NNI SDK Documentation

Technical reference for the NNI compiled memory-mapped interconnect stack.

1. Quickstart (Python FFI)

Installation

pip install nni-sdk

Establish a named shared memory block backed by the system paging file. Both writing and reading processes must connect to the exact same string namespace identifier.

1_init.py

import nni_sdk
# Initialize the named RAM clipboard segment
slate = nni_sdk.SharedSlate(name="nni_shared_slate")

2. Hardware Architecture & Masking

Compress high-dimensional model weight tensors into an optimized, 64-byte cache-line aligned packet. Values below the importance threshold are automatically zero-skipped.

2_compress.py

# Compresses elements, packing up to 9 active float values
packet = nni_sdk.Compressor.compress_sparse(
    floats=my_raw_floats,
    threshold=0.5
)

3. Secure Enclave (TEE) Encryption

Encrypt and sign your compressed packet inside your processor's isolated Trusted Execution Environment using ChaCha20-Poly1305 AEAD cryptography.

3_encrypt.py

# Returns secure (nonce, ciphertext, tag) envelope
nonce, ciphertext, tag = nni_sdk.SecureEnclave.encrypt(packet)

4. Framework Integrations

PyTorch Distributed Hook

Drop-in replacement for torch.distributed.send targeting multi-GPU clusters.

Hugging Face Transformers Bridge

Direct integration with transformers.pipeline for localized KV-Cache streaming.

TensorRT Optimization

NPU-Direct API pushing verified activations straight to NVIDIA hardware registers.

Knowledge Base