Commit Graph

29 Commits

Author SHA1 Message Date
Bruce MacDonald
1a2c413225 move mask 2025-05-12 13:49:44 -07:00
Bruce MacDonald
57279f89a2 calculate block mask once, rather than in attention 2025-05-12 13:49:44 -07:00
Bruce MacDonald
9876c8453a update exported functions for tests 2025-05-12 13:49:43 -07:00
Bruce MacDonald
16b13e0cfc Revert "ropeTheta should be 1e5"
This reverts commit cc1638b26763eae7daddd44e3975a885671ef9d3.

This reverts commit
b32385591307e2d33a8f43ce1626b529d2dac83e.
2025-05-12 13:49:43 -07:00
Bruce MacDonald
45f96e898d ropeTheta should be 1e5 2025-05-12 13:49:43 -07:00
Bruce MacDonald
39ee6d2bd0 ranges for lint 2025-05-12 13:49:43 -07:00
Michael Yang
698a92aa4a reverse window 2025-05-12 13:49:43 -07:00
Michael Yang
150c499cae use silu 2025-05-12 13:49:43 -07:00
Bruce MacDonald
b68af0370f move sdpa to model forward pass 2025-05-12 13:49:43 -07:00
Bruce MacDonald
b3da8a319e Update model_vision.go 2025-05-12 13:49:42 -07:00
Bruce MacDonald
359e1d5b19 full attention layers 2025-05-12 13:49:42 -07:00
Bruce MacDonald
ff1f74534b block attention 2025-05-12 13:49:42 -07:00
Bruce MacDonald
104f802df1 remove todos 2025-05-12 13:49:42 -07:00
Bruce MacDonald
eed0ac2948 clean up vision model forward pass 2025-05-12 13:49:42 -07:00
Bruce MacDonald
fcfad744ff fix patch merger 2025-05-12 13:49:42 -07:00
Michael Yang
fb3c16f2a2 window index 2025-05-12 13:49:42 -07:00
Michael Yang
88b231f903 use maxgridsize 2025-05-12 13:49:42 -07:00
Michael Yang
7e920c8d75 fix: patch merger and convert
convert:
- split patch embedding
- split qkv

remove duplicate PatchMerger
2025-05-12 13:49:42 -07:00
Bruce MacDonald
2af76d0e7a default to 32 for vision block count 2025-05-12 13:49:42 -07:00
Bruce MacDonald
8d901825f0 reshape cos and sin 2025-05-12 13:49:41 -07:00
Bruce MacDonald
04936b719f Update model_vision.go 2025-05-12 13:49:41 -07:00
Bruce MacDonald
0f0136d419 simplify by doing operations in Go rather than with tensors
Co-Authored-By: Michael Yang <2372640+mxyng@users.noreply.github.com>
2025-05-12 13:49:41 -07:00
Bruce MacDonald
80498f76de fix build 2025-05-12 13:49:41 -07:00
Bruce MacDonald
5ff0d538b0 wip: implementing rope 2025-05-12 13:49:41 -07:00
Bruce MacDonald
eedc969c35 grid refactor 2025-05-12 13:49:41 -07:00
Bruce MacDonald
963531215e update convert 2025-05-12 13:49:41 -07:00
Bruce MacDonald
3fe090f447 get patch embedding vals from config 2025-05-12 13:49:41 -07:00
Bruce MacDonald
1704072746 patch embeddings 2025-05-12 13:49:41 -07:00
Bruce MacDonald
c1f9bcb4dd restructure
image processing

Update model.go

Update model.go

Update model.go

no projector

no projector

vision model scaffold

...

...

wip

...

rebase

fix patch merger

tidy

...

Update model_vision.go

server: do not attempt to parse offset file as gguf

This logic was causing issues for me when importing a gguf that had some padding at the end of the file. The valid gguf would be read, but then it would try to read the offset as a different gguf file. This does not seem right.

Update process_image_test.go

apply norm

prompt processing

prompt processing

fix post tokenize

fix gguf padding + populate the split patch embeddings

...

...

another shot at patch embeddings

...

patch embedding

Update model_vision.go

split pixels
2025-05-12 13:49:41 -07:00