Skip to main content

​
What compilation does

thx.compile(model, "output.thx") performs three things:
  1. Resolution. Runtime hyperparameters (tau_mem, dt) are resolved to discrete constants (alpha = exp(-dt / tau_mem)). Once resolved, the constant is in the artifact; the hyperparameter is not.
  2. Validation. The compiler checks the architecture against the target’s capability descriptor. It rejects models that cannot be represented: missing required parameters, unsupported layer types, weight shapes inconsistent with the declared architecture.
  3. Sealing. The artifact is written with a CRC32 integrity hash. Any post-compilation modification of the .thx file will cause thrindex run and thrindex doctor to report [E0004] artifact CRC32 mismatch.

​
What is resolved at compile time

The Python model stores tau_mem. The artifact stores alpha. You can reconstruct tau_mem from alpha and dt_ms, but there is no round-trip guarantee for tau_syn if it was not explicitly set (the default means no synaptic filter, which is distinct from tau_syn = dt).

​
Compile targets

By default, thx.compile targets the behavioral simulator (sim). To compile for a specific hardware backend:
The compiler loads the loihi2 capability descriptor and applies target-specific transformations:
  • Retiming: if the target’s native dt differs from the model’s dt, weights and time constants are adjusted.
  • Quantization advisory: if the target is fixed-point, the compiler reports the expected quantization error bracket based on CONFORMANCE_ENVELOPE_V0.
  • Delay negotiation: if the model specifies delays and the target cannot represent them natively, the compiler inserts buffer stages or rejects.

​
Compile-time advisories

The compiler prints advisories — not errors — for decisions that succeed but require attention:
Advisories are informational. The compilation succeeds. They appear in thrindex compile --verbose.

​
What is not locked at compile time

  • Input size (T and n_features) is locked in the artifact. You cannot feed a compiled 100×700 model a 50×700 input without recompiling.
  • Batch size is not locked — the artifact runs on any batch.
  • Seed is not locked — thrindex run --seed N controls the RNG at inference time.

​
The format version

Every artifact carries a format version (m3 as of THRINDEX 0.2). The version controls the deserializer: a m3 artifact is rejected by an engine that only supports m2. The CLI prints the format version in the run header. Format versions are additive: m3 adds fields that m2 did not have, but any valid m2 artifact can be represented as a m3 artifact by filling new fields with defaults. See Concepts: Artifact Format for the complete field layout.