open-agents-ai 0.185.31 → 0.185.32

package/dist/index.js

@@ -41281,7 +41281,7 @@ __export(personaplex_exports, {
   startPersonaPlexDaemon: () => startPersonaPlexDaemon,
   stopPersonaPlex: () => stopPersonaPlex
 });
-import { existsSync as existsSync37, writeFileSync as writeFileSync16, readFileSync as readFileSync28, mkdirSync as mkdirSync15, copyFileSync as copyFileSync2, readdirSync as readdirSync11 } from "node:fs";
+import { existsSync as existsSync37, writeFileSync as writeFileSync16, readFileSync as readFileSync28, mkdirSync as mkdirSync15, copyFileSync as copyFileSync2, readdirSync as readdirSync11, statSync as statSync13 } from "node:fs";
 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";
@@ -41293,31 +41293,56 @@ function selectWeightTier(vramGB) {
     return "nf4";
   return "turbo2bit";
 }
+function detectJetson() {
+  try {
+    const model = readFileSync28("/proc/device-tree/model", "utf8").replace(/\0/g, "").trim();
+    if (/jetson|orin|tegra/i.test(model)) {
+      const memInfo = execSync27("grep MemTotal /proc/meminfo", { encoding: "utf8", timeout: 3e3, stdio: "pipe" });
+      const memKB = parseInt(memInfo.match(/(\d+)/)?.[1] ?? "0", 10);
+      return { isJetson: true, model, totalMemGB: memKB / 1024 / 1024 };
+    }
+  } catch {
+  }
+  return { isJetson: false, model: "", totalMemGB: 0 };
+}
 function detectPersonaPlexCapability() {
+  const fail = (reason) => ({
+    supported: false,
+    reason,
+    gpuName: "",
+    vramGB: 0,
+    weightTier: "turbo2bit",
+    needsHfToken: false
+  });
+  const jetson = detectJetson();
+  if (jetson.isJetson) {
+    const vramGB = jetson.totalMemGB;
+    if (vramGB < 8)
+      return { ...fail(`Jetson has ${vramGB.toFixed(0)}GB unified memory (need \u22658GB)`), gpuName: jetson.model, 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: `Jetson ${jetson.model} \u2014 ${effectiveTier} weights (${WEIGHT_REPOS[effectiveTier].sizeGB}GB)`,
+      gpuName: jetson.model,
+      vramGB,
+      weightTier: effectiveTier,
+      needsHfToken: WEIGHT_REPOS[effectiveTier].needsToken
+    };
+  }
   try {
     const nvsmi = execSync27("nvidia-smi --query-gpu=name,memory.total --format=csv,noheader,nounits", {
       encoding: "utf8",
       timeout: 5e3,
       stdio: "pipe"
     }).trim();
-    if (!nvsmi) {
-      return { supported: false, reason: "No NVIDIA GPU detected", gpuName: "", vramGB: 0, weightTier: "turbo2bit", needsHfToken: false };
-    }
+    if (!nvsmi)
+      return fail("No NVIDIA GPU detected");
     const [gpuName, vramMB] = nvsmi.split("\n")[0].split(", ");
-    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 {
-      }
-    }
+    const vramGB = parseInt(vramMB ?? "0", 10) / 1024;
     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 };
+      return { ...fail(`GPU has ${vramGB.toFixed(1)}GB VRAM (need \u22658GB)`), gpuName: gpuName ?? "", vramGB };
     }
     try {
       execSync27('python3 -c "import torch; assert torch.cuda.is_available()"', {
@@ -41326,7 +41351,7 @@ function detectPersonaPlexCapability() {
       });
     } catch {
       const tier2 = selectWeightTier(vramGB);
-      return { supported: false, reason: "PyTorch CUDA not available", gpuName: gpuName ?? "", vramGB, weightTier: tier2, needsHfToken: WEIGHT_REPOS[tier2].needsToken };
+      return { ...fail("PyTorch CUDA not available"), gpuName: gpuName ?? "", vramGB, weightTier: tier2, needsHfToken: WEIGHT_REPOS[tier2].needsToken };
     }
     const tier = selectWeightTier(vramGB);
