open-agents-ai 0.185.30 → 0.185.31

package/dist/index.js

@@ -41270,6 +41270,8 @@ __export(personaplex_exports, {
   clonePersonaPlexVoice: () => clonePersonaPlexVoice,
   detectPersonaPlexCapability: () => detectPersonaPlexCapability,
   getPersonaPlexWSUrl: () => getPersonaPlexWSUrl,
+  getWeightRepoInfo: () => getWeightRepoInfo,
+  getWeightTier: () => getWeightTier,
   installPersonaPlex: () => installPersonaPlex,
   isPersonaPlexInstalled: () => isPersonaPlexInstalled,
   isPersonaPlexRunning: () => isPersonaPlexRunning,
@@ -41284,6 +41286,13 @@ import { join as join54, dirname as dirname18 } from "node:path";
 import { homedir as homedir13 } from "node:os";
 import { execSync as execSync27, spawn as spawn19 } from "node:child_process";
 import { fileURLToPath as fileURLToPath11 } from "node:url";
+function selectWeightTier(vramGB) {
+  if (vramGB >= 48)
+    return "original";
+  if (vramGB >= 16)
+    return "nf4";
+  return "turbo2bit";
+}
 function detectPersonaPlexCapability() {
   try {
     const nvsmi = execSync27("nvidia-smi --query-gpu=name,memory.total --format=csv,noheader,nounits", {
@@ -41292,24 +41301,23 @@ function detectPersonaPlexCapability() {
       stdio: "pipe"
     }).trim();
     if (!nvsmi) {
-      return { supported: false, reason: "No NVIDIA GPU detected", gpuName: "", vramGB: 0 };
+      return { supported: false, reason: "No NVIDIA GPU detected", gpuName: "", vramGB: 0, weightTier: "turbo2bit", needsHfToken: false };
     }
     const [gpuName, vramMB] = nvsmi.split("\n")[0].split(", ");
-    const vramGB = parseInt(vramMB ?? "0", 10) / 1024;
-    if (vramGB < 16) {
-      const isJetson = /orin|tegra|jetson/i.test(gpuName ?? "");
-      if (isJetson) {
-        try {
-          const memInfo = execSync27("grep MemTotal /proc/meminfo", { encoding: "utf8", timeout: 3e3, stdio: "pipe" });
-          const memKB = parseInt(memInfo.match(/(\d+)/)?.[1] ?? "0", 10);
-          const totalGB = memKB / 1024 / 1024;
-          if (totalGB >= 32) {
-            return { supported: true, reason: `Jetson unified memory (${totalGB.toFixed(0)}GB total)`, gpuName: gpuName ?? "", vramGB: totalGB };
-          }
-        } catch {
-        }
+    let vramGB = parseInt(vramMB ?? "0", 10) / 1024;
+    const isJetson = /orin|tegra|jetson/i.test(gpuName ?? "");
+    if (isJetson) {
+      try {
+        const memInfo = execSync27("grep MemTotal /proc/meminfo", { encoding: "utf8", timeout: 3e3, stdio: "pipe" });
+        const memKB = parseInt(memInfo.match(/(\d+)/)?.[1] ?? "0", 10);
+        const totalGB = memKB / 1024 / 1024;
+        if (totalGB > vramGB)
+          vramGB = totalGB;
+      } catch {
       }
-      return { supported: false, reason: `GPU has ${vramGB.toFixed(1)}GB VRAM (need \u226516GB)`, gpuName: gpuName ?? "", vramGB };
+    }
+    if (vramGB < 8) {
+      return { supported: false, reason: `GPU has ${vramGB.toFixed(1)}GB VRAM (need \u22658GB for 2-bit weights)`, gpuName: gpuName ?? "", vramGB, weightTier: "turbo2bit", needsHfToken: false };
     }
     try {
       execSync27('python3 -c "import torch; assert torch.cuda.is_available()"', {
@@ -41317,11 +41325,22 @@ function detectPersonaPlexCapability() {
         stdio: "pipe"
       });
     } catch {
-      return { supported: false, reason: "PyTorch CUDA not available", gpuName: gpuName ?? "", vramGB };
+      const tier2 = selectWeightTier(vramGB);
+      return { supported: false, reason: "PyTorch CUDA not available", gpuName: gpuName ?? "", vramGB, weightTier: tier2, needsHfToken: WEIGHT_REPOS[tier2].needsToken };
