u(t?.fit?.totalRequiredBytes),o=d(t?.fit?.fitsInGpu),s=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${r} | ${i} | ${a} | ${o} | ${s} |`)}),n.push(`
26
26
  ---`),n.push(`
27
- ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${v.hostname()}`),n.push(`- **OS:** ${v.type()} ${v.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${v.cpus().length}`),n.push(`- **Total System Memory:** ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
27
+ ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${b.hostname()}`),n.push(`- **OS:** ${b.type()} ${b.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${b.cpus().length}`),n.push(`- **Total System Memory:** ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
28
28
  ### Command Used`);let p=process.argv.slice(2).join(` `);if(n.push(`\`\`\`bash\n${process.argv[0]} ${process.argv[1]} ${p}\n\`\`\``),n.push(`
29
- ### Package Information`),n.push(`- **Name:** ${Y}`),n.push(`- **Version:** ${X}`),jr.description&&n.push(`- **Description:** ${jr.description}`),t&&Object.keys(t).length>0){n.push(`
30
- ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=A.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-stt-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=_.join(process.cwd(),m);k.writeFileSync(h,n.join(`
31
- `),`utf8`),console.log(`\nSTT model capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate STT model table:`,e.message),process.exit(1)}}async function Ir({modelId:e=null,defaultConfig:t=null}={}){console.log(`${Y} v${X}`),console.log(`Testing capabilities for backend: ggml-stt`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{repoId:n,filename:r}=Mr(e),{server:i,generators:a=[],...o}=t||{},s=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=s(n[e]||{},t):n[e]=t}),n},c=e=>s(structuredClone(o),e||{}),l=(e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return c(n)}}return Object.keys(o).length>0?c({}):null})(n);n&&(l={...l||{},model:{...o.runtime,...l?.model||{},repo_id:n,...r&&{filename:r}}});let u=await J(`ggml-stt`,null,{config:l,includeBreakdown:!0}),d=u.buttress?.selected||null,f=u.modelConfig||null;if(n||f?.repoId){console.log(`
29
+ ### Package Information`),n.push(`- **Name:** ${Y}`),n.push(`- **Version:** ${X}`),Pr.description&&n.push(`- **Description:** ${Pr.description}`),t&&Object.keys(t).length>0){n.push(`
30
+ ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=j.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-stt-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=y.join(process.cwd(),m);A.writeFileSync(h,n.join(`
31
+ `),`utf8`),console.log(`\nSTT model capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate STT model table:`,e.message),process.exit(1)}}async function zr({modelId:e=null,defaultConfig:t=null}={}){console.log(`${Y} v${X}`),console.log(`Testing capabilities for backend: ggml-stt`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{repoId:n,filename:r}=Fr(e),{server:i,generators:a=[],...o}=t||{},s=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=s(n[e]||{},t):n[e]=t}),n},c=e=>s(structuredClone(o),e||{}),l=(e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return c(n)}}return Object.keys(o).length>0?c({}):null})(n);n&&(l={...l||{},model:{...o.runtime,...l?.model||{},repo_id:n,...r&&{filename:r}}});let u=await J(`ggml-stt`,null,{config:l,includeBreakdown:!0}),d=u.buttress?.selected||null,f=u.modelConfig||null;if(n||f?.repoId){console.log(`
32
32
  === Model Information ===`);let e=n||f?.repoId;console.log(`Repository ID: ${e}`),r&&console.log(`Filename: ${r}`),d?.modelBytes&&console.log(`Model Size: ${(d.modelBytes/1024/1024).toFixed(1)} MB`);let t=d?.processingBytes||d?.kvCacheBytes;if(t&&console.log(`Processing Buffer: ${(t/1024/1024).toFixed(1)} MB`),d?.modelBytes&&t)console.log(`Total Required Memory: ${((d.modelBytes+t)/1024/1024).toFixed(1)} MB`);else if(u.buttress?.selected?.fit){let{totalRequiredBytes:e}=u.buttress.selected.fit;console.log(`Total Required Memory: ${(e/1024/1024).toFixed(1)} MB`)}}if(u.buttress?.selected){let{selected:e}=u.buttress;console.log(`
33
- === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${v.hostname()}`),console.log(`OS: ${v.type()} ${v.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${v.cpus().length}`),console.log(`Total System Memory: ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
33
+ === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${b.hostname()}`),console.log(`OS: ${b.type()} ${b.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${b.cpus().length}`),console.log(`Total System Memory: ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
34
34
  --- GPU Details ---`),e.devices.filter(e=>e.type===`gpu`).forEach(t=>{console.log(`GPU Backend: ${t.backend}`),console.log(`GPU Name: ${t.deviceName}`),console.log(`GPU Total Memory: ${(t.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let n=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;console.log(`GPU Usable Memory: ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),t.metadata&&(t.metadata.hasBFloat16&&console.log(`Supports BFloat16: Yes`),t.metadata.hasUnifiedMemory&&console.log(`Unified Memory: Yes`))})):console.log(`GPU: Not available`),console.log(`\nBackend Variant: ${e.variant}`),console.log(`Performance Score: ${e.score}`),e.fit&&(console.log(`
35
35
  --- Model Fit Analysis ---`),console.log(`Fits in GPU: ${e.fit.fitsInGpu?`Yes`:`No`}`),console.log(`Fits in CPU: ${e.fit.fitsInCpu?`Yes`:`No`}`),console.log(`Limiting Factor: ${e.fit.limiting}`))}console.log(`
36
- === Full Capabilities JSON ===`),console.log(JSON.stringify(u,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}var Lr=e({finalizeGenerator:()=>Vr,generatorRegistry:()=>Z,getCapabilities:()=>J,getModelIdentifier:()=>Gr,ggmlLlm:()=>Hr,ggmlStt:()=>Wr,globalDownloadManager:()=>Rr,mlxLlm:()=>Ur,showModelsTable:()=>Nr,showSttModelsTable:()=>Fr,startGenerator:()=>Br,startModelDownload:()=>qr,status:()=>Kr,testGgmlLlmCapabilities:()=>Pr,testGgmlSttCapabilities:()=>Ir});const Z=new Map,Rr={downloads:new Map,getDownload(e){return this.downloads.get(e)||null},setDownload(e,t){this.downloads.set(e,t)},deleteDownload(e){this.downloads.delete(e)},isDownloading(e){return this.downloads.has(e)},getActiveDownloads(){return Array.from(this.downloads.entries()).map(([e,t])=>({localPath:e,promise:t}))}},zr=e=>{let t=Z.get(e);if(!t)throw Error(`Unknown generator id "${e}"`);return t},Q=(e,t)=>{let n=zr(e);if(n.type!==t)throw Error(`Generator "${e}" does not support ${t} backend`);return n.instance};async function Br(e,t){let n={"ggml-llm":{create:nn,getId:rn},"ggml-stt":{create:Gn,getId:Kn},"mlx-llm":{create:Dr,getId:Or}}[e];if(!n)throw Error(`Unsupported backend type: ${e}`);let r=n.getId(t);if(!r)throw Error(`Buttress generator config missing repo identifier`);let i=`${e}:${r}`,a=Z.get(i);if(a)return a.refCount+=1,a.instance.resetFinalized?.(),{id:a.id,info:a.instance.info};let o=await n.create(i,t,{globalDownloadManager:Rr}),s={id:i,type:o.type,instance:o,refCount:1};return Z.set(i,s),{id:i,info:o.info}}async function Vr(e){let t=Z.get(e);return t?(--t.refCount,t.refCount<=0&&(await t.instance.finalize(),(t.instance.hasPendingReleases?.()??!1)||Z.delete(e)),!0):!1}const Hr={async initContext(e,t){return Q(e,`ggml-llm`).initContext(t)},async completion(e,t){return Q(e,`ggml-llm`).completion(t)},async tokenize(e,t){return Q(e,`ggml-llm`).tokenize(t)},async detokenize(e,t){return Q(e,`ggml-llm`).detokenize(t)},async applyChatTemplate(e,t){return Q(e,`ggml-llm`).applyChatTemplate(t)},async releaseContext(e,t){let n=Z.get(e);if(!n)return{released:!0,alreadyReleased:!0};if(n.type!==`ggml-llm`)throw Error(`Generator "${e}" does not support ggml-llm backend`);return n.instance.releaseContext(t)}},Ur={async initContext(e,t){return Q(e,`mlx-llm`).initContext(t)},async completion(e,t){return Q(e,`mlx-llm`).completion(t)},async tokenize(e,t){return Q(e,`mlx-llm`).tokenize(t)},async detokenize(e,t){return Q(e,`mlx-llm`).detokenize(t)},async applyChatTemplate(e,t){return Q(e,`mlx-llm`).applyChatTemplate(t)},async releaseContext(e,t){let n=Z.get(e);if(!n)return{released:!0,alreadyReleased:!0};if(n.type!==`mlx-llm`)throw Error(`Generator "${e}" does not support mlx-llm backend`);return n.instance.releaseContext(t)}},Wr={async initContext(e,t){return Q(e,`ggml-stt`).initContext(t)},async transcribe(e,t){return Q(e,`ggml-stt`).transcribe(t)},async transcribeData(e,t){return Q(e,`ggml-stt`).transcribeData(t)},async releaseContext(e,t){let n=Z.get(e);if(!n)return{released:!0,alreadyReleased:!0};if(n.type!==`ggml-stt`)throw Error(`Generator "${e}" does not support ggml-stt backend`);return n.instance.releaseContext(t)}};function Gr(e,t){return e===`ggml-llm`?rn(t):e===`ggml-stt`?Kn(t):e===`mlx-llm`?Or(t):null}const Kr={getFullStatus:()=>et(Z),getGgmlLlmStatus:()=>Ze(Z),getGgmlSttStatus:()=>Qe(Z),getMlxLlmStatus:()=>$e(Z),subscribeToStatus:Ye,subscribeToStatusWithId:Xe,llmStatusTracker:H,sttStatusTracker:U,statusEmitter:V};async function qr(e,t,n={}){let r={"ggml-llm":an,"ggml-stt":qn,"mlx-llm":Ar}[e];return r?r(t,Rr,n):{started:!1,localPath:null,repoId:null,error:`Unknown backend type: ${e}`}}var Jr=`@fugood/buttress-server`,Yr=`2.24.1`,Xr={name:Jr,version:Yr,main:`lib/index.mjs`,types:`lib/index.d.mts`,type:`module`,bin:{"bricks-buttress":`./bin/bricks-buttress`},files:[`lib`,`bin`,`config`,`public`],scripts:{typecheck:`tsc --noEmit`,build:`tsdown -c rolldown.config.js`,prepublish:`bun run build`,dev:`bun src/index.ts`,start:`bun lib/index.mjs`,"start-with-node":`node lib/index.mjs`},keywords:[`BRICKS`,`buttress`,`server`],license:`MIT`,dependencies:{"@elysiajs/cors":`^1.1.1`,"@elysiajs/node":`^1.4.2`,"@fugood/llama.node":`^1.7.0-rc.11`,"@fugood/whisper.node":`^1.0.18`,"@huggingface/gguf":`^0.3.2`,"@iarna/toml":`^3.0.0`,bytes:`^3.1.0`,elysia:`^1.4.19`,ms:`^2.1.1`,"node-machine-id":`^1.1.12`,zod:`^3.25.76`},devDependencies:{tsdown:`^0.20.1`,typescript:`^5.9.3`},gitHead:`984a440ed04862f12c65f3cf62bdc70a938fcdd6`};const Zr=()=>({version:Yr,name:Jr,description:Xr.description}),Qr=typeof process<`u`&&process.versions&&process.versions.node,$r=e=>new n({adapter:Qr?t():void 0,...e}),ei=a.Object({id:a.String(),name:a.String(),version:a.String(),generators:a.Array(a.Object({type:a.String()})),authentication:a.Object({required:a.Boolean(),type:a.Literal(`device-group`)})}),ti=({store:{serverInfo:e}})=>({id:e.id,name:e.name,version:e.version,generators:e.generators,authentication:e.authentication});var ni=e=>{let t=$r(),n=e.autodiscover.http?.path??`/buttress/info`;return t.get(n,ti,{response:ei}),t};const ri=typeof process<`u`&&process.versions!=null&&process.versions.node!=null;var ii=$r().post(`/buttress/upload`,async({body:{file:e},store:{config:t}})=>{let n=`${Date.now()}-${e.name.replace(/[^\dA-Za-z]/g,`_`)}`,r=_.join(t.server.temp_file_dir,n);try{return ri?await g(r,await e.stream()):await g(r,await e.arrayBuffer()),{ok:!0,filename:n}}catch(e){return{ok:!1,error:String(e)}}},{body:a.Object({file:a.File()}),response:a.Object({ok:a.Boolean(),filename:a.Optional(a.String()),error:a.Optional(a.String())})}).get(`/buttress/download/:filename`,async({params:{filename:e},store:{config:t},status:n})=>{let i=_.join(t.server.temp_file_dir,e);return _.relative(t.server.temp_file_dir,i).includes(`..`)?(n(400),`Invalid file path`):r(i)},{params:a.Object({filename:a.String()})});const ai=_.dirname(D(import.meta.url)),oi=async()=>{let e=[_.join(ai,`..`,`public`,`status.html`),_.join(ai,`..`,`..`,`public`,`status.html`)];return(await Promise.all(e.map(e=>c.access(e).then(()=>e,()=>null)))).find(e=>e!==null)??null},si=e=>{let{status:t}=e;return t?.getFullStatus?t.getFullStatus():{timestamp:new Date().toISOString(),ggmlLlm:{generators:[],history:{}},ggmlStt:{generators:[],history:{}}}},ci=async()=>{let e=await oi();if(!e)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 t=await c.readFile(e,`utf-8`);return new Response(t,{headers:{"Content-Type":`text/html; charset=utf-8`}})}catch(e){return console.error(`[Status] Failed to serve status page:`,e),new Response(`Status page not found`,{status:404,headers:{"Content-Type":`text/plain`}})}};var li=$r().get(`/status`,ci).get(`/status/`,ci).get(`/buttress/status`,({store:{backend:e}})=>si(e));const ui=[`ggml-llm`,`mlx-llm`],di=new Map;function fi(e,t){return t===`mlx-llm`?e.mlxLlm:e.ggmlLlm}async function pi(e,t,n,r=`[LLM]`){let i=(t.generators||[]).filter(e=>ui.includes(e.type));if(i.length===0)throw Error(`No LLM generator configured. Add a [[generators]] with type = "ggml-llm" or "mlx-llm" to your config.`);let a=i[0],o=n||a.model?.repo_id;if(n){let e=i.find(e=>e.model?.repo_id===n);e&&(a=e)}else o=a.model?.repo_id;let s=a.type||`ggml-llm`,c=o,l=di.get(c);if(l?.initialized)return l;let{generators:u,server:d,...f}=t.global||{},p={...f,...a,model:{...a.model,repo_id:o}};console.log(`${r} Creating ${s} generator for ${c}`);let{id:m}=await e.startGenerator(s,p),h={id:m,type:s,config:p,repoId:o,initialized:!1};return di.set(c,h),await fi(e,s).initContext(m,{}),h.initialized=!0,console.log(`${r} Generator ready: ${c}`),h}function mi(e){let t=e.timings||{},n=t.prompt_n??t.promptN??0,r=t.cache_n??t.cacheN??0,i=t.predicted_n??t.predictedN??0;return{promptTokens:n||e.prompt_tokens||e.promptTokens||0,cachedTokens:r,completionTokens:i||e.tokens_evaluated||e.tokensEvaluated||e.tokens_predicted||e.tokensPredicted||0}}function hi(e){let{promptTokens:t,cachedTokens:n,completionTokens:r}=mi(e),i=t+n;return{prompt_tokens:i,completion_tokens:r,total_tokens:i+r}}const gi=()=>`chatcmpl-${Date.now()}-${Math.random().toString(36).slice(2,9)}`;async function _i(e,t,n,r){let i=e.getReader(),a=``,o=null,s=null,c=`stop`,l={prompt_tokens:0,completion_tokens:0,total_tokens:0};try{let e=!1;for(;!e;){let n=await i.read();if({done:e}=n,e)break;let{event:r,data:u}=n.value;if(r===`token`)u.content!=null&&(a=u.content),u.reasoning_content!=null&&(o=u.reasoning_content);else if(r===`result`)u.content==null?u.text&&(a=u.text):a=u.content,u.reasoning_content!=null&&(o=u.reasoning_content),u.tool_calls?.length>0?(s=u.tool_calls.map((e,n)=>({id:e.id||`call_${t}_${n}`,type:`function`,function:{name:e.function?.name||``,arguments:e.function?.arguments||``}})),c=`tool_calls`):c=u.interrupted?`length`:`stop`,l=hi(u);else if(r===`error`)throw Error(u.message)}}finally{i.cancel().catch(()=>{})}let u={role:`assistant`,content:a||null};return o&&(u.reasoning_content=o),s&&(u.tool_calls=s),{id:t,object:`chat.completion`,created:n,model:r,choices:[{index:0,message:u,finish_reason:c}],usage:l}}function vi({global:e}){let t=$r({prefix:`/oai-compat`});return t.use(M({origin:e?.openai_compat?.cors_allowed_origins??!1,methods:[`GET`,`POST`,`OPTIONS`],allowedHeaders:[`Content-Type`,`Authorization`],maxAge:86400,preflight:!0})),t.get(`/v1/models`,({store:e})=>{let{config:t}=e,n=(t.generators||[]).filter(e=>ui.includes(e.type)).map(e=>({id:e.model?.repo_id||e.type,object:`model`,created:Math.floor(Date.now()/1e3),owned_by:`local`}));return n.length===0&&n.push({id:`ggml-llm`,object:`model`,created:Math.floor(Date.now()/1e3),owned_by:`local`}),{object:`list`,data:n}}),t.post(`/v1/chat/completions`,async function*({body:e,set:t,store:n}){let{config:r,backend:a}=n,{messages:o=[],stream:s=!1,model:c,tools:l,temperature:u,stop:d,top_p:f,max_tokens:p,presence_penalty:m,frequency_penalty:h,tool_choice:g,stream_options:_,enable_thinking:v}=e;if(!o||o.length===0)return t.status=400,{error:{message:`messages is required and must not be empty`,type:`invalid_request_error`}};try{let e=await pi(a,r,c,`[OpenAI]`),t=gi(),n=Math.floor(Date.now()/1e3),y=e.repoId||`ggml-llm`,b={reasoning_format:`auto`,messages:o,jinja:!0,add_generation_prompt:!0};u!=null&&(b.temperature=u),f!=null&&(b.top_p=f),p!=null&&(b.n_predict=p),d!=null&&(b.stop=Array.isArray(d)?d:[d]),m!=null&&(b.presence_penalty=m),h!=null&&(b.frequency_penalty=h),l!=null&&(b.tools=l),g!=null&&(b.tool_choice=g),b.enable_thinking=v??!1;let x=await fi(a,e.type).completion(e.id,{options:b});if(!s)return await _i(x,t,n,y);let S=_?.include_usage===!0,C=x.getReader(),w=``,T=``,E=new Map,D=new Map;try{let e=!1;for(;!e;){let r=await C.read();if({done:e}=r,e)break;let{event:a,data:o}=r.value;if(a===`token`){let e={};if(o.content!=null){let t=o.content;t.length>w.length&&(e.content=t.slice(w.length),w=t)}if(o.reasoning_content!