     const hasHfToken = !!(process.env["HF_TOKEN"] || process.env["HUGGING_FACE_HUB_TOKEN"]);
@@ -41340,7 +41365,7 @@ function detectPersonaPlexCapability() {
       needsHfToken: WEIGHT_REPOS[effectiveTier].needsToken
     };
   } catch {
-    return { supported: false, reason: "nvidia-smi not found", gpuName: "", vramGB: 0, weightTier: "turbo2bit", needsHfToken: false };
+    return fail("No NVIDIA GPU detected (nvidia-smi not found)");
   }
 }
 function isPersonaPlexRunning() {
@@ -41383,11 +41408,20 @@ async function installPersonaPlex(onInfo, weightTier) {
   const log = onInfo ?? (() => {
   });
   mkdirSync15(PERSONAPLEX_DIR, { recursive: true });
+  let arch2 = "";
+  try {
+    arch2 = execSync27("uname -m", { encoding: "utf8", timeout: 3e3, stdio: "pipe" }).trim();
+  } catch {
+  }
+  const isAarch64 = arch2 === "aarch64" || arch2 === "arm64";
+  if (isAarch64)
+    log(`Detected ARM64 platform (${arch2}) \u2014 Jetson/ARM install path`);
   const venvDir = join54(PERSONAPLEX_DIR, "venv");
   if (!existsSync37(venvDir)) {
     log("Creating Python virtual environment...");
     try {
-      execSync27(`python3 -m venv "${venvDir}"`, { timeout: 6e4, stdio: "pipe" });
+      const ssp = isAarch64 ? " --system-site-packages" : "";
+      execSync27(`python3 -m venv${ssp} "${venvDir}"`, { timeout: 6e4, stdio: "pipe" });
     } catch (err) {
       log(`Failed to create venv: ${err instanceof Error ? err.message : String(err)}`);
       return false;
@@ -41395,14 +41429,6 @@ async function installPersonaPlex(onInfo, weightTier) {
   }
   const pip = process.platform === "win32" ? join54(venvDir, "Scripts", "pip.exe") : join54(venvDir, "bin", "pip");
   const python = process.platform === "win32" ? join54(venvDir, "Scripts", "python.exe") : join54(venvDir, "bin", "python3");
-  let arch2 = "";
-  try {
-    arch2 = execSync27("uname -m", { encoding: "utf8", timeout: 3e3, stdio: "pipe" }).trim();
-  } catch {
-  }
-  const isAarch64 = arch2 === "aarch64" || arch2 === "arm64";
-  if (isAarch64)
-    log(`Detected ARM64 platform (${arch2}) \u2014 Jetson/ARM install path`);
   log("Checking system dependencies (libopus)...");
   try {
     if (process.platform === "linux") {
@@ -41566,7 +41592,39 @@ async function startPersonaPlexDaemon(onInfo) {
   const venvPython2 = process.platform === "win32" ? join54(PERSONAPLEX_DIR, "venv", "Scripts", "python.exe") : join54(PERSONAPLEX_DIR, "venv", "bin", "python3");
   const sslDir = join54(PERSONAPLEX_DIR, "ssl");
   mkdirSync15(sslDir, { recursive: true });
-  log("Starting PersonaPlex daemon (loading ~7B model)...");
+  const tier = getWeightTier();
+  const repoInfo = WEIGHT_REPOS[tier];
+  const extraArgs = [];
+  if (tier !== "original") {
+    log(`Weight tier: ${tier} (${repoInfo.sizeGB}GB) \u2014 dequantizing to bf16 cache...`);
+    const dequantScript = join54(PERSONAPLEX_DIR, "dequant-loader.py");
+    const cachedBf16 = join54(PERSONAPLEX_DIR, "model-bf16-cache.safetensors");
+    if (!existsSync37(dequantScript)) {
+      const shipped = getShippedVoicesDir();
+      if (shipped) {
+        const src = join54(shipped, "dequant-loader.py");
+        if (existsSync37(src))
+          copyFileSync2(src, dequantScript);
+      }
+    }
+    try {
+      const weightPath = execSync27(`"${venvPython2}" -c "from huggingface_hub import hf_hub_download; print(hf_hub_download('${repoInfo.repo}', '${repoInfo.file}'${repoInfo.needsToken ? "" : ", token=False"}))"`, { encoding: "utf8", timeout: 3e4, stdio: "pipe" }).trim();
+      if (existsSync37(dequantScript) && existsSync37(weightPath)) {