     }
-    return { supported: true, reason: "OK", gpuName: gpuName ?? "", vramGB };
+    const tier = selectWeightTier(vramGB);
+    const hasHfToken = !!(process.env["HF_TOKEN"] || process.env["HUGGING_FACE_HUB_TOKEN"]);
+    const effectiveTier = tier === "original" && !hasHfToken ? "nf4" : tier;
+    return {
+      supported: true,
+      reason: `OK \u2014 ${effectiveTier} weights (${WEIGHT_REPOS[effectiveTier].sizeGB}GB)`,
+      gpuName: gpuName ?? "",
+      vramGB,
+      weightTier: effectiveTier,
+      needsHfToken: WEIGHT_REPOS[effectiveTier].needsToken
+    };
   } catch {
-    return { supported: false, reason: "nvidia-smi not found", gpuName: "", vramGB: 0 };
+    return { supported: false, reason: "nvidia-smi not found", gpuName: "", vramGB: 0, weightTier: "turbo2bit", needsHfToken: false };
   }
 }
 function isPersonaPlexRunning() {
@@ -41348,7 +41367,19 @@ function getPersonaPlexWSUrl() {
 function isPersonaPlexInstalled() {
   return existsSync37(join54(PERSONAPLEX_DIR, "model_ready"));
 }
-async function installPersonaPlex(onInfo) {
+function getWeightTier() {
+  const tierFile = join54(PERSONAPLEX_DIR, "weight_tier");
+  if (existsSync37(tierFile)) {
+    const saved = readFileSync28(tierFile, "utf8").trim();
+    if (saved in WEIGHT_REPOS)
+      return saved;
+  }
+  return detectPersonaPlexCapability().weightTier;
+}
+function getWeightRepoInfo(tier) {
+  return WEIGHT_REPOS[tier];
+}
+async function installPersonaPlex(onInfo, weightTier) {
   const log = onInfo ?? (() => {
   });
   mkdirSync15(PERSONAPLEX_DIR, { recursive: true });
@@ -41456,12 +41487,64 @@ async function installPersonaPlex(onInfo) {
     execSync27(`"${pip}" install --quiet pyloudnorm noisereduce torchaudio`, { timeout: 12e4, stdio: "pipe" });
   } catch {
   }
-  log("PersonaPlex installed. Model will download on first launch (~14GB).");
-  if (isAarch64) {
-    log("ARM64: On first run, weights will load in INT4 mode for real-time performance.");
+  const tier = weightTier ?? detectPersonaPlexCapability().weightTier;
+  const repoInfo = WEIGHT_REPOS[tier];
+  log(`Weight tier: ${tier} (${repoInfo.sizeGB}GB) \u2014 ${repoInfo.needsToken ? "requires HF_TOKEN" : "public, no token needed"}`);
+  log(`Downloading PersonaPlex weights (${repoInfo.sizeGB}GB)...`);
+  try {
+    const tokenArg = repoInfo.needsToken ? "" : "--token ''";
+    const dlCmd = `"${python}" -c "from huggingface_hub import hf_hub_download; f=hf_hub_download('${repoInfo.repo}', '${repoInfo.file}'${repoInfo.needsToken ? "" : ", token=False"}); print(f)"`;
+    const weightPath = execSync27(dlCmd, {
+      encoding: "utf8",
+      timeout: 6e5,
+      stdio: "pipe",
+      env: { ...process.env }
+    }).trim();
+    log(`Weights downloaded: ${repoInfo.file}`);
+    if (tier !== "original") {
+      log("Downloading Mimi codec and tokenizer...");
+      try {
+        const hasToken = !!(process.env["HF_TOKEN"] || process.env["HUGGING_FACE_HUB_TOKEN"]);
+        if (hasToken) {
+          execSync27(`"${python}" -c "from huggingface_hub import hf_hub_download; hf_hub_download('nvidia/personaplex-7b-v1', 'tokenizer_spm_32k_3.model'); hf_hub_download('nvidia/personaplex-7b-v1', 'tokenizer-e351c8d8-checkpoint125.safetensors')"`, {
+            timeout: 3e5,
+            stdio: "pipe"
+          });
+          log("Codec + tokenizer downloaded.");
+        } else {
