DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2025-12-12 13:28:37 +00:00
Code Issues Packages Projects Releases Wiki Activity
stable-diffusion.cpp/wan.hpp

2307 lines
106 KiB
C++
Raw Blame History

#ifndef __WAN_HPP__
#define __WAN_HPP__
#include <map>
#include <memory>
#include <utility>
#include "common.hpp"
#include "flux.hpp"
#include "ggml_extend.hpp"
#include "rope.hpp"
#include "vae.hpp"
namespace WAN {
constexpr int CACHE_T = 2;
constexpr int WAN_GRAPH_SIZE = 10240;
class CausalConv3d : public GGMLBlock {
protected:
int64_t in_channels;
int64_t out_channels;
std::tuple<int, int, int> kernel_size;
std::tuple<int, int, int> stride;
std::tuple<int, int, int> padding;
std::tuple<int, int, int> dilation;
bool bias;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
params["weight"] = ggml_new_tensor_4d(ctx,
GGML_TYPE_F16,
std::get<2>(kernel_size),
std::get<1>(kernel_size),
std::get<0>(kernel_size),
in_channels * out_channels);
if (bias) {
params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
}
}
public:
CausalConv3d(int64_t in_channels,
int64_t out_channels,
std::tuple<int, int, int> kernel_size,
std::tuple<int, int, int> stride = {1, 1, 1},
std::tuple<int, int, int> padding = {0, 0, 0},
std::tuple<int, int, int> dilation = {1, 1, 1},
bool bias = true)
: in_channels(in_channels),
out_channels(out_channels),
kernel_size(std::move(kernel_size)),
stride(std::move(stride)),
padding(std::move(padding)),
dilation(std::move(dilation)),
bias(bias) {}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* cache_x = nullptr) {
// x: [N*IC, ID, IH, IW]
// result: x: [N*OC, ID, IH, IW]
struct ggml_tensor* w = params["weight"];
struct ggml_tensor* b = nullptr;
if (bias) {
b = params["bias"];
}
int lp0 = std::get<2>(padding);
int rp0 = std::get<2>(padding);
int lp1 = std::get<1>(padding);
int rp1 = std::get<1>(padding);
int lp2 = 2 * std::get<0>(padding);
int rp2 = 0;
if (cache_x != nullptr && lp2 > 0) {
x = ggml_concat(ctx, cache_x, x, 2);
lp2 -= (int)cache_x->ne[2];
}
x = ggml_pad_ext(ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, 0, 0);
return ggml_ext_conv_3d(ctx, x, w, b, in_channels,
std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
0, 0, 0,
std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation));
}
};
class RMS_norm : public UnaryBlock {
protected:
int64_t dim;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
ggml_type wtype = GGML_TYPE_F32;
params["gamma"] = ggml_new_tensor_1d(ctx, wtype, dim);
}
public:
RMS_norm(int64_t dim)
: dim(dim) {}
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
// x: [N*IC, ID, IH, IW], IC == dim
// assert N == 1
struct ggml_tensor* w = params["gamma"];
auto h = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 3, 0, 1, 2)); // [ID, IH, IW, N*IC]
h = ggml_rms_norm(ctx, h, 1e-12);
h = ggml_mul(ctx, h, w);
h = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, h, 1, 2, 3, 0));
return h;
}
};
class Resample : public GGMLBlock {
protected:
int64_t dim;
std::string mode;
public:
Resample(int64_t dim, const std::string& mode, bool wan2_2 = false)
: dim(dim), mode(mode) {
if (mode == "upsample2d") {
if (wan2_2) {
blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {3, 3}, {1, 1}, {1, 1}));
} else {
blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim / 2, {3, 3}, {1, 1}, {1, 1}));
}
} else if (mode == "upsample3d") {
if (wan2_2) {
blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {3, 3}, {1, 1}, {1, 1}));
} else {
blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim / 2, {3, 3}, {1, 1}, {1, 1}));
}
blocks["time_conv"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(dim, dim * 2, {3, 1, 1}, {1, 1, 1}, {1, 0, 0}));
} else if (mode == "downsample2d") {
blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {3, 3}, {2, 2}));
} else if (mode == "downsample3d") {
blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {3, 3}, {2, 2}));
blocks["time_conv"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(dim, dim, {3, 1, 1}, {2, 1, 1}, {0, 0, 0}));
} else if (mode == "none") {
// nn.Identity()
} else {
GGML_ASSERT(false && "invalid mode");
}
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b,
