This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3

Developer Guides

These guides explain how Neuron Agentic Development skills work internally. Read them to understand the architecture, decision logic, and validation strategies that the agents use under the hood.

Deep dive: Port HuggingFace models to Neuron with the Autoport skill

How the Autoport skill analyzes model architectures, generates NxD Inference implementations, compiles to NEFF, and validates accuracy against the original model.

Deep dive: Port GPU-compatible models to the vLLM Neuron backend with the Autoport skill

How the Autoport vLLM Neuron skill researches model architectures, generates vLLM Neuron implementations for the Trainium2 backend, registers them, and validates logits and equivalence against the GPU-compatible reference implementation.

Deep dive: Validate model ports with the Equivalence skill

How the Equivalence skill compares numerical output between a reference model and its port. Covers both NxD Inference and vLLM Neuron targets.

Deep dive: Capture device and system profiles from a vLLM Neuron server with the Profiling skill

How the Profiling skill isolates the prefill and decode phases into separate device traces, manages the server lifecycle so a capture does not orphan device-holding workers, and records a reproducible input and output contract for the analysis skill.

Deep dive: Localize performance bottlenecks with the Profile Analysis skill

How the Profile Analysis skill reads a capture’s system and device traces together: session workload timing, roofline bounds, a collective breakdown, and a per-component attribution, and which of those outputs are measured rather than approximate.

Deep dive: Plan inference parallelism with the Parallelism Planner skill

How the Parallelism Planner skill turns a model config and hardware constants into a TP, CP, and DP recommendation using an analytic entitlement model, before you run on hardware.

Deep dive: Find the binding operator with the Hardware Entitlement skill

How the Hardware Entitlement skill computes per-operator compute and HBM floors for a chosen config, and attributes a deployed graph’s resident HBM back to those floors.

This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3