ponder-lab / Hybridize-Functions-Refactoring / #2271

80%

main: 80%

Build Type

Pull #712

github

Committed by khatchad

Commit Message

De-hybridize barren hybrid functions (issue 709 hybrid-to-eager direction)

The eager-to-hybrid barren block landed earlier; its hybrid-to-eager
counterpart was deferred. A hybrid function with a tensor parameter that
performs no tensor computation gains nothing from graph execution, only
tracing overhead, so it is de-hybridized — the symmetric completion of the
benefit signal and a peer of the P2/P3 de-hybridizations.

`Function.check()` gains a Table 2 precondition (new `PreconditionSuccess`
P6): a barren, side-effect-free hybrid function with no primitive parameter
converts to eager; a barren function with side-effects cannot be safely
de-hybridized and fails on `HAS_PYTHON_SIDE_EFFECTS`, mirroring the existing
de-hybridization side-effect gate. Barren takes precedence over the
reconfigure path. `computeTensorComputation` now runs for hybrid
tensor-parameter functions too (the `!isHybrid` gate is dropped), mirroring
the recursion check.

`testBarrenHybridDehybridizes` pins the new behavior. Three existing hybrid
fixtures (`testSpeculativeAnalysis`, `testSpeculativeAnalysis3`,
`testRecursion5`) were incidentally barren and now correctly de-hybridize;
they gain tensor operations so they still exercise speculative inference and
recursion, mirroring the eager-side fixture updates.

The residual false positives (#714, #716) must be resolved before this
direction runs over the corpus, since de-hybridizing a false-barren function
de-optimizes real tensor code.

Refs #709.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FuhM8SsTRB6BraLt3Ph7ZC

Pull Request Pull Request #712: Add a benefit precondition for barren eager functions

Coverage Stats

118 of 124 new or added lines in 5 files covered. (95.16%)

64 existing lines in 2 files now uncovered.

1753 of 2194 relevant lines covered (79.9%)

0.8 hits per line