llama-cpp-pro 0.2.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (310) hide show
  1. package/CHANGELOG.md +295 -0
  2. package/LICENSE +21 -0
  3. package/LlamaCpp.podspec +33 -0
  4. package/LlamaCppCapacitor.podspec +33 -0
  5. package/Package.swift +29 -0
  6. package/README.md +93 -0
  7. package/android/build.gradle +88 -0
  8. package/android/src/main/AndroidManifest.xml +4 -0
  9. package/android/src/main/CMakeLists-arm64.txt +138 -0
  10. package/android/src/main/CMakeLists-x86_64.txt +141 -0
  11. package/android/src/main/CMakeLists.txt +138 -0
  12. package/android/src/main/java/ai/annadata/plugin/capacitor/LlamaCpp.java +1340 -0
  13. package/android/src/main/java/ai/annadata/plugin/capacitor/LlamaCppPlugin.java +814 -0
  14. package/android/src/main/java/ai/annadata/plugin/capacitor/ModelAdmissionController.java +214 -0
  15. package/android/src/main/jni-chat-session.cpp +261 -0
  16. package/android/src/main/jni-lora.cpp +197 -0
  17. package/android/src/main/jni-multimodal.cpp +167 -0
  18. package/android/src/main/jni-tts.cpp +290 -0
  19. package/android/src/main/jni-utils.h +148 -0
  20. package/android/src/main/jni.cpp +1808 -0
  21. package/android/src/main/jniLibs/arm64-v8a/libllama-cpp-arm64.so +0 -0
  22. package/android/src/main/res/.gitkeep +0 -0
  23. package/build-native.sh +280 -0
  24. package/cmake/desktop-metal-embed.cmake +30 -0
  25. package/cmake/desktop-sources.cmake +100 -0
  26. package/cmake/ggml-backends.cmake +44 -0
  27. package/cpp/LICENSE +21 -0
  28. package/cpp/README.md +15 -0
  29. package/cpp/anyascii.c +22223 -0
  30. package/cpp/anyascii.h +42 -0
  31. package/cpp/cap-completion.cpp +942 -0
  32. package/cpp/cap-completion.h +127 -0
  33. package/cpp/cap-embedding.cpp +193 -0
  34. package/cpp/cap-embedding.h +35 -0
  35. package/cpp/cap-ios-bridge.cpp +1810 -0
  36. package/cpp/cap-ios-bridge.h +61 -0
  37. package/cpp/cap-llama.cpp +412 -0
  38. package/cpp/cap-llama.h +161 -0
  39. package/cpp/cap-mtmd.hpp +602 -0
  40. package/cpp/cap-native-server.cpp +1095 -0
  41. package/cpp/cap-native-server.h +40 -0
  42. package/cpp/cap-tts.cpp +591 -0
  43. package/cpp/cap-tts.h +59 -0
  44. package/cpp/cap-wasm-fs.cpp +156 -0
  45. package/cpp/cap-wasm-jspi.cpp +19 -0
  46. package/cpp/cap-wasm-jspi.h +30 -0
  47. package/cpp/cap-wasm-vfs.cpp +22 -0
  48. package/cpp/chat-parser.cpp +393 -0
  49. package/cpp/chat-parser.h +120 -0
  50. package/cpp/chat.cpp +2315 -0
  51. package/cpp/chat.h +221 -0
  52. package/cpp/common.cpp +1664 -0
  53. package/cpp/common.h +744 -0
  54. package/cpp/ggml-alloc.c +1028 -0
  55. package/cpp/ggml-alloc.h +76 -0
  56. package/cpp/ggml-backend-impl.h +255 -0
  57. package/cpp/ggml-backend-reg.cpp +600 -0
  58. package/cpp/ggml-backend.cpp +2121 -0
  59. package/cpp/ggml-backend.h +354 -0
  60. package/cpp/ggml-common.h +1878 -0
  61. package/cpp/ggml-cpp.h +39 -0
  62. package/cpp/ggml-cpu/amx/amx.cpp +221 -0
  63. package/cpp/ggml-cpu/amx/amx.h +8 -0
  64. package/cpp/ggml-cpu/amx/common.h +91 -0
  65. package/cpp/ggml-cpu/amx/mmq.cpp +2512 -0
  66. package/cpp/ggml-cpu/amx/mmq.h +10 -0
  67. package/cpp/ggml-cpu/arch/arm/cpu-feats.cpp +94 -0
  68. package/cpp/ggml-cpu/arch/arm/quants.c +3650 -0
  69. package/cpp/ggml-cpu/arch/arm/repack.cpp +1891 -0
  70. package/cpp/ggml-cpu/arch/x86/cpu-feats.cpp +327 -0