=null){let t=o.reasoning_content;t.length>T.length&&(e.reasoning_content=t.slice(T.length),T=t)}if(o.tool_calls?.length>0){let n=[];o.tool_calls.forEach((e,r)=>{let i={index:r};D.has(r)||(D.set(r,e.id||`call_${t}_${r}`),i.id=D.get(r),i.type=`function`);let a=e.function?.arguments||``,o=E.get(r)||``,s={};!E.has(r)&&e.function?.name&&(s.name=e.function.name),a.length>o.length&&(s.arguments=a.slice(o.length),E.set(r,a)),Object.keys(s).length>0?(i.function=s,n.push(i)):i.id&&(i.function={name:e.function?.name||``,arguments:``},n.push(i))}),n.length>0&&(e.tool_calls=n)}Object.keys(e).length>0&&(yield i({data:JSON.stringify({id:t,object:`chat.completion.chunk`,created:n,model:y,choices:[{index:0,delta:e,finish_reason:null}]})}))}else if(a===`result`){let e=`stop`;o.tool_calls?.length>0||D.size>0?e=`tool_calls`:o.interrupted&&(e=`length`);let r={id:t,object:`chat.completion.chunk`,created:n,model:y,choices:[{index:0,delta:{},finish_reason:e}]};S&&(r.usage=hi(o)),yield i({data:JSON.stringify(r)})}else a===`error`&&(yield i({data:JSON.stringify({error:{message:o.message,type:`server_error`}})}))}yield i({data:`[DONE]`})}finally{C.cancel().catch(()=>{})}}catch(e){return console.error(`[OpenAI] Chat completion error:`,e),t.status=500,{error:{message:e.message||`Internal server error`,type:`server_error`}}}},{body:a.Object({model:a.Optional(a.String()),messages:a.Array(a.Any()),stream:a.Optional(a.Boolean()),temperature:a.Optional(a.Number()),top_p:a.Optional(a.Number()),max_tokens:a.Optional(a.Number()),stop:a.Optional(a.Union([a.String(),a.Array(a.String())])),presence_penalty:a.Optional(a.Number()),frequency_penalty:a.Optional(a.Number()),tools:a.Optional(a.Array(a.Any())),tool_choice:a.Optional(a.Any()),stream_options:a.Optional(a.Object({include_usage:a.Optional(a.Boolean())})),enable_thinking:a.Optional(a.Boolean())})}),t}const yi=()=>`msg_${Date.now()}${Math.random().toString(36).slice(2,11)}`;function bi(e){let t={},n=[];if(e.system!=null){let t=``;if(typeof e.system==`string`)t=e.system;else if(Array.isArray(e.system))for(let n of e.system)n?.type===`text`&&typeof n.text==`string`&&(t+=n.text);t&&n.push({role:`system`,content:t})}if(!Array.isArray(e.messages))throw Error(`'messages' is required and must be an array`);for(let t of e.messages){let e=t?.role||`user`;if(t?.content==null){if(e===`assistant`)continue;n.push(t);continue}if(typeof t.content==`string`){n.push({role:e,content:t.content});continue}if(!Array.isArray(t.content)){n.push(t);continue}let r=[],i=[],a=[],o=``,s=!1;for(let e of t.content){let t=e?.type||``;if(t===`text`)i.push({type:`text`,text:e.text||``});else if(t===`thinking`)o+=e.thinking||``;else if(t===`image`){let t=e.source||{};if(t.type===`base64`){let e=t.media_type||`image/jpeg`,n=t.data||``;i.push({type:`image_url`,image_url:{url:`data:${e};base64,${n}`}})}else t.type===`url`&&i.push({type:`image_url`,image_url:{url:t.url||``}})}else if(t===`tool_use`)r.push({id:e.id||``,type:`function`,function:{name:e.name||``,arguments:JSON.stringify(e.input??{})}}),s=!0;else if(t===`tool_result`){let t=e.tool_use_id||``,n=``,r=e.content;if(typeof r==`string`)n=r;else if(Array.isArray(r))for(let e of r)e?.type===`text`&&(n+=e.text||``);a.push({role:`tool`,tool_call_id:t,content:n})}}if(i.length>0||s||o){let t={role:e};i.length>0?t.content=i:(s||o)&&(t.content=``),r.length>0&&(t.tool_calls=r),o&&(t.reasoning_content=o),n.push(t)}for(let e of a)n.push(e)}if(t.messages=n,Array.isArray(e.tools)&&(t.tools=e.tools.map(e=>({type:`function`,function:{name:e.name||``,description:e.description||``,parameters:e.input_schema||{}}}))),e.tool_choice&&typeof e.tool_choice==`object`){let n=e.tool_choice.type;n===`auto`?t.tool_choice=`auto`:n===`any`||n===`tool`?t.tool_choice=`required`:n===`none`&&(t.tool_choice=`none`)}else Array.isArray(t.tools)&&t.tools.length>0&&(t.tool_choice=`auto`);e.stop_sequences!=null&&(t.stop=Array.isArray(e.stop_sequences)?e.stop_sequences:[e.stop_sequences]),t.max_tokens=e.max_tokens??4096;for(let n of[`temperature`,`top_p`,`top_k`,`stream`])e[n]!=null&&(t[n]=e[n]);return e.thinking&&typeof e.thinking==`object`&&e.thinking.type===`enabled`&&(t.enable_thinking=!0,e.thinking.budget_tokens!=null&&(t.thinking_budget_tokens=e.thinking.budget_tokens)),t}function xi(e,t){return t?`tool_use`:e.stopping_word||e.stoppingWord?`stop_sequence`:e.interrupted||e.truncated?`max_tokens`:`end_turn`}function Si(e){let{promptTokens:t,cachedTokens:n,completionTokens:r}=mi(e);return{cache_read_input_tokens:n,input_tokens:t,output_tokens:r}}async function Ci(e,t,n){let r=e.getReader(),i=``,a=``,o=[],s={cache_read_input_tokens:0,input_tokens:0,output_tokens:0},c=`end_turn`,l=null;try{let e=!1;for(;!e;){let t=await r.read();if({done:e}=t,e)break;let{event:n,data:u}=t.value;if(n===`token`)u.content!=null&&(i=u.content),u.reasoning_content!=null&&(a=u.reasoning_content);else if(n===`result`)u.content==null?u.text&&u.reasoning_content==null&&(i=u.text):i=u.content,u.reasoning_content!=null&&(a=u.reasoning_content),Array.isArray(u.tool_calls)&&(o=u.tool_calls),s=Si(u),l=u.stopping_word||u.stoppingWord||null,c=xi(u,o.length>0);else if(n===`error`)throw Error(u.message||`completion error`)}}finally{r.cancel().catch(()=>{})}let u=[];a&&u.push({type:`thinking`,thinking:a,signature:``}),i&&u.push({type:`text`,text:i});for(let e of o){let t={};try{t=JSON.parse(e.function?.arguments||`{}`)}catch{t={}}u.push({type:`tool_use`,id:e.id||`toolu_${Math.random().toString(36).slice(2,11)}`,name:e.function?.name||``,input:t})}return{id:t,type:`message`,role:`assistant`,content:u,model:n,stop_reason:c,stop_sequence:l,usage:s}}function wi({global:e}){let t=$r({prefix:`/anthropic-messages`});return t.use(M({origin:e?.anthropic_messages?.cors_allowed_origins??!1,methods:[`GET`,`POST`,`OPTIONS`],allowedHeaders:[`Content-Type`,`Authorization`,`x-api-key`,`anthropic-version`],maxAge:86400,preflight:!0})),t.post(`/v1/messages`,async function*({body:e,set:t,store:n}){let{config:r,backend:a}=n,o=e;try{if(!Array.isArray(o.messages)||o.messages.length===0)return t.status=400,{type:`error`,error:{type:`invalid_request_error`,message:`messages is required and must not be empty`}};let e=bi(o),n=await pi(a,r,o.model,`[Anthropic]`),s=yi(),c=n.repoId||`ggml-llm`,l={reasoning_format:`auto`,messages:e.messages,jinja:!0,add_generation_prompt:!0};e.temperature!=null&&(l.temperature=e.temperature),e.top_p!=null&&(l.top_p=e.top_p),e.top_k!=null&&(l.top_k=e.top_k),e.max_tokens!=null&&(l.n_predict=e.max_tokens),e.stop!=null&&(l.stop=e.stop),e.tools!=null&&(l.tools=e.tools),e.tool_choice!=null&&(l.tool_choice=e.tool_choice),l.enable_thinking=e.enable_thinking??!1,e.thinking_budget_tokens!=null&&(l.thinking_budget_tokens=e.thinking_budget_tokens);let u=await fi(a,n.type).completion(n.id,{options:l});if(!o.stream)return await Ci(u,s,c);let d=u.getReader(),f=``,p=``,m=new Map,h=new Map,g=new Map,_=new Set,v=!1,y=!1,b=0,x=0,S={cache_read_input_tokens:0,input_tokens:0,output_tokens:0},C=`end_turn`,w=null,T=!1,E=e=>(v?1:0)+(y?1:0)+e;try{let e=!1;for(;!e;){let t=await d.read();if({done:e}=t,e)break;let{event:n,data:r}=t.value;if(n===`token`){if(!T){let e=Si(r);yield i({event:`message_start`,data:JSON.stringify({type:`message_start`,message:{id:s,type:`message`,role:`assistant`,content:[],model:c,stop_reason:null,stop_sequence:null,usage:e}})}),T=!0}if(r.reasoning_content!=null){let e=r.reasoning_content;e.length>p.length&&(v||(yield i({event:`content_block_start`,data:JSON.stringify({type:`content_block_start`,index:0,content_block:{type:`thinking`,thinking:``}})}),v=!0,b=1),yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:0,delta:{type:`thinking_delta`,thinking:e.slice(p.length)}})}),p=e)}if(r.content!=null){let e=r.content;e.length>f.length&&(y||=(yield i({event:`content_block_start`,data:JSON.stringify({type:`content_block_start`,index:b,content_block:{type:`text`,text:``}})}),!0),yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:b,delta:{type:`text_delta`,text:e.slice(f.length)}})}),f=e)}if(Array.isArray(r.tool_calls)&&r.tool_calls.length>0){for(let e=0;e<r.tool_calls.length;e+=1){let t=r.tool_calls[e],n=E(e),a=t?.function?.arguments||``,o=m.get(e)||``;if(!_.has(e)){let r=t?.id||`toolu_${s}_${e}`,a=t?.function?.name||g.get(e)||``;h.set(e,r),g.set(e,a),_.add(e),yield i({event:`content_block_start`,data:JSON.stringify({type:`content_block_start`,index:n,content_block:{type:`tool_use`,id:r,name:a,input:{}}})})}a.length>o.length&&(yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:n,delta:{type:`input_json_delta`,partial_json:a.slice(o.length)}})}),m.set(e,a))}x=r.tool_calls.length}}else if(n===`result`){if(!T){let e=Si(r);yield i({event:`message_start`,data:JSON.stringify({type:`message_start`,message:{id:s,type:`message`,role:`assistant`,content:[],model:c,stop_reason:null,stop_sequence:null,usage:e}})}),T=!0}Array.isArray(r.tool_calls)&&(x=Math.max(x,r.tool_calls.length)),S=Si(r),w=r.stopping_word||r.stoppingWord||null,C=xi(r,_.size>0)}else if(n===`error`){yield i({event:`error`,data:JSON.stringify({type:`error`,error:{type:`api_error`,message:r.message||`completion error`}})});return}}v&&(yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:0,delta:{type:`signature_delta`,signature:``}})}),yield i({event:`content_block_stop`,data:JSON.stringify({type:`content_block_stop`,index:0})})),y&&(yield i({event:`content_block_stop`,data:JSON.stringify({type:`content_block_stop`,index:b})}));for(let e of[..._].sort((e,t)=>e-t))yield i({event:`content_block_stop`,data:JSON.stringify({type:`content_block_stop`,index:E(e)})});yield i({event:`message_delta`,data:JSON.stringify({type:`message_delta`,delta:{stop_reason:C,stop_sequence:w},usage:{output_tokens:S.output_tokens}})}),yield i({event:`message_stop`,data:JSON.stringify({type:`message_stop`})})}finally{d.cancel().catch(()=>{})}}catch(e){return console.error(`[Anthropic] Messages error:`,e),t.status=500,{type:`error`,error:{type:`api_error`,message:e?.message||`Internal server error`}}}},{body:a.Object({model:a.Optional(a.String()),messages:a.Array(a.Any()),system:a.Optional(a.Any()),max_tokens:a.Optional(a.Number()),stream:a.Optional(a.Boolean()),temperature:a.Optional(a.Number()),top_p:a.Optional(a.Number()),top_k:a.Optional(a.Number()),stop_sequences:a.Optional(a.Union([a.String(),a.Array(a.String())])),tools:a.Optional(a.Array(a.Any())),tool_choice:a.Optional(a.Any()),thinking:a.Optional(a.Any()),metadata:a.Optional(a.Any())})}),t.post(`/v1/messages/count_tokens`,async({body:e,set:t,store:n})=>{let{config:r,backend:i}=n,a=e;try{let e=bi(a),t=await pi(i,r,a.model,`[Anthropic]`),n=fi(i,t.type),o={messages:e.messages,add_generation_prompt:!0,jinja:!0};e.tools!=null&&(o.tools=e.tools);let s=await n.applyChatTemplate(t.id,o),c=typeof s==`string`?s:s?.prompt||``,l=await n.tokenize(t.id,{text:c,add_special:!0,parse_special:!0});return{input_tokens:(Array.isArray(l)?l:l?.tokens||[]).length}}catch(e){return console.error(`[Anthropic] count_tokens error:`,e),t.status=500,{type:`error`,error:{type:`api_error`,message:e?.message||`Internal server error`}}}},{body:a.Object({model:a.Optional(a.String()),messages:a.Array(a.Any()),system:a.Optional(a.Any()),tools:a.Optional(a.Array(a.Any())),tool_choice:a.Optional(a.Any())})}),t.get(`/v1/models`,({store:e})=>{let{config:t}=e,n=(t.generators||[]).filter(e=>ui.includes(e.type)).map(e=>{let t=e.model?.repo_id||e.type;return{id:t,type:`model`,display_name:t,created_at:new Date().toISOString()}});return n.length===0&&n.push({id:`ggml-llm`,type:`model`,display_name:`ggml-llm`,created_at:new Date().toISOString()}),{data:n,has_more:!1,first_id:n[0]?.id,last_id:n.at(-1)?.id}}),t}const Ti=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=Ti(n[e]||{},t):n[e]=t}),n},Ei=e=>e&&typeof e==`object`?structuredClone(e):null,Di=(e,t)=>Ti(Ei(e)||{},Ei(t)||{}),Oi=(e,t)=>Ti(structuredClone(e.global),t||{}),ki=(e,t,n,r)=>{if(e.generators.length>0){let i=e.generators.filter(e=>e?.type===n);if(i.length>0&&r){let a=i.find(e=>t.getModelIdentifier(n,e)===r);if(a)return Oi(e,a)}}return Object.keys(e.global).length>0?Oi(e,{}):null},Ai={udp:{port:8089,announcements:{enabled:!0,interval:5e3},requests:{enabled:!0,responseDelay:100}},http:{enabled:!0,path:`/buttress/info`,cors:!0}},ji=e=>e?e===!0?{...Ai}:Ti(Ai,e):null,Mi=(e,t)=>{if(!e.generators||e.generators.length===0)return t.map(e=>({type:e}));let n=new Set;return e.generators.forEach(e=>{e.type&&n.add(e.type)}),n.size===0?t.map(e=>({type:e})):Array.from(n).map(e=>({type:e}))},Ni=(e,t,n)=>e===void 0?n:typeof e==`number`?e:t(e)??n,Pi=6e4,Fi=1024*1024*50,Ii=e=>{let t=N.machineIdSync(),n=Ti({server:{id:`buttress-${t}`,name:`Buttress Server (${t.slice(-8)})`,port:2080,temp_file_dir:_.join(v.tmpdir(),`.buttress`),session_timeout:Pi,max_body_size:Fi},autodiscover:!1},Ei(e)||{}),r=Array.isArray(n.generators)?n.generators:[],{server:i,generators:a,autodiscover:o,...s}=n;return{autodiscover:ji(o),server:{id:i.id,name:i.name,port:i.port,log_level:i.log_level,temp_file_dir:i.temp_file_dir,max_body_size:Ni(i.max_body_size,w.parse,Fi),session_timeout:Ni(i.session_timeout,P,Pi)},global:s,generators:r}},Li={getCapabilities:j.tuple([j.object({type:j.string().optional().default(`ggml-llm`),config:j.any().optional(),currentClientCapabilities:j.any().optional(),options:j.any().optional()}).nullable().optional()]),startGenerator:j.tuple([j.string(),j.any().optional()]),finalizeGenerator:j.tuple([j.string()])};var Ri={async getCapabilities({backend:e,config:t},n=null){console.log(`[Server] Get Capabilities:`,n);let{type:r=`ggml-llm`,config:i,currentClientCapabilities:a=null,options:o={}}=n||{type:`ggml-llm`},s=Ei(i),c=Di(ki(t,e,r,e.getModelIdentifier(r,s)),i);if(Object.keys(c).length===0)throw Error(`Buttress server missing generator configuration`);return c.backend=c.backend||{},c.backend.type||(c.backend.type=r),e.getCapabilities(r,a,{...o,config:c})},async startGenerator({backend:e,config:t,session:n},r,i){console.log(`[Server] Start Generator:`,r,i);let a=Ei(i),o=Di(ki(t,e,r,e.getModelIdentifier(r,a)),i);if(Object.keys(o).length===0)throw Error(`Buttress server missing generator configuration`);o.backend=o.backend||{},o.backend.type||(o.backend.type=r);let s=await e.startGenerator(r,o);return n.generators.add(s.id),s},async finalizeGenerator({backend:e,session:t},n){return console.log(`[Server] Finalize Generator:`,n),t.generators.delete(n),e.finalizeGenerator(n)}};const zi={initContext:j.tuple([j.string(),j.any().optional()]),completion:j.tuple([j.string(),j.any().optional()]),tokenize:j.tuple([j.string(),j.any()]),detokenize:j.tuple([j.string(),j.any()]),applyChatTemplate:j.tuple([j.string(),j.any()]),releaseContext:j.tuple([j.string()])};function Bi(e){return function({backend:t,session:n},r,i){return new s({async start(a){try{let o=await e(t).initContext(r,{...i,onProgress:e=>{a.enqueue({event:`progress`,data:{progress:e}})}});n.initializedContexts.add(r),await new Promise(e=>setTimeout(e,1e3));let{download:s,...c}=o||{};a.enqueue({event:`result`,data:{result:c}}),a.close()}catch(e){a.error(e)}}})}}function Vi(e,t){return async function({backend:n,session:r},i,a){return console.log(`[Server] ${t}:`,{id:i,force:a}),r.initializedContexts.has(i)?(r.initializedContexts.delete(i),e(n).releaseContext(i,{force:a})):(console.log(`[Server] ${t} skipped - not initialized by this session:`,{id:i}),{released:!1,skipped:!0})}}function Hi(e,t){return{initContext:Bi(e),completion({backend:n},r,i){return console.log(`[Server] ${t}Completion:`,{id:r,property:i}),e(n).completion(r,i)},async tokenize({backend:n},r,i){return console.log(`[Server] ${t}Tokenize:`,{id:r,property:i}),e(n).tokenize(r,i)},async detokenize({backend:n},r,i){return console.log(`[Server] ${t}Detokenize:`,{id:r,property:i}),e(n).detokenize(r,i)},async applyChatTemplate({backend:n},r,i){return console.log(`[Server] ${t}Apply Chat Template:`,{id:r,property:i}),e(n).applyChatTemplate(r,i)},releaseContext:Vi(e,`${t}Release Context`)}}var Ui=Hi(e=>e.ggmlLlm,``);const Wi={initContext:j.tuple([j.string(),j.any().optional()]),transcribe:j.tuple([j.string(),j.string(),j.any().optional()]),transcribeData:j.tuple([j.string(),j.union([j.instanceof(Buffer),j.instanceof(Uint8Array)]),j.any().optional()]),releaseContext:j.tuple([j.string()])},Gi=e=>e.ggmlStt,Ki={common:Ri,ggmlLlm:Ui,ggmlStt:{initContext:Bi(Gi),async transcribe({backend:e,config:{server:t}},n,r,i){return console.log(`[Server] Transcribe:`,{id:n,audioPath:r,options:i}),e.ggmlStt.transcribe(n,{audioPath:_.join(t.temp_file_dir,r),options:i})},async transcribeData({backend:e},t,n,r){return console.log(`[Server] Transcribe Data:`,{id:t,audioDataLength:n?.length||0,options:r}),e.ggmlStt.transcribeData(t,{audioData:n,options:r})},releaseContext:Vi(Gi,`Release STT Context`)},mlxLlm:Hi(e=>e.mlxLlm,`MLX `)},qi={common:Li,ggmlLlm:zi,ggmlStt:Wi,mlxLlm:zi};var Ji=Ki;const Yi=e=>{try{return JSON.parse(e,(e,t)=>{if(!t)return t;if(t?.type===`Buffer`&&t?.data)return F.from(t.data,`base64`);if(t?.type===`Uint8Array`&&t?.data){let e=F.from(t.data,`base64`);return e.buffer.slice(e.byteOffset,e.byteOffset+e.byteLength)}return t?.type===`Error`&&t?.name&&t?.message?Error(t.name,t.message):t})}catch{return e}},Xi=e=>{try{return JSON.stringify(e,(e,t)=>t instanceof Error?