nanoflow-llm 0.6.3

NanoFlow LLM Universal Bindings (Self-Contained)

NanoFlow: The "Blind" Inference Engine 🧅🚀

NanoFlow is a lightweight, high-performance LLM inference engine optimized for Zero-Copy interoperability and Layer-Wise execution.

🌟 Key Features

• Universal Format Support: Decoupled Compute Engine. Load weights in Python (Safetensors, PyTorch, GGUF, ONNX) and pass raw pointers to C++.
• Zero-Copy Adapter: No data duplication. C++ operates directly on Python-allocated memory.
• "Onion System" Architecture:
  • Memory Isolation: Fixed "Hot Buffer" for active layer weights.
  • Temporal Segmentation: Streams layers just-in-time, allowing 70B models to run on consumer hardware.
• Cross-Platform: Native support for Windows (MSVC) and Linux (GCC/Clang).
• Powered by GGML: Utilizes the raw math performance of the GGML library.

📦 Installation

pip install nanoflow-llm

🚀 Usage (Universal Adapter)

import torch
from safetensors.torch import load_file
import nanoflow_ext

# 1. Load Data (Python)
tensors = load_file("model.safetensors")
tensor_data = tensors["layer1.weight"]

# 2. Pass Pointer (Zero-Copy)
engine = nanoflow_ext.NanoFlowEngine(ram_limit_mb=256)
engine.set_weights(
    name="layer1.weight", 
    raw_pointer=tensor_data.data_ptr(), 
    rows=tensor_data.shape[0], 
    cols=tensor_data.shape[1], 
    ggml_type_id=0 # F32
)

# 3. Compute (C++)
engine.compute("layer1.weight")