Phase 0: Architecture Map
Freeze the first implementation strategy before code changes: keep trainable support hls4ml-native, centered on config, graph/flow passes, layer attributes, backend templates, and narrow writer hooks.
Linked Tasksโ
- HLS4ML-001: trainable
HLSConfigschema. - HLS4ML-002: trainable
ModelGraphflow and validation. - HLS4ML-011: writer hooks.
- HLS4ML-012: ENABOL bridge.
Decisionโ
The implementation should not be a large Vivado-writer patch that reconstructs training behavior from generated C++ strings. The trainable path should be represented before writing:
ENABOL config export
-> HLSConfig Model.Training
-> ModelGraph + trainable flow passes
-> layer attributes and endpoint attributes
-> backend templates
-> writer emits already-resolved trainable code
This keeps the upstream patch smaller and makes the generated firmware easier to reason about.
Relevant hls4ml Objectsโ
HLSConfig is the normalized configuration access layer. It should own trainable schema accessors and validation-facing helpers.
ModelGraph is the compiled graph that carries layers, variables, attributes, and applied flows. It should remain the central object, not be replaced by a separate training graph.
Flows and optimizer passes are the right place to validate trainable mode, resolve loss endpoints, compute backward traversal order, and attach template-ready attributes.
Backend templates are the right place to generate Dense backpass, loss, update, and controller C++ snippets.
The Vivado writer should only add top-level IO, static trainable assets, and precomputed snippets. It should not infer weight, gradient, or loss semantics.
Exit Criteriaโ
Phase 0 is done when the roadmap, reference pages, and task stream agree on the same architecture. The remaining work should be executable as phase-specific hls4ml changes rather than open-ended reverse engineering.