{type:`Error`,name:t.name,message:t.message}:t instanceof F?{type:`Buffer`,data:t.toString(`base64`)}:t instanceof Uint8Array?{type:`Uint8Array`,data:F.from(t).toString(`base64`)}:t)}catch{return e}};var Zi=class{name=`udp`;socket=null;announcementTimer=null;config;getServerInfo;port;constructor(e,t){this.config=e,this.getServerInfo=t,this.port=e.port??8089}async start(){if(this.socket=re.createSocket({type:`udp4`,reuseAddr:!0}),this.socket.on(`message`,(e,t)=>{this.handleMessage(e,t)}),this.socket.on(`error`,e=>{console.error(`[Autodiscover UDP] Socket error:`,e.message)}),await new Promise((e,t)=>{this.socket.bind(this.port,()=>{this.socket.setBroadcast(!0),console.log(`[Autodiscover UDP] Listening on port ${this.port}`),e()}),this.socket.once(`error`,t)}),this.config.announcements.enabled){let e=this.config.announcements.interval??5e3;this.announcementTimer=setInterval(()=>{this.sendAnnouncement()},e),this.sendAnnouncement()}}async stop(){this.announcementTimer&&=(clearInterval(this.announcementTimer),null),this.socket&&=(await new Promise(e=>{this.socket.close(()=>e())}),null)}handleMessage(e,t){try{let n=JSON.parse(e.toString());if(n.t===`QUERY`&&this.config.requests.enabled){let e=n.d,r=this.config.requests.responseDelay??0,i=r>0?Math.random()*r:0;setTimeout(()=>{this.sendResponse(e.id,t)},i)}}catch{}}sendAnnouncement(){if(!this.socket)return;let e={t:`ANNOUNCE`,v:`1.0`,d:{info:this.getServerInfo()}},t=Buffer.from(JSON.stringify(e));this.socket.send(t,0,t.length,this.port,`255.255.255.255`,e=>{e&&console.error(`[Autodiscover UDP] Announcement error:`,e.message)})}sendResponse(e,t){if(!this.socket)return;let n={t:`RESPONSE`,v:`1.0`,d:{request_id:e,info:this.getServerInfo()}},r=Buffer.from(JSON.stringify(n));this.socket.send(r,0,r.length,t.port,t.address,e=>{e&&console.error(`[Autodiscover UDP] Response error:`,e.message)})}},Qi=class{transports=[];started=!1;constructor(e,t){this.config=e,this.getServerInfo=t,(e.udp?.announcements?.enabled||e.udp?.requests?.enabled)&&this.transports.push(new Zi(e.udp,t))}async start(){this.started||=((await Promise.allSettled(this.transports.map(e=>e.start()))).forEach((e,t)=>{e.status===`rejected`&&console.error(`[Autodiscover] Failed to start ${this.transports[t].name}:`,e.reason)}),!0)}async stop(){this.started&&=(await Promise.allSettled(this.transports.map(e=>e.stop())),!1)}};const $i=()=>{let e=v.networkInterfaces();return Object.values(e).flat().find(e=>e?.family===`IPv4`&&!e?.internal)?.address||null},$=Zr(),ea=e=>{if(!e)return{repoId:null,filename:null};let[t,n]=e.split(`:`);return{repoId:t,filename:n||null}};async function ta({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${$.name} v${$.version}`),console.log(`Generating model capabilities comparison...
36
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(u,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}var Br=e({finalizeGenerator:()=>Wr,generatorRegistry:()=>Z,getCapabilities:()=>J,getModelIdentifier:()=>Jr,ggmlLlm:()=>Gr,ggmlStt:()=>qr,globalDownloadManager:()=>Vr,mlxLlm:()=>Kr,showModelsTable:()=>Ir,showSttModelsTable:()=>Rr,startGenerator:()=>Ur,startModelDownload:()=>Xr,status:()=>Yr,testGgmlLlmCapabilities:()=>Lr,testGgmlSttCapabilities:()=>zr});const Z=new Map,Vr={downloads:new Map,getDownload(e){return this.downloads.get(e)||null},setDownload(e,t){this.downloads.set(e,t)},deleteDownload(e){this.downloads.delete(e)},isDownloading(e){return this.downloads.has(e)},getActiveDownloads(){return Array.from(this.downloads.entries()).map(([e,t])=>({localPath:e,promise:t}))}},Hr=e=>{let t=Z.get(e);if(!t)throw Error(`Unknown generator id "${e}"`);return t},Q=(e,t)=>{let n=Hr(e);if(n.type!==t)throw Error(`Generator "${e}" does not support ${t} backend`);return n.instance};async function Ur(e,t){let n={"ggml-llm":{create:on,getId:sn},"ggml-stt":{create:Jn,getId:Yn},"mlx-llm":{create:Ar,getId:jr}}[e];if(!n)throw Error(`Unsupported backend type: ${e}`);let r=n.getId(t);if(!r)throw Error(`Buttress generator config missing repo identifier`);let i=`${e}:${r}`,a=Z.get(i);if(a)return a.refCount+=1,a.instance.resetFinalized?.(),{id:a.id,info:a.instance.info};let o=await n.create(i,t,{globalDownloadManager:Vr}),s={id:i,type:o.type,instance:o,refCount:1};return Z.set(i,s),{id:i,info:o.info}}async function Wr(e){let t=Z.get(e);return t?(--t.refCount,t.refCount<=0&&(await t.instance.finalize(),(t.instance.hasPendingReleases?.()??!1)||Z.delete(e)),!0):!1}const Gr={async initContext(e,t){return Q(e,`ggml-llm`).initContext(t)},async completion(e,t){return Q(e,`ggml-llm`).completion(t)},async tokenize(e,t){return Q(e,`ggml-llm`).tokenize(t)},async detokenize(e,t){return Q(e,`ggml-llm`).detokenize(t)},async applyChatTemplate(e,t){return Q(e,`ggml-llm`).applyChatTemplate(t)},async releaseContext(e,t){let n=Z.get(e);if(!n)return{released:!0,alreadyReleased:!0};if(n.type!==`ggml-llm`)throw Error(`Generator "${e}" does not support ggml-llm backend`);return n.instance.releaseContext(t)}},Kr={async initContext(e,t){return Q(e,`mlx-llm`).initContext(t)},async completion(e,t){return Q(e,`mlx-llm`).completion(t)},async tokenize(e,t){return Q(e,`mlx-llm`).tokenize(t)},async detokenize(e,t){return Q(e,`mlx-llm`).detokenize(t)},async applyChatTemplate(e,t){return Q(e,`mlx-llm`).applyChatTemplate(t)},async releaseContext(e,t){let n=Z.get(e);if(!n)return{released:!0,alreadyReleased:!0};if(n.type!==`mlx-llm`)throw Error(`Generator "${e}" does not support mlx-llm backend`);return n.instance.releaseContext(t)}},qr={async initContext(e,t){return Q(e,`ggml-stt`).initContext(t)},async transcribe(e,t){return Q(e,`ggml-stt`).transcribe(t)},async transcribeData(e,t){return Q(e,`ggml-stt`).transcribeData(t)},async releaseContext(e,t){let n=Z.get(e);if(!n)return{released:!0,alreadyReleased:!0};if(n.type!==`ggml-stt`)throw Error(`Generator "${e}" does not support ggml-stt backend`);return n.instance.releaseContext(t)}};function Jr(e,t){return e===`ggml-llm`?sn(t):e===`ggml-stt`?Yn(t):e===`mlx-llm`?jr(t):null}const Yr={getFullStatus:()=>rt(Z),getGgmlLlmStatus:()=>et(Z),getGgmlSttStatus:()=>tt(Z),getMlxLlmStatus:()=>nt(Z),subscribeToStatus:Qe,subscribeToStatusWithId:$e,llmStatusTracker:H,sttStatusTracker:U,statusEmitter:V};async function Xr(e,t,n={}){let r={"ggml-llm":cn,"ggml-stt":Xn,"mlx-llm":Nr}[e];return r?r(t,Vr,n):{started:!1,localPath:null,repoId:null,error:`Unknown backend type: ${e}`}}var Zr=`@fugood/buttress-server`,Qr=`2.24.2`,$r={name:Zr,version:Qr,main:`lib/index.mjs`,types:`lib/index.d.mts`,type:`module`,bin:{"bricks-buttress":`./bin/bricks-buttress`},files:[`lib`,`bin`,`config`,`public`],scripts:{typecheck:`tsc --noEmit`,build:`tsdown -c rolldown.config.js`,prepublish:`bun run build`,dev:`bun src/index.ts`,start:`bun lib/index.mjs`,"start-with-node":`node lib/index.mjs`},keywords:[`BRICKS`,`buttress`,`server`],license:`MIT`,dependencies:{"@elysiajs/cors":`^1.1.1`,"@elysiajs/node":`^1.4.2`,"@fugood/llama.node":`^1.7.0`,"@fugood/whisper.node":`^1.0.18`,"@huggingface/gguf":`^0.3.2`,"@iarna/toml":`^3.0.0`,bytes:`^3.1.0`,elysia:`^1.4.19`,jose:`^5.9.6`,ms:`^2.1.1`,"node-machine-id":`^1.1.12`,zod:`^3.25.76`},devDependencies:{tsdown:`^0.20.1`,typescript:`^5.9.3`},gitHead:`984a440ed04862f12c65f3cf62bdc70a938fcdd6`};const ei=()=>({version:Qr,name:Zr,description:$r.description}),ti=typeof process<`u`&&process.versions&&process.versions.node,ni=e=>new n({adapter:ti?t():void 0,...e}),ri=a.Object({id:a.String(),name:a.String(),version:a.String(),generators:a.Array(a.Object({type:a.String(),score:a.Optional(a.Number()),hasGpu:a.Optional(a.Boolean()),usableBytes:a.Optional(a.Number())})),authentication:a.Object({required:a.Boolean(),type:a.String(),kid:a.Optional(a.String()),bound:a.Optional(a.Boolean())}),workspace:a.Optional(a.Object({id:a.String(),name:a.Optional(a.String())}))}),ii=({store:{serverInfo:e}})=>({id:e.id,name:e.name,version:e.version,generators:e.generators,authentication:e.authentication,workspace:e.workspace});var ai=e=>{let t=ni(),n=e.autodiscover?.http?.path??`/buttress/info`;return t.get(n,ii,{response:ri}),t};let oi=null;const si=async e=>{if(oi&&oi.kid===e.kid)return oi.key;let t=await P(e.issuerPublicKey,`EdDSA`);return oi={kid:e.kid,key:t},t},ci=/^Bearer\s+(.+)$/i,li=(e,t)=>{if(e){let t=e.authorization||e.Authorization;if(t){let e=t.match(ci);return e?e[1].trim():t.trim()}}if(t){let e=t.access_token??t.token;if(typeof e==`string`&&e)return e}return null},ui=async(e,t)=>{if(!e||!t)return null;try{let{payload:n}=await F(e,await si(t),{algorithms:[`EdDSA`]}),r=n;return r.k!==`ba`||r.w_id!==t.id||r.st!==`ws`&&r.st!==`dev`||!r.sid||!r.exp?null:{workspaceId:r.w_id,subjectType:r.st,subjectId:r.sid,jti:r.jti,exp:r.exp}}catch{return null}},di=async({headers:e,query:t,set:n,store:r})=>{let i=r.workspaceState?.workspace;if(i&&!await ui(li(e,t),i))return n.status=401,n.headers&&(n.headers[`WWW-Authenticate`]=`Bearer`),{error:{code:`UNAUTHORIZED`,message:`Invalid or missing workspace access token`}}},fi=typeof process<`u`&&process.versions!=null&&process.versions.node!=null;var pi=ni().onBeforeHandle(di).post(`/buttress/upload`,async({body:{file:e},store:{config:t}})=>{let n=`${Date.now()}-${e.name.replace(/[^\dA-Za-z]/g,`_`)}`,r=y.join(t.server.temp_file_dir,n);try{return fi?await v(r,await e.stream()):await v(r,await e.arrayBuffer()),{ok:!0,filename:n}}catch(e){return{ok:!1,error:String(e)}}},{body:a.Object({file:a.File()}),response:a.Object({ok:a.Boolean(),filename:a.Optional(a.String()),error:a.Optional(a.String())})}).get(`/buttress/download/:filename`,async({params:{filename:e},store:{config:t},status:n})=>{let i=y.join(t.server.temp_file_dir,e);return y.relative(t.server.temp_file_dir,i).includes(`..`)?(n(400),`Invalid file path`):r(i)},{params:a.Object({filename:a.String()})});const mi=y.dirname(ee(import.meta.url)),hi=async()=>{let e=[y.join(mi,`..`,`public`,`status.html`),y.join(mi,`..`,`..`,`public`,`status.html`)];return(await Promise.all(e.map(e=>u.access(e).then(()=>e,()=>null)))).find(e=>e!==null)??null},gi=e=>{let{status:t}=e;return t?.getFullStatus?t.getFullStatus():{timestamp:new Date().toISOString(),ggmlLlm:{generators:[],history:{}},ggmlStt:{generators:[],history:{}}}},_i=async()=>{let e=await hi();if(!e)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 t=await u.readFile(e,`utf-8`);return new Response(t,{headers:{"Content-Type":`text/html; charset=utf-8`}})}catch(e){return console.error(`[Status] Failed to serve status page:`,e),new Response(`Status page not found`,{status:404,headers:{"Content-Type":`text/plain`}})}};var vi=ni().get(`/status`,_i).get(`/status/`,_i).get(`/buttress/status`,({store:{backend:e}})=>gi(e));const yi=[`ggml-llm`,`mlx-llm`],bi=new Map;function xi(e,t){return t===`mlx-llm`?e.mlxLlm:e.ggmlLlm}async function Si(e,t,n,r=`[LLM]`){let i=(t.generators||[]).filter(e=>yi.includes(e.type));if(i.length===0)throw Error(`No LLM generator configured. Add a [[generators]] with type = "ggml-llm" or "mlx-llm" to your config.`);let a=i[0],o=n||a.model?.repo_id;if(n){let e=i.find(e=>e.model?.repo_id===n);e&&(a=e)}else o=a.model?.repo_id;let s=a.type||`ggml-llm`,c=o,l=bi.get(c);if(l?.initialized)return l;let{generators:u,server:d,...f}=t.global||{},p={...f,...a,model:{...a.model,repo_id:o}};console.log(`${r} Creating ${s} generator for ${c}`);let{id:m}=await e.startGenerator(s,p),h={id:m,type:s,config:p,repoId:o,initialized:!1};return bi.set(c,h),await xi(e,s).initContext(m,{}),h.initialized=!0,console.log(`${r} Generator ready: ${c}`),h}function Ci(e){let t=e.timings||{},n=t.prompt_n??t.promptN??0,r=t.cache_n??t.cacheN??0,i=t.predicted_n??t.predictedN??0;return{promptTokens:n||e.prompt_tokens||e.promptTokens||0,cachedTokens:r,completionTokens:i||e.tokens_evaluated||e.tokensEvaluated||e.tokens_predicted||e.tokensPredicted||0}}function wi(e){let{promptTokens:t,cachedTokens:n,completionTokens:r}=Ci(e),i=t+n;return{prompt_tokens:i,completion_tokens:r,total_tokens:i+r}}const Ti=()=>`chatcmpl-${Date.now()}-${Math.random().toString(36).slice(2,9)}`;async function Ei(e,t,n,r){let i=e.getReader(),a=``,o=null,s=null,c=`stop`,l={prompt_tokens:0,completion_tokens:0,total_tokens:0};try{let e=!1;for(;!e;){let n=await i.read();if({done:e}=n,e)break;let{event:r,data:u}=n.value;if(r===`token`)u.content!=null&&(a=u.content),u.reasoning_content!=null&&(o=u.reasoning_content);else if(r===`result`)u.content==null?u.text&&(a=u.text):a=u.content,u.reasoning_content!=null&&(o=u.reasoning_content),u.tool_calls?.length>0?(s=u.tool_calls.map((e,n)=>({id:e.id||`call_${t}_${n}`,type:`function`,function:{name:e.function?.name||``,arguments:e.function?.arguments||``}})),c=`tool_calls`):c=u.interrupted?`length`:`stop`,l=wi(u);else if(r===`error`)throw Error(u.message)}}finally{i.cancel().catch(()=>{})}let u={role:`assistant`,content:a||null};return o&&(u.reasoning_content=o),s&&(u.tool_calls=s),{id:t,object:`chat.completion`,created:n,model:r,choices:[{index:0,message:u,finish_reason:c}],usage:l}}function Di({global:e}){let t=ni({prefix:`/oai-compat`});return t.use(I({origin:e?.openai_compat?.cors_allowed_origins??!1,methods:[`GET`,`POST`,`OPTIONS`],allowedHeaders:[`Content-Type`,`Authorization`],maxAge:86400,preflight:!0})),t.onBeforeHandle(di),t.get(`/v1/models`,({store:e})=>{let{config:t}=e,n=(t.generators||[]).filter(e=>yi.includes(e.type)).map(e=>({id:e.model?.repo_id||e.type,object:`model`,created:Math.floor(Date.now()/1e3),owned_by:`local`}));return n.length===0&&n.push({id:`ggml-llm`,object:`model`,created:Math.floor(Date.now()/1e3),owned_by:`local`}),{object:`list`,data:n}}),t.post(`/v1/chat/completions`,async function*({body:e,set:t,store:n}){let{config:r,backend:a}=n,{messages:o=[],stream:s=!1,model:c,tools:l,temperature:u,stop:d,top_p:f,max_tokens:p,presence_penalty:m,frequency_penalty:h,tool_choice:g,stream_options:_,enable_thinking:v}=e;if(!o||o.length===0)return t.status=400,{error:{message:`messages is required and must not be empty`,type:`invalid_request_error`}};try{let e=await Si(a,r,c,`[OpenAI]`),t=Ti(),n=Math.floor(Date.now()/1e3),y=e.repoId||`ggml-llm`,b={reasoning_format:`auto`,messages:o,jinja:!0,add_generation_prompt:!0};u!=null&&(b.temperature=u),f!=null&&(b.top_p=f),p!=null&&(b.n_predict=p),d!=null&&(b.stop=Array.isArray(d)?d:[d]),m!=null&&(b.presence_penalty=m),h!=null&&(b.frequency_penalty=h),l!=null&&(b.tools=l),g!=null&&(b.tool_choice=g),b.enable_thinking=v??!1;let x=await xi(a,e.type).completion(e.id,{options:b});if(!s)return await Ei(x,t,n,y);let S=_?.include_usage===!0,C=x.getReader(),w=``,T=``,E=new Map,D=new Map;try{let e=!1;for(;!e;){let r=await C.read();if({done:e}=r,e)break;let{event:a,data:o}=r.value;if(a===`token`){let e={};if(o.content!=null){let t=o.content;t.length>w.length&&(e.content=t.slice(w.length),w=t)}if(o.reasoning_content!=null){let t=o.reasoning_content;t.length>T.length&&(e.reasoning_content=t.slice(T.length),T=t)}if(o.tool_calls?.length>0){let n=[];o.tool_calls.forEach((e,r)=>{let i={index:r};D.has(r)||(D.set(r,e.id||`call_${t}_${r}`),i.id=D.get(r),i.type=`function`);let a=e.function?.arguments||``,o=E.get(r)||``,s={};!E.has(r)&&e.function?.name&&(s.name=e.function.name),a.length>o.length&&(s.arguments=a.slice(o.length),E.set(r,a)),Object.keys(s).length>0?