Skip to content

ggml: add ops for WAN video model (cuda && cpu) #15669

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 31 commits into from
Sep 4, 2025
+754 −85
Merged

ggml: add ops for WAN video model (cuda && cpu) #15669

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
31 commits
Select commit Hold shift + click to select a range
c92f9b4
add conv3d support
leejet Jun 27, 2025
93c7e77
add ggml_pad_ext for cpu & cuda backend
leejet Jul 25, 2025
f7a12f9
cuda/cpu: add im2col_3d support
leejet Aug 1, 2025
85c8e1e
cuda: make im2col a little faster
leejet Aug 1, 2025
ae47cac
fix cuda pad/scale/im2col3d
leejet Aug 12, 2025
dd745ba
make im2col_3d faster
leejet Aug 12, 2025
d8377a0
gguf: support loading tensors which n_dims > GGML_MAX_DIMS
leejet Aug 15, 2025
d30e07d
fix cuda get_rows
leejet Aug 23, 2025
df05913
avoid ggml_conv_3d conflict
leejet Aug 29, 2025
9d035c4
correct GGML_OP_COUNT assertion
leejet Aug 29, 2025
d11a729
avoid build failure
leejet Aug 29, 2025
f6a874c
avoid build failure on MacOS
leejet Aug 29, 2025
f6278c8
cuda: remove unnecessary MIN define
leejet Aug 29, 2025
c9b9fab
fix cpu im2col_3d
leejet Aug 30, 2025
131ae2d
adjust the code style
leejet Aug 30, 2025
0d5eb51
cuda: use simpler loop in get_rows
leejet Aug 30, 2025
aafa79a
add test_im2col_3d to test-backend-ops
leejet Aug 30, 2025
3f901e3
test-backend-ops.cpp: remove trailing whitespace
leejet Aug 30, 2025
e66bf6e
cpu: im2col_3d support non continuous src
leejet Aug 31, 2025
b4c50be
fix test_im2col_3d
leejet Aug 31, 2025
8f5e7b0
remove unused variables
leejet Aug 31, 2025
21e9338
cuda: get_rows: dfloat2 -> float2
leejet Aug 31, 2025
36f2215
add test_pad_ext to test-backend-ops.cpp
leejet Aug 31, 2025
d9f1d13
add gguf_init_from_file_ext impl
leejet Sep 2, 2025
6b71242
Merge branch 'master' into wan
leejet Sep 2, 2025
9b365e8
Revert "gguf: support loading tensors which n_dims > GGML_MAX_DIMS"
leejet Sep 2, 2025
6b6eede
Revert "add gguf_init_from_file_ext impl"
leejet Sep 2, 2025
2412bb0
update ggml_backend_vk_device_supports_op
leejet Sep 2, 2025
b38bfbb
fix ggml_backend_vk_device_supports_op
leejet Sep 2, 2025
457f186
update other backend supports op for ggml_pad_ext
leejet Sep 2, 2025
1618844
metal/opencl/sycl/vulkan: fix GGML_OP_PAD check in supports_op
leejet Sep 2, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
51 changes: 50 additions & 1 deletion ggml/include/ggml.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -511,6 +511,7 @@ extern "C" {
GGML_OP_CONV_TRANSPOSE_1D,
GGML_OP_IM2COL,
GGML_OP_IM2COL_BACK,
GGML_OP_IM2COL_3D,
GGML_OP_CONV_2D,
GGML_OP_CONV_3D,
GGML_OP_CONV_2D_DW,
Expand Down Expand Up @@ -1870,6 +1871,41 @@ extern "C" {
int d0, // dilation dimension 0
int d1); // dilation dimension 1

GGML_API struct ggml_tensor * ggml_im2col_3d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int64_t IC,
int s0, // stride width
int s1, // stride height
int s2, // stride depth
int p0, // padding width
int p1, // padding height
int p2, // padding depth
int d0, // dilation width
int d1, // dilation height
int d2, // dilation depth
enum ggml_type dst_type);

// a: [OC*IC, KD, KH, KW]
// b: [N*IC, ID, IH, IW]
// result: [N*OC, OD, OH, OW]
GGML_API struct ggml_tensor * ggml_conv_3d(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
int64_t IC,
int s0, // stride width
int s1, // stride height
int s2, // stride depth
int p0, // padding width
int p1, // padding height
int p2, // padding depth
int d0, // dilation width
int d1, // dilation height
int d2 // dilation depth
);

// kernel size is a->ne[0] x a->ne[1]
// stride is equal to kernel size
// padding is zero
Expand Down Expand Up @@ -1941,7 +1977,7 @@ extern "C" {
int d0, // dilation dimension 0
int d1); // dilation dimension 1

GGML_API struct ggml_tensor * ggml_conv_3d(
GGML_API struct ggml_tensor * ggml_conv_3d_direct(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think ideally there should be only one conv_3d operation, and the backend can decide on an implementation (direct, im2col, tiled im2col, or choose one depending on inputs & hardware).

