2f068b5d87
dflash: use draft context as capacity contract ( #2341 )
...
dflash: account selector drafts in telemetry
2026-08-26 18:26:50 +02:00
28fbe34ce9
Dflash 2 speculative decoding ( #2345 )
...
* Add DFlash2 speculative decoding support
* Fix DFlash2 requantized output selection
* Fix legacy DFlash output row contract
* return llm_build_norm to apply also for dflash 2
* Fix DFlash2 GGUF conversion
* capture dflash states in any type
* remove duplicated layer_rows_raw
2026-08-26 17:09:34 +02:00
66b2f50ce3
Allow dspark to draft more that the amount of block size ( #2323 )
2026-08-24 09:55:26 +02:00
c46ffaa566
fix dspark: seed draft block at id_last's true position (+1 off-by-one) ( #2296 )
...
The draft block was seeded at last_target_pos (the newest committed
feature row = id_last's predecessor), placing the whole block one
position early vs mainline's [id_last @ n_past, ...] convention and
colliding the seed with the newest cross-KV row. Shift the common-side
batch and the graph-side SWA mask base coherently.
2026-08-12 18:21:30 +02:00
7ebbb906d2
Initial implementation of DSpark ( #2280 )
...
* Implement initial arch for DSpark
* feat: Add Dspark architecture support
* avoid to many splits in graph and improve rope logic
2026-08-10 08:46:03 +02:00
6cae8c7ba2
clean logs
2026-06-14 21:07:57 -03:00
0d75eee35a
remove duplicated code and unnecesary refactor
2026-06-14 16:02:02 -03:00
Samuel Oliveira Alves
rumas77
SamuelOliveirads