(i.function=s,n.push(i)):i.id&&(i.function={name:e.function?.name||``,arguments:``},n.push(i))}),n.length>0&&(e.tool_calls=n)}Object.keys(e).length>0&&(yield i({data:JSON.stringify({id:t,object:`chat.completion.chunk`,created:n,model:y,choices:[{index:0,delta:e,finish_reason:null}]})}))}else if(a===`result`){let e=`stop`;o.tool_calls?.length>0||D.size>0?e=`tool_calls`:o.interrupted&&(e=`length`);let r={id:t,object:`chat.completion.chunk`,created:n,model:y,choices:[{index:0,delta:{},finish_reason:e}]};S&&(r.usage=wi(o)),yield i({data:JSON.stringify(r)})}else a===`error`&&(yield i({data:JSON.stringify({error:{message:o.message,type:`server_error`}})}))}yield i({data:`[DONE]`})}finally{C.cancel().catch(()=>{})}}catch(e){return console.error(`[OpenAI] Chat completion error:`,e),t.status=500,{error:{message:e.message||`Internal server error`,type:`server_error`}}}},{body:a.Object({model:a.Optional(a.String()),messages:a.Array(a.Any()),stream:a.Optional(a.Boolean()),temperature:a.Optional(a.Number()),top_p:a.Optional(a.Number()),max_tokens:a.Optional(a.Number()),stop:a.Optional(a.Union([a.String(),a.Array(a.String())])),presence_penalty:a.Optional(a.Number()),frequency_penalty:a.Optional(a.Number()),tools:a.Optional(a.Array(a.Any())),tool_choice:a.Optional(a.Any()),stream_options:a.Optional(a.Object({include_usage:a.Optional(a.Boolean())})),enable_thinking:a.Optional(a.Boolean())})}),t}const Oi=()=>`msg_${Date.now()}${Math.random().toString(36).slice(2,11)}`;function ki(e){let t={},n=[];if(e.system!=null){let t=``;if(typeof e.system==`string`)t=e.system;else if(Array.isArray(e.system))for(let n of e.system)n?.type===`text`&&typeof n.text==`string`&&(t+=n.text);t&&n.push({role:`system`,content:t})}if(!Array.isArray(e.messages))throw Error(`'messages' is required and must be an array`);for(let t of e.messages){let e=t?.role||`user`;if(t?.content==null){if(e===`assistant`)continue;n.push(t);continue}if(typeof t.content==`string`){n.push({role:e,content:t.content});continue}if(!Array.isArray(t.content)){n.push(t);continue}let r=[],i=[],a=[],o=``,s=!1;for(let e of t.content){let t=e?.type||``;if(t===`text`)i.push({type:`text`,text:e.text||``});else if(t===`thinking`)o+=e.thinking||``;else if(t===`image`){let t=e.source||{};if(t.type===`base64`){let e=t.media_type||`image/jpeg`,n=t.data||``;i.push({type:`image_url`,image_url:{url:`data:${e};base64,${n}`}})}else t.type===`url`&&i.push({type:`image_url`,image_url:{url:t.url||``}})}else if(t===`tool_use`)r.push({id:e.id||``,type:`function`,function:{name:e.name||``,arguments:JSON.stringify(e.input??{})}}),s=!0;else if(t===`tool_result`){let t=e.tool_use_id||``,n=``,r=e.content;if(typeof r==`string`)n=r;else if(Array.isArray(r))for(let e of r)e?.type===`text`&&(n+=e.text||``);a.push({role:`tool`,tool_call_id:t,content:n})}}if(i.length>0||s||o){let t={role:e};i.length>0?t.content=i:(s||o)&&(t.content=``),r.length>0&&(t.tool_calls=r),o&&(t.reasoning_content=o),n.push(t)}for(let e of a)n.push(e)}if(t.messages=n,Array.isArray(e.tools)&&(t.tools=e.tools.map(e=>({type:`function`,function:{name:e.name||``,description:e.description||``,parameters:e.input_schema||{}}}))),e.tool_choice&&typeof e.tool_choice==`object`){let n=e.tool_choice.type;n===`auto`?t.tool_choice=`auto`:n===`any`||n===`tool`?t.tool_choice=`required`:n===`none`&&(t.tool_choice=`none`)}else Array.isArray(t.tools)&&t.tools.length>0&&(t.tool_choice=`auto`);e.stop_sequences!=null&&(t.stop=Array.isArray(e.stop_sequences)?e.stop_sequences:[e.stop_sequences]),t.max_tokens=e.max_tokens??4096;for(let n of[`temperature`,`top_p`,`top_k`,`stream`])e[n]!=null&&(t[n]=e[n]);return e.thinking&&typeof e.thinking==`object`&&e.thinking.type===`enabled`&&(t.enable_thinking=!0,e.thinking.budget_tokens!=null&&(t.thinking_budget_tokens=e.thinking.budget_tokens)),t}function Ai(e,t){return t?`tool_use`:e.stopping_word||e.stoppingWord?`stop_sequence`:e.interrupted||e.truncated?`max_tokens`:`end_turn`}function ji(e){let{promptTokens:t,cachedTokens:n,completionTokens:r}=Ci(e);return{cache_read_input_tokens:n,input_tokens:t,output_tokens:r}}async function Mi(e,t,n){let r=e.getReader(),i=``,a=``,o=[],s={cache_read_input_tokens:0,input_tokens:0,output_tokens:0},c=`end_turn`,l=null;try{let e=!1;for(;!e;){let t=await r.read();if({done:e}=t,e)break;let{event:n,data:u}=t.value;if(n===`token`)u.content!=null&&(i=u.content),u.reasoning_content!=null&&(a=u.reasoning_content);else if(n===`result`)u.content==null?u.text&&u.reasoning_content==null&&(i=u.text):i=u.content,u.reasoning_content!=null&&(a=u.reasoning_content),Array.isArray(u.tool_calls)&&(o=u.tool_calls),s=ji(u),l=u.stopping_word||u.stoppingWord||null,c=Ai(u,o.length>0);else if(n===`error`)throw Error(u.message||`completion error`)}}finally{r.cancel().catch(()=>{})}let u=[];a&&u.push({type:`thinking`,thinking:a,signature:``}),i&&u.push({type:`text`,text:i});for(let e of o){let t={};try{t=JSON.parse(e.function?.arguments||`{}`)}catch{t={}}u.push({type:`tool_use`,id:e.id||`toolu_${Math.random().toString(36).slice(2,11)}`,name:e.function?.name||``,input:t})}return{id:t,type:`message`,role:`assistant`,content:u,model:n,stop_reason:c,stop_sequence:l,usage:s}}function Ni({global:e}){let t=ni({prefix:`/anthropic-messages`});return t.use(I({origin:e?.anthropic_messages?.cors_allowed_origins??!1,methods:[`GET`,`POST`,`OPTIONS`],allowedHeaders:[`Content-Type`,`Authorization`,`x-api-key`,`anthropic-version`],maxAge:86400,preflight:!0})),t.onBeforeHandle(di),t.post(`/v1/messages`,async function*({body:e,set:t,store:n}){let{config:r,backend:a}=n,o=e;try{if(!Array.isArray(o.messages)||o.messages.length===0)return t.status=400,{type:`error`,error:{type:`invalid_request_error`,message:`messages is required and must not be empty`}};let e=ki(o),n=await Si(a,r,o.model,`[Anthropic]`),s=Oi(),c=n.repoId||`ggml-llm`,l={reasoning_format:`auto`,messages:e.messages,jinja:!0,add_generation_prompt:!0};e.temperature!=null&&(l.temperature=e.temperature),e.top_p!=null&&(l.top_p=e.top_p),e.top_k!=null&&(l.top_k=e.top_k),e.max_tokens!=null&&(l.n_predict=e.max_tokens),e.stop!=null&&(l.stop=e.stop),e.tools!=null&&(l.tools=e.tools),e.tool_choice!=null&&(l.tool_choice=e.tool_choice),l.enable_thinking=e.enable_thinking??!1,e.thinking_budget_tokens!=null&&(l.thinking_budget_tokens=e.thinking_budget_tokens);let u=await xi(a,n.type).completion(n.id,{options:l});if(!o.stream)return await Mi(u,s,c);let d=u.getReader(),f=``,p=``,m=new Map,h=new Map,g=new Map,_=new Set,v=!1,y=!1,b=0,x=0,S={cache_read_input_tokens:0,input_tokens:0,output_tokens:0},C=`end_turn`,w=null,T=!1,E=e=>(v?1:0)+(y?1:0)+e;try{let e=!1;for(;!e;){let t=await d.read();if({done:e}=t,e)break;let{event:n,data:r}=t.value;if(n===`token`){if(!T){let e=ji(r);yield i({event:`message_start`,data:JSON.stringify({type:`message_start`,message:{id:s,type:`message`,role:`assistant`,content:[],model:c,stop_reason:null,stop_sequence:null,usage:e}})}),T=!0}if(r.reasoning_content!=null){let e=r.reasoning_content;e.length>p.length&&(v||(yield i({event:`content_block_start`,data:JSON.stringify({type:`content_block_start`,index:0,content_block:{type:`thinking`,thinking:``}})}),v=!0,b=1),yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:0,delta:{type:`thinking_delta`,thinking:e.slice(p.length)}})}),p=e)}if(r.content!=null){let e=r.content;e.length>f.length&&(y||=(yield i({event:`content_block_start`,data:JSON.stringify({type:`content_block_start`,index:b,content_block:{type:`text`,text:``}})}),!0),yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:b,delta:{type:`text_delta`,text:e.slice(f.length)}})}),f=e)}if(Array.isArray(r.tool_calls)&&r.tool_calls.length>0){for(let e=0;e<r.tool_calls.length;e+=1){let t=r.tool_calls[e],n=E(e),a=t?.function?.arguments||``,o=m.get(e)||``;if(!_.has(e)){let r=t?.id||`toolu_${s}_${e}`,a=t?.function?.name||g.get(e)||``;h.set(e,r),g.set(e,a),_.add(e),yield i({event:`content_block_start`,data:JSON.stringify({type:`content_block_start`,index:n,content_block:{type:`tool_use`,id:r,name:a,input:{}}})})}a.length>o.length&&(yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:n,delta:{type:`input_json_delta`,partial_json:a.slice(o.length)}})}),m.set(e,a))}x=r.tool_calls.length}}else if(n===`result`){if(!T){let e=ji(r);yield i({event:`message_start`,data:JSON.stringify({type:`message_start`,message:{id:s,type:`message`,role:`assistant`,content:[],model:c,stop_reason:null,stop_sequence:null,usage:e}})}),T=!0}Array.isArray(r.tool_calls)&&(x=Math.max(x,r.tool_calls.length)),S=ji(r),w=r.stopping_word||r.stoppingWord||null,C=Ai(r,_.size>0)}else if(n===`error`){yield i({event:`error`,data:JSON.stringify({type:`error`,error:{type:`api_error`,message:r.message||`completion error`}})});return}}v&&(yield i({event:`content_block_delta`,data:JSON.stringify({type:`content_block_delta`,index:0,delta:{type:`signature_delta`,signature:``}})}),yield i({event:`content_block_stop`,data:JSON.stringify({type:`content_block_stop`,index:0})})),y&&(yield i({event:`content_block_stop`,data:JSON.stringify({type:`content_block_stop`,index:b})}));for(let e of[..._].sort((e,t)=>e-t))yield i({event:`content_block_stop`,data:JSON.stringify({type:`content_block_stop`,index:E(e)})});yield i({event:`message_delta`,data:JSON.stringify({type:`message_delta`,delta:{stop_reason:C,stop_sequence:w},usage:{output_tokens:S.output_tokens}})}),yield i({event:`message_stop`,data:JSON.stringify({type:`message_stop`})})}finally{d.cancel().catch(()=>{})}}catch(e){return console.error(`[Anthropic] Messages error:`,e),t.status=500,{type:`error`,error:{type:`api_error`,message:e?.message||`Internal server error`}}}},{body:a.Object({model:a.Optional(a.String()),messages:a.Array(a.Any()),system:a.Optional(a.Any()),max_tokens:a.Optional(a.Number()),stream:a.Optional(a.Boolean()),temperature:a.Optional(a.Number()),top_p:a.Optional(a.Number()),top_k:a.Optional(a.Number()),stop_sequences:a.Optional(a.Union([a.String(),a.Array(a.String())])),tools:a.Optional(a.Array(a.Any())),tool_choice:a.Optional(a.Any()),thinking:a.Optional(a.Any()),metadata:a.Optional(a.Any())})}),t.post(`/v1/messages/count_tokens`,async({body:e,set:t,store:n})=>{let{config:r,backend:i}=n,a=e;try{let e=ki(a),t=await Si(i,r,a.model,`[Anthropic]`),n=xi(i,t.type),o={messages:e.messages,add_generation_prompt:!0,jinja:!0};e.tools!=null&&(o.tools=e.tools);let s=await n.applyChatTemplate(t.id,o),c=typeof s==`string`?s:s?.prompt||``,l=await n.tokenize(t.id,{text:c,add_special:!0,parse_special:!0});return{input_tokens:(Array.isArray(l)?l:l?.tokens||[]).length}}catch(e){return console.error(`[Anthropic] count_tokens error:`,e),t.status=500,{type:`error`,error:{type:`api_error`,message:e?.message||`Internal server error`}}}},{body:a.Object({model:a.Optional(a.String()),messages:a.Array(a.Any()),system:a.Optional(a.Any()),tools:a.Optional(a.Array(a.Any())),tool_choice:a.Optional(a.Any())})}),t.get(`/v1/models`,({store:e})=>{let{config:t}=e,n=(t.generators||[]).filter(e=>yi.includes(e.type)).map(e=>{let t=e.model?.repo_id||e.type;return{id:t,type:`model`,display_name:t,created_at:new Date().toISOString()}});return n.length===0&&n.push({id:`ggml-llm`,type:`model`,display_name:`ggml-llm`,created_at:new Date().toISOString()}),{data:n,has_more:!1,first_id:n[0]?.id,last_id:n.at(-1)?.id}}),t}const Pi=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=Pi(n[e]||{},t):n[e]=t}),n},Fi=e=>e&&typeof e==`object`?structuredClone(e):null,Ii=(e,t)=>Pi(Fi(e)||{},Fi(t)||{}),Li=(e,t)=>Pi(structuredClone(e.global),t||{}),Ri=(e,t,n,r)=>{if(e.generators.length>0){let i=e.generators.filter(e=>e?.type===n);if(i.length>0&&r){let a=i.find(e=>t.getModelIdentifier(n,e)===r);if(a)return Li(e,a)}}return Object.keys(e.global).length>0?Li(e,{}):null},zi={udp:{port:8089,announcements:{enabled:!0,interval:5e3},requests:{enabled:!0,responseDelay:100}},http:{enabled:!0,path:`/buttress/info`,cors:!0}},Bi=e=>e?e===!0?{...zi}:Pi(zi,e):null,Vi=(e,t)=>{if(!e.generators||e.generators.length===0)return t.map(e=>({type:e}));let n=new Set;return e.generators.forEach(e=>{e.type&&n.add(e.type)}),n.size===0?t.map(e=>({type:e})):Array.from(n).map(e=>({type:e}))},Hi=(e,t,n)=>e===void 0?n:typeof e==`number`?e:t(e)??n,Ui=6e4,Wi=1024*1024*50,Gi=e=>{let t=ne.machineIdSync(),n=Pi({server:{id:`buttress-${t}`,name:`Buttress Server (${t.slice(-8)})`,port:2080,temp_file_dir:y.join(b.tmpdir(),`.buttress`),session_timeout:Ui,max_body_size:Wi},autodiscover:!1},Fi(e)||{}),r=Array.isArray(n.generators)?n.generators:[],{server:i,generators:a,autodiscover:o,...s}=n;return{autodiscover:Bi(o),server:{id:i.id,name:i.name,port:i.port,log_level:i.log_level,temp_file_dir:i.temp_file_dir,max_body_size:Hi(i.max_body_size,T.parse,Wi),session_timeout:Hi(i.session_timeout,re,Ui)},global:s,generators:r}},Ki={getCapabilities:N.tuple([N.object({type:N.string().optional().default(`ggml-llm`),config:N.any().optional(),currentClientCapabilities:N.any().optional(),options:N.any().optional()}).nullable().optional()]),startGenerator:N.tuple([N.string(),N.any().optional()]),finalizeGenerator:N.tuple([N.string()])};var qi={async getCapabilities({backend:e,config:t},n=null){console.log(`[Server] Get Capabilities:`,n);let{type:r=`ggml-llm`,config:i,currentClientCapabilities:a=null,options:o={}}=n||{type:`ggml-llm`},s=Fi(i),c=Ii(Ri(t,e,r,e.getModelIdentifier(r,s)),i);if(Object.keys(c).length===0)throw Error(`Buttress server missing generator configuration`);return c.backend=c.backend||{},c.backend.type||(c.backend.type=r),e.getCapabilities(r,a,{...o,config:c})},async startGenerator({backend:e,config:t,session:n},r,i){console.log(`[Server] Start Generator:`,r,i);let a=Fi(i),o=Ii(Ri(t,e,r,e.getModelIdentifier(r,a)),i);if(Object.keys(o).length===0)throw Error(`Buttress server missing generator configuration`);o.backend=o.backend||{},o.backend.type||(o.backend.type=r);let s=await e.startGenerator(r,o);return n.generators.add(s.id),s},async finalizeGenerator({backend:e,session:t},n){return console.log(`[Server] Finalize Generator:`,n),t.generators.delete(n),e.finalizeGenerator(n)}};const Ji={initContext:N.tuple([N.string(),N.any().optional()]),completion:N.tuple([N.string(),N.any().optional()]),tokenize:N.tuple([N.string(),N.any()]),detokenize:N.tuple([N.string(),N.any()]),applyChatTemplate:N.tuple([N.string(),N.any()]),releaseContext:N.tuple([N.string()])};function Yi(e){return function({backend:t,session:n},r,i){return new l({async start(a){try{let o=await e(t).initContext(r,{...i,onProgress:e=>{a.enqueue({event:`progress`,data:{progress:e}})}});n.initializedContexts.add(r),await new Promise(e=>setTimeout(e,1e3));let{download:s,...c}=o||{};a.enqueue({event:`result`,data:{result:c}}),a.close()}catch(e){a.error(e)}}})}}function Xi(e,t){return async function({backend:n,session:r},i,a){return console.log(`[Server] ${t}:`,{id:i,force:a}),r.initializedContexts.has(i)?(r.initializedContexts.delete(i),e(n).releaseContext(i,{force:a})):(console.log(`[Server] ${t} skipped - not initialized by this session:`,{id:i}),{released:!1,skipped:!0})}}function Zi(e,t){return{initContext:Yi(e),completion({backend:n},r,i){return console.log(`[Server] ${t}Completion:`,{id:r,property:i}),e(n).completion(r,i)},async tokenize({backend:n},r,i){return console.log(`[Server] ${t}Tokenize:`,{id:r,property:i}),e(n).tokenize(r,i)},async detokenize({backend:n},r,i){return console.log(`[Server] ${t}Detokenize:`,{id:r,property:i}),e(n).detokenize(r,i)},async applyChatTemplate({backend:n},r,i){return console.log(`[Server] ${t}Apply Chat Template:`,{id:r,property:i}),e(n).applyChatTemplate(r,i)},releaseContext:Xi(e,`${t}Release Context`)}}var Qi=Zi(e=>e.ggmlLlm,``);const $i={initContext:N.tuple([N.string(),N.any().optional()]),transcribe:N.tuple([N.string(),N.string(),N.any().optional()]),transcribeData:N.tuple([N.string(),N.union([N.instanceof(Buffer),N.instanceof(Uint8Array)]),N.any().optional()]),releaseContext:N.tuple([N.string()])},ea=e=>e.ggmlStt,ta={common:qi,ggmlLlm:Qi,ggmlStt:{initContext:Yi(ea),async transcribe({backend:e,config:{server:t}},n,r,i){return console.log(`[Server] Transcribe:`,{id:n,audioPath:r,options:i}),e.ggmlStt.transcribe(n,{audioPath:y.join(t.temp_file_dir,r),options:i})},async transcribeData({backend:e},t,n,r){return console.log(`[Server] Transcribe Data:`,{id:t,audioDataLength:n?.length||0,options:r}),e.ggmlStt.transcribeData(t,{audioData:n,options:r})},releaseContext:Xi(ea,`Release STT Context`)},mlxLlm:Zi(e=>e.mlxLlm,`MLX `)},na={common:Ki,ggmlLlm:Ji,ggmlStt:$i,mlxLlm:Ji};var ra=ta;const ia=e=>{try{return JSON.parse(e,(e,t)=>{if(!t)return t;if(t?.type===`Buffer`&&t?.data)return L.from(t.data,`base64`);if(t?.type===`Uint8Array`&&t?.data){let e=L.from(t.data,`base64`);return e.buffer.slice(e.byteOffset,e.byteOffset+e.byteLength)}return t?.type===`Error`&&t?.name&&t?.message?Error(t.name,t.message):t})}catch{return e}},aa=e=>{try{return JSON.stringify(e,(e,t)=>t instanceof Error?