+        try {
+          execSync27(`"${venvPython2}" "${dequantScript}" --input "${weightPath}" --output "${cachedBf16}"`, { timeout: 3e5, stdio: "pipe" });
+          if (existsSync37(cachedBf16)) {
+            extraArgs.push("--moshi-weight", cachedBf16);
+            log(`Using dequantized cache: ${(statSync13(cachedBf16).size / 1024 ** 3).toFixed(1)}GB`);
+          }
+        } catch (e) {
+          log(`Dequantization failed \u2014 server will try to load original weights`);
+        }
+      }
+    } catch {
+      log(`Weight file not found \u2014 server will download on first run`);
+    }
+  }
+  log(`Starting PersonaPlex daemon (${tier} tier)...`);
   const child = spawn19(venvPython2, [
     "-m",
     "moshi.server",
@@ -41577,7 +41635,8 @@ async function startPersonaPlexDaemon(onInfo) {
     "--ssl",
     sslDir,
     "--device",
-    "cuda"
+    "cuda",
+    ...extraArgs
   ], {
     stdio: ["ignore", "pipe", "pipe"],
     detached: true,
@@ -41664,8 +41723,7 @@ function listPersonaPlexVoices() {
   }
   if (existsSync37(CUSTOM_VOICES_DIR)) {
     try {
-      const { readdirSync: readdirSync24 } = __require("node:fs");
-      for (const f of readdirSync24(CUSTOM_VOICES_DIR)) {
+      for (const f of readdirSync11(CUSTOM_VOICES_DIR)) {
         if (f.endsWith(".pt")) {
           const name = f.replace(/\.pt$/, "");
           voices.push({ name, type: "custom", path: join54(CUSTOM_VOICES_DIR, f) });
@@ -45416,7 +45474,7 @@ __export(voice_exports, {
   registerCustomOnnxModel: () => registerCustomOnnxModel,
   resetNarrationContext: () => resetNarrationContext
 });
-import { existsSync as existsSync42, mkdirSync as mkdirSync18, writeFileSync as writeFileSync19, readFileSync as readFileSync31, unlinkSync as unlinkSync9, readdirSync as readdirSync12, renameSync, statSync as statSync13 } from "node:fs";
+import { existsSync as existsSync42, mkdirSync as mkdirSync18, writeFileSync as writeFileSync19, readFileSync as readFileSync31, unlinkSync as unlinkSync9, readdirSync as readdirSync12, renameSync, statSync as statSync14 } from "node:fs";
 import { join as join58, dirname as dirname19 } from "node:path";
 import { homedir as homedir15, tmpdir as tmpdir9, platform as platform3 } from "node:os";
 import { execSync as execSync30, spawn as nodeSpawn } from "node:child_process";
@@ -46550,7 +46608,7 @@ var init_voice = __esm({
           const p = join58(dir, f);
           let size = 0;
           try {
-            size = statSync13(p).size;
+            size = statSync14(p).size;
           } catch {
           }
           return {
@@ -48166,7 +48224,7 @@ Error: ${err instanceof Error ? err.message : String(err)}`);
 // packages/cli/dist/tui/commands.js
 import * as nodeOs from "node:os";
 import { execSync as nodeExecSync } from "node:child_process";
-import { existsSync as existsSync43, readFileSync as readFileSync32, writeFileSync as writeFileSync20, mkdirSync as mkdirSync19, readdirSync as readdirSync13, statSync as statSync14, rmSync } from "node:fs";
+import { existsSync as existsSync43, readFileSync as readFileSync32, writeFileSync as writeFileSync20, mkdirSync as mkdirSync19, readdirSync as readdirSync13, statSync as statSync15, rmSync } from "node:fs";
 import { join as join59 } from "node:path";
 function safeLog(text) {
   if (isNeovimActive()) {
@@ -48979,7 +49037,7 @@ async function handleSlashCommand(input, ctx) {
           ipfsFiles = files.length;
           for (const f of files) {
             try {