  71. package/cpp/ggml-cpu/arch/x86/quants.c +3820 -0
  72. package/cpp/ggml-cpu/arch/x86/repack.cpp +6307 -0
  73. package/cpp/ggml-cpu/arch-fallback.h +215 -0
  74. package/cpp/ggml-cpu/binary-ops.cpp +158 -0
  75. package/cpp/ggml-cpu/binary-ops.h +16 -0
  76. package/cpp/ggml-cpu/common.h +73 -0
  77. package/cpp/ggml-cpu/ggml-cpu-impl.h +559 -0
  78. package/cpp/ggml-cpu/ggml-cpu.c +3578 -0
  79. package/cpp/ggml-cpu/ggml-cpu.cpp +672 -0
  80. package/cpp/ggml-cpu/ops.cpp +10587 -0
  81. package/cpp/ggml-cpu/ops.h +114 -0
  82. package/cpp/ggml-cpu/quants.c +1193 -0
  83. package/cpp/ggml-cpu/quants.h +97 -0
  84. package/cpp/ggml-cpu/repack.cpp +1982 -0
  85. package/cpp/ggml-cpu/repack.h +120 -0
  86. package/cpp/ggml-cpu/simd-mappings.h +1184 -0
  87. package/cpp/ggml-cpu/traits.cpp +36 -0
  88. package/cpp/ggml-cpu/traits.h +38 -0
  89. package/cpp/ggml-cpu/unary-ops.cpp +186 -0
  90. package/cpp/ggml-cpu/unary-ops.h +28 -0
  91. package/cpp/ggml-cpu/vec.cpp +348 -0
  92. package/cpp/ggml-cpu/vec.h +1121 -0
  93. package/cpp/ggml-cpu.h +145 -0
  94. package/cpp/ggml-impl.h +622 -0
  95. package/cpp/ggml-metal-impl.h +688 -0
  96. package/cpp/ggml-metal.h +66 -0
  97. package/cpp/ggml-metal.m +6833 -0
  98. package/cpp/ggml-metal.metal +10754 -0
  99. package/cpp/ggml-opt.cpp +1093 -0
  100. package/cpp/ggml-opt.h +256 -0
  101. package/cpp/ggml-quants.c +5324 -0
  102. package/cpp/ggml-quants.h +106 -0
  103. package/cpp/ggml-threading.cpp +12 -0
  104. package/cpp/ggml-threading.h +14 -0
  105. package/cpp/ggml.c +7108 -0
  106. package/cpp/ggml.h +2492 -0
  107. package/cpp/gguf.cpp +1358 -0
  108. package/cpp/gguf.h +202 -0
  109. package/cpp/json-partial.cpp +256 -0
  110. package/cpp/json-partial.h +38 -0
  111. package/cpp/json-schema-to-grammar.cpp +985 -0
  112. package/cpp/json-schema-to-grammar.h +21 -0
  113. package/cpp/llama-adapter.cpp +388 -0
  114. package/cpp/llama-adapter.h +76 -0
  115. package/cpp/llama-arch.cpp +2355 -0
  116. package/cpp/llama-arch.h +499 -0
  117. package/cpp/llama-batch.cpp +875 -0
  118. package/cpp/llama-batch.h +160 -0
  119. package/cpp/llama-chat.cpp +783 -0
  120. package/cpp/llama-chat.h +65 -0
  121. package/cpp/llama-context.cpp +2788 -0
  122. package/cpp/llama-context.h +306 -0
  123. package/cpp/llama-cparams.cpp +5 -0
  124. package/cpp/llama-cparams.h +41 -0
  125. package/cpp/llama-cpp.h +30 -0
  126. package/cpp/llama-grammar.cpp +1229 -0
  127. package/cpp/llama-grammar.h +173 -0
  128. package/cpp/llama-graph.cpp +1891 -0
  129. package/cpp/llama-graph.h +810 -0
  130. package/cpp/llama-hparams.cpp +180 -0
  131. package/cpp/llama-hparams.h +233 -0
  132. package/cpp/llama-impl.cpp +167 -0
  133. package/cpp/llama-impl.h +61 -0
  134. package/cpp/llama-io.cpp +15 -0
  135. package/cpp/llama-io.h +35 -0
  136. package/cpp/llama-kv-cache-iswa.cpp +318 -0
  137. package/cpp/llama-kv-cache-iswa.h +135 -0
  138. package/cpp/llama-kv-cache.cpp +2059 -0
  139. package/cpp/llama-kv-cache.h +374 -0
  140. package/cpp/llama-kv-cells.h +491 -0
  141. package/cpp/llama-memory-hybrid.cpp +258 -0
  142. package/cpp/llama-memory-hybrid.h +137 -0
  143. package/cpp/llama-memory-recurrent.cpp +1146 -0
  144. package/cpp/llama-memory-recurrent.h +179 -0
  145. package/cpp/llama-memory.cpp +59 -0
  146. package/cpp/llama-memory.h +119 -0