{type:`Error`,name:t.name,message:t.message}:t instanceof L?{type:`Buffer`,data:t.toString(`base64`)}:t instanceof Uint8Array?{type:`Uint8Array`,data:L.from(t).toString(`base64`)}:t)}catch{return e}},oa={score:0,hasGpu:!1,usableBytes:0},sa=e=>e?{score:Number(e.score)||0,hasGpu:!!e.hasGpu,usableBytes:Number(e.gpuUsableBytes||e.cpuUsableBytes||0)}:oa,ca=async e=>{let t=[];for(let n of e){let e=oa;try{e=sa((await J(n.type,null,{}))?.buttress?.selected)}catch(e){console.warn(`[Caps] Failed to detect capabilities for "${n.type}":`,e instanceof Error?e.message:e)}t.push({...n,...e})}return t},la=()=>{let e=b.networkInterfaces();return Object.values(e).flat().find(e=>e?.family===`IPv4`&&!e?.internal)?.address||null},ua=e=>{let t=e.split(`.`).map(Number);return t.length!==4||t.some(e=>Number.isNaN(e))?0:(t[0]<<24|t[1]<<16|t[2]<<8|t[3])>>>0},da=e=>[e>>>24&255,e>>>16&255,e>>>8&255,e&255].join(`.`),fa=()=>{let e=[],t=new Set;for(let n of Object.values(b.networkInterfaces()))for(let r of n??[]){if(r.family!==`IPv4`||r.internal||!r.address||!r.netmask)continue;let n=ua(r.address),i=ua(r.netmask);if(!n||!i||i===4294967295)continue;let a=da((n&i|~i>>>0)>>>0);if(a===r.address)continue;let o=`${r.address}->${a}`;t.has(o)||(t.add(o),e.push({address:r.address,broadcast:a}))}return e},pa=e=>{if(!e)return!1;let t=e;return t.code===`ENOTSUP`||/Failed to bind socket/i.test(t.message??``)};var ma=class e{name=`udp`;receiver=null;senders=[];announcementTimer=null;config;getServerInfo;port;signer;constructor(e,t,n){this.config=e,this.getServerInfo=t,this.port=e.port??8089,this.signer=n}async start(){this.receiver=await this.bindReceiver(!0).catch(e=>{if(!pa(e))throw e;return console.warn(`[Autodiscover UDP] SO_REUSEPORT not supported by this runtime; falling back to REUSEADDR only (multiple buttress instances on one host will not coexist on the discovery port).`),this.bindReceiver(!1)}),this.receiver.on(`message`,(e,t)=>{this.handleMessage(e,t)}),this.receiver.on(`error`,e=>{console.error(`[Autodiscover UDP] Receiver error:`,e.message)}),this.receiver.setBroadcast(!0),this.senders=await this.createSenders();let e=this.senders.map(e=>`${e.address}->${e.broadcast}`).join(`, `)||`<none>`;if(console.log(`[Autodiscover UDP] Listening on port ${this.port}; announce interfaces: ${e}`),this.config.announcements.enabled){let e=this.config.announcements.interval??5e3;this.announcementTimer=setInterval(()=>{this.sendAnnouncement()},e),this.sendAnnouncement()}}async stop(){this.announcementTimer&&=(clearInterval(this.announcementTimer),null),await Promise.all(this.senders.map(({socket:e})=>new Promise(t=>{e.close(()=>t())}))),this.senders=[],this.receiver&&=(await new Promise(e=>{this.receiver.close(()=>e())}),null)}async bindReceiver(e){let t=R.createSocket(e?{type:`udp4`,reuseAddr:!0,reusePort:!0}:{type:`udp4`,reuseAddr:!0});return new Promise((e,n)=>{let r=e=>{t.close(),n(e)};t.once(`error`,r),t.bind(this.port,()=>{t.off(`error`,r),e(t)})})}async createSenders(){let e=fa();return(await Promise.all(e.map(async e=>{try{let t=R.createSocket({type:`udp4`});return await new Promise((n,r)=>{let i=e=>{t.close(),r(e)};t.once(`error`,i),t.bind({port:0,address:e.address},()=>{t.off(`error`,i),t.setBroadcast(!0),n()})}),t.on(`error`,t=>{console.error(`[Autodiscover UDP] Sender ${e.address} error:`,t.message)}),{...e,socket:t}}catch(t){return console.warn(`[Autodiscover UDP] Failed to bind sender on ${e.address}:`,t.message),null}}))).filter(e=>e!==null)}handleMessage(e,t){try{let n=JSON.parse(e.toString());if(n.t===`QUERY`&&this.config.requests.enabled){let e=n.d,r=this.config.requests.responseDelay??0,i=r>0?Math.random()*r:0;setTimeout(()=>{this.sendResponse(e.id,t)},i)}}catch{}}static canonicalBytes(e,t,n){return Buffer.from(JSON.stringify({t:e,d:t,ts:n}),`utf8`)}signEnvelope(t,n){if(!this.signer)return null;let r=Math.floor(Date.now()/1e3),i=o.sign(null,e.canonicalBytes(t,n,r),this.signer.privateKey).toString(`base64`);return{t,v:`2.0`,d:n,ts:r,kid:this.signer.kid,sig:i}}sendAnnouncement(){if(this.senders.length===0)return;let e=this.getServerInfo(),t=this.signEnvelope(`ANNOUNCE`,{info:e});if(!t){console.warn("[Autodiscover UDP] no per-server keypair; skipping announcement. Run `bricks buttress bind` to register a key.");return}let n=Buffer.from(JSON.stringify(t));for(let{broadcast:e,socket:t,address:r}of this.senders)t.send(n,0,n.length,this.port,e,t=>{t&&console.error(`[Autodiscover UDP] Announcement ${r}->${e} error:`,t.message)})}sendResponse(e,t){if(!this.receiver)return;let n=this.getServerInfo(),r=this.signEnvelope(`RESPONSE`,{request_id:e,info:n});if(!r)return;let i=Buffer.from(JSON.stringify(r));this.receiver.send(i,0,i.length,t.port,t.address,e=>{e&&console.error(`[Autodiscover UDP] Response error:`,e.message)})}},ha=class{transports=[];started=!1;constructor(e,t,n){this.config=e,this.getServerInfo=t,this.signer=n,(e.udp?.announcements?.enabled||e.udp?.requests?.enabled)&&this.transports.push(new ma(e.udp,t,n))}async start(){this.started||=((await Promise.allSettled(this.transports.map(e=>e.start()))).forEach((e,t)=>{e.status===`rejected`&&console.error(`[Autodiscover] Failed to start ${this.transports[t].name}:`,e.reason)}),!0)}async stop(){this.started&&=(await Promise.allSettled(this.transports.map(e=>e.stop())),!1)}};const ga=()=>process.env.BRICKS_BUTTRESS_STATE_DIR||y.join(b.homedir(),`.bricks-cli`,`buttress`),_a=()=>y.join(ga(),`state.json`),va=e=>{if(!e||typeof e!=`object`)return!1;let t=e;return typeof t.id==`string`&&typeof t.serverId==`string`&&typeof t.issuerPublicKey==`string`&&typeof t.kid==`string`},ya=e=>{if(!e||typeof e!=`object`)return!1;let t=e;return typeof t.publicKeySpki==`string`&&typeof t.privateKeyPkcs8==`string`&&typeof t.kid==`string`},ba=()=>{let e=_a();try{let t=A.readFileSync(e,`utf8`),n=JSON.parse(t);return{workspace:va(n?.workspace)?n.workspace:null,serverKeyPair:ya(n?.serverKeyPair)?n.serverKeyPair:null}}catch(e){return e.code!==`ENOENT`&&console.warn(`[Buttress] Failed to read workspace state:`,e.message),{workspace:null,serverKeyPair:null}}},$=ei(),xa=e=>{if(!e)return{repoId:null,filename:null};let[t,n]=e.split(`:`);return{repoId:t,filename:n||null}};async function Sa({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${$.name} v${$.version}`),console.log(`Generating model capabilities comparison...
37
37
  `),n.push(`${$.name} v${$.version}`),n.push(`## Model Capabilities Comparison
38
- `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t];console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let r=c(n);r={...r||{},model:{...o.runtime,...r?.model||{},repo_id:n}};let i=await J(`ggml-llm`,null,{config:r,includeBreakdown:!0});l.push({modelId:n,capabilities:i,modelInfo:i.buttress?.selected||null,modelConfig:i.modelConfig||null})}let u=e=>e?(e/1024/1024/1024).toFixed(2):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model ID | Size (GB) | Context Size | KV Cache Size (GB) | Recurrent Mem (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |`),n.push(`|----------|-----------|--------------|--------------------|--------------------|----------------------------|-----------------|-----------------|`),l.forEach(({modelId:e,modelInfo:t,modelConfig:r})=>{let i=u(t?.modelBytes),a=r?.nCtx||t?.kvInfo?.nCtxTrain||`N/A`,o=Ue(t),s=Number(a),c=t?.kvCacheBytes||(o&&Number.isFinite(s)&&s>0?o(s):o&&o(t?.kvInfo?.nCtxTrain||0))||null,l=u(c),f=t?.recurrentMemoryBytes||0,p=f>0?u(f):`-`,m=u(t?.modelBytes&&(c!=null||f>0)?t.modelBytes+(c||0)+f:t?.fit?.totalRequiredBytes),h=d(t?.fit?.fitsInGpu),g=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${i} | ${a} | ${l} | ${p} | ${m} | ${h} | ${g} |`);let _=t?.memoryLimitedCtx!=null||t?.limitedFit!=null,v=!t?.fit?.fitsInGpu||!t?.fit?.fitsInCpu;if(_&&v){let e=t?.memoryLimitedCtx||a,r=Number(e),s=t?.limitedKvCacheBytes||o&&Number.isFinite(r)&&r>0&&o(r)||null,c=u(s),h=u(t?.modelBytes&&(s!=null||f>0)?t.modelBytes+(s||0)+f:t?.limitedFit?.totalRequiredBytes),g=d(t?.limitedFit?.fitsInGpu),_=d(t?.limitedFit?.fitsInCpu);(e!==a||c!==l||h!==m)&&n.push(`| ↳ Limited | ${i} | ${e} | ${c} | ${p} | ${h} | ${g} | ${_} |`)}}),n.push(`
38
+ `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t];console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let r=c(n);r={...r||{},model:{...o.runtime,...r?.model||{},repo_id:n}};let i=await J(`ggml-llm`,null,{config:r,includeBreakdown:!0});l.push({modelId:n,capabilities:i,modelInfo:i.buttress?.selected||null,modelConfig:i.modelConfig||null})}let u=e=>e?(e/1024/1024/1024).toFixed(2):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model ID | Size (GB) | Context Size | KV Cache Size (GB) | Recurrent Mem (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |`),n.push(`|----------|-----------|--------------|--------------------|--------------------|----------------------------|-----------------|-----------------|`),l.forEach(({modelId:e,modelInfo:t,modelConfig:r})=>{let i=u(t?.modelBytes),a=r?.nCtx||t?.kvInfo?.nCtxTrain||`N/A`,o=Ke(t),s=Number(a),c=t?.kvCacheBytes||(o&&Number.isFinite(s)&&s>0?o(s):o&&o(t?.kvInfo?.nCtxTrain||0))||null,l=u(c),f=t?.recurrentMemoryBytes||0,p=f>0?u(f):`-`,m=u(t?.modelBytes&&(c!=null||f>0)?t.modelBytes+(c||0)+f:t?.fit?.totalRequiredBytes),h=d(t?.fit?.fitsInGpu),g=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${i} | ${a} | ${l} | ${p} | ${m} | ${h} | ${g} |`);let _=t?.memoryLimitedCtx!=null||t?.limitedFit!=null,v=!t?.fit?.fitsInGpu||!t?.fit?.fitsInCpu;if(_&&v){let e=t?.memoryLimitedCtx||a,r=Number(e),s=t?.limitedKvCacheBytes||o&&Number.isFinite(r)&&r>0&&o(r)||null,c=u(s),h=u(t?.modelBytes&&(s!=null||f>0)?t.modelBytes+(s||0)+f:t?.limitedFit?.totalRequiredBytes),g=d(t?.limitedFit?.fitsInGpu),_=d(t?.limitedFit?.fitsInCpu);(e!==a||c!==l||h!==m)&&n.push(`| ↳ Limited | ${i} | ${e} | ${c} | ${p} | ${h} | ${g} | ${_} |`)}}),n.push(`
39
39
  ---`),n.push(`
40
- ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${v.hostname()}`),n.push(`- **OS:** ${v.type()} ${v.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${v.cpus().length}`),n.push(`- **Total System Memory:** ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
40
+ ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${b.hostname()}`),n.push(`- **OS:** ${b.type()} ${b.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${b.cpus().length}`),n.push(`- **Total System Memory:** ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
41
41
  ### Command Used`);let p=process.argv.slice(2).join(` `);if(n.push(`\`\`\`bash\n${process.argv[0]} ${process.argv[1]} ${p}\n\`\`\``),n.push(`
42
42
  ### Package Information`),n.push(`- **Name:** ${$.name}`),n.push(`- **Version:** ${$.version}`),$.description&&n.push(`- **Description:** ${$.description}`),t&&Object.keys(t).length>0){n.push(`
43
- ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=A.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=_.join(process.cwd(),m);k.writeFileSync(h,n.join(`
44
- `),`utf8`),console.log(`\nModel capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate model table:`,e.message),process.exit(1)}}async function na({modelId:e=null,defaultConfig:t=null}={}){console.log(`${$.name} v${$.version}`),console.log(`Testing capabilities for backend: ggml-llm`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{server:n,generators:r=[],...i}=t||{},a=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=a(n[e]||{},t):n[e]=t}),n},o=e=>a(structuredClone(i),e||{}),s=(e=>{if(Array.isArray(r)&&r.length>0){let t=r.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return o(n)}}return Object.keys(i).length>0?o({}):null})(e);e&&(s={...s||{},model:{...s?.model||{},repo_id:e}});let c=await J(`ggml-llm`,null,{config:s,includeBreakdown:!0}),l=c.buttress?.selected||null,u=c.modelConfig||null;if(e||u?.repoId){console.log(`
43
+ ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=j.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=y.join(process.cwd(),m);A.writeFileSync(h,n.join(`
44
+ `),`utf8`),console.log(`\nModel capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate model table:`,e.message),process.exit(1)}}async function Ca({modelId:e=null,defaultConfig:t=null}={}){console.log(`${$.name} v${$.version}`),console.log(`Testing capabilities for backend: ggml-llm`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{server:n,generators:r=[],...i}=t||{},a=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=a(n[e]||{},t):n[e]=t}),n},o=e=>a(structuredClone(i),e||{}),s=(e=>{if(Array.isArray(r)&&r.length>0){let t=r.filter(e=>e?.type===`ggml-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return o(n)}}return Object.keys(i).length>0?o({}):null})(e);e&&(s={...s||{},model:{...s?.model||{},repo_id:e}});let c=await J(`ggml-llm`,null,{config:s,includeBreakdown:!0}),l=c.buttress?.selected||null,u=c.modelConfig||null;if(e||u?.repoId){console.log(`
45
45
  === Model Information ===`);let t=e||u?.repoId;if(console.log(`Repository ID: ${t}`),u?.quantization&&console.log(`Quantization: ${u.quantization}`),u?.nCtx&&console.log(`Context Length: ${u.nCtx}`),l?.quantization){let{fileType:e}=l.quantization;e!=null&&console.log(`Model File Type (GGUF): ${e}`)}let n=u?.cache_type_k||`f16`,r=u?.cache_type_v||`f16`;if(console.log(`KV Cache Type: K=${n}, V=${r}`),l?.modelBytes&&l?.kvCacheBytes!=null){console.log(`Model Size: ${(l.modelBytes/1024/1024/1024).toFixed(2)} GB`),l.kvInfo?console.log(`KV Cache Size: ${(l.kvCacheBytes/1024/1024/1024).toFixed(2)} GB (KV info: ${JSON.stringify(l.kvInfo)})`):console.log(`KV Cache Size: ${(l.kvCacheBytes/1024/1024/1024).toFixed(2)} GB`);let e=l.recurrentMemoryBytes||0;e>0&&console.log(`Recurrent Memory: ${(e/1024/1024/1024).toFixed(2)} GB`);let t=l.modelBytes+l.kvCacheBytes+e;if(console.log(`Total Required Memory: ${(t/1024/1024/1024).toFixed(2)} GB`),l.memoryLimitedCtx!=null){let e=l.memoryLimitedCtx,t=l.kvInfo?.nCtxTrain;t?console.log(`\nMemory-Limited Context: ${e} (Train: ${t})`):console.log(`\nMemory-Limited Context: ${e}`),l.limitedKvCacheBytes!=null&&console.log(`Limited KV Cache Size: ${(l.limitedKvCacheBytes/1024/1024/1024).toFixed(2)} GB`)}}else if(c.buttress?.selected?.fit){let{totalRequiredBytes:e}=c.buttress.selected.fit;console.log(`Total Required Memory: ${(e/1024/1024/1024).toFixed(2)} GB`)}}if(c.buttress?.selected){let{selected:e}=c.buttress;console.log(`
46
- === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${v.hostname()}`),console.log(`OS: ${v.type()} ${v.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${v.cpus().length}`),console.log(`Total System Memory: ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
46
+ === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${b.hostname()}`),console.log(`OS: ${b.type()} ${b.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${b.cpus().length}`),console.log(`Total System Memory: ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
47
47
  --- GPU Details ---`),e.devices.filter(e=>e.type===`gpu`).forEach(t=>{console.log(`GPU Backend: ${t.backend}`),console.log(`GPU Name: ${t.deviceName}`),console.log(`GPU Total Memory: ${(t.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let n=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;console.log(`GPU Usable Memory: ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),t.metadata&&(t.metadata.hasBFloat16&&console.log(`Supports BFloat16: Yes`),t.metadata.hasUnifiedMemory&&console.log(`Unified Memory: Yes`))})):console.log(`GPU: Not available`),console.log(`\nBackend Variant: ${e.variant}`),console.log(`Performance Score: ${e.score}`),e.fit&&(console.log(`
48
48
  --- Model Fit Analysis ---`),console.log(`Fits in GPU: ${e.fit.fitsInGpu?`Yes`:`No`}`),console.log(`Fits in CPU: ${e.fit.fitsInCpu?`Yes`:`No`}`),console.log(`Limiting Factor: ${e.fit.limiting}`),e.limitedFit&&(console.log(`
49
49
  --- Memory-Limited Fit Analysis ---`),console.log(`Limited Total Required: ${(e.limitedFit.totalRequiredBytes/1024/1024/1024).toFixed(2)} GB`),console.log(`Fits in GPU (Limited): ${e.limitedFit.fitsInGpu?`Yes`:`No`}`),console.log(`Fits in CPU (Limited): ${e.limitedFit.fitsInCpu?`Yes`:`No`}`),console.log(`Limiting Factor (Limited): ${e.limitedFit.limiting}`)))}console.log(`
50
- === Full Capabilities JSON ===`),console.log(JSON.stringify(c,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}async function ra({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${$.name} v${$.version}`),console.log(`Generating STT model capabilities comparison...
