4ae23deb50
Revert "Update 0007-add-unpad-operator.patch"
...
This reverts commit 340359fd087dd93c99bed4b9c87ccea3e759e184.
Update 0007-add-unpad-operator.patch
2025-05-12 13:51:04 -07:00
5d3da85a16
remove out of date comments
2025-05-12 13:51:04 -07:00
8b64b456c1
Update 0007-add-unpad-operator.patch
2025-05-12 13:51:04 -07:00
684f0d9291
set default values for vision model in config
2025-05-12 13:51:04 -07:00
3308bff137
add i32 copy and argsort for cuda
2025-05-12 13:51:04 -07:00
bf1929a3bc
Delete 0017-add-ollama-vocab-for-grammar-support.patch
2025-05-12 13:49:44 -07:00
1a2c413225
move mask
2025-05-12 13:49:44 -07:00
57279f89a2
calculate block mask once, rather than in attention
2025-05-12 13:49:44 -07:00
9ceee25d8b
chunk vision outputs
2025-05-12 13:49:44 -07:00
661bf04696
add picture prefix
2025-05-12 13:49:44 -07:00
2521a55ae6
fixes after rebase
2025-05-12 13:49:44 -07:00
32948ec952
increase rope base
2025-05-12 13:49:43 -07:00
9876c8453a
update exported functions for tests
2025-05-12 13:49:43 -07:00
919b3d6e21
require new engine for qwen25vl arch
2025-05-12 13:49:43 -07:00
16b13e0cfc
Revert "ropeTheta should be 1e5"
...
This reverts commit cc1638b26763eae7daddd44e3975a885671ef9d3.
This reverts commit
b32385591307e2d33a8f43ce1626b529d2dac83e.
2025-05-12 13:49:43 -07:00
75441c56f3
add comment explaining rope theta
2025-05-12 13:49:43 -07:00
45f96e898d
ropeTheta should be 1e5
2025-05-12 13:49:43 -07:00
7c555d394c
simplify patch creation
2025-05-12 13:49:43 -07:00
39ee6d2bd0
ranges for lint
2025-05-12 13:49:43 -07:00
47705b5168
simplify rope changes
2025-05-12 13:49:43 -07:00
698a92aa4a
reverse window
2025-05-12 13:49:43 -07:00
150c499cae
use silu
2025-05-12 13:49:43 -07:00
f1257a7de4
update vision rope theta default
2025-05-12 13:49:43 -07:00
b68af0370f
move sdpa to model forward pass
2025-05-12 13:49:43 -07:00
ca981c8a49
full attn block indexes should be []int32
2025-05-12 13:49:43 -07:00
b3da8a319e
Update model_vision.go
2025-05-12 13:49:42 -07:00
359e1d5b19
full attention layers
2025-05-12 13:49:42 -07:00
bde6b46ce9
fix padding
...
padding was being added to offset but not to the running count
2025-05-12 13:49:42 -07:00
ff1f74534b
block attention
2025-05-12 13:49:42 -07:00
104f802df1
remove todos
2025-05-12 13:49:42 -07:00
eed0ac2948
clean up vision model forward pass
2025-05-12 13:49:42 -07:00
fcfad744ff
fix patch merger
2025-05-12 13:49:42 -07:00
fb3c16f2a2
window index
2025-05-12 13:49:42 -07:00
ee869f35e4
fix image processing
...
python built-in `round()` rounds to the nearest even number if the value
is in the middle
https://docs.python.org/3/library/functions.html#round
2025-05-12 13:49:42 -07:00
ff5d1a3dc0
duplicate input embeddings
2025-05-12 13:49:42 -07:00
88b231f903
use maxgridsize
2025-05-12 13:49:42 -07:00
7e920c8d75
fix: patch merger and convert
...
convert:
- split patch embedding
- split qkv
remove duplicate PatchMerger
2025-05-12 13:49:42 -07:00
dd8c619fba
fixes after rebase
2025-05-12 13:49:42 -07:00
2af76d0e7a
default to 32 for vision block count
2025-05-12 13:49:42 -07:00
8d901825f0
reshape cos and sin
2025-05-12 13:49:41 -07:00
04936b719f
Update model_vision.go
2025-05-12 13:49:41 -07:00
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
80498f76de
fix build
2025-05-12 13:49:41 -07:00
f8b48aa784
Delete model_external_test.go
2025-05-12 13:49:41 -07:00
5ff0d538b0
wip: implementing rope
2025-05-12 13:49:41 -07:00
eedc969c35
grid refactor
2025-05-12 13:49:41 -07:00
963531215e
update convert
2025-05-12 13:49:41 -07:00
3fe090f447
get patch embedding vals from config
2025-05-12 13:49:41 -07:00
1704072746
patch embeddings
2025-05-12 13:49:41 -07:00
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
Bruce MacDonald
jmorganca
Michael Yang