-              ipfsBytes += statSync14(join59(ipfsLocalDir, f)).size;
+              ipfsBytes += statSync15(join59(ipfsLocalDir, f)).size;
             } catch {
             }
           }
@@ -48993,7 +49051,7 @@ async function handleSlashCommand(input, ctx) {
               else {
                 heliaBlocks++;
                 try {
-                  heliaBytes += statSync14(join59(dir, entry.name)).size;
+                  heliaBytes += statSync15(join59(dir, entry.name)).size;
                 } catch {
                 }
               }
@@ -49086,7 +49144,7 @@ async function handleSlashCommand(input, ctx) {
           const count = memStore.count();
           lines.push(`
   ${c2.bold("Structured Memory (SQLite)")}`);
-          lines.push(`    Memories: ${c2.bold(String(count))}   DB: ${c2.dim(formatFileSize(statSync14(dbPath).size))}`);
+          lines.push(`    Memories: ${c2.bold(String(count))}   DB: ${c2.dim(formatFileSize(statSync15(dbPath).size))}`);
           cDb(db);
         }
       } catch {
@@ -49117,7 +49175,7 @@ async function handleSlashCommand(input, ctx) {
               walkStorage(full, subCat);
             } else {
               try {
-                const sz = statSync14(full).size;
+                const sz = statSync15(full).size;
                 totalBytes += sz;
                 if (!categories[category])
                   categories[category] = { files: 0, bytes: 0 };
@@ -49418,7 +49476,7 @@ async function handleSlashCommand(input, ctx) {
           const caps = detectPersonaPlexCapability2();
           if (!caps.supported) {
             renderWarning(`PersonaPlex not available: ${caps.reason}`);
-            renderInfo("Requirements: NVIDIA GPU with \u226516GB VRAM (RTX 3090/4090/A100+), CUDA 12.1+, PyTorch");
+            renderInfo("Requirements: NVIDIA GPU with \u22658GB VRAM (RTX 3060+, Jetson AGX Orin), CUDA, PyTorch");
             return "handled";
           }
           renderInfo(`GPU: ${caps.gpuName} (${caps.vramGB.toFixed(0)}GB VRAM) \u2014 PersonaPlex compatible \u2713`);
@@ -51087,7 +51145,7 @@ async function showCohereDashboard(ctx) {
                 const snapItems = snaps.slice(0, 20).map((f) => ({
                   key: f,
                   label: f.replace(".json", ""),
-                  detail: `${formatFileSize(statSync14(join59(snapDir, f)).size)}`
+                  detail: `${formatFileSize(statSync15(join59(snapDir, f)).size)}`
                 }));
                 if (snapItems.length > 0) {
                   await tuiSelect({
@@ -59364,7 +59422,7 @@ var init_tool_policy = __esm({
 });
 
 // packages/cli/dist/tui/telegram-bridge.js
-import { mkdirSync as mkdirSync25, existsSync as existsSync51, unlinkSync as unlinkSync11, readdirSync as readdirSync19, statSync as statSync15 } from "node:fs";
+import { mkdirSync as mkdirSync25, existsSync as existsSync51, unlinkSync as unlinkSync11, readdirSync as readdirSync19, statSync as statSync16 } from "node:fs";
 import { join as join68, resolve as resolve30 } from "node:path";
 import { writeFile as writeFileAsync } from "node:fs/promises";
 function convertMarkdownToTelegramHTML(md) {
@@ -71366,7 +71424,7 @@ __export(index_repo_exports, {
   indexRepoCommand: () => indexRepoCommand
 });
 import { resolve as resolve34 } from "node:path";
-import { existsSync as existsSync56, statSync as statSync16 } from "node:fs";
+import { existsSync as existsSync56, statSync as statSync17 } from "node:fs";
 import { cwd as cwd2 } from "node:process";
 async function indexRepoCommand(opts, _config) {
