open-agents-ai 0.185.29 → 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", {
@@ -41292,12 +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) {
-      return { supported: false, reason: `GPU has ${vramGB.toFixed(1)}GB VRAM (need \u226516GB)`, gpuName: gpuName ?? "", vramGB };
+    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 {
+      }
+    }
+    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()"', {
@@ -41305,11 +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() {
@@ -41336,7 +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 });
@@ -41352,6 +41395,14 @@ async function installPersonaPlex(onInfo) {
   }
   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") {
@@ -41361,12 +41412,43 @@ async function installPersonaPlex(onInfo) {
     }
   } catch {
   }
+  if (isAarch64) {
+    log("ARM64: Checking Rust toolchain for sphn build...");
+    try {
+      execSync27("rustc --version", { timeout: 5e3, stdio: "pipe" });
+    } catch {
+      log("ARM64: Installing Rust toolchain (needed for sphn audio codec)...");
+      try {
+        execSync27("curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y", { timeout: 12e4, stdio: "pipe" });
+      } catch (e) {
+        log(`Rust install failed: ${e instanceof Error ? e.message : String(e)}`);
+        log("Install Rust manually: curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh");
+        return false;
+      }
+    }
+    try {
+      execSync27(`"${pip}" install --quiet maturin`, { timeout: 6e4, stdio: "pipe" });
+    } catch {
+    }
+  }
   log("Installing PersonaPlex (moshi package)...");
   const repoDir = join54(PERSONAPLEX_DIR, "personaplex-repo");
   try {
     if (!existsSync37(repoDir)) {
       execSync27(`git clone https://github.com/NVIDIA/personaplex.git "${repoDir}"`, { timeout: 12e4, stdio: "pipe" });
     }
+    if (isAarch64) {
+      log("ARM64: Building sphn from source (Opus codec bindings)...");
+      try {
+        const rustEnv = `export PATH="$HOME/.cargo/bin:$PATH" &&`;
+        execSync27(`${rustEnv} "${pip}" install --quiet --no-binary sphn sphn`, { timeout: 3e5, stdio: "pipe", shell: "/bin/bash" });
+        log("ARM64: sphn built successfully");
+      } catch (e) {
+        log(`ARM64: sphn build failed \u2014 ${e instanceof Error ? e.message : String(e)}`);
+        log("Ensure Rust, libopus-dev, and cmake are installed.");
+        return false;
+      }
+    }
     execSync27(`"${pip}" install --quiet "${join54(repoDir, "moshi")}/."`, { timeout: 3e5, stdio: "pipe" });
   } catch (err) {
     log(`Moshi install failed: ${err instanceof Error ? err.message : String(err)}`);
@@ -41394,9 +41476,75 @@ async function installPersonaPlex(onInfo) {
     }
   } catch {
   }
-  log("PersonaPlex installed. Model will download on first launch (~14GB).");
+  if (isAarch64) {
+    log("ARM64: Installing bitsandbytes for INT4 inference...");
+    try {
+      execSync27(`"${pip}" install --quiet bitsandbytes`, { timeout: 12e4, stdio: "pipe" });
+    } catch {
+    }
+  }
+  try {
+    execSync27(`"${pip}" install --quiet pyloudnorm noisereduce torchaudio`, { timeout: 12e4, stdio: "pipe" });
+  } catch {
+  }
+  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) {
@@ -41719,10 +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;
@@ -41742,11 +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": "0.185.29",
+  "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",

package/voices/personaplex/quantize-weights.py

@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+"""
+quantize-weights.py — Quantize PersonaPlex 7B weights to INT4 (NF4) for edge devices.
+
+Creates a ~3.5GB quantized checkpoint from the ~14GB bf16 weights.
+The quantized model runs 3-4x faster on memory-bandwidth-limited devices
+like Jetson AGX Orin while maintaining voice quality.
+
+Usage:
+  python quantize-weights.py [--device cuda] [--output personaplex-7b-nf4.safetensors]
+
+Requirements:
+  pip install bitsandbytes safetensors torch
+"""
+
+import argparse
+import os
+import sys
+import logging
+
+logging.basicConfig(level=logging.INFO, format="%(message)s")
+log = logging.getLogger(__name__)
+
+
+def quantize_model(device: str = "cuda", output_path: str = None):
+    """Quantize PersonaPlex 7B to NF4 (4-bit Normal Float)"""
+    import torch
+    from huggingface_hub import hf_hub_download
+    from safetensors.torch import load_file, save_file
+
+    hf_repo = "nvidia/personaplex-7b-v1"
+
+    # 1) Download original weights
+    log.info("Downloading PersonaPlex 7B weights...")
+    weight_path = hf_hub_download(hf_repo, "model.safetensors")
+    log.info(f"  Weights: {weight_path}")
+    log.info(f"  Size: {os.path.getsize(weight_path) / 1024**3:.1f} GB")
+
+    # 2) Load state dict
+    log.info("Loading state dict...")
