neat-pcie-application-builder

Neat PCIe Application Builder

Overview

Build native host applications against the installed Neat PCIe application APIs. Treat the installed PCIe header, Python module, and packaged tutorials as the source of truth. This is a separate API from the Neat Library used inside the SDK or directly on a DevKit.

The expected input is an MPK model archive or a prepared LLiMa model directory and the requested C++ or Python host application behavior. Follow explicit user requirements when they are compatible with the installed public API; treat the defaults below as guidance.

Choose References

• Read references/source-of-truth.md when generating or reviewing API use, inspecting an installed release, or resolving version differences.
• For MPK model archives, use pcie::Model. Read references/model-lifecycle.md when implementing build, inference, pipelining, cleanup, or multiple-model behavior.
• For pcie::Model, read references/tensors-and-images.md when constructing inputs, consuming outputs, handling multiple inputs, or choosing tensor mode versus image mode.
• For pcie::Model, read references/model-options.md when configuring a card, queue, preprocessing, MLA-only mode, or object decode behavior.
• For prepared LLiMa model directories (LLM, VLM, or ASR), use pcie::genai::GenAIModel and read references/genai.md. Do not apply the MPK build or queue workflow.
• Read references/build-and-validation.md when building, testing, or reporting validation. Do not claim connected-card execution from compile or import checks alone.

Defaults

• For pcie::Model in C++, include <simaai/neat/pcie/Model.h> and use simaai::neat::pcie. For GenAI, include <simaai/neat/pcie/genai/GenAIModel.h> and use simaai::neat::pcie::genai.
• In Python, use pyneatpcie as pcie for Model or from pyneatpcie import genai for GenAI from the PCIe host Python environment.
• In pcie::Model tensor mode, inspect model.info() before allocating or naming model-ready inputs. In image mode, treat that information as the card-side preprocessing output contract, not the submitted image contract.
• For pcie::Model, enable mla_only only when the application owns the model's dtype conversion; then model.info() is the MLA's own contract and, for INT8 tensors, its quant parameters are the only valid source for the conversion.
• Use run() for ordinary request/response inference. Use pcie::Model push()/pull() for bounded pipelining, or GenAI stream() for incremental output and cancellation.
• For performance-sensitive pcie::Model tensor pipelines, prefer a bounded ring of reusable contiguous input buffers. Wrap C++ storage with Tensor::from_external() or Python arrays with Tensor.from_numpy(..., copy=False), and recycle a buffer only after its matching result is pulled. Use simpler owning or copying constructors when reuse and staging overhead do not matter.
• Use finite startup and inference timeouts in applications that must fail predictably.
• Close every opened model on normal and error paths. Prefer a Python context manager.