std::vector<struct ggml_tensor*>& feat_cache,
int& feat_idx,
int chunk_idx) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
int64_t c = x->ne[3] / b;
int64_t t = x->ne[2];
int64_t h = x->ne[1];
int64_t w = x->ne[0];
if (mode == "upsample3d") {
if (feat_cache.size() > 0) {
int idx = feat_idx;
feat_idx += 1;
if (chunk_idx == 0) {
// feat_cache[idx] == nullptr, pass
} else {
auto time_conv = std::dynamic_pointer_cast<CausalConv3d>(blocks["time_conv"]);
auto cache_x = ggml_ext_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) { // chunk_idx >= 2
// cache last frame of last two chunk
cache_x = ggml_concat(ctx,
ggml_ext_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
cache_x,
2);
}
if (chunk_idx == 1 && cache_x->ne[2] < 2) { // Rep
cache_x = ggml_pad_ext(ctx, cache_x, 0, 0, 0, 0, (int)cache_x->ne[2], 0, 0, 0);
// aka cache_x = torch.cat([torch.zeros_like(cache_x).to(cache_x.device),cache_x],dim=2)
}
if (chunk_idx == 1) {
x = time_conv->forward(ctx, x);
} else {
x = time_conv->forward(ctx, x, feat_cache[idx]);
}
feat_cache[idx] = cache_x;
x = ggml_reshape_4d(ctx, x, w * h, t, c, 2); // (2, c, t, h*w)
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 3, 1, 2)); // (c, t, 2, h*w)
x = ggml_reshape_4d(ctx, x, w, h, 2 * t, c); // (c, t*2, h, w)
}
}
}
t = x->ne[2];
if (mode != "none") {
auto resample_1 = std::dynamic_pointer_cast<Conv2d>(blocks["resample.1"]);
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2)); // (t, c, h, w)
if (mode == "upsample2d") {
x = ggml_upscale(ctx, x, 2, GGML_SCALE_MODE_NEAREST);
} else if (mode == "upsample3d") {
x = ggml_upscale(ctx, x, 2, GGML_SCALE_MODE_NEAREST);
} else if (mode == "downsample2d") {
x = ggml_pad(ctx, x, 1, 1, 0, 0);
} else if (mode == "downsample3d") {
x = ggml_pad(ctx, x, 1, 1, 0, 0);
}
x = resample_1->forward(ctx, x);
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2)); // (c, t, h, w)
}
if (mode == "downsample3d") {
if (feat_cache.size() > 0) {
int idx = feat_idx;
if (feat_cache[idx] == nullptr) {
feat_cache[idx] = x;
feat_idx += 1;
} else {
auto time_conv = std::dynamic_pointer_cast<CausalConv3d>(blocks["time_conv"]);
auto cache_x = ggml_ext_slice(ctx, x, 2, -1, x->ne[2]);
x = ggml_concat(ctx,
ggml_ext_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
x,
2);
x = time_conv->forward(ctx, x);
feat_cache[idx] = cache_x;
feat_idx += 1;
}
}
}
return x;
}
};
class AvgDown3D : public GGMLBlock {
protected:
int64_t in_channels;
int64_t out_channels;
int64_t factor_t;
int64_t factor_s;
int64_t factor;
int64_t group_size;
public:
AvgDown3D(int64_t in_channels, int64_t out_channels, int64_t factor_t, int64_t factor_s = 1)
: in_channels(in_channels), out_channels(out_channels), factor_t(factor_t), factor_s(factor_s) {
factor = factor_t * factor_s * factor_s;
GGML_ASSERT(in_channels * factor % out_channels == 0);
group_size = in_channels * factor / out_channels;
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t B = 1) {
// x: [B*IC, T, H, W]
// return: [B*OC, T/factor_t, H/factor_s, W/factor_s]
GGML_ASSERT(B == 1);
int64_t C = x->ne[3];
int64_t T = x->ne[2];
int64_t H = x->ne[1];
int64_t W = x->ne[0];
int64_t pad_t = (factor_t - T % factor_t) % factor_t;
x = ggml_pad_ext(ctx, x, 0, 0, 0, 0, pad_t, 0, 0, 0);
T = x->ne[2];
x = ggml_reshape_4d(ctx, x, W * H, factor_t, T / factor_t, C); // [C, T/factor_t, factor_t, H*W]
x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [C, factor_t, T/factor_t, H*W]
x = ggml_reshape_4d(ctx, x, W, factor_s, (H / factor_s) * (T / factor_t), factor_t * C); // [C*factor_t, T/factor_t*H/factor_s, factor_s, W]
x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [C*factor_t, factor_s, T/factor_t*H/factor_s, W]
x = ggml_reshape_4d(ctx, x, factor_s, W / factor_s, (H / factor_s) * (T / factor_t), factor_s * factor_t * C); // [C*factor_t*factor_s, T/factor_t*H/factor_s, W/factor_s, factor_s]
x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3)); // [C*factor_t*factor_s, factor_s, T/factor_t*H/factor_s, W/factor_s]
x = ggml_reshape_3d(ctx, x, (W / factor_s) * (H / factor_s) * (T / factor_t), group_size, out_channels); // [out_channels, group_size, T/factor_t*H/factor_s*W/factor_s]