  147. package/cpp/llama-mmap.cpp +609 -0
  148. package/cpp/llama-mmap.h +68 -0
  149. package/cpp/llama-model-loader.cpp +1166 -0
  150. package/cpp/llama-model-loader.h +170 -0
  151. package/cpp/llama-model-saver.cpp +282 -0
  152. package/cpp/llama-model-saver.h +37 -0
  153. package/cpp/llama-model.cpp +19061 -0
  154. package/cpp/llama-model.h +491 -0
  155. package/cpp/llama-sampling.cpp +2575 -0
  156. package/cpp/llama-sampling.h +32 -0
  157. package/cpp/llama-vocab.cpp +3792 -0
  158. package/cpp/llama-vocab.h +176 -0
  159. package/cpp/llama.cpp +358 -0
  160. package/cpp/llama.h +1373 -0
  161. package/cpp/log.cpp +428 -0
  162. package/cpp/log.h +103 -0
  163. package/cpp/minja/chat-template.hpp +550 -0
  164. package/cpp/minja/minja.hpp +3009 -0
  165. package/cpp/nlohmann/json.hpp +25526 -0
  166. package/cpp/nlohmann/json_fwd.hpp +187 -0
  167. package/cpp/regex-partial.cpp +204 -0
  168. package/cpp/regex-partial.h +56 -0
  169. package/cpp/sampling.cpp +579 -0
  170. package/cpp/sampling.h +107 -0
  171. package/cpp/tools/mtmd/clip-impl.h +473 -0
  172. package/cpp/tools/mtmd/clip.cpp +4322 -0
  173. package/cpp/tools/mtmd/clip.h +106 -0
  174. package/cpp/tools/mtmd/miniaudio/miniaudio.h +93468 -0
  175. package/cpp/tools/mtmd/mtmd-audio.cpp +769 -0
  176. package/cpp/tools/mtmd/mtmd-audio.h +47 -0
  177. package/cpp/tools/mtmd/mtmd-helper.cpp +460 -0
  178. package/cpp/tools/mtmd/mtmd-helper.h +95 -0
  179. package/cpp/tools/mtmd/mtmd.cpp +1066 -0
  180. package/cpp/tools/mtmd/mtmd.h +302 -0
  181. package/cpp/tools/mtmd/stb/stb_image.h +7988 -0
  182. package/cpp/unicode-data.cpp +7034 -0
  183. package/cpp/unicode-data.h +20 -0
  184. package/cpp/unicode.cpp +1061 -0
  185. package/cpp/unicode.h +68 -0
  186. package/cpp/vendor/cpp-httplib/httplib.cpp +16509 -0
  187. package/cpp/vendor/cpp-httplib/httplib.h +3883 -0
  188. package/desktop/electron-builder.config.cjs +157 -0
  189. package/desktop/entitlements.mac.plist +12 -0
  190. package/desktop/package.json +11 -0
  191. package/desktop/resolve-package-root.cjs +88 -0
  192. package/desktop/src/main/backend-selector.cjs +148 -0
  193. package/desktop/src/main/gpu-probe.cjs +142 -0
  194. package/desktop/src/main/index.cjs +58 -0
  195. package/desktop/src/main/ipc-handlers.cjs +212 -0
  196. package/desktop/src/main/model-store.cjs +50 -0
  197. package/desktop/src/main/preload.cjs +39 -0
  198. package/desktop/src/main/sidecar-client.cjs +76 -0
  199. package/desktop/src/main/sidecar-manager.cjs +364 -0
  200. package/dist/docs.json +14357 -0
  201. package/dist/esm/definitions.d.ts +731 -0
  202. package/dist/esm/definitions.js +2 -0
  203. package/dist/esm/definitions.js.map +1 -0
  204. package/dist/esm/desktop.d.ts +37 -0
  205. package/dist/esm/desktop.js +133 -0
  206. package/dist/esm/desktop.js.map +1 -0
  207. package/dist/esm/index.d.ts +200 -0
  208. package/dist/esm/index.js +612 -0
  209. package/dist/esm/index.js.map +1 -0
  210. package/dist/esm/isomorphic/desktop.runtime.d.ts +37 -0
  211. package/dist/esm/isomorphic/desktop.runtime.js +36 -0
  212. package/dist/esm/isomorphic/desktop.runtime.js.map +1 -0
  213. package/dist/esm/isomorphic/errors.d.ts +6 -0
  214. package/dist/esm/isomorphic/errors.js +9 -0
  215. package/dist/esm/isomorphic/errors.js.map +1 -0
  216. package/dist/esm/isomorphic/model.admission.d.ts +17 -0
