clm-0104

ggml_flash_attn_ext on this HIP/gfx1151 build does not fully honour an additive F16 block-causal mask: a committed prefix's per-layer K/V changed by several logits when a later masked block entered the attention window (compounding through layers), while the manual ggml_soft_max_ext path with the identical mask did not. The boundary tracks 64 keys (= 2x the Wave Size 32), pointing at a tiling/reduction issue in the flash kernel's handling of masked lanes. Inferred so far via a downstream KV-cache oracle; an independent generation-free minimal repro is drafted but NOT yet run, so this is not yet an upstream-filed bug.

Cited by — computed at build time, never stored