   const repoRoot = resolve34(opts.repoPath ?? cwd2());
@@ -71376,7 +71434,7 @@ async function indexRepoCommand(opts, _config) {
     printError(`Path does not exist: ${repoRoot}`);
     process.exit(1);
   }
-  const stat5 = statSync16(repoRoot);
+  const stat5 = statSync17(repoRoot);
   if (!stat5.isDirectory()) {
     printError(`Path is not a directory: ${repoRoot}`);
     process.exit(1);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.185.31",
+  "version": "0.185.32",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",

package/voices/personaplex/dequant-loader.py

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+"""
+dequant-loader.py — Pre-dequantize quantized PersonaPlex weights to bf16 cache.
+
+For NF4 (INT4) or TurboQuant 2-bit weights, dequantizes to a temporary
+bf16 safetensors file that moshi.server can load natively.
+
+Usage:
+  python dequant-loader.py --input model-nf4.safetensors --output /tmp/model-bf16.safetensors
+  python dequant-loader.py --input model-turbo2bit.safetensors --output /tmp/model-bf16.safetensors
+
+The output file can then be passed to moshi.server via --moshi-weight.
+"""
+
+import os, sys, math, time
+import torch
+from safetensors.torch import load_file, save_file
+
+NF2_CENTROIDS = torch.tensor([-1.5104, -0.4528, 0.4528, 1.5104])
+
+
+def fast_wht(x):
+    """Vectorized Walsh-Hadamard Transform."""
+    n = x.shape[-1]
+    h = 1
+    while h < n:
+        x_view = x.view(*x.shape[:-1], -1, 2, h)
+        a = x_view[..., 0, :].clone()
+        b = x_view[..., 1, :].clone()
+        x_view[..., 0, :] = a + b
+        x_view[..., 1, :] = a - b
+        x = x_view.reshape(*x.shape)
+        h *= 2
+    return x / math.sqrt(n)
+
+
+def detect_format(state):
+    """Detect if weights are NF4 (INT4), TurboQuant 2-bit, or plain."""
+    has_scales = any(k.endswith(".__scales__") for k in state)
+    has_packed = any(k.endswith(".packed") for k in state)
+    if has_packed:
+        return "turbo2bit"
+    if has_scales:
+        return "nf4"
+    return "plain"
+
+
+def dequant_nf4(state):
+    """Dequantize INT4 NF4 weights."""
+    result = {}
+    processed = set()
+
+    for name in list(state.keys()):
+        if name.endswith(".__scales__") or name.endswith(".__shape__") or name.endswith(".__numel__"):
+            continue
+        if name in processed:
+            continue
+
+        scales_key = f"{name}.__scales__"
+        if scales_key in state:
+            packed = state[name]
+            scales = state[scales_key].float()
+            shape = state[f"{name}.__shape__"].tolist()
+            numel = state[f"{name}.__numel__"].item()
+            group_size = 64
+
+            lo = (packed & 0x0F).to(torch.int8) - 8
+            hi = ((packed >> 4) & 0x0F).to(torch.int8) - 8
+            unpacked = torch.zeros(packed.numel() * 2, dtype=torch.float32)
+            unpacked[0::2] = lo.float()
+            unpacked[1::2] = hi.float()
+
+            n_groups = scales.numel()
+            groups = unpacked[:n_groups * group_size].reshape(n_groups, group_size)
+            deq = (groups * scales.unsqueeze(1)).reshape(-1)[:numel]
+
+            orig_shape = [s for s in shape if s > 0]
+            result[name] = deq.reshape(orig_shape).to(torch.bfloat16)
+            processed.add(name)
+        else:
+            result[name] = state[name].to(torch.bfloat16)
+            processed.add(name)
+
+    return result
+
+
+def dequant_turbo2bit(state):
+    """Dequantize TurboQuant 2-bit (NF2 + WHT) weights."""