50
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(c,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}async function wa({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${$.name} v${$.version}`),console.log(`Generating STT model capabilities comparison...
51
51
  `),n.push(`${$.name} v${$.version}`),n.push(`## STT Model Capabilities Comparison
52
- `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t],{repoId:r,filename:i}=ea(n);console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let a=c(r);a={...a||{},model:{...a?.model||{},repo_id:r,...i&&{filename:i}}};let o=await J(`ggml-stt`,null,{config:a,includeBreakdown:!0});l.push({modelId:n,repoId:r,filename:i,capabilities:o,modelInfo:o.buttress?.selected||null,modelConfig:o.modelConfig||null})}let u=e=>e?(e/1024/1024).toFixed(1):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model | Size (MB) | Processing Buffer (MB) | Total Required (MB) | Fits GPU | Fits CPU |`),n.push(`|-------|-----------|------------------------|---------------------|----------|----------|`),l.forEach(({modelId:e,modelInfo:t})=>{let r=u(t?.modelBytes),i=u(t?.processingBytes||t?.kvCacheBytes),a=u(t?.fit?.totalRequiredBytes),o=d(t?.fit?.fitsInGpu),s=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${r} | ${i} | ${a} | ${o} | ${s} |`)}),n.push(`
52
+ `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t],{repoId:r,filename:i}=xa(n);console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let a=c(r);a={...a||{},model:{...a?.model||{},repo_id:r,...i&&{filename:i}}};let o=await J(`ggml-stt`,null,{config:a,includeBreakdown:!0});l.push({modelId:n,repoId:r,filename:i,capabilities:o,modelInfo:o.buttress?.selected||null,modelConfig:o.modelConfig||null})}let u=e=>e?(e/1024/1024).toFixed(1):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model | Size (MB) | Processing Buffer (MB) | Total Required (MB) | Fits GPU | Fits CPU |`),n.push(`|-------|-----------|------------------------|---------------------|----------|----------|`),l.forEach(({modelId:e,modelInfo:t})=>{let r=u(t?.modelBytes),i=u(t?.processingBytes||t?.kvCacheBytes),a=u(t?.fit?.totalRequiredBytes),o=d(t?.fit?.fitsInGpu),s=d(t?.fit?.fitsInCpu);n.push(`| ${e} | ${r} | ${i} | ${a} | ${o} | ${s} |`)}),n.push(`
53
53
  ---`),n.push(`
54
- ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${v.hostname()}`),n.push(`- **OS:** ${v.type()} ${v.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${v.cpus().length}`),n.push(`- **Total System Memory:** ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
54
+ ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${b.hostname()}`),n.push(`- **OS:** ${b.type()} ${b.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **CPU Cores:** ${b.cpus().length}`),n.push(`- **Total System Memory:** ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;if(e){let t=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;if(n.push(`- **Usable CPU Memory:** ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${t}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu){let t=e.devices.filter(e=>e.type===`gpu`);if(t.length>0){let r=t[0];n.push(`- **GPU Backend:** ${r.backend}`),n.push(`- **GPU Name:** ${r.deviceName}`),n.push(`- **GPU Total Memory:** ${(r.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let i=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;n.push(`- **GPU Usable Memory:** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${i}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else n.push(`- **GPU:** Not available`)}}n.push(`
55
55
  ### Command Used`);let p=process.argv.slice(2).join(` `);if(n.push(`\`\`\`bash\n${process.argv[0]} ${process.argv[1]} ${p}\n\`\`\``),n.push(`
56
56
  ### Package Information`),n.push(`- **Name:** ${$.name}`),n.push(`- **Version:** ${$.version}`),$.description&&n.push(`- **Description:** ${$.description}`),t&&Object.keys(t).length>0){n.push(`
57
- ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=A.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-stt-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=_.join(process.cwd(),m);k.writeFileSync(h,n.join(`
58
- `),`utf8`),console.log(`\nSTT model capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate STT model table:`,e.message),process.exit(1)}}async function ia({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${$.name} v${$.version}`),console.log(`Generating MLX model capabilities comparison...
57
+ ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=j.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`ggml-stt-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=y.join(process.cwd(),m);A.writeFileSync(h,n.join(`
58
+ `),`utf8`),console.log(`\nSTT model capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate STT model table:`,e.message),process.exit(1)}}async function Ta({modelIds:e=[],defaultConfig:t=null}={}){let n=[];console.log(`${$.name} v${$.version}`),console.log(`Generating MLX model capabilities comparison...
59
59
  `),n.push(`${$.name} v${$.version}`),n.push(`## MLX Model Capabilities Comparison
60
60
  `),(!e||e.length===0)&&(console.error(`Error: No model IDs provided`),process.exit(1));try{let r=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=r(n[e]||{},t):n[e]=t}),n},{server:i,generators:a=[],...o}=t||{},s=e=>r(structuredClone(o),e||{}),c=e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`mlx-llm`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return s(n)}}return Object.keys(o).length>0?s({}):null},l=[];for(let t=0;t<e.length;t+=1){let n=e[t];console.log(`[${t+1}/${e.length}] Analyzing ${n}...`);let r=c(n);r={...r||{},model:{...r?.model||{},repo_id:n}};let i=await J(`mlx-llm`,null,{config:r,includeBreakdown:!0});l.push({modelId:n,capabilities:i,modelInfo:i.buttress?.selected||null,modelConfig:i.modelConfig||null})}let u=e=>e?(e/1024/1024/1024).toFixed(2):`N/A`,d=e=>e?`✅`:`🚫`;n.push(`| Model ID | Quant | Size (GB) | Context | KV Cache (GB) | Total Required (GB) | Fits Unified Memory |`),n.push(`|----------|-------|-----------|---------|---------------|---------------------|---------------------|`),l.forEach(({modelId:e,modelInfo:t,modelConfig:r})=>{let i=t?.quantization?.bits||r?.quantBits||`N/A`,a=typeof i==`number`?`${i}bit`:i,o=u(t?.modelBytes),s=r?.nCtx||t?.kvInfo?.nCtxTrain||`N/A`,c=u(t?.kvCacheBytes),l=u(t?.fit?.totalRequiredBytes),f=d(t?.fit?.fitsInGpu);n.push(`| ${e} | ${a} | ${o} | ${s} | ${c} | ${l} | ${f} |`);let p=t?.limitedFit!=null&&t?.memoryLimitedCtx!=null,m=!t?.fit?.fitsInGpu;if(p&&m){let e=t.memoryLimitedCtx,r=u(t.limitedKvCacheBytes),i=u(t.limitedFit.totalRequiredBytes),s=d(t.limitedFit.fitsInGpu);n.push(`| ↳ Limited | ${a} | ${o} | ${e} | ${r} | ${i} | ${s} |`)}}),n.push(`
61
61
  ---`),n.push(`
62
- ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${v.hostname()}`),n.push(`- **OS:** ${v.type()} ${v.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **Architecture:** ${v.arch()}`),n.push(`- **CPU Cores:** ${v.cpus().length}`),n.push(`- **Total System Memory:** ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;e&&(n.push(`- **Unified Memory (Usable):** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0)}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.devices?.length>0&&n.push(`- **Metal Device:** ${e.devices[0].deviceName}`),n.push(`- **Performance Score:** ${e.score}`))}if(l.length>0){let e=l[0].capabilities;n.push(`- **Python3:** ${e.python?.available?`Available`:`Not found`}`),n.push(`- **MLX (system):** ${e.mlx?.systemAvailable?`Available`:`Not installed (will use venv)`}`)}n.push(`
62
+ ### System Information`);let f=null;if(process.platform!==`win32`)try{f=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}if(f?n.push(`- **System:** ${f}`):(n.push(`- **Hostname:** ${b.hostname()}`),n.push(`- **OS:** ${b.type()} ${b.release()}`)),n.push(`- **Platform:** ${process.platform}`),n.push(`- **Architecture:** ${b.arch()}`),n.push(`- **CPU Cores:** ${b.cpus().length}`),n.push(`- **Total System Memory:** ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`),l.length>0){let e=l[0].capabilities.buttress?.selected;e&&(n.push(`- **Unified Memory (Usable):** ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0)}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.devices?.length>0&&n.push(`- **Metal Device:** ${e.devices[0].deviceName}`),n.push(`- **Performance Score:** ${e.score}`))}if(l.length>0){let e=l[0].capabilities;n.push(`- **Python3:** ${e.python?.available?`Available`:`Not found`}`),n.push(`- **MLX (system):** ${e.mlx?.systemAvailable?`Available`:`Not installed (will use venv)`}`)}n.push(`
63
63
  ### Command Used`);let p=process.argv.slice(2).join(` `);if(n.push(`\`\`\`bash\n${process.argv[0]} ${process.argv[1]} ${p}\n\`\`\``),n.push(`
64
64
  ### Package Information`),n.push(`- **Name:** ${$.name}`),n.push(`- **Version:** ${$.version}`),$.description&&n.push(`- **Description:** ${$.description}`),t&&Object.keys(t).length>0){n.push(`
65
- ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=A.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`mlx-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=_.join(process.cwd(),m);k.writeFileSync(h,n.join(`
66
- `),`utf8`),console.log(`\nMLX model capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate MLX model table:`,e.message),process.exit(1)}}async function aa({modelId:e=null,defaultConfig:t=null}={}){console.log(`${$.name} v${$.version}`),console.log(`Testing capabilities for backend: mlx-llm`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{server:n,generators:r=[],...i}=t||{},a=Object.keys(i).length>0?{...i}:null;e&&(a={...a||{},model:{...a?.model||{},repo_id:e}});let o=await J(`mlx-llm`,null,{config:a,includeBreakdown:!0});console.log(`
65
+ ### Configuration`),n.push(`<details>`),n.push(`<summary>Click to expand TOML configuration</summary>`),n.push("\n```toml");try{let e=j.stringify(t);n.push(e)}catch{n.push(`# Error serializing config`),n.push(JSON.stringify(t,null,2))}n.push("```"),n.push(`</details>`)}let m=`mlx-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,`-`).split(`T`)[0]}.md`,h=y.join(process.cwd(),m);A.writeFileSync(h,n.join(`
66
+ `),`utf8`),console.log(`\nMLX model capabilities table saved to: ${h}`),process.exit(0)}catch(e){console.error(`Failed to generate MLX model table:`,e.message),process.exit(1)}}async function Ea({modelId:e=null,defaultConfig:t=null}={}){console.log(`${$.name} v${$.version}`),console.log(`Testing capabilities for backend: mlx-llm`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{server:n,generators:r=[],...i}=t||{},a=Object.keys(i).length>0?{...i}:null;e&&(a={...a||{},model:{...a?.model||{},repo_id:e}});let o=await J(`mlx-llm`,null,{config:a,includeBreakdown:!0});console.log(`
67
67
  === Platform Information ===`),console.log(`Available: ${o.available?`Yes`:`No`}`),console.log(`OS: ${o.platform.os} (${o.platform.arch})`),o.platform.errors?.length>0&&console.log(`Errors: ${o.platform.errors.join(`; `)}`),console.log(`Python3: ${o.python.available?`Available`:`Not found`}`),console.log(`MLX (system): ${o.mlx.systemAvailable?`Available`:`Not installed (will use venv)`}`);let s=o.buttress?.selected,c=o.modelConfig;if(e||c?.repoId){console.log(`
68
68
  === Model Information ===`);let t=e||c?.repoId;if(console.log(`Repository ID: ${t}`),c?.architecture&&console.log(`Architecture: ${c.architecture}`),c?.quantBits&&console.log(`Quantization: ${c.quantBits}bit (group_size=${c.quantGroupSize||`N/A`})`),c?.nCtx&&console.log(`Max Context Length: ${c.nCtx}`),s?.modelBytes&&console.log(`Model Size: ${(s.modelBytes/1024/1024/1024).toFixed(2)} GB`),s?.kvCacheBytes!=null&&console.log(`KV Cache Size (full context): ${(s.kvCacheBytes/1024/1024/1024).toFixed(2)} GB`),s?.kvInfo&&console.log(`KV Info: ${s.kvInfo.nLayer} layers, ${s.kvInfo.nHeadKv} KV heads, ${s.kvInfo.headDim} head dim`),s?.modelBytes&&s?.kvCacheBytes!=null){let e=s.modelBytes+s.kvCacheBytes;console.log(`Total Required Memory: ${(e/1024/1024/1024).toFixed(2)} GB`)}s?.memoryLimitedCtx!=null&&(console.log(`\nMemory-Limited Context: ${s.memoryLimitedCtx}`),s.limitedKvCacheBytes!=null&&console.log(`Limited KV Cache Size: ${(s.limitedKvCacheBytes/1024/1024/1024).toFixed(2)} GB`))}if(o.buttress?.selected){let{selected:e}=o.buttress;console.log(`
69
- === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${v.hostname()}`),console.log(`OS: ${v.type()} ${v.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${v.cpus().length}`),console.log(`Total System Memory: ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`),console.log(`
69
+ === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${b.hostname()}`),console.log(`OS: ${b.type()} ${b.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${b.cpus().length}`),console.log(`Total System Memory: ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`),console.log(`
70
70
  --- Unified Memory (Metal) ---`),e.devices?.length>0&&console.log(`Device: ${e.devices[0].deviceName}`),console.log(`Total Memory: ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB`);let n=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;console.log(`Usable Memory: ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}%)`),console.log(`Performance Score: ${e.score}`),e.fit&&(console.log(`
71
71
  --- Model Fit Analysis ---`),console.log(`Fits in Unified Memory: ${e.fit.fitsInGpu?`Yes`:`No`}`),console.log(`Limiting Factor: ${e.fit.limiting}`),console.log(`Total Required: ${(e.fit.totalRequiredBytes/1024/1024/1024).toFixed(2)} GB`),e.limitedFit&&(console.log(`
72
72
  --- Memory-Limited Fit Analysis ---`),console.log(`Limited Total Required: ${(e.limitedFit.totalRequiredBytes/1024/1024/1024).toFixed(2)} GB`),console.log(`Fits (Limited): ${e.limitedFit.fitsInGpu?`Yes`:`No`}`),console.log(`Limiting Factor (Limited): ${e.limitedFit.limiting}`)))}console.log(`
73
- === Full Capabilities JSON ===`),console.log(JSON.stringify(o,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}async function oa({modelId:e=null,defaultConfig:t=null}={}){console.log(`${$.name} v${$.version}`),console.log(`Testing capabilities for backend: ggml-stt`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{repoId:n,filename:r}=ea(e),{server:i,generators:a=[],...o}=t||{},s=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=s(n[e]||{},t):n[e]=t}),n},c=e=>s(structuredClone(o),e||{}),l=(e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return c(n)}}return Object.keys(o).length>0?c({}):null})(n);n&&(l={...l||{},model:{...o.runtime,...l?.model||{},repo_id:n,...r&&{filename:r}}});let u=await J(`ggml-stt`,null,{config:l,includeBreakdown:!0}),d=u.buttress?.selected||null,f=u.modelConfig||null;if(n||f?.repoId){console.log(`
73
