pLiner for floating point drift
toolingComments
To Marie's point, the real challenge is often the Fused Multiply-Add (FMA) instructions, which combine a multiplication and addition into one rounding step. If pLiner can differentiate between standard IEEE 754 sequences and FMA optimizations, it solves a legitimate architectural headache.
I'm skeptical about the "traceable" part if the drift happens because of different CPU architectures. A tool that works on a dev machine is one thing, but seeing it consistently across mixed hardware is where these things usually fall apart.
Does this handle the weirdness of FP16 or BF16 used in those AI models we've been seeing... especially if the project is targeting WASM for browser execution? That would change the debugging game entirely...
Even if it doesn't solve cross-platform parity perfectly, having a tool that flags the exact operation where precision drops is a huge step forward. It saves hours of manual print-statement hunting in large math libraries.