  217. package/dist/esm/isomorphic/model.admission.js +27 -0
  218. package/dist/esm/isomorphic/model.admission.js.map +1 -0
  219. package/dist/esm/isomorphic/model.scheduler.d.ts +31 -0
  220. package/dist/esm/isomorphic/model.scheduler.js +95 -0
  221. package/dist/esm/isomorphic/model.scheduler.js.map +1 -0
  222. package/dist/esm/isomorphic/provider.desktop.d.ts +40 -0
  223. package/dist/esm/isomorphic/provider.desktop.js +411 -0
  224. package/dist/esm/isomorphic/provider.desktop.js.map +1 -0
  225. package/dist/esm/isomorphic/provider.factory.d.ts +2 -0
  226. package/dist/esm/isomorphic/provider.factory.js +16 -0
  227. package/dist/esm/isomorphic/provider.factory.js.map +1 -0
  228. package/dist/esm/isomorphic/provider.interface.d.ts +76 -0
  229. package/dist/esm/isomorphic/provider.interface.js +2 -0
  230. package/dist/esm/isomorphic/provider.interface.js.map +1 -0
  231. package/dist/esm/isomorphic/provider.native.d.ts +18 -0
  232. package/dist/esm/isomorphic/provider.native.js +173 -0
  233. package/dist/esm/isomorphic/provider.native.js.map +1 -0
  234. package/dist/esm/isomorphic/provider.web.d.ts +88 -0
  235. package/dist/esm/isomorphic/provider.web.js +573 -0
  236. package/dist/esm/isomorphic/provider.web.js.map +1 -0
  237. package/dist/esm/isomorphic/sidecar-sse.d.ts +14 -0
  238. package/dist/esm/isomorphic/sidecar-sse.js +76 -0
  239. package/dist/esm/isomorphic/sidecar-sse.js.map +1 -0
  240. package/dist/esm/isomorphic/wasmMemoryCalibration.d.ts +26 -0
  241. package/dist/esm/isomorphic/wasmMemoryCalibration.js +60 -0
  242. package/dist/esm/isomorphic/wasmMemoryCalibration.js.map +1 -0
  243. package/dist/esm/isomorphic/wasmMemoryPolicy.d.ts +55 -0
  244. package/dist/esm/isomorphic/wasmMemoryPolicy.js +93 -0
  245. package/dist/esm/isomorphic/wasmMemoryPolicy.js.map +1 -0
  246. package/dist/esm/storage/manifest.d.ts +13 -0
  247. package/dist/esm/storage/manifest.js +87 -0
  248. package/dist/esm/storage/manifest.js.map +1 -0
  249. package/dist/esm/storage/opfs.store.d.ts +31 -0
  250. package/dist/esm/storage/opfs.store.js +241 -0
  251. package/dist/esm/storage/opfs.store.js.map +1 -0
  252. package/dist/esm/web.d.ts +206 -0
  253. package/dist/esm/web.js +521 -0
  254. package/dist/esm/web.js.map +1 -0
  255. package/dist/esm/workers/async-file.d.ts +13 -0
  256. package/dist/esm/workers/async-file.js +12 -0
  257. package/dist/esm/workers/async-file.js.map +1 -0
  258. package/dist/esm/workers/heapfs.d.ts +55 -0
  259. package/dist/esm/workers/heapfs.js +115 -0
  260. package/dist/esm/workers/heapfs.js.map +1 -0
  261. package/dist/esm/workers/wasm.engine.d.ts +86 -0
  262. package/dist/esm/workers/wasm.engine.js +504 -0
  263. package/dist/esm/workers/wasm.engine.js.map +1 -0
  264. package/dist/esm/workers/worker.protocol.d.ts +160 -0
  265. package/dist/esm/workers/worker.protocol.js +2 -0
  266. package/dist/esm/workers/worker.protocol.js.map +1 -0
  267. package/dist/plugin.cjs +3179 -0
  268. package/dist/plugin.cjs.map +1 -0
  269. package/dist/plugin.js +3181 -0
  270. package/dist/plugin.js.map +1 -0
  271. package/dist/wasm/llama_engine.d.ts +74 -0
  272. package/dist/wasm/llama_engine.js +1829 -0
  273. package/dist/wasm/llama_engine.wasm +0 -0
  274. package/dist/wasm/llama_engine_emscripten.mjs +2 -0
  275. package/dist/wasm/package.json +16 -0
