Commit History
9 days ago
-
Card: package size and parameters per part, one pass per question
Verified signaturec1f651c
12 days ago
-
Decision Index: board numbers of 2026-10-28
Verified signature1e13154 -
Decision Index: on the board since 28 Oct 2026
Verified signaturef8f8cb5 -
Text encoder: attend without dense masks on long rows on GPUs
Verified signature`text/modeling_ultrabert.py` is the file vela3-fast shipped before its long-row fix. Under SDPA, `UltrabertModel.forward` builds two `[B, 1, L, L]` float masks per forward, one for the ±128 local layers and one for the global layers, although the local layers read only 257 keys per query. On a GPU they cost memory and time with the square of the row width. `text/modeling_ultrabert.py` is now the same file as vela3-fast's `modeling_ultrabert.py`: on a GPU under SDPA, rows of 4,096 tokens or more skip those masks (`blocked_attention`). Global layers read a `[B, 1, 1, L]` key mask, or none when no row is padded. Local layers attend in 128-query blocks over the keys within the window. Shorter rows, the CPU, eager attention and FlashAttention 2 run as before. The vision side, which uses transformers' own UltraBERT, is not changed. SHA256SUMS is updated. Tested on one MI300X, text side only, with four router questions, the PII label set and an NER question over Apache-2.0 licence text, and…219e264
13 days ago
-
Card: state the size as 0.3B, as in the blog and the other cards
Verified signature5f1c4b9
14 days ago
-
Card: Decision Index 37.5 with no state cut; size per call
Verified signaturee8aeb88 -
No state cut and no dropped images in decisions; add the Decision Index vision engine and the /v1/systemone server; size is 0.25B
Verified signatured279d5f -
Add the GroundLM 2026 (EMNLP) citation
Verified signature274738c
15 days ago
-
vela3-fast-vision: release package (hybrid2) and card
e2b12d6