x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 0, 2, 3)); // [out_channels, T/factor_t*H/factor_s*W/factor_s, group_size]
x = ggml_mean(ctx, x); // [out_channels, T/factor_t*H/factor_s*W/factor_s, 1]
x = ggml_reshape_4d(ctx, x, W / factor_s, H / factor_s, T / factor_t, out_channels);
return x;
}
};
class DupUp3D : public GGMLBlock {
protected:
int64_t in_channels;
int64_t out_channels;
int64_t factor_t;
int64_t factor_s;
int64_t factor;
int64_t repeats;
public:
DupUp3D(int64_t in_channels, int64_t out_channels, int64_t factor_t, int64_t factor_s = 1)
: in_channels(in_channels), out_channels(out_channels), factor_t(factor_t), factor_s(factor_s) {
factor = factor_t * factor_s * factor_s;
GGML_ASSERT(out_channels * factor % in_channels == 0);
repeats = out_channels * factor / in_channels;
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
bool first_chunk = false,
int64_t B = 1) {
// x: [B*IC, T, H, W]
// return: [B*OC, T/factor_t, H/factor_s, W/factor_s]
GGML_ASSERT(B == 1);
int64_t C = x->ne[3];
int64_t T = x->ne[2];
int64_t H = x->ne[1];
int64_t W = x->ne[0];
auto x_ = x;
for (int64_t i = 1; i < repeats; i++) {
x = ggml_concat(ctx, x, x_, 2);
}
C = out_channels;
x = ggml_reshape_4d(ctx, x, W, H * T, factor_s, factor_s * factor_t * C); // [C*factor_t*factor_s, factor_s, T*H, W]
x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3)); // [C*factor_t*factor_s, T*H, W, factor_s]
x = ggml_reshape_4d(ctx, x, factor_s * W, H * T, factor_s, factor_t * C); // [C*factor_t, factor_s, T*H, W*factor_s]
x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [C*factor_t, T*H, factor_s, W*factor_s]
x = ggml_reshape_4d(ctx, x, factor_s * W * factor_s * H, T, factor_t, C); // [C, factor_t, T, H*factor_s*W*factor_s]
x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [C, T, factor_t, H*factor_s*W*factor_s]
x = ggml_reshape_4d(ctx, x, factor_s * W, factor_s * H, factor_t * T, C); // [C, T*factor_t, H*factor_s, W*factor_s]
if (first_chunk) {
x = ggml_ext_slice(ctx, x, 2, factor_t - 1, x->ne[2]);
}
return x;
}
};
class ResidualBlock : public GGMLBlock {
protected:
int64_t in_dim;
int64_t out_dim;
public:
ResidualBlock(int64_t in_dim, int64_t out_dim)
: in_dim(in_dim), out_dim(out_dim) {
blocks["residual.0"] = std::shared_ptr<GGMLBlock>(new RMS_norm(in_dim));
// residual.1 is nn.SiLU()
blocks["residual.2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(in_dim, out_dim, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
blocks["residual.3"] = std::shared_ptr<GGMLBlock>(new RMS_norm(out_dim));
// residual.4 is nn.SiLU()
// residual.5 is nn.Dropout()
blocks["residual.6"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(out_dim, out_dim, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
if (in_dim != out_dim) {
blocks["shortcut"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(in_dim, out_dim, {1, 1, 1}));
}
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b,
std::vector<struct ggml_tensor*>& feat_cache,
int& feat_idx) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
struct ggml_tensor* h = x;
if (in_dim != out_dim) {
auto shortcut = std::dynamic_pointer_cast<CausalConv3d>(blocks["shortcut"]);
h = shortcut->forward(ctx, x);
}
for (int i = 0; i < 7; i++) {
if (i == 0 || i == 3) { // RMS_norm
auto layer = std::dynamic_pointer_cast<RMS_norm>(blocks["residual." + std::to_string(i)]);
x = layer->forward(ctx, x);
} else if (i == 2 || i == 6) { // CausalConv3d
auto layer = std::dynamic_pointer_cast<CausalConv3d>(blocks["residual." + std::to_string(i)]);
if (feat_cache.size() > 0) {
int idx = feat_idx;
auto cache_x = ggml_ext_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
// cache last frame of last two chunk
cache_x = ggml_concat(ctx,
ggml_ext_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
cache_x,
2);
}
x = layer->forward(ctx, x, feat_cache[idx]);
feat_cache[idx] = cache_x;
feat_idx += 1;
}
} else if (i == 1 || i == 4) {
x = ggml_silu(ctx, x);
} else { // i == 5
// nn.Dropout(), ignore
}
}
x = ggml_add(ctx, x, h);
return x;
}
};
class Down_ResidualBlock : public GGMLBlock {
protected:
int mult;
bool down_flag;
public:
Down_ResidualBlock(int64_t in_dim,
int