  276. package/dist/workers/llm.worker.js +1226 -0
  277. package/dist/workers/llm.worker.js.map +7 -0
  278. package/extraResources/llama-wasm/llama_engine.d.ts +74 -0
  279. package/extraResources/llama-wasm/llama_engine.js +1829 -0
  280. package/extraResources/llama-wasm/llama_engine.wasm +0 -0
  281. package/extraResources/llama-wasm/llama_engine_emscripten.mjs +2 -0
  282. package/extraResources/llama-wasm/package.json +16 -0
  283. package/extraResources/sidecar/README.md +11 -0
  284. package/extraResources/sidecar/darwin-arm64 +0 -0
  285. package/extraResources/sidecar/darwin-x64 +0 -0
  286. package/ios/CMakeLists-arm64.txt +161 -0
  287. package/ios/CMakeLists-x86_64.txt +188 -0
  288. package/ios/CMakeLists.txt +161 -0
  289. package/ios/Frameworks/llama-cpp.framework/Info.plist +28 -0
  290. package/ios/Frameworks/llama-cpp.framework/llama-cpp +0 -0
  291. package/ios/Sources/LlamaCppCapacitor/LlamaCpp.swift +1365 -0
  292. package/ios/Sources/LlamaCppCapacitor/LlamaCppPlugin.swift +694 -0
  293. package/ios/Sources/LlamaCppCapacitor/LlamaNativeBridge.swift +349 -0
  294. package/ios/Sources/LlamaCppCapacitor/ModelAdmissionController.swift +177 -0
  295. package/ios/embed-metal-shaders.sh +24 -0
  296. package/ios/metal-embed.cmake +29 -0
  297. package/package.json +281 -0
  298. package/scripts/build-js-preserve-wasm.cjs +53 -0
  299. package/scripts/build-sidecar-linux.sh +6 -0
  300. package/scripts/build-sidecar-win.bat +19 -0
  301. package/scripts/build-sidecar.sh +184 -0
  302. package/scripts/embed-llama-ios-app-framework.sh +63 -0
  303. package/scripts/ensure-desktop-sidecar-bundle.cjs +103 -0
  304. package/scripts/ensure-llama-ios-xcframework.sh +108 -0
  305. package/scripts/fix-esm-extensions.cjs +45 -0
  306. package/scripts/prepare-js-dist.cjs +28 -0
  307. package/scripts/stage-desktop-resources.cjs +116 -0
  308. package/sidecar/CMakeLists.txt +192 -0
  309. package/sidecar/cap-sidecar-main.cpp +68 -0
  310. package/types/llama-cpp-pro.d.ts +441 -0
@@ -0,0 +1,364 @@
1
+ /**
2
+ * Sidecar process lifecycle manager for llama-cpp desktop.
3
+ * Supports NVIDIA (CUDA), AMD (ROCm), Intel (OpenVINO), Apple (Metal), Vulkan, and CPU.
4
+ */
5
+
6
+ const path = require('path');
7
+ const os = require('os');
8
+ const fs = require('fs');
9
+ const { spawn } = require('child_process');
10
+ const { createSidecarClient } = require('./sidecar-client.cjs');
11
+ const { backendToVariant } = require('./backend-selector.cjs');
12
+
13
+ const MAX_RESTART_ATTEMPTS = 2;
14
+ const HEALTH_INTERVAL_MS = 500;
15
+ const HEALTH_MAX_ATTEMPTS = 30;
16
+ const SHUTDOWN_GRACE_MS = 500;
17
+
18
+ function getRandomPort() {
19
+ return 18080 + Math.floor(Math.random() * 2000);
20
+ }
21
+
22
+ function vulkanStderrImpliesGpuFailure(buf) {
23
+ if (!buf || buf.length < 12) return false;
24
+ return (
25
+ /VK_ERROR_[A-Z0-9_]+/i.test(buf)
26
+ || (/ggml_vulkan/i.test(buf) && /error|fail/i.test(buf))
27
+ || /Vulkan.*(fail|error)/i.test(buf)
28
+ );
29
+ }
30
+
31
+ function rocmStderrImpliesGpuFailure(buf) {
32
+ if (!buf || buf.length < 12) return false;
33
+ return (
34
+ /rocblas\s*error/i.test(buf)
35
+ || /tensilelibrary/i.test(buf)
36
+ || /ROCm error/i.test(buf)
37
+ || /hipblas/i.test(buf)
38
+ );
39
+ }
40
+
41
+ /**
42
+ * Resolve sidecar binary for platform, arch, and GPU backend variant.