+          log("Note: Mimi codec needs HF_TOKEN on first run (set HF_TOKEN env var).");
+          log("Weights themselves are public \u2014 no token needed for the model.");
+        }
+      } catch {
+      }
+    }
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    if (repoInfo.needsToken && /401|403|gated|unauthorized/i.test(msg)) {
+      log(`HF_TOKEN required for ${tier} weights. Set HF_TOKEN or accept license at https://huggingface.co/${repoInfo.repo}`);
+      if (tier === "original") {
+        log("Auto-downgrading to INT4 weights (no token required)...");
+        const nf4 = WEIGHT_REPOS["nf4"];
+        try {
+          execSync27(`"${python}" -c "from huggingface_hub import hf_hub_download; hf_hub_download('${nf4.repo}', '${nf4.file}', token=False)"`, {
+            timeout: 6e5,
+            stdio: "pipe"
+          });
+          writeFileSync16(join54(PERSONAPLEX_DIR, "weight_tier"), "nf4");
+          log(`Downloaded INT4 weights instead (${nf4.sizeGB}GB, public).`);
+        } catch {
+          log("Weight download failed.");
+          return false;
+        }
+      }
+    } else {
+      log(`Weight download failed: ${msg}`);
+      log("Weights will download on first server launch.");
+    }
   }
+  writeFileSync16(join54(PERSONAPLEX_DIR, "weight_tier"), tier);
   writeFileSync16(join54(PERSONAPLEX_DIR, "model_ready"), (/* @__PURE__ */ new Date()).toISOString());
-  log("PersonaPlex installed successfully.");
+  log(`PersonaPlex installed (${tier} tier). Use /call to start voice session.`);
   return true;
 }
 async function startPersonaPlexDaemon(onInfo) {
@@ -41784,10 +41867,11 @@ async function autoSetupPersonaPlex(onInfo) {
     log(`PersonaPlex not available: ${caps.reason}`);
     return null;
   }
-  log(`GPU: ${caps.gpuName} (${caps.vramGB.toFixed(0)}GB) \u2014 PersonaPlex compatible`);
+  const tierInfo = WEIGHT_REPOS[caps.weightTier];
+  log(`GPU: ${caps.gpuName} (${caps.vramGB.toFixed(0)}GB) \u2192 ${caps.weightTier} weights (${tierInfo.sizeGB}GB${caps.needsHfToken ? "" : ", no HF token needed"})`);
   if (!isPersonaPlexInstalled()) {
     log("Installing PersonaPlex (first time setup)...");
-    const ok = await installPersonaPlex(log);
+    const ok = await installPersonaPlex(log, caps.weightTier);
     if (!ok) {
       log("PersonaPlex installation failed.");
       return null;
@@ -41807,11 +41891,16 @@ async function autoSetupPersonaPlex(onInfo) {
   }
   return await startPersonaPlexDaemon(log);
 }
-var PERSONAPLEX_DIR, PID_FILE, PORT_FILE, LOG_FILE, CUSTOM_VOICES_DIR;
+var WEIGHT_REPOS, PERSONAPLEX_DIR, PID_FILE, PORT_FILE, LOG_FILE, CUSTOM_VOICES_DIR;
 var init_personaplex = __esm({
   "packages/cli/dist/tui/personaplex.js"() {
     "use strict";
     init_render();
+    WEIGHT_REPOS = {
+      original: { repo: "nvidia/personaplex-7b-v1", file: "model.safetensors", sizeGB: 15.6, needsToken: true },
+      nf4: { repo: "cudabenchmarktest/personaplex-7b-nf4", file: "model-nf4.safetensors", sizeGB: 4.1, needsToken: false },
+      turbo2bit: { repo: "cudabenchmarktest/personaplex-7b-turbo2bit", file: "model-turbo2bit.safetensors", sizeGB: 2.1, needsToken: false }
+    };
     PERSONAPLEX_DIR = join54(homedir13(), ".open-agents", "voice", "personaplex");
     PID_FILE = join54(PERSONAPLEX_DIR, "daemon.pid");
     PORT_FILE = join54(PERSONAPLEX_DIR, "daemon.port");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.185.30",
+  "version": "0.185.31",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",