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(o,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}async function Da({modelId:e=null,defaultConfig:t=null}={}){console.log(`${$.name} v${$.version}`),console.log(`Testing capabilities for backend: ggml-stt`),e&&console.log(`Model: ${e}`),console.log(`--------------------------------`);try{let{repoId:n,filename:r}=xa(e),{server:i,generators:a=[],...o}=t||{},s=(e={},t={})=>{let n=Array.isArray(e)?[...e]:{...e};return Object.entries(t||{}).forEach(([e,t])=>{t&&typeof t==`object`&&!Array.isArray(t)?n[e]=s(n[e]||{},t):n[e]=t}),n},c=e=>s(structuredClone(o),e||{}),l=(e=>{if(Array.isArray(a)&&a.length>0){let t=a.filter(e=>e?.type===`ggml-stt`);if(t.length>0&&e){let n=t.find(t=>t.model?.repo_id===e);if(n)return c(n)}}return Object.keys(o).length>0?c({}):null})(n);n&&(l={...l||{},model:{...o.runtime,...l?.model||{},repo_id:n,...r&&{filename:r}}});let u=await J(`ggml-stt`,null,{config:l,includeBreakdown:!0}),d=u.buttress?.selected||null,f=u.modelConfig||null;if(n||f?.repoId){console.log(`
74
74
  === Model Information ===`);let e=n||f?.repoId;console.log(`Repository ID: ${e}`),r&&console.log(`Filename: ${r}`),d?.modelBytes&&console.log(`Model Size: ${(d.modelBytes/1024/1024).toFixed(1)} MB`);let t=d?.processingBytes||d?.kvCacheBytes;if(t&&console.log(`Processing Buffer: ${(t/1024/1024).toFixed(1)} MB`),d?.modelBytes&&t)console.log(`Total Required Memory: ${((d.modelBytes+t)/1024/1024).toFixed(1)} MB`);else if(u.buttress?.selected?.fit){let{totalRequiredBytes:e}=u.buttress.selected.fit;console.log(`Total Required Memory: ${(e/1024/1024).toFixed(1)} MB`)}}if(u.buttress?.selected){let{selected:e}=u.buttress;console.log(`
75
- === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${v.hostname()}`),console.log(`OS: ${v.type()} ${v.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${v.cpus().length}`),console.log(`Total System Memory: ${(v.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
75
+ === Hardware Information ===`);let t=null;if(process.platform!==`win32`)try{t=O(`uname -a`,{encoding:`utf8`}).trim()}catch{}t?console.log(`System: ${t}`):(console.log(`Hostname: ${b.hostname()}`),console.log(`OS: ${b.type()} ${b.release()}`)),console.log(`Platform: ${e.platform}`),console.log(`CPU Cores: ${b.cpus().length}`),console.log(`Total System Memory: ${(b.totalmem()/1024/1024/1024).toFixed(2)} GB`);let n=e.cpuTotalBytes>0?(e.cpuUsableBytes/e.cpuTotalBytes*100).toFixed(0):0;console.log(`Usable CPU Memory: ${(e.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),e.hasGpu?(console.log(`
76
76
  --- GPU Details ---`),e.devices.filter(e=>e.type===`gpu`).forEach(t=>{console.log(`GPU Backend: ${t.backend}`),console.log(`GPU Name: ${t.deviceName}`),console.log(`GPU Total Memory: ${(t.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let n=e.gpuTotalBytes>0?(e.gpuUsableBytes/e.gpuTotalBytes*100).toFixed(0):0;console.log(`GPU Usable Memory: ${(e.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${n}% of ${(e.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),t.metadata&&(t.metadata.hasBFloat16&&console.log(`Supports BFloat16: Yes`),t.metadata.hasUnifiedMemory&&console.log(`Unified Memory: Yes`))})):console.log(`GPU: Not available`),console.log(`\nBackend Variant: ${e.variant}`),console.log(`Performance Score: ${e.score}`),e.fit&&(console.log(`
77
77
  --- Model Fit Analysis ---`),console.log(`Fits in GPU: ${e.fit.fitsInGpu?`Yes`:`No`}`),console.log(`Fits in CPU: ${e.fit.fitsInCpu?`Yes`:`No`}`),console.log(`Limiting Factor: ${e.fit.limiting}`))}console.log(`
78
- === Full Capabilities JSON ===`),console.log(JSON.stringify(u,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}const sa=Zr();var ca=async()=>{(process.argv.includes(`--version`)||process.argv.includes(`-v`))&&(console.log(sa.version),process.exit(0)),(process.argv.includes(`--help`)||process.argv.includes(`-h`))&&(console.log(`
79
- bricks-buttress v${sa.version}
78
+ === Full Capabilities JSON ===`),console.log(JSON.stringify(u,null,2)),process.exit(0)}catch(e){console.error(`Failed to get capabilities:`,e.message),process.exit(1)}}const Oa=ei();var ka=async()=>{(process.argv.includes(`--version`)||process.argv.includes(`-v`))&&(console.log(Oa.version),process.exit(0)),(process.argv.includes(`--help`)||process.argv.includes(`-h`))&&(console.log(`
79
+ bricks-buttress v${Oa.version}
80
80
 
81
81
  Buttress server for remote inference with GGML backends.
82
82
 
@@ -99,9 +99,7 @@ Testing Options:
99
99
  file (e.g., ggml-llm-model-capabilities-YYYY-MM-DD.md)
100
100
 
101
101
  Environment Variables:
102
- NODE_ENV Set to 'development' for dev mode
103
- ENABLE_OPENAI_COMPAT_ENDPOINT Set to '1' to enable OpenAI-compatible API
104
- ENABLE_ANTHROPIC_MESSAGES_ENDPOINT Set to '1' to enable Anthropic Messages API
102
+ NODE_ENV Set to 'development' for dev mode
105
103
 
106
104
  Examples:
107
105
  bricks-buttress
@@ -111,4 +109,4 @@ Examples:
111
109
  bricks-buttress --test-caps ggml-stt --test-caps-model-id BricksDisplay/whisper-ggml:ggml-small.bin
112
110
  bricks-buttress --test-caps mlx-llm --test-models-default
113
111
  `),process.exit(0));let e=process.argv.findIndex(e=>e===`--port`||e===`-p`),t=e>=0?Number(process.argv[e+1]):void 0,n=process.argv.findIndex(e=>e===`--config`||e===`-c`),r=n>=0?process.argv[n+1]:null,i=null;if(r){let e;if(r.includes(`
114
- `))e=r;else{let t=_.resolve(r);try{e=k.readFileSync(t,`utf8`)}catch(e){console.error(`Failed to read Buttress config at ${t}:`,e),process.exit(1)}}try{let t=A.parse(e);t.env&&typeof t.env==`object`&&(Object.entries(t.env).forEach(([e,t])=>{process.env[e]===void 0&&(process.env[e]=String(t))}),delete t.env),i=t}catch(e){console.error(`Failed to parse TOML config:`,e),process.exit(1)}}async function a(e){if(!e?.generators||!Array.isArray(e.generators))return;let t=e.generators.filter(e=>{if(!e.model?.download)return!1;let{type:t}=e;return!t||t!==`ggml-llm`&&t!==`ggml-stt`&&t!==`mlx-llm`?(console.warn(`[Download] Skipping unknown generator type: ${t}`),!1):!0});if(t.length===0)return;let{server:n,generators:r,...i}=e,a=t.map(e=>{let{type:t}=e,n=e.model?.repo_id;return console.log(`[Download] Starting pre-download for ${t}: ${n}`),qr(t,{...i,backend:e.backend||{},model:e.model||{},runtime:{...i.runtime,...e.runtime||{}}},{onProgress:()=>{},onComplete:({repoId:e,alreadyExists:t})=>{t?console.log(`[Download] Pre-download complete (already exists): ${e}`):console.log(`[Download] Pre-download complete: ${e}`)},onError:e=>{console.error(`[Download] Pre-download failed for ${n}:`,e.message)}})}),o=await Promise.all(a),s=o.filter(e=>e.started).length,c=o.filter(e=>e.alreadyExists).length,l=o.filter(e=>e.alreadyDownloading).length;console.log(`[Download] Pre-download summary: ${s} started, ${c} already exist, ${l} already downloading`)}let o=[`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`,`unsloth/Qwen3.5-27B-GGUF`,`unsloth/gemma-4-26B-A4B-it-GGUF`,`unsloth/gemma-4-31B-it-GGUF`,`unsloth/GLM-4.7-Flash-GGUF`,`DevQuasar/MiniMaxAI.MiniMax-M2.5-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`],s=[`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`],c=[`mlx-community/Qwen3.5-27B-8bit`,`mlx-community/Qwen3.5-27B-4bit`,`mlx-community/Qwen3.5-35B-A3B-8bit`,`mlx-community/Qwen3.5-35B-A3B-4bit`,`mlx-community/Qwen3-235B-A22B-8bit`,`mlx-community/Qwen3-235B-A22B-4bit`,`mlx-community/GLM-4.7-Flash-8bit`,`mlx-community/GLM-4.7-Flash-4bit`,`mlx-community/MiniMax-M2.5-4bit`,`mlx-community/gpt-oss-120b-4bit`,`mlx-community/gemma-4-26b-a4b-it-8bit`,`mlx-community/gemma-4-26b-a4b-it-4bit`,`mlx-community/gemma-4-31b-it-8bit`,`mlx-community/gemma-4-31b-it-4bit`],l=process.argv.findIndex(e=>e===`--test-caps`);if(l>=0){let e=process.argv[l+1]||`ggml-llm`;e!==`ggml-llm`&&e!==`ggml-stt`&&e!==`mlx-llm`&&(console.error(`Only ggml-llm, ggml-stt, and mlx-llm backends are supported for testing capabilities`),process.exit(1));let t=process.argv.findIndex(e=>e===`--test-models`),n=process.argv.includes(`--test-models-default`);if(e===`mlx-llm`)if(t>=0){let e=process.argv[t+1];e||(console.error(`Error: --test-models requires a comma-separated list of model IDs`),process.exit(1)),await ia({modelIds:e.split(`,`).map(e=>e.trim()),defaultConfig:i})}else if(n)await ia({modelIds:c,defaultConfig:i});else{let e=process.argv.findIndex(e=>e===`--test-caps-model-id`);await aa({modelId:e>=0?process.argv[e+1]:null,defaultConfig:i})}else if(e===`ggml-stt`)if(t>=0){let e=process.argv[t+1];e||(console.error(`Error: --test-models requires a comma-separated list of model IDs`),process.exit(1)),await ra({modelIds:e.split(`,`).map(e=>e.trim()),defaultConfig:i})}else if(n)await ra({modelIds:s,defaultConfig:i});else{let e=process.argv.findIndex(e=>e===`--test-caps-model-id`);await oa({modelId:e>=0?process.argv[e+1]:null,defaultConfig:i})}else if(t>=0){let e=process.argv[t+1];e||(console.error(`Error: --test-models requires a comma-separated list of model IDs`),process.exit(1)),await ta({modelIds:e.split(`,`).map(e=>e.trim()),defaultConfig:i})}else if(n)await ta({modelIds:o,defaultConfig:i});else{let e=process.argv.findIndex(e=>e===`--test-caps-model-id`);await na({modelId:e>=0?process.argv[e+1]:null,defaultConfig:i})}}let u=Ii(i);t&&(u.server.port=t),u.server.port||(u.server.port=2080),_a({config:u,enableOpenAICompat:process.env.ENABLE_OPENAI_COMPAT_ENDPOINT===`1`,enableAnthropicMessages:process.env.ENABLE_ANTHROPIC_MESSAGES_ENDPOINT===`1`}).then(async({port:e,openaiEnabled:t,anthropicMessagesEnabled:n,autoDiscover:r})=>{let o=$i();console.log(`Buttress server listening on port ${e}`),console.log(`--------------------------------`),await ma(),console.log();let s={"ggml-llm":`LLM (GGML)`,"ggml-stt":`STT (GGML)`,"mlx-llm":`LLM (MLX)`};console.log(`Current supported Generators:`);let c=new Set((u?.generators||[]).map(e=>e.type).filter(Boolean));if(c.size===0)console.log(`- LLM (GGML)`),console.log(`- STT (GGML)`);else for(let e of c)console.log(`- ${s[e]||e}`);console.log(),console.log("Please configure `Buttress (Remote Inference)` in the Generator to connect to this server."),console.log(),console.log(`- Use http://${o}:${e} to connect to this server via LAN.`),console.log(`- Visit http://${o}:${e}/status to see status via LAN.`),console.log(),t?(console.log(`OpenAI-compatible API [EXPERIMENTAL]:`),console.log(`- Base URL: http://${o}:${e}/oai-compat/v1`),console.log(`- Chat completions: POST http://${o}:${e}/oai-compat/v1/chat/completions`),console.log(`- Models: GET http://${o}:${e}/oai-compat/v1/models`),console.log()):(console.log(`OpenAI-compatible API [EXPERIMENTAL]: disabled`),console.log(` Set ENABLE_OPENAI_COMPAT_ENDPOINT=1 to enable`),console.log()),n?(console.log(`Anthropic Messages API [EXPERIMENTAL]:`),console.log(`- Base URL: http://${o}:${e}/anthropic-messages`),console.log(`- Messages: POST http://${o}:${e}/anthropic-messages/v1/messages`),console.log(`- Count tokens: POST http://${o}:${e}/anthropic-messages/v1/messages/count_tokens`),console.log()):(console.log(`Anthropic Messages API [EXPERIMENTAL]: disabled`),console.log(` Set ENABLE_ANTHROPIC_MESSAGES_ENDPOINT=1 to enable`),console.log()),r&&(console.log(`Auto-discover enabled`),console.log()),i&&await a(i)}).catch(e=>{console.error(`Failed to start Buttress server:`,e),process.exitCode=1})};const{version:la,name:ua}=Zr(),da=async()=>{let e=`https://registry.npmjs.org/${ua}/latest`;try{let t=new AbortController,n=setTimeout(()=>t.abort(),3e3),r=await fetch(e,{headers:{Accept:`application/json`},signal:t.signal});return clearTimeout(n),r.ok&&(await r.json()).version||null}catch{return null}},fa=(e,t)=>{if(!t)return!1;let n=e.split(/[.-]/),r=t.split(/[.-]/);for(let e=0;e<Math.max(n.length,r.length);e+=1){let t=parseInt(n[e])||0,i=parseInt(r[e])||0;if(i>t)return!0;if(i<t)return!1}return!1},pa=e=>{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`,la.padEnd(12),e.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`,ua.padEnd(27)),console.log(`\x1B[33m╰─────────────────────────────────────────────────╯\x1B[0m`),console.log(``)},ma=async()=>{try{let e=await da();e&&fa(la,e)&&pa(e)}catch{}},ha=typeof process<`u`&&process.versions&&process.versions.node,ga=async({backend:e,router:r,config:i,enableOpenAICompat:o,enableAnthropicMessages:l})=>{try{await c.mkdir(i.server.temp_file_dir,{recursive:!0})}catch{}let u=$i()||`0.0.0.0`,d={id:i.server.id,name:i.server.name,version:la,address:u,port:i.server.port,url:`http://${u}:${i.server.port}`,generators:Mi(i,i.generators.map(e=>e.type)),authentication:{required:!0,type:`device-group`}},f=new n({serve:{maxRequestBodySize:i.server.max_body_size},websocket:{idleTimeout:Math.ceil(i.server.session_timeout/1e3)},adapter:ha?t():void 0}).state({sessions:new Map,backend:e||Lr,config:i,serverInfo:d});r&&f.use(r),i.autodiscover?.http?.enabled&&f.use(ni(i)),f.use(ii),f.use(li),o&&f.use(vi(i)),l&&f.use(wi(i));let p={INVALID_REQUEST:-32600,INVALID_PARAMS:-32602,METHOD_NOT_FOUND:-32601,INTERNAL_ERROR:-32603},m=e=>e.id??e.raw?.id??e.remoteAddress;return f.ws(`/buttress/rpc`,{parse:(e,t)=>{if(typeof t==`string`)try{return JSON.parse(t)}catch{return e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:p.INVALID_REQUEST,message:`Invalid request`},id:null})),null}return t},body:a.Object({jsonrpc:a.String(),method:a.String(),params:a.String(),id:a.String()}),open(e){let t=m(e);if(console.log(`[Request] New connection: ${t}`),!e.data.store.sessions.has(t))e.data.store.sessions.set(t,{streamReaders:new Map,generators:new Set,initializedContexts:new Set,timeout:null});else{let n=e.data.store.sessions.get(t);n&&(clearTimeout(n.timeout),n.timeout=null)}},async message(e,{id:t,method:n,params:r}){let i=m(e);console.log(`[Request] Received request from ${i}: ${n}`);let a=e.data.store.sessions.get(i),[o,c]=n.split(`.`),l=Ji[o]?.[c];if(!l){e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:p.METHOD_NOT_FOUND,message:`Method not found`},id:t}));return}try{if(n===`cancel`){a.streamReaders.has(t)&&(a.streamReaders.get(t)?.cancel(),a.streamReaders.delete(t));return}if(n===`ping`){e.send(JSON.stringify({jsonrpc:`2.0`,result:`pong`,id:t}));return}let u=Yi(r),d=qi[o]?.[c],f=d?d.parse(u):u,p=await l({...e.data.store,peerId:i,session:a},...f);if(p instanceof s){let r=p.getReader();a.streamReaders.set(t,r),e.send(JSON.stringify({jsonrpc:`2.0`,result:{type:`stream`},id:t}));try{for(;;){let{value:n,done:i}=await r.read();if(i)break;let{event:a,data:o}=n;e.send(JSON.stringify({jsonrpc:`2.0`,method:`notification/${a}`,params:Xi(o),id:t}))}e.send(JSON.stringify({jsonrpc:`2.0`,method:`notification/_end`,id:t}))}catch(r){console.error(`[RPC] Stream error for ${n}:`,r),e.send(JSON.stringify({jsonrpc:`2.0`,method:`notification/_error`,params:Xi(r),id:t}))}a.streamReaders.delete(t)}else e.send(JSON.stringify({jsonrpc:`2.0`,result:Xi(p),id:t}))}catch(r){if(r instanceof ne){e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:p.INVALID_PARAMS,message:`Invalid params`,data:r.issues},id:t}));return}console.error(`[RPC] Handler error for ${n}:`,r),e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:p.INTERNAL_ERROR,message:String(r)},id:t}))}},async close(e){let t=m(e);console.log(`[Request] Connection closed: ${t}`);let{backend:n,sessions:r}=e.data.store,a=r.get(t);a&&(a.streamReaders.forEach(e=>e.cancel()),a.streamReaders.clear(),a.timeout=setTimeout(()=>{r.delete(t),console.log(`[Request] Session timed out: ${t}`),a.generators.forEach(e=>{n.finalizeGenerator(e)})},i.server.session_timeout))}}),{app:f,config:i}},_a=async({backend:e,router:t,config:n,enableOpenAICompat:r=!1,enableAnthropicMessages:i=!1})=>{let{app:a,config:o}=await ga({backend:e,router:t,config:n,enableOpenAICompat:r,enableAnthropicMessages:i}),{server:{port:s}}=o,c=[new Promise(e=>a.listen(s,e))],l=null;return o.autodiscover&&(l=new Qi(o.autodiscover,()=>a.store.serverInfo),c.push(l.start())),await Promise.all(c),{app:a,port:s,openaiEnabled:r,anthropicMessagesEnabled:i,autoDiscover:l}},va=[new URL(`index.mjs`,import.meta.url).pathname,new URL(`index.ts`,import.meta.url).pathname];(process.argv[1]?.endsWith(`/bricks-buttress`)||va.includes(process.argv[1]))&&await ca();export{ma as checkAndNotifyUpdates,da as checkForUpdates,fa as compareVersions,ga as createServer,pa as logUpdateMessage,Ii as processConfig,qr as startModelDownload,_a as startServer};
