functiongemma: chunk prefill-with-past graph gemma_prefill_with_past — an utterance against the cache in one call (#410) - #417

Merged

michalharakal merged 2 commits into

functiongemma-chunk-prefill

Sep 4, 2026

Commits on Sep 4, 2026

functiongemma: chunk prefill-with-past graph gemma_prefill_with_past — a fixed 64-token chunk against the dynamic cache in one call (#410)
michalharakal
committed
bf9b299
functiongemma: chunk masks are per-head [1, nHeads, C, past+C] — a broadcast over heads to a dynamic shape is not expressible in static StableHLO
michalharakal
committed
ee1c4d3