43
+ */
44
+ function resolveBinaryPath(deps) {
45
+ const platform = (deps && deps.platform) || process.platform;
46
+ const arch = (deps && deps.arch) || process.arch;
47
+ const gpuBackend = (deps && deps.gpuBackend) || null;
48
+ const variant = (deps && deps.variant) || (gpuBackend ? backendToVariant(gpuBackend, platform) : null);
49
+ const resourcesPath = (deps && deps.resourcesPath) || process.resourcesPath;
50
+ const _fs = (deps && deps.fs) || fs;
51
+
52
+ const ext = platform === 'win32' ? '.exe' : '';
53
+ const names = [];
54
+
55
+ if (variant) {
56
+ names.push(`${platform}-${arch}-${variant}${ext}`);
57
+ }
58
+ if (gpuBackend === 'rocm') {
59
+ names.push(`${platform}-${arch}-rocm${ext}`);
60
+ }
61
+ if (gpuBackend === 'cuda') {
62
+ names.push(`${platform}-${arch}-cuda${ext}`);
63
+ }
64
+ if (gpuBackend === 'vulkan') {
65
+ names.push(`${platform}-${arch}-vulkan${ext}`);
66
+ }
67
+ names.push(`${platform}-${arch}${ext}`);
68
+
69
+ const searchDirs = [];
70
+ if (resourcesPath) {
71
+ searchDirs.push(path.join(resourcesPath, 'sidecar'));
72
+ }
73
+ searchDirs.push(
74
+ path.join(process.cwd(), 'extraResources', 'sidecar'),
75
+ path.join(process.cwd(), 'sidecar', 'bin'),
76
+ path.join(process.cwd(), 'sidecar', 'build', 'bin'),
77
+ );
78
+
79
+ for (const dir of searchDirs) {
80
+ for (const name of names) {
81
+ const p = path.join(dir, name);
82
+ try {
83
+ _fs.accessSync(p);
84
+ return p;
85
+ } catch (_) { /* next */ }
86
+ }
87
+ }
88
+
89
+ return path.join(searchDirs[0], names[0]);
90
+ }
91
+
92
+ function resolveBackendPluginDir(binaryPath, deps) {
93
+ const platform = (deps && deps.platform) || process.platform;
94
+ const arch = (deps && deps.arch) || process.arch;
95
+ const resourcesPath = (deps && deps.resourcesPath) || process.resourcesPath;
96
+ const _fs = (deps && deps.fs) || fs;
97
+
98
+ const candidates = [
99
+ path.join(path.dirname(binaryPath), 'ggml-plugins', `${platform}-${arch}`),
100
+ path.join(path.dirname(binaryPath), 'runtime-libs', `${platform}-${arch}`),
101
+ path.dirname(binaryPath),
102
+ ];
103
+ if (resourcesPath) {
104
+ candidates.unshift(
105
+ path.join(resourcesPath, 'sidecar', 'ggml-plugins', `${platform}-${arch}`),
106
+ path.join(resourcesPath, 'sidecar', 'runtime-libs', `${platform}-${arch}`),
107
+ );
108
+ }
109
+
110
+ for (const c of candidates) {
111
+ try {
112
+ if (_fs.existsSync(c)) return c;
113
+ } catch (_) { /* skip */ }
114
+ }
115
+ return path.dirname(binaryPath);
116
+ }
117
+
118
+ function buildSpawnEnv(binaryPath, backendDir, forceCpu, deps) {
119
+ const platform = (deps && deps.platform) || process.platform;
120
+ const env = { ...process.env };
121
+
122
+ if (backendDir) {
123
+ if (platform === 'win32') {
124
+ env.PATH = `${backendDir};${env.PATH || ''}`;
125
+ } else {
126
+ env.LD_LIBRARY_PATH = `${backendDir}:${env.LD_LIBRARY_PATH || ''}`;
127
+ env.DYLD_LIBRARY_PATH = `${backendDir}:${env.DYLD_LIBRARY_PATH || ''}`;
128
+ }
129
+ }
130
+
131
+ if (forceCpu) {
132
+ env.LLAMA_NO_GPU = '1';
133
+ }
134
+
135
+ return env;
136
+ }
137
+
138
+ function createSidecarManager(deps) {
139
+ const _fs = (deps && deps.fs) || fs;
140
+ const _spawn = (deps && deps.spawn) || spawn;
141
+ const _createClient = (deps && deps.createClient) || createSidecarClient;
142
