BrianHarrisonAMD f0545b5c15 [rocm-libraries] ROCm/rocm-libraries#8132 (commit 57d21a1) ...

[CK dispatcher] - LGBM predict data_type FLOAT32->FLOAT64 in
 ml_heuristic (#8132)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

## Summary
`ml_heuristic.hpp` calls `LGBM_BoosterPredictForMat(...,
/*data_type=*/0, ...)` (`C_API_DTYPE_FLOAT32`) against a
`std::array<double, NUM_FEATURES>` feature buffer. LightGBM reinterprets
the 8-byte doubles as 4-byte floats → invalid predictions → the
heuristic's argmax always tie-breaks to the first/smallest enumerated
config.

**Fix:** `data_type 0 → 1` (`C_API_DTYPE_FLOAT64`), matching the
`double` buffer. After the fix, predictions vary and track real TFLOPS
(the model correctly prefers larger tiles).

## Verification
- The feature buffer `f` is `std::array<double, NUM_FEATURES>`
(NUM_FEATURES = 72) → `f.data()` is a `double*`.
- The changed `0` is the 3rd positional `data_type` argument (not
`nrow`/`ncol`/`is_row_major`).

One-line functional change.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

2026-06-10 18:57:31 +00:00

..

ck_tile

[rocm-libraries] ROCm/rocm-libraries#8132 (commit 57d21a1)

2026-06-10 18:57:31 +00:00