node-qnn-llm

Provides native Node.js bindings to Qualcomm's QNN (Qualcomm Neural Network) SDK, enabling LLM inference execution directly on Snapdragon NPUs (Neural Processing Units) rather than CPU or GPU. The binding wraps QNN's C++ runtime APIs, allowing developers to load quantized LLM models (particularly Llama variants) and execute forward passes with hardware acceleration on compatible Snapdragon processors. This approach offloads computation to specialized silicon, reducing power consumption and latency compared to CPU-only inference.

Implements Llama-specific model loading logic that parses Llama weights, initializes the QNN computation graph, and provides tokenization via integrated or external tokenizer bindings. The capability handles model state initialization, weight quantization validation, and token encoding/decoding for Llama architectures specifically, bridging the gap between Llama model artifacts and QNN's generic tensor execution layer. Supports streaming token generation with proper context management.

Provides token-by-token generation with support for multiple sampling methods (temperature, top-k, top-p) to control output diversity and coherence. The implementation iteratively calls the QNN inference engine, applies sampling logic to the output logits, and yields tokens as they are generated, enabling real-time streaming responses. Supports early stopping conditions (EOS token detection, max length) and allows fine-grained control over generation parameters.

Handles allocation and lifecycle management of NPU memory buffers for model weights and inference activations, including validation that loaded models match QNN's quantization requirements (typically INT8 or lower precision). The binding tracks memory usage, prevents buffer overflows, and provides diagnostics for out-of-memory conditions. Includes utilities to verify model compatibility before attempting inference and to estimate memory footprint based on model size and quantization level.

Supports processing multiple prompts in a single inference batch to improve throughput and hardware utilization. The implementation groups prompts, pads sequences to uniform length, executes a single QNN forward pass over the batch, and unpacks results back to individual prompts. Enables efficient processing of multiple requests without sequential per-prompt overhead, though with latency-throughput tradeoffs depending on batch size and sequence length variance.

Implements in-memory model caching to avoid reloading weights from disk on every inference call, and provides hot-reload capability to swap model versions without stopping the inference service. The binding maintains a model registry, tracks reference counts, and coordinates transitions between model versions to ensure in-flight requests complete before unloading old models. Enables A/B testing different model versions and rapid iteration without service interruption.