The reason there are two versions for conv_2d is because historically the im2col version was there first, and it coded im2col into the graph. This means the backend cannot optimize it or choose a better implementation, and the huge memory requirements of full tensor im2col or baked into the graph. This was difficult to change without breaking backends, hence the "direct" workaround. But IMO we should avoid this situation for newly introduced conv ops.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, this is the ultimate goal. However, currently the support for each backend is not yet complete, and conv_2d_direct cannot fully replace im2col + gemm at present. In some backends such as CUDA, conv_2d_direct is much slower than im2col + gemm.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree there's no solution for conv_2d right away. My point is, this PR now creates the same problem for 3d. I don't think new code should be introduced that repeats the issue and bakes im2col as the default way to do conv_3d.

My suggestion would be to either implement im2col+mul_mat behind OP_CONV_3D in the CUDA backend. Or only expose im2col_3d on the API, and move im2col+mul_mat calls into the application (simpler but not as nice).

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For CUDA one of my long-term goals is to write convolution kernels, particularly ones that can make use of quantized data. But I should stress that I have so many other things I want to do that realistically I would start working on it in a year at the absolute earliest.

struct ggml_context * ctx,
struct ggml_tensor * a, // kernel [KW, KH, KD, IC * OC]
struct ggml_tensor * b, // input [W, H, D, C * N]
Expand Down Expand Up @@ -2048,6 +2084,19 @@ extern "C" {
int p2,
int p3);

GGML_API struct ggml_tensor * ggml_pad_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
int lp0,
int rp0,
int lp1,
int rp1,
int lp2,
int rp2,
int lp3,
int rp3
);

// pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
struct ggml_context * ctx,
Expand Down
13 changes: 10 additions & 3 deletions ggml/src/ggml-cann/aclnn_ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -589,9 +589,16 @@ void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
// the position of elements in the array means which dirction to padding,
// each position means: [dim0.front, dim0.behind, dim1.front, dim1.behind,
// dim2.front, dim2.behind, dim3.front, dim3.behind]
int64_t paddings[] = {
0, dst->ne[0] - src->ne[0], 0, dst->ne[1] - src->ne[1],
0, dst->ne[2] - src->ne[2], 0, dst->ne[3] - src->ne[3]};
const int32_t lp0 = ggml_get_op_params_i32(dst, 0);
const int32_t rp0 = ggml_get_op_params_i32(dst, 1);
const int32_t lp1 = ggml_get_op_params_i32(dst, 2);
const int32_t rp1 = ggml_get_op_params_i32(dst, 3);
const int32_t lp2 = ggml_get_op_params_i32(dst, 4);
const int32_t rp2 = ggml_get_op_params_i32(dst, 5);
const int32_t lp3 = ggml_get_op_params_i32(dst, 6);
const int32_t rp3 = ggml_get_op_params_i32(dst, 7);

int64_t paddings[] = {lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3};
aclnn_pad(ctx, acl_src, acl_dst, paddings);
ggml_cann_release_resources(ctx, acl_src, acl_dst);
}
Expand Down
5 changes: 5 additions & 0 deletions ggml/src/ggml-cpu/ggml-cpu.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1876,6 +1876,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_im2col_back_f32(params, tensor);
} break;
case GGML_OP_IM2COL_3D:
{
ggml_compute_forward_im2col_3d(params, tensor);
} break;
case GGML_OP_CONV_2D:
{
ggml_compute_forward_conv_2d(params, tensor);
Expand Down Expand Up @@ -2255,6 +2259,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
} break;
case GGML_OP_IM2COL:
case GGML_OP_IM2COL_BACK:
case GGML_OP_IM2COL_3D:
case GGML_OP_CONV_2D:
case GGML_OP_CONV_3D:
case GGML_OP_CONV_2D_DW:
Expand Down
222 changes: 218 additions & 4 deletions ggml/src/ggml-cpu/ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7027,6 +7027,209 @@ void ggml_compute_forward_im2col_back_f32(
}
}


// ggml_compute_forward_im2col_3d_f16
// src0: kernel [OC*IC, KD, KH, KW]
// src1: image [N*IC, ID, IH, IW]
// dst: result [N*OD, OH, OW, IC * KD * KH * KW]
static void ggml_compute_forward_im2col_3d_f16(
const ggml_compute_params * params,
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];

GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F16);

GGML_TENSOR_BINARY_OP_LOCALS;

const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
const int32_t s2 = ((const int32_t *)(dst->op_params))[2];
const int32_t p0 = ((const int32_t *)(dst->op_params))[3];
const int32_t p1 = ((const int32_t *)(dst->op_params))[4];
const int32_t p2 = ((const int32_t *)(dst->op_params))[5];
const int32_t d0 = ((const int32_t *)(dst->op_params))[6];
const int32_t d1 = ((const int32_t *)(dst->op_params))[7];
const int32_t d2 = ((const int32_t *)(dst->op_params))[8];
const int32_t IC = ((const int32_t *)(dst->op_params))[9];


const int ith = params->ith;
const int nth = params->nth;

const int64_t N = ne13 / IC;
const int64_t ID = ne12;
const int64_t IH = ne11;
const int64_t IW = ne10;

const int64_t OC = ne03 / IC;
GGML_UNUSED(OC);
const int64_t KD = ne02;
const int64_t KH = ne01;
const int64_t KW = ne00;

const int64_t OD = ne3 / N;
const int64_t OH = ne2;
const int64_t OW = ne1;
const int64_t OH_OW = OH*OW;
const int64_t KD_KH_KW = KD*KH*KW;
const int64_t KH_KW = KH*KW;
const int64_t IC_KD_KH_KW = IC*KD*KH*KW;

GGML_ASSERT(nb10 == sizeof(float));

// im2col: [N*IC, ID, IH, IW] => [N*OD, OH, OW, IC * KD * KH * KW]
{
ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data;

for (int64_t in = 0; in < N; in++) {
for (int64_t iod = 0; iod < OD; iod++) {
for (int64_t ioh = 0; ioh < OH; ioh++) {
for (int64_t iow = 0; iow < OW; iow++) {
for (int64_t iic = ith; iic < IC; iic += nth) {

// micro kernel
ggml_fp16_t * dst_data = wdata + (in*OD*OH_OW + iod*OH_OW + ioh*OW + iow)*IC_KD_KH_KW; // [IC, KD, KH, KW]
const float * const src_data = (const float *) ((const char *)src1->data + (in*IC + iic)*nb13); // [ID, IH, IW]

for (int64_t ikd = 0; ikd < KD; ikd++) {
for (int64_t ikh = 0; ikh < KH; ikh++) {
for (int64_t ikw = 0; ikw < KW; ikw++) {
const int64_t iiw = iow*s0 + ikw*d0 - p0;
const int64_t iih = ioh*s1 + ikh*d1 - p1;
const int64_t iid = iod*s2 + ikd*d2 - p2;

if (iid < 0 || iid >= ID || iih < 0 || iih >= IH || iiw < 0 || iiw >= IW || iid < 0 || iid >= ID) {
dst_data[iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw] = 0;
} else {
const float * const s = (const float *) ((const char *)src_data + iid*nb12 + iih*nb11 + iiw*nb10); // [ID, IH, IW]
dst_data[iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(*s);
}
}
}
}
}
}
}
}
}
}
}

// ggml_compute_forward_im2col_3d_f32
// src0: kernel [OC*IC, KD, KH, KW]
// src1: image [N*IC, ID, IH, IW]
// dst: result [N*OD, OH, OW, IC * KD * KH * KW]
static void ggml_compute_forward_im2col_3d_f32(
const ggml_compute_params * params,
ggml_tensor * dst) {

const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];

GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);

GGML_TENSOR_BINARY_OP_LOCALS;

const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
const int32_t s2 = ((const int32_t *)(dst->op_params))[2];
const int32_t p0 = ((const int32_t *)(dst->op_params))[3];
const int32_t p1 = ((const int32_t *)(dst->op_params))[4];
const int32_t p2 = ((const int32_t *)(dst->op_params))[5];
const int32_t d0 = ((const int32_t *)(dst->op_params))[6];
const int32_t d1 = ((const int32_t *)(dst->op_params))[7];
const int32_t d2 = ((const int32_t *)(dst->op_params))[8];
const int32_t IC = ((const int32_t *)(dst->op_params))[9];


const int ith = params->ith;
const int nth = params->nth;

const int64_t N = ne13 / IC;
const int64_t ID = ne12;
const int64_t IH = ne11;
const int64_t IW = ne10;

const int64_t OC = ne03 / IC;
GGML_UNUSED(OC);
const int64_t KD = ne02;
const int64_t KH = ne01;
const int64_t KW = ne00;

const int64_t OD = ne3 / N;
const int64_t OH = ne2;
const int64_t OW = ne1;

const int64_t OH_OW = OH*OW;
const int64_t KD_KH_KW = KD*KH*KW;
const int64_t KH_KW = KH*KW;
const int64_t IC_KD_KH_KW = IC*KD*KH*KW;

