@qvac/translation-nmtcpp 0.17.1 → 0.18.1

package/README.md

@@ -288,7 +288,7 @@ const backendSettings = {
   use_gpu: true,              // Enable GPU inference (default: false → CPU-only)
   gpu_backend: 'vulkan',      // Optional: pick a specific backend by name substring
   gpu_device: 0,              // Optional: ordinal within matching devices (default: 0)
-  backendsDir: './prebuilds', // Optional: dir where libqvac-ggml-*.so / *.dll live
+  backendsDir: './prebuilds', // Optional override of the fabric backends root
   openclCacheDir: '/path/ok'  // Optional (Android only): OpenCL kernel-cache dir
 }
 ```
@@ -300,7 +300,7 @@ The three GPU control keys each accept a camelCase alias alongside the snake_cas
 | `use_gpu` | `useGPU` | boolean | Enable GPU inference. When `false` (default), only the CPU backend is used. Bergamot is CPU-only by design — this flag is effectively a no-op for Bergamot. |
 | `gpu_backend` | `gpuBackend` | string | Case-insensitive **substring** match against eligible Vulkan, Metal, OpenCL, or CUDA device names (e.g. `"vulkan"`, `"vulkan0"`, `"opencl"`, `"metal"`). When unset, the default gated selection runs (see [Backends](#backends)). Any explicit selector that resolves to OpenCL bypasses the build-time `USE_OPENCL` guard — an informed opt-in. |
 | `gpu_device` | `gpuDevice` | int | Ordinal within the matching devices. Defaults to `0` (first match). Example: `{gpu_backend: "vulkan", gpu_device: 1}` picks the second Vulkan adapter. |
-| `backendsDir` | — | string | Path to the directory containing the runtime backend shared libraries (`libqvac-ggml-vulkan.so`, etc.). Defaults to `<package>/prebuilds` when unset, which is where `npm install` places the shipped prebuilds. Must be an absolute path; paths with `..` segments or unresolvable symlinks are rejected with a warning and fall back to the default prebuilds directory. |
+| `backendsDir` | — | string | Path to the directory containing the runtime backend shared libraries (`libqvac-ggml-vulkan.so`, etc.). Defaults to the root `@qvac/fabric/backends` resolves on desktop, falling back to this package's `prebuilds/` on mobile, where the package tree isn't resolvable from the packed worklet. Must be an absolute path when set; paths with `..` segments or unresolvable symlinks are rejected with a warning and fall back to that default. |
 | `openclCacheDir` | — | string | **Android only.** Writable directory the OpenCL backend uses for its JIT kernel cache (forwarded via `GGML_OPENCL_CACHE_DIR`). Must be an absolute path; paths with `..` segments are rejected. The OpenCL backend falls back to a non-writable relative path if this is unset, which `ggml_abort()`s during init inside the app sandbox — always provide an app-writable path when exercising OpenCL on Android. |
 
 GPU execution is limited to Vulkan, Metal/MTL, OpenCL, and CUDA devices reported as GPU or integrated GPU. RPC, ROCm/HIP, SYCL, MUSA, and unknown backend families fall back to CPU — RPC because translation has no split-mode support and cannot execute over it. CUDA is admitted ahead of its planned Fabric build so that build does not require another selector change. An explicit selector only narrows this eligible set; it cannot enable an unsupported family.
@@ -389,9 +389,11 @@ try {
     })
     .await() // Wait for translation to complete
 
-  // Access performance statistics (if enabled with opts.stats)
+  // Access performance statistics (if enabled with opts.stats). The counters
+  // accumulate over the life of the loaded model; difference consecutive
+  // readings for per-run figures.
   if (response.stats) {
-    console.log('Translation completed in:', response.stats.totalTime, 's')
+    console.log('Total translation time so far:', response.stats.totalTime, 's')
   }
 } catch (error) {
   console.error('Translation failed:', error)
@@ -601,6 +603,15 @@ Populated on `response.stats` when the model was constructed with
 > Pivot translations may emit keys prefixed with the model name
 > (e.g. `"BERGAMOT : ->TPS"`); the table above models the non-pivot shape.
 
+`totalTime`, `decodeTime`, `encodeTime` and `TTFT` accumulate over the life
+of the loaded model rather than describing the run that returned them.
+`totalTokens` accumulates on Bergamot and on GGML greedy decoding
+(`beamsize: 1`), but GGML beam search reports the run's own count. `TPS`
+divides these readings as they stand. The counters restart at zero on
+`unload()`/`load()` and on `cancel()`. For per-run figures, difference
+consecutive readings — jobs are serialized, so nothing lands between them —
+and read a reading lower than the last as that run's own figure.
+
 ### Errors
 
 Native-facing failures reject with a structured `QvacErrorAddonMarian`

package/index.d.ts

@@ -147,8 +147,8 @@ declare namespace TranslationNmtcpp {
         gpuDevice?: number;
         /**
          * Path to the directory containing backend shared libraries
-         * (libqvac-ggml-vulkan.so, etc.). Defaults to `@qvac/fabric`'s `prebuilds/`
-         * on desktop, falling back to this package's `prebuilds/` on mobile where
+         * (libqvac-ggml-vulkan.so, etc.). Defaults to the root `@qvac/fabric/backends`
+         * resolves on desktop, falling back to this package's `prebuilds/` on mobile where
          * the package tree isn't resolvable from the packed worklet.