112
+ `))e=r;else{let t=y.resolve(r);try{e=A.readFileSync(t,`utf8`)}catch(e){console.error(`Failed to read Buttress config at ${t}:`,e),process.exit(1)}}try{let t=j.parse(e);t.env&&typeof t.env==`object`&&(Object.entries(t.env).forEach(([e,t])=>{process.env[e]===void 0&&(process.env[e]=String(t))}),delete t.env),i=t}catch(e){console.error(`Failed to parse TOML config:`,e),process.exit(1)}}async function a(e){if(!e?.generators||!Array.isArray(e.generators))return;let t=e.generators.filter(e=>{if(!e.model?.download)return!1;let{type:t}=e;return!t||t!==`ggml-llm`&&t!==`ggml-stt`&&t!==`mlx-llm`?(console.warn(`[Download] Skipping unknown generator type: ${t}`),!1):!0});if(t.length===0)return;let{server:n,generators:r,...i}=e,a=t.map(e=>{let{type:t}=e,n=e.model?.repo_id;return console.log(`[Download] Starting pre-download for ${t}: ${n}`),Xr(t,{...i,backend:e.backend||{},model:e.model||{},runtime:{...i.runtime,...e.runtime||{}}},{onProgress:()=>{},onComplete:({repoId:e,alreadyExists:t})=>{t?console.log(`[Download] Pre-download complete (already exists): ${e}`):console.log(`[Download] Pre-download complete: ${e}`)},onError:e=>{console.error(`[Download] Pre-download failed for ${n}:`,e.message)}})}),o=await Promise.all(a),s=o.filter(e=>e.started).length,c=o.filter(e=>e.alreadyExists).length,l=o.filter(e=>e.alreadyDownloading).length;console.log(`[Download] Pre-download summary: ${s} started, ${c} already exist, ${l} already downloading`)}let o=[`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`,`unsloth/Qwen3.5-27B-GGUF`,`unsloth/gemma-4-26B-A4B-it-GGUF`,`unsloth/gemma-4-31B-it-GGUF`,`unsloth/GLM-4.7-Flash-GGUF`,`DevQuasar/MiniMaxAI.MiniMax-M2.5-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`],s=[`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`],c=[`mlx-community/Qwen3.5-27B-8bit`,`mlx-community/Qwen3.5-27B-4bit`,`mlx-community/Qwen3.5-35B-A3B-8bit`,`mlx-community/Qwen3.5-35B-A3B-4bit`,`mlx-community/Qwen3-235B-A22B-8bit`,`mlx-community/Qwen3-235B-A22B-4bit`,`mlx-community/GLM-4.7-Flash-8bit`,`mlx-community/GLM-4.7-Flash-4bit`,`mlx-community/MiniMax-M2.5-4bit`,`mlx-community/gpt-oss-120b-4bit`,`mlx-community/gemma-4-26b-a4b-it-8bit`,`mlx-community/gemma-4-26b-a4b-it-4bit`,`mlx-community/gemma-4-31b-it-8bit`,`mlx-community/gemma-4-31b-it-4bit`],l=process.argv.findIndex(e=>e===`--test-caps`);if(l>=0){let e=process.argv[l+1]||`ggml-llm`;e!==`ggml-llm`&&e!==`ggml-stt`&&e!==`mlx-llm`&&(console.error(`Only ggml-llm, ggml-stt, and mlx-llm backends are supported for testing capabilities`),process.exit(1));let t=process.argv.findIndex(e=>e===`--test-models`),n=process.argv.includes(`--test-models-default`);if(e===`mlx-llm`)if(t>=0){let e=process.argv[t+1];e||(console.error(`Error: --test-models requires a comma-separated list of model IDs`),process.exit(1)),await Ta({modelIds:e.split(`,`).map(e=>e.trim()),defaultConfig:i})}else if(n)await Ta({modelIds:c,defaultConfig:i});else{let e=process.argv.findIndex(e=>e===`--test-caps-model-id`);await Ea({modelId:e>=0?process.argv[e+1]:null,defaultConfig:i})}else if(e===`ggml-stt`)if(t>=0){let e=process.argv[t+1];e||(console.error(`Error: --test-models requires a comma-separated list of model IDs`),process.exit(1)),await wa({modelIds:e.split(`,`).map(e=>e.trim()),defaultConfig:i})}else if(n)await wa({modelIds:s,defaultConfig:i});else{let e=process.argv.findIndex(e=>e===`--test-caps-model-id`);await Da({modelId:e>=0?process.argv[e+1]:null,defaultConfig:i})}else if(t>=0){let e=process.argv[t+1];e||(console.error(`Error: --test-models requires a comma-separated list of model IDs`),process.exit(1)),await Sa({modelIds:e.split(`,`).map(e=>e.trim()),defaultConfig:i})}else if(n)await Sa({modelIds:o,defaultConfig:i});else{let e=process.argv.findIndex(e=>e===`--test-caps-model-id`);await Ca({modelId:e>=0?process.argv[e+1]:null,defaultConfig:i})}}let u=Gi(i);t&&(u.server.port=t),u.server.port||(u.server.port=2080),Ra({config:u,enableOpenAICompat:process.env.ENABLE_OPENAI_COMPAT_ENDPOINT===`1`||u.global.openai_compat?.enabled===!0,enableAnthropicMessages:process.env.ENABLE_ANTHROPIC_MESSAGES_ENDPOINT===`1`||u.global.anthropic_messages?.enabled===!0}).then(async({app:e,port:t,openaiEnabled:n,anthropicMessagesEnabled:r,autoDiscover:o})=>{let s=la();console.log(`Buttress server listening on port ${t}`),console.log(`--------------------------------`),await Fa(),console.log();let c=e.store.workspaceState.workspace;if(c){let e=c.name?`${c.name} (${c.id})`:c.id;console.log(`Workspace: ${e}`),console.log(`- Server ID: ${c.serverId}`),console.log(`- Issuer kid: ${c.kid}`),console.log(`- Bound at: ${c.boundAt}`)}else console.log(`Workspace: not bound`),console.log(`- State file: ${_a()}`),console.log("- Run `bricks buttress bind` from a workspace-authed CLI to pair.");console.log();let l={"ggml-llm":`LLM (GGML)`,"ggml-stt":`STT (GGML)`,"mlx-llm":`LLM (MLX)`};console.log(`Current supported Generators:`);let d=new Set((u?.generators||[]).map(e=>e.type).filter(Boolean));if(d.size===0)console.log(`- LLM (GGML)`),console.log(`- STT (GGML)`);else for(let e of d)console.log(`- ${l[e]||e}`);console.log(),console.log("Please configure `Buttress (Remote Inference)` in the Generator to connect to this server."),console.log(),console.log(`- Use http://${s}:${t} to connect to this server via LAN.`),console.log(`- Visit http://${s}:${t}/status to see status via LAN.`),console.log(),n?(console.log(`OpenAI-compatible API [EXPERIMENTAL]:`),console.log(`- Base URL: http://${s}:${t}/oai-compat/v1`),console.log(`- Chat completions: POST http://${s}:${t}/oai-compat/v1/chat/completions`),console.log(`- Models: GET http://${s}:${t}/oai-compat/v1/models`),console.log()):(console.log(`OpenAI-compatible API [EXPERIMENTAL]: disabled`),console.log(` Set [openai_compat] enabled = true in config to enable`),console.log()),r?(console.log(`Anthropic Messages API [EXPERIMENTAL]:`),console.log(`- Base URL: http://${s}:${t}/anthropic-messages`),console.log(`- Messages: POST http://${s}:${t}/anthropic-messages/v1/messages`),console.log(`- Count tokens: POST http://${s}:${t}/anthropic-messages/v1/messages/count_tokens`),console.log()):(console.log(`Anthropic Messages API [EXPERIMENTAL]: disabled`),console.log(` Set [anthropic_messages] enabled = true in config to enable`),console.log()),o&&(console.log(`Auto-discover enabled`),console.log()),i&&await a(i)}).catch(e=>{console.error(`Failed to start Buttress server:`,e),process.exitCode=1})};const{version:Aa,name:ja}=ei(),Ma=async()=>{let e=`https://registry.npmjs.org/${ja}/latest`;try{let t=new AbortController,n=setTimeout(()=>t.abort(),3e3),r=await fetch(e,{headers:{Accept:`application/json`},signal:t.signal});return clearTimeout(n),r.ok&&(await r.json()).version||null}catch{return null}},Na=(e,t)=>{if(!t)return!1;let n=e.split(/[.-]/),r=t.split(/[.-]/);for(let e=0;e<Math.max(n.length,r.length);e+=1){let t=parseInt(n[e])||0,i=parseInt(r[e])||0;if(i>t)return!0;if(i<t)return!1}return!1},Pa=e=>{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`,Aa.padEnd(12),e.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`,ja.padEnd(27)),console.log(`\x1B[33m╰─────────────────────────────────────────────────╯\x1B[0m`),console.log(``)},Fa=async()=>{try{let e=await Ma();e&&Na(Aa,e)&&Pa(e)}catch{}},Ia=typeof process<`u`&&process.versions&&process.versions.node,La=async({backend:e,router:r,config:i,enableOpenAICompat:o,enableAnthropicMessages:s})=>{try{await u.mkdir(i.server.temp_file_dir,{recursive:!0})}catch{}let c=la()||`0.0.0.0`,d=ba(),f=d.workspace!=null,p=await ca(Vi(i,i.generators.map(e=>e.type))),m={id:i.server.id,name:i.server.name,version:Aa,address:c,port:i.server.port,url:`http://${c}:${i.server.port}`,generators:p,authentication:f?{required:!0,type:`workspace-jwt`,kid:d.workspace.kid,bound:!0}:{required:!1,type:`workspace-jwt`,bound:!1},...f?{workspace:{id:d.workspace.id,name:d.workspace.name}}:{}},h=new n({serve:{maxRequestBodySize:i.server.max_body_size},websocket:{idleTimeout:Math.ceil(i.server.session_timeout/1e3)},adapter:Ia?t():void 0}).state({sessions:new Map,backend:e||Br,config:i,serverInfo:m,workspaceState:d});r&&h.use(r),i.autodiscover?.http?.enabled&&h.use(ai(i)),h.use(pi),h.use(vi),o&&h.use(Di(i)),s&&h.use(Ni(i));let g=(e,t)=>ui(li(e,t),d.workspace),_={INVALID_REQUEST:-32600,INVALID_PARAMS:-32602,METHOD_NOT_FOUND:-32601,INTERNAL_ERROR:-32603},v=e=>e.id??e.raw?.id??e.remoteAddress;return h.ws(`/buttress/rpc`,{parse:(e,t)=>{if(typeof t==`string`)try{return JSON.parse(t)}catch{return e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.INVALID_REQUEST,message:`Invalid request`},id:null})),null}return t},body:a.Object({jsonrpc:a.String(),method:a.String(),params:a.String(),id:a.String()}),async open(e){let t=v(e),n=e.data?.headers,r=e.data?.query,i,a=new Promise(e=>{i=e}),o=e.data.store.sessions.get(t);if(o?(clearTimeout(o.timeout),o.timeout=null,o.identity=null,o.ready=a):e.data.store.sessions.set(t,{streamReaders:new Map,generators:new Set,initializedContexts:new Set,timeout:null,identity:null,ready:a}),!f){console.log(`[Request] New connection: ${t} (unbound, no auth)`),i(!0);return}let s=await g(n||{},r);if(!s){console.warn(`[Auth] Rejecting WS ${t}: invalid or missing workspace-access token`),e.data.store.sessions.delete(t),i(!1),e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.INVALID_REQUEST,message:`Unauthorized`},id:null})),e.close(1008,`UNAUTHORIZED`);return}console.log(`[Request] New connection: ${t} (subject=${s.subjectType}:${s.subjectId})`);let c=e.data.store.sessions.get(t);c&&(c.identity=s),i(!0)},async message(e,{id:t,method:n,params:r}){let i=v(e);console.log(`[Request] Received request from ${i}: ${n}`);let a=e.data.store.sessions.get(i);if(a?.ready){if(!await a.ready){e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.INVALID_REQUEST,message:`Unauthorized`},id:t}));return}a=e.data.store.sessions.get(i)}if(!a){e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.INTERNAL_ERROR,message:`Session not found`},id:t}));return}if(f&&!a.identity){e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.INVALID_REQUEST,message:`Unauthorized`},id:t}));return}let[o,s]=n.split(`.`),c=ra[o]?.[s];if(!c){e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.METHOD_NOT_FOUND,message:`Method not found`},id:t}));return}try{if(n===`cancel`){a.streamReaders.has(t)&&(a.streamReaders.get(t)?.cancel(),a.streamReaders.delete(t));return}if(n===`ping`){e.send(JSON.stringify({jsonrpc:`2.0`,result:`pong`,id:t}));return}let u=ia(r),d=na[o]?.[s],f=d?d.parse(u):u,p=await c({...e.data.store,peerId:i,session:a},...f);if(p instanceof l){let r=p.getReader();a.streamReaders.set(t,r),e.send(JSON.stringify({jsonrpc:`2.0`,result:{type:`stream`},id:t}));try{for(;;){let{value:n,done:i}=await r.read();if(i)break;let{event:a,data:o}=n;e.send(JSON.stringify({jsonrpc:`2.0`,method:`notification/${a}`,params:aa(o),id:t}))}e.send(JSON.stringify({jsonrpc:`2.0`,method:`notification/_end`,id:t}))}catch(r){console.error(`[RPC] Stream error for ${n}:`,r),e.send(JSON.stringify({jsonrpc:`2.0`,method:`notification/_error`,params:aa(r),id:t}))}a.streamReaders.delete(t)}else e.send(JSON.stringify({jsonrpc:`2.0`,result:aa(p),id:t}))}catch(r){if(r instanceof M){e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.INVALID_PARAMS,message:`Invalid params`,data:r.issues},id:t}));return}console.error(`[RPC] Handler error for ${n}:`,r),e.send(JSON.stringify({jsonrpc:`2.0`,error:{code:_.INTERNAL_ERROR,message:String(r)},id:t}))}},async close(e){let t=v(e);console.log(`[Request] Connection closed: ${t}`);let{backend:n,sessions:r}=e.data.store,a=r.get(t);a&&(a.streamReaders.forEach(e=>e.cancel()),a.streamReaders.clear(),a.timeout=setTimeout(()=>{r.delete(t),console.log(`[Request] Session timed out: ${t}`),a.generators.forEach(e=>{n.finalizeGenerator(e)})},i.server.session_timeout))}}),{app:h,config:i}},Ra=async({backend:e,router:t,config:n,enableOpenAICompat:r=!1,enableAnthropicMessages:i=!1})=>{let{app:a,config:s}=await La({backend:e,router:t,config:n,enableOpenAICompat:r,enableAnthropicMessages:i}),{server:{port:c}}=s,l=[new Promise(e=>a.listen(c,e))],u=null;if(s.autodiscover){let{workspace:e,serverKeyPair:t}=a.store.workspaceState;if(e&&t){let e={kid:t.kid,privateKey:o.createPrivateKey({key:Buffer.from(t.privateKeyPkcs8,`base64`),format:`der`,type:`pkcs8`})};u=new ha(s.autodiscover,()=>a.store.serverInfo,e),l.push(u.start())}else e?console.warn("[Autodiscover] disabled: bound to a workspace but state.json is missing serverKeyPair. Re-run `bricks buttress bind` to register a per-server announce key (required for v2.0 signed UDP discovery)."):console.warn("[Autodiscover] disabled: buttress-server is not bound to a workspace. Run `bricks buttress bind` from a workspace-authed CLI to pair.")}return await Promise.all(l),{app:a,port:c,openaiEnabled:r,anthropicMessagesEnabled:i,autoDiscover:u}},za=[new URL(`index.mjs`,import.meta.url).pathname,new URL(`index.ts`,import.meta.url).pathname];(process.argv[1]?.endsWith(`/bricks-buttress`)||za.includes(process.argv[1]))&&await ka();export{Fa as checkAndNotifyUpdates,Ma as checkForUpdates,Na as compareVersions,La as createServer,Pa as logUpdateMessage,Gi as processConfig,Xr as startModelDownload,Ra as startServer};
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@fugood/buttress-server",
3
- "version": "2.24.2",
3
+ "version": "2.24.3",
4
4
  "main": "lib/index.mjs",
5
5
  "types": "lib/index.d.mts",
6
6
  "type": "module",
@@ -30,12 +30,13 @@
30
30
  "dependencies": {
31
31
  "@elysiajs/cors": "^1.1.1",
32
32
  "@elysiajs/node": "^1.4.2",
33
- "@fugood/llama.node": "^1.7.0-rc.11",
33
+ "@fugood/llama.node": "^1.7.0",
34
34
  "@fugood/whisper.node": "^1.0.18",
35
35
  "@huggingface/gguf": "^0.3.2",
36
36
  "@iarna/toml": "^3.0.0",
37
37
  "bytes": "^3.1.0",
38
38
  "elysia": "^1.4.19",
39
+ "jose": "^5.9.6",
39
40
  "ms": "^2.1.1",
40
41
  "node-machine-id": "^1.1.12",
41
42
  "zod": "^3.25.76"