GGML_ASSERT(nb10 == sizeof(float));

// im2col: [N*IC, ID, IH, IW] => [N*OD, OH, OW, IC * KD * KH * KW]
{
float * const wdata = (float *) dst->data;

for (int64_t in = 0; in < N; in++) {
for (int64_t iod = 0; iod < OD; iod++) {
for (int64_t ioh = 0; ioh < OH; ioh++) {
for (int64_t iow = 0; iow < OW; iow++) {
for (int64_t iic = ith; iic < IC; iic += nth) {

// micro kernel
float * dst_data = wdata + (in*OD*OH_OW + iod*OH_OW + ioh*OW + iow)*IC_KD_KH_KW; // [IC, KD, KH, KW]
const float * const src_data = (const float *) ((const char *)src1->data + (in*IC + iic)*nb13); // [ID, IH, IW]

for (int64_t ikd = 0; ikd < KD; ikd++) {
for (int64_t ikh = 0; ikh < KH; ikh++) {
for (int64_t ikw = 0; ikw < KW; ikw++) {
const int64_t iiw = iow*s0 + ikw*d0 - p0;
const int64_t iih = ioh*s1 + ikh*d1 - p1;
const int64_t iid = iod*s2 + ikd*d2 - p2;

if (iid < 0 || iid >= ID || iih < 0 || iih >= IH || iiw < 0 || iiw >= IW || iid < 0 || iid >= ID) {
dst_data[iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw] = 0;
} else {
const float * const s = (const float *) ((const char *)src_data + iid*nb12 + iih*nb11 + iiw*nb10); // [ID, IH, IW]
dst_data[iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw] = *s;
}
}
}
}
}
}
}
}
}
}
}


void ggml_compute_forward_im2col_3d(
const ggml_compute_params * params,
ggml_tensor * dst) {
switch (dst->type) {
case GGML_TYPE_F16:
{
ggml_compute_forward_im2col_3d_f16(params, dst);
} break;
case GGML_TYPE_F32:
{
ggml_compute_forward_im2col_3d_f32(params, dst);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}

static void ggml_call_mul_mat(ggml_type type, const ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
void * a, void * b, float * c) {
const ggml_type_traits * traits = ggml_get_type_traits(type);
Expand Down Expand Up @@ -8014,6 +8217,15 @@ static void ggml_compute_forward_pad_f32(
GGML_TENSOR_UNARY_OP_LOCALS

float * dst_ptr = (float *) dst->data;
const int32_t lp0 = ggml_get_op_params_i32(dst, 0);
const int32_t rp0 = ggml_get_op_params_i32(dst, 1);
const int32_t lp1 = ggml_get_op_params_i32(dst, 2);
const int32_t rp1 = ggml_get_op_params_i32(dst, 3);
const int32_t lp2 = ggml_get_op_params_i32(dst, 4);
const int32_t rp2 = ggml_get_op_params_i32(dst, 5);
const int32_t lp3 = ggml_get_op_params_i32(dst, 6);
const int32_t rp3 = ggml_get_op_params_i32(dst, 7);


// TODO: optimize

Expand All @@ -8022,10 +8234,12 @@ static void ggml_compute_forward_pad_f32(
for (int64_t i0 = 0; i0 < ne0; ++i0) {
for (int64_t i3 = 0; i3 < ne3; ++i3) {
const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;

const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);

if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
if ((i0 >= lp0 && i0 < ne0 - rp0) \
&& (i1 >= lp1 && i1 < ne1 - rp1) \
&& (i2 >= lp2 && i2 < ne2 - rp2) \
&& (i3 >= lp3 && i3 < ne3 - rp3)) {
const int64_t src_idx = (i3 - lp3)*nb03 + (i2 - lp2)*nb02 + (i1 - lp1)*nb01 + (i0 - lp0)*nb00;
const float * src_ptr = (const float *)((char *) src0->data + src_idx);
dst_ptr[dst_idx] = *src_ptr;
} else {
dst_ptr[dst_idx] = 0;
Expand Down
1 change: 1 addition & 0 deletions ggml/src/ggml-cpu/ops.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,6 +69,7 @@ void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struc
void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_im2col_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
Expand Down
Loading
Loading