+ const _resolveBinaryPath = (deps && deps.resolveBinaryPath) || resolveBinaryPath;
143
+ const _getRandomPort = (deps && deps.getRandomPort) || getRandomPort;
144
+ const _log = (deps && deps.log) || console;
145
+
146
+ let status = {
147
+ running: false,
148
+ port: null,
149
+ pid: null,
150
+ backend: null,
151
+ variant: null,
152
+ binaryPath: null,
153
+ restartCount: 0,
154
+ permanentWasmFallback: false,
155
+ forceCpu: false,
156
+ };
157
+
158
+ let childProcess = null;
159
+ let client = null;
160
+ let intentionalStop = false;
161
+ let accelStderr = '';
162
+ const listeners = [];
163
+
164
+ function emit() {
165
+ const snap = { ...status };
166
+ for (const fn of listeners) {
167
+ try { fn(snap); } catch (_) { /* ignore */ }
168
+ }
169
+ }
170
+
171
+ function buildSpawnArgs(modelPath, port, opts) {
172
+ const args = [
173
+ '--host', '127.0.0.1',
174
+ '--port', String(port),
175
+ ];
176
+ if (modelPath) {
177
+ args.push('--model', modelPath);
178
+ if (opts && opts.modelId) {
179
+ args.push('--model-id', String(opts.modelId));
180
+ }
181
+ }
182
+ if (opts && opts.n_ctx) {
183
+ args.push('--ctx-size', String(opts.n_ctx));
184
+ }
185
+ if (opts && opts.n_threads) {
186
+ args.push('--threads', String(opts.n_threads));
187
+ }
188
+ const ngl = opts && opts.n_gpu_layers;
189
+ if (status.forceCpu || opts?.forceCpu) {
190
+ args.push('--no-gpu');
191
+ } else if (ngl != null && ngl >= 0) {
192
+ args.push('--n-gpu-layers', String(ngl));
193
+ } else if (status.backend && status.backend !== 'cpu') {
194
+ args.push('--n-gpu-layers', '99');
195
+ }
196
+ const backendDir = resolveBackendPluginDir(status.binaryPath || '', deps);
197
+ if (backendDir) {
198
+ args.push('--backend-dir', backendDir);
199
+ }
200
+ return args;
201
+ }
202
+
203
+ async function waitForHealth(port) {
204
+ const c = _createClient(port);
205
+ for (let i = 0; i < HEALTH_MAX_ATTEMPTS; i++) {
206
+ try {
207
+ const h = await c.health();
208
+ if (h && h.status === 'ok') return true;
209
+ } catch (_) { /* retry */ }
210
+ await new Promise((r) => setTimeout(r, HEALTH_INTERVAL_MS));
211
+ }
212
+ return false;
213
+ }
214
+
215
+ async function start(options) {
216
+ if (status.permanentWasmFallback) {
217
+ return { ok: false, reason: 'permanent-wasm-fallback' };
218
+ }
219
+
220
+ const selection = options && options.selection;
221
+ const modelPath = options && options.modelPath;
222
+
223
+ if (status.running && status.port) {
224
+ return { ok: true, port: status.port };
225
+ }
226
+
227
+ const gpuBackend = selection && selection.gpuBackend;
228
+ const variant = (selection && selection.variant) || null;
229
+ status.variant = variant;
230
+ status.backend = gpuBackend || (selection && selection.type) || 'cpu';
231
+ status.binaryPath = _resolveBinaryPath({ ...deps, variant, gpuBackend });
232
+
233
+ try {
234
+ _fs.accessSync(status.binaryPath);
235
+ } catch (_) {
236
+ _log.warn(`[sidecar-manager] Binary not found: ${status.binaryPath}`);
237
+ status.permanentWasmFallback = true;
238
+ emit();
239
+ return { ok: false, reason: 'binary-missing', path: status.binaryPath };
240
+ }
241
+
242
+ const port = options?.port || _getRandomPort();
243
+ const backendDir = resolveBackendPluginDir(status.binaryPath, deps);
244
+ const args = buildSpawnArgs(modelPath, port, options);