+    result = {}
+    processed = set()
+
+    for name in list(state.keys()):
+        if any(name.endswith(f".{s}") for s in ["packed", "scales", "shape", "numel", "gs", "np2"]):
+            continue
+        if name in processed:
+            continue
+
+        packed_key = f"{name}.packed"
+        if packed_key in state:
+            gs = state[f"{name}.gs"].item()
+            gs_pow2 = state[f"{name}.np2"].item()
+            numel = state[f"{name}.numel"].item()
+            shape = [s for s in state[f"{name}.shape"].tolist() if s > 0]
+            scales = state[f"{name}.scales"].float()
+            packed = state[packed_key]
+            n_groups = scales.numel()
+
+            # Unpack 2-bit
+            p = packed.reshape(n_groups, gs // 4)
+            codes = torch.zeros(n_groups, gs, dtype=torch.long)
+            for i in range(4):
+                codes[:, i::4] = (p >> (2 * i)) & 0x03
+
+            dequant = NF2_CENTROIDS[codes]
+
+            # Inverse WHT
+            if gs_pow2 > gs:
+                dequant = torch.cat([dequant, torch.zeros(n_groups, gs_pow2 - gs)], dim=1)
+            dequant = fast_wht(dequant)
+            dequant = dequant[:, :gs]
+
+            dequant = dequant * scales.unsqueeze(1)
+            result[name] = dequant.reshape(-1)[:numel].reshape(shape).to(torch.bfloat16)
+            processed.add(name)
+        else:
+            result[name] = state[name].to(torch.bfloat16)
+            processed.add(name)
+
+    return result
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="Dequantize PersonaPlex weights to bf16")
+    parser.add_argument("--input", "-i", required=True, help="Quantized safetensors file")
+    parser.add_argument("--output", "-o", required=True, help="Output bf16 safetensors file")
+    parser.add_argument("--device", "-d", default="cpu", help="Device for dequantization")
+    args = parser.parse_args()
+
+    if not os.path.exists(args.input):
+        print(f"Error: {args.input} not found")
+        sys.exit(1)
+
+    # Skip if output already exists and is newer than input
+    if os.path.exists(args.output) and os.path.getmtime(args.output) > os.path.getmtime(args.input):
+        print(f"Cached: {args.output} is up to date")
+        sys.exit(0)
+
+    print(f"Loading {args.input}...")
+    t0 = time.time()
+    state = load_file(args.input, device=args.device)
+
+    fmt = detect_format(state)
+    print(f"Format: {fmt}")
+
+    if fmt == "nf4":
+        result = dequant_nf4(state)
+    elif fmt == "turbo2bit":
+        result = dequant_turbo2bit(state)
+    else:
+        print("Already plain bf16/fp16 — copying")
+        result = {k: v.to(torch.bfloat16) for k, v in state.items()}
+
+    t1 = time.time()
+    print(f"Dequantized {len(result)} tensors in {t1-t0:.1f}s")
+
+    print(f"Saving to {args.output}...")
+    save_file(result, args.output)
+    size_gb = os.path.getsize(args.output) / 1024**3
+    print(f"Done: {size_gb:.2f} GB")
+
+
+if __name__ == "__main__":
+    main()