omnius 1.0.0

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()