          */
         backendsDir?: string;

package/index.js

@@ -1,8 +1,9 @@
 "use strict";
 /* eslint-disable @typescript-eslint/no-require-imports -- Bare modules and @qvac/logging expose CommonJS export shapes. */
-const path = require("bare-path");
 const fs = require("bare-fs");
+const path = require("bare-path");
 const QvacLogger = require("@qvac/logging");
+const fabricBackends = require("@qvac/fabric/backends");
 /* eslint-enable @typescript-eslint/no-require-imports */
 const infer_base_1 = require("@qvac/infer-base");
 const marian_1 = require("./marian");
@@ -18,24 +19,20 @@ const indic_processor_1 = require("./third-party/indic-processor");
 const BERGAMOT_TARGET_TOKEN_BY_PAIR = {
     "en:pt": ">>por<<",
 };
-// The ggml compute backends (GGML_BACKEND_DL modules) ship exactly once, in the
-// @qvac/fabric dependency (prebuilds/<host>/qvac__fabric). We deliberately do
-// not copy them into this addon to avoid duplicating tens of MB per fabric
-// consumer. On desktop, resolve the single @qvac/fabric install and load the
-// backends from there. On mobile the package tree isn't resolvable at runtime
-// (the worklet runs from a packed bundle), so fall back to this addon's own
-// prebuilds, where the mobile packaging stages the backends. The native side
+// The ggml compute backends (GGML_BACKEND_DL modules) ship exactly once, next to
+// the @qvac/fabric runtime (<root>/<host>/qvac__fabric). We deliberately do not
+// copy them into this addon to avoid duplicating tens of MB per fabric
+// consumer. On desktop, @qvac/fabric/backends resolves that root in whichever
+// package holds the runtime. On mobile the package tree isn't resolvable at
+// runtime (the worklet runs from a packed bundle), so fall back to this addon's
+// own prebuilds, where the mobile packaging stages the backends. The native side
 // appends BACKENDS_SUBDIR ("<host>/qvac__fabric") to whichever root we return.
 function resolveBackendsDir() {
-    try {
-        const fabricPkg = require.resolve("@qvac/fabric/package");
-        const fabricPrebuilds = path.join(path.dirname(fabricPkg), "prebuilds");
-        if (fs.existsSync(fabricPrebuilds))
-            return fabricPrebuilds;
-    }
-    catch {
-        // Mobile worklets cannot resolve the @qvac/fabric package tree.
-    }
+    // fabric's resolver only checks that the platform package resolves, not that
+    // its prebuilds are on disk.
+    const fabricRoot = fabricBackends.resolveBackendsDir();
+    if (fabricRoot !== null && fs.existsSync(fabricRoot))
+        return fabricRoot;
     return path.join(__dirname, "prebuilds");
 }
 class QvacIndicTransResponse extends infer_base_1.QvacResponse {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/translation-nmtcpp",
-  "version": "0.17.1",
+  "version": "0.18.1",
   "description": "translation addon for qvac",
   "addon": true,
   "engines": {
@@ -78,6 +78,8 @@
   "bugs": "https://github.com/tetherto/qvac/issues",
   "homepage": "https://qvac.tether.io",
   "devDependencies": {
+    "@qvac/fabric-android-arm64": "^0.19.0",
+    "@qvac/fabric-ios": "^0.19.0",
     "@qvac/registry-client": "^0.6.1",
     "@types/node": "^24.2.1",
     "bare-fetch": "^3.0.1",
@@ -96,7 +98,7 @@
   },
   "dependencies": {
     "@qvac/error": "^0.1.0",
-    "@qvac/fabric": "^0.17.0",
+    "@qvac/fabric": "^0.19.0",
     "@qvac/infer-base": "^0.6.2",
     "@qvac/logging": "^0.1.0",
     "bare-fs": "^4.5.1",

package/prebuilds/darwin-arm64/qvac__translation-nmtcpp.bare.exports

@@ -7,7 +7,7 @@
     ],
     "current_versions": [
       {
-        "version": "0.17.1"
+        "version": "0.18.1"
       }
     ],
     "exported_symbols": [

package/prebuilds/darwin-x64/qvac__translation-nmtcpp.bare.exports

@@ -7,7 +7,7 @@
     ],
     "current_versions": [
       {
-        "version": "0.17.1"
+        "version": "0.18.1"
       }
     ],
     "exported_symbols": [

package/prebuilds/ios-arm64/qvac__translation-nmtcpp.bare.exports

@@ -7,7 +7,7 @@
     ],
     "current_versions": [
       {
-        "version": "0.17.1"
+        "version": "0.18.1"
       }
     ],
     "exported_symbols": [

package/prebuilds/ios-arm64-simulator/qvac__translation-nmtcpp.bare.exports

@@ -7,7 +7,7 @@
     ],
     "current_versions": [
       {
-        "version": "0.17.1"
+        "version": "0.18.1"
       }
     ],
     "exported_symbols": [

package/prebuilds/ios-x64-simulator/qvac__translation-nmtcpp.bare.exports

@@ -7,7 +7,7 @@
     ],
     "current_versions": [
       {
-        "version": "0.17.1"
+        "version": "0.18.1"
       }
     ],
     "exported_symbols": [