+    state_dict = load_file(weight_path, device="cpu")
+    log.info(f"  Loaded {len(state_dict)} tensors")
+
+    # 3) Quantize each weight tensor to INT4 using block-wise NF4
+    try:
+        import bitsandbytes as bnb
+        from bitsandbytes.functional import quantize_nf4, dequantize_nf4
+        HAS_BNB = True
+    except ImportError:
+        HAS_BNB = False
+        log.info("  bitsandbytes not available — using manual INT4 quantization")
+
+    quantized_state = {}
+    quant_meta = {}  # Store quantization parameters for dequantization
+    total_original = 0
+    total_quantized = 0
+    skipped = 0
+
+    for name, tensor in state_dict.items():
+        original_bytes = tensor.numel() * tensor.element_size()
+        total_original += original_bytes
+
+        # Only quantize large weight matrices (≥1024 elements, 2D)
+        # Skip biases, norms, embeddings, small tensors
+        should_quantize = (
+            tensor.ndim >= 2
+            and tensor.numel() >= 1024
+            and not any(skip in name for skip in [
+                "norm", "bias", "embed", "positional", "rope",
+                "depformer_emb", "depformer_in",
+            ])
+        )
+
+        if not should_quantize:
+            quantized_state[name] = tensor.to(torch.float16).contiguous()
+            total_quantized += tensor.numel() * 2  # fp16
+            skipped += 1
+            continue
+
+        # Reshape to 2D for quantization
+        orig_shape = tensor.shape
+        flat = tensor.reshape(-1).float()
+
+        if HAS_BNB:
+            # Use bitsandbytes NF4 quantization
+            quant_tensor, quant_state = bnb.functional.quantize_4bit(
+                flat, quant_type="nf4", compress_statistics=True,
+            )
+            # Store the quantized bytes + metadata for reconstruction
+            quantized_state[name] = quant_tensor.contiguous()
+            quant_meta[f"{name}.__quant_state__"] = torch.tensor(
+                list(orig_shape) + [0] * (4 - len(orig_shape)),
+                dtype=torch.int64,
+            )
+            # Store absmax for dequantization
+            if hasattr(quant_state, 'absmax'):
+                quantized_state[f"{name}.__absmax__"] = quant_state.absmax.contiguous()
+            if hasattr(quant_state, 'quant_map'):
+                quantized_state[f"{name}.__quant_map__"] = quant_state.quant_map.contiguous()
+            total_quantized += quant_tensor.numel()
+        else:
+            # Manual symmetric INT4 quantization (no bitsandbytes)
+            # Block size 64 for good accuracy
+            block_size = 64
+            n_blocks = (flat.numel() + block_size - 1) // block_size
+            padded = torch.zeros(n_blocks * block_size)
+            padded[:flat.numel()] = flat
+
+            blocks = padded.reshape(n_blocks, block_size)
+            scales = blocks.abs().max(dim=1).values / 7.0  # INT4 range: -8 to 7
+            scales = scales.clamp(min=1e-8)
+
+            # Quantize to INT4 (stored as INT8 pairs)
+            quantized_blocks = torch.round(blocks / scales.unsqueeze(1)).clamp(-8, 7).to(torch.int8)
+
+            # Pack two INT4 values into one INT8
+            packed = torch.zeros(n_blocks, block_size // 2, dtype=torch.uint8)
+            for i in range(block_size // 2):
+                low = (quantized_blocks[:, 2 * i] + 8).to(torch.uint8)
+                high = (quantized_blocks[:, 2 * i + 1] + 8).to(torch.uint8)
+                packed[:, i] = low | (high << 4)
+
+            quantized_state[name] = packed.reshape(-1).contiguous()
+            quantized_state[f"{name}.__scales__"] = scales.to(torch.float16).contiguous()
+            quant_meta[f"{name}.__quant_state__"] = torch.tensor(
+                list(orig_shape) + [0] * (4 - len(orig_shape)) + [block_size, flat.numel()],
+                dtype=torch.int64,
+            )
+            total_quantized += packed.numel() + scales.numel() * 2
+
+    # Add metadata tensors
+    quantized_state.update(quant_meta)
+
+    # 4) Save quantized weights
+    if output_path is None:
+        output_path = os.path.join(os.path.dirname(weight_path), "model-nf4.safetensors")
+
+    log.info(f"\nSaving quantized weights to: {output_path}")
+    save_file(quantized_state, output_path)
+
+    final_size = os.path.getsize(output_path)
+    compression = total_original / max(final_size, 1)
+
+    log.info(f"\nQuantization complete!")
+    log.info(f"  Original: {total_original / 1024**3:.1f} GB (bf16)")
+    log.info(f"  Quantized: {final_size / 1024**3:.1f} GB (NF4)")
+    log.info(f"  Compression: {compression:.1f}x")
+    log.info(f"  Tensors quantized: {len(state_dict) - skipped}/{len(state_dict)}")
+    log.info(f"  Tensors kept fp16: {skipped} (norms, biases, embeddings)")
+    log.info(f"\nUse --quantized flag with PersonaPlex server for INT4 inference")
+
+    return output_path
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Quantize PersonaPlex 7B to INT4 NF4")
+    parser.add_argument("--device", default="cuda", help="Device for quantization")
+    parser.add_argument("--output", "-o", default=None, help="Output path for quantized weights")
+    args = parser.parse_args()
+
+    import torch
+    with torch.no_grad():
+        quantize_model(device=args.device, output_path=args.output)
+
+
+if __name__ == "__main__":
+    main()