245
+ const env = buildSpawnEnv(status.binaryPath, backendDir, status.forceCpu, deps);
246
+
247
+ intentionalStop = false;
248
+ accelStderr = '';
249
+
250
+ childProcess = _spawn(status.binaryPath, args, {
251
+ stdio: ['ignore', 'pipe', 'pipe'],
252
+ windowsHide: true,
253
+ env,
254
+ });
255
+
256
+ status.pid = childProcess.pid;
257
+ status.port = port;
258
+
259
+ childProcess.stderr.on('data', (chunk) => {
260
+ const text = String(chunk);
261
+ accelStderr = (accelStderr + text).slice(-65536);
262
+ if (process.env.LLAMA_SIDECAR_VERBOSE === '1') {
263
+ _log.info('[sidecar]', text.trim());
264
+ }
265
+ });
266
+
267
+ childProcess.on('exit', (code) => {
268
+ if (intentionalStop) {
269
+ status.running = false;
270
+ status.pid = null;
271
+ emit();
272
+ return;
273
+ }
274
+
275
+ const gpuFail =
276
+ vulkanStderrImpliesGpuFailure(accelStderr)
277
+ || rocmStderrImpliesGpuFailure(accelStderr);
278
+
279
+ if (gpuFail && !status.forceCpu && options?.retryCpu !== false) {
280
+ _log.warn('[sidecar-manager] GPU init failed — retrying CPU-only');
281
+ status.forceCpu = true;
282
+ status.restartCount += 1;
283
+ status.running = false;
284
+ emit();
285
+ start({ ...options, forceCpu: true, retryCpu: false }).catch(() => {});
286
+ return;
287
+ }
288
+
289
+ if (status.restartCount < MAX_RESTART_ATTEMPTS) {
290
+ status.restartCount += 1;
291
+ _log.warn(`[sidecar-manager] Unexpected exit (${code}), restart ${status.restartCount}`);
292
+ status.running = false;
293
+ emit();
294
+ start(options).catch(() => {});
295
+ return;
296
+ }
297
+
298
+ status.permanentWasmFallback = true;
299
+ _log.warn('[sidecar-manager] Max restarts exceeded — WASM fallback');
300
+ status.running = false;
301
+ status.pid = null;
302
+ emit();
303
+ });
304
+
305
+ const healthy = await waitForHealth(port);
306
+ if (!healthy) {
307
+ await stop();
308
+ if (!status.forceCpu && options?.retryCpu !== false) {
309
+ status.forceCpu = true;
310
+ return start({ ...options, forceCpu: true, retryCpu: false });
311
+ }
312
+ status.permanentWasmFallback = true;
313
+ emit();
314
+ return { ok: false, reason: 'health-check-failed' };
315
+ }
316
+
317
+ client = _createClient(port);
318
+ status.running = true;
319
+ emit();
320
+ return { ok: true, port };
321
+ }
322
+
323
+ async function stop() {
324
+ intentionalStop = true;
325
+ if (childProcess && !childProcess.killed) {
326
+ childProcess.kill('SIGTERM');
327
+ await new Promise((r) => setTimeout(r, SHUTDOWN_GRACE_MS));
328
+ if (!childProcess.killed) childProcess.kill('SIGKILL');
329
+ }
330
+ childProcess = null;
331
+ client = null;
332
+ status.running = false;
333
+ status.port = null;
334
+ status.pid = null;
335
+ emit();
336
+ }
337
+
338
+ function getClient() {
339
+ return client;
340
+ }
341
+
342
+ function getStatus() {
343
+ return { ...status };
344
+ }
345
+
346
+ function onStatusChange(fn) {
347
+ listeners.push(fn);
348
+ return () => {
349
+ const i = listeners.indexOf(fn);
350
+ if (i >= 0) listeners.splice(i, 1);
351
+ };
352
+ }
353
+
354
+ return {
355
+ start, stop, getClient, getStatus, onStatusChange, resolveBinaryPath,
356
+ };
357
+ }
358
+
359
+ module.exports = {
360
+ createSidecarManager,
361
+ resolveBinaryPath,
362
+ resolveBackendPluginDir,
363
+ getRandomPort,
364
+ };