ggml: add ops for WAN video model (cuda && cpu) #15669
JohannesGaessler
merged
31 commits into
ggml-org:master
from
leejet:wan
add conv3d support
c92f9b4a
add ggml_pad_ext for cpu & cuda backend
93c7e775
cuda/cpu: add im2col_3d support
f7a12f9e
cuda: make im2col a little faster
85c8e1e5
fix cuda pad/scale/im2col3d
ae47caca
make im2col_3d faster
dd745ba3
gguf: support loading tensors which n_dims > GGML_MAX_DIMS
d8377a0a
fix cuda get_rows
d30e07db
avoid ggml_conv_3d conflict
df05913b
correct GGML_OP_COUNT assertion
9d035c4c
avoid build failure
d11a7298
avoid build failure on MacOS
f6a874c0
cuda: remove unnecessary MIN define
f6278c83
fix cpu im2col_3d
c9b9fabe
Acly
commented
on 2025-08-30
JohannesGaessler
commented
on 2025-08-30
adjust the code style
131ae2d5
cuda: use simpler loop in get_rows
0d5eb512
add test_im2col_3d to test-backend-ops
aafa79ae
test-backend-ops.cpp: remove trailing whitespace
3f901e31
cpu: im2col_3d support non continuous src
e66bf6e5
fix test_im2col_3d
b4c50bec
remove unused variables
8f5e7b0c
cuda: get_rows: dfloat2 -> float2
21e93380
JohannesGaessler
commented
on 2025-08-31
add test_pad_ext to test-backend-ops.cpp
36f2215e
JohannesGaessler
commented
on 2025-09-02
add gguf_init_from_file_ext impl
d9f1d132
Merge branch 'master' into wan
6b71242a
Revert "gguf: support loading tensors which n_dims > GGML_MAX_DIMS"
9b365e83
Revert "add gguf_init_from_file_ext impl"
6b6eeded
update ggml_backend_vk_device_supports_op
2412bb0b
fix ggml_backend_vk_device_supports_op
b38bfbb5
update other backend supports op for ggml_pad_ext
457f1864
metal/opencl/sycl/vulkan: fix GGML_OP_PAD check in supports_op
1618844e
JohannesGaessler
approved these changes
on 2025-09-04
Reviewers
Assignees
No one assigned
Labels
testing
Nvidia GPU
Vulkan
ggml
SYCL
Apple Metal
Ascend NPU
OpenCL
Milestone
No milestone
Login to write a write a comment.
Login via GitHub