@a9i5k4/dsh-auto-memory 2.1.7 → 2.1.8

package/lib/client.js

@@ -1223,6 +1223,17 @@ window.__ModuleLoader__.load({
       var pollPair = useState(null)
       var poll = pollPair[0]
       var setPoll = pollPair[1]
+      var gpuPair = useState(function () { try { return localStorage.getItem('dsh-auto-memory.pyGpu') === '1' } catch (eG) { return false } })
+      var gpuPref = gpuPair[0]
+      var setGpuPref = function (v) {
+        gpuPair[1](v)
+        try { localStorage.setItem('dsh-auto-memory.pyGpu', v ? '1' : '0') } catch (eG) {}
+      }
+      var actGpu = function () {
+        if (busy) return
+        setBusy(true)
+        apiPost(API.pyDeps, { gpu: gpuPref }).then(function (d) { setData(d); setBusy(false); setPoll(true); var ph = d && d.phase; if (ph === 'ready' || ph === 'error' || ph === 'idle') setPoll(false) }).catch(function () { setBusy(false); setPoll(false) })
+      }
       var refresh = function () { apiGet(API.pyStatus).then(function (d) { if (d) setData(d) }).catch(function () {}) }
       useEffect(function () {
         var alive = true
@@ -1267,7 +1278,11 @@ window.__ModuleLoader__.load({
         stepRow(t('pyWizStep2'), data.venvOk ? true : 'run',
           (!data.venvOk && (data.pythons || []).some(function (p2) { return p2.status === 'ok' && !p2.isVenv })) ? h('button', Object.assign({}, btnCls, { onClick: function () { act(API.pyVenv, true) } }), t('pyWizCreate')) : null),
         stepRow(t('pyWizStep3'), data.depsOk ? true : 'run',
-          (data.venvOk && !data.depsOk) ? h('button', Object.assign({}, btnCls, { onClick: function () { act(API.pyDeps, true) } }), t('pyWizInstall')) : null),
+          (data.venvOk && !data.depsOk) ? h('span', { style: { display: 'flex', gap: '8px', alignItems: 'center' } },
+            h('label', { style: { display: 'flex', gap: '4px', alignItems: 'center', fontSize: 'calc(10.5px * var(--dam-scale))' } },
+              h('input', { type: 'checkbox', checked: gpuPref, onChange: function (e) { setGpuPref(e.target.checked) } }),
+              locale === 'zh' ? 'GPU 推理(NVIDIA,需已装 CUDA;否则自动回退 CPU)' : 'GPU inference (NVIDIA + CUDA; auto-falls back to CPU)'),
+            h('button', Object.assign({}, btnCls, { onClick: function () { actGpu() } }), t('pyWizInstall'))) : null),
         stepRow(t('pyWizStep4'), data.modelReady ? true : 'run',
           (!data.modelReady && !dlActive) ? h('button', Object.assign({}, btnCls, { onClick: function () { act(API.pyModel, true) } }), t('pyWizDownload')) : null,
           dlActive ? h('button', { 'data-dam-btn': '', onClick: function () { apiPost(API.pyCancel, {}).then(refresh).catch(function () {}) } }, t('pyWizCancel')) : null),
@@ -3421,7 +3436,7 @@ window.__ModuleLoader__.load({
                   h('option', { value: 'auto' }, t('mAuto')),
                   h('option', { value: 'cn' }, t('mCn')),
                   h('option', { value: 'intl' }, t('mIntl'))) : null,
-                ready ? h('button', { 'data-dam-btn': '', onClick: function () { setGuide(''); var n2 = Object.assign({}, cfg); n2.semanticEngineMode = guide; if (guide === 'js') { n2.activationSource = 'js'; n2.contextSinkMode = 'null' } else if (guide === 'python') { n2.activationSource = 'python'; n2.contextSinkMode = 'python' } setCfg(n2); setDirty(true) } }, locale === 'zh' ? '启用并继续' : 'Enable & continue') : null,
+                ready ? h('button', { 'data-dam-btn': '', onClick: function () { setGuide(''); var n2 = Object.assign({}, cfg); n2.semanticEngineMode = guide; if (guide === 'js') { n2.activationSource = 'js'; n2.contextSinkMode = 'null' } else if (guide === 'python') { n2.activationSource = 'python'; n2.contextSinkMode = 'python'; try { var gp = localStorage.getItem('dsh-auto-memory.pyGpu'); if (gp === '1') n2.pythonGpu = true } catch (eL) {} } setCfg(n2); setDirty(true) } }, locale === 'zh' ? '启用并继续' : 'Enable & continue') : null,
                 !ready && isJs && !dlActive ? h('button', { 'data-dam-btn': '', onClick: startDl }, (dl.phase === 'error' || dl.phase === 'cancelled') ? t('semDlRetry') : t('semDlStart')) : null,
                 !ready && isJs && dlActive ? h('button', { 'data-dam-btn': '', onClick: cancelDl }, t('semDlCancel')) : null,
                 h('button', { 'data-dam-btn': '', onClick: function () { setGuide('') } }, t('gotIt'))))

package/lib/index.js

@@ -5458,7 +5458,8 @@ export function apply(ctx, config) {
       handler: async (req, res) => {
         if (!isLoopbackRequest(req)) return writeJson(res, 403, { error: 'forbidden: loopback-only' })
         if ((req.method || 'POST') !== 'POST') return writeJson(res, 405, { error: 'method not allowed' })
-        try { writeJson(res, 200, await engine._pythonSetup.ensureDeps()) } catch (e) { writeJson(res, 500, { error: String(e && e.message || e) }) }
+        const body = await readJsonBody(req).catch(() => ({}))
+        try { writeJson(res, 200, await engine._pythonSetup.ensureDeps({ gpu: !!(body && body.gpu) })) } catch (e) { writeJson(res, 500, { error: String(e && e.message || e) }) }
       },
     },
     {

package/lib/python-setup.js

@@ -35,8 +35,9 @@ const MODEL_SPEC = {
   ],
 }
 
-/** venv 内 pip 依赖:与 bench 校准环境同源(transformers AutoTokenizer + onnxruntime;torch 为张量构造所需;池化/精度契约冻结自 R@5 0.925 基线,fastembed 注册表无 BGE-M3 且池化契约不同,不可用)。 */
-const PIP_DEPS = ['transformers', 'onnxruntime', 'torch']
+/** venv 内 pip 依赖:#19 实测(int8 档 encode_ids 全程 numpy+onnxruntime,tokenizer 走 Rust 快速路径)——基础集无 torch,venv 体积 ~400MB。GPU 推理开关追加 onnxruntime-gpu(CUDAExecutionProvider)。池化/精度契约冻结自 R@5 0.925 基线,fastembed 注册表无 BGE-M3 且池化契约不同,不可用。 */
+const PIP_DEPS_CPU = ['transformers', 'onnxruntime']
+const PIP_DEPS_GPU_EXTRA = ['onnxruntime-gpu']
 
 export function createPythonSetupPre(opts = {}) {
   const dshHomeOf = typeof opts.dshHome === 'function' ? opts.dshHome : () => opts.dshHome || path.join(homedir(), '.dsh')
@@ -110,7 +111,7 @@ export function createPythonSetupPre(opts = {}) {
     return {
       phase: st.phase, error: st.error,
       pythons: st.pythons, chosenPython: st.chosenPython,
-      venvOk: st.venvOk, depsOk: st.depsOk, modelReady: st.modelReady,
+      venvOk: st.venvOk, depsOk: st.depsOk, modelReady: st.modelReady, wantGpu: !!st.wantGpu,
       modelPath: modelPath(), venvPython: venvPython(),
       modelBytes: st.modelReady ? statSync(modelPath()).size : 0, modelExpectedBytes: MODEL_SPEC.bytes,
       dl: { ...st.dl },
@@ -136,14 +137,17 @@ export function createPythonSetupPre(opts = {}) {
     return snapshot()
   }
 
-  /** ③deps:venv 内 pip 装 transformers+onnxruntime+torch;官方源失败切清华镜像。 */
-  async function ensureDeps() {
+  /** ③deps:venv 内 pip 装 transformers+onnxruntime(CPU 基础集);opts.gpu=true 追加 onnxruntime-gpu(用户选 GPU 推理时)。官方源失败切清华镜像。 */
+  async function ensureDeps(opts2) {
+    const wantGpu = !!(opts2 && opts2.gpu)
     if (!st.venvOk) { st.error = 'venv 未就绪,先执行 venv 步骤'; st.phase = 'error'; return snapshot() }
     st.phase = 'deps'
+    st.wantGpu = wantGpu
+    const deps = PIP_DEPS_CPU.concat(wantGpu ? PIP_DEPS_GPU_EXTRA : [])
     const runPip = async (extra) => {
       try {
-        const r = await execFileP(venvPython(), ['-m', 'pip', 'install', '--quiet'].concat(PIP_DEPS).concat(extra || []),
-          { timeout: 600000, windowsHide: true, maxBuffer: 4 * 1024 * 1024 })
+        const r = await execFileP(venvPython(), ['-m', 'pip', 'install', '--quiet'].concat(deps).concat(extra || []),
+          { timeout: 900000, windowsHide: true, maxBuffer: 4 * 1024 * 1024 })
         return { ok: true }
       } catch (e) { return { ok: false, tail: String((e && (e.stderr || e.message)) || e).slice(-300) } }
     }
@@ -151,7 +155,18 @@ export function createPythonSetupPre(opts = {}) {
     if (!r.ok) r = await runPip(['-i', 'https://pypi.tuna.tsinghua.edu.cn/simple'])
     st.depsOk = !!r.ok
     if (!r.ok) { st.error = '依赖安装失败: ' + (r.tail || ''); st.phase = 'error'; return snapshot() }
-    diagOf('python-setup: deps installed into venv')
+    // GPU 偏好写入 embedding-config.json(worker 读 config['gpu'] 选 CUDA/CPU provider;read-modify-write 不覆盖既有键)
+    try {
+      const cfgDir = path.join(dshHomeOf(), 'memory', 'semantic')
+      mkdirSync(cfgDir, { recursive: true })
+      const cfgPath = path.join(cfgDir, 'embedding-config.json')
+      let cfg = {}
+      try { cfg = JSON.parse(await readFile(cfgPath, 'utf8')) } catch (_) {}
+      if (!cfg || typeof cfg !== 'object') cfg = {}
+      cfg.gpu = wantGpu
+      await writeFile(cfgPath, JSON.stringify(cfg, null, 2), 'utf8')
+    } catch (eCfg) { diagOf('python-setup: gpu pref write failed: ' + String(eCfg && eCfg.message || eCfg)) }
+    diagOf('python-setup: deps installed into venv (gpu=' + wantGpu + ')')
     return snapshot()
   }
 

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@a9i5k4/dsh-auto-memory",
   "description": "Proactive associative memory for DSH: zero-prompt recall injected before the model speaks, three-layer auto-consolidation, skill crystallization, and Astra-style context management - handoff ledgers, PLAN whiteboard, water-level sensing. Local-first, model-agnostic, zero deps. 主动联想记忆+Astra 式上下文管理:自动唤回/自动沉淀/技能固化/交接账本与白板跨窗口续命/水位感知。",
-  "version": "2.1.7",
+  "version": "2.1.8",
   "type": "module",
   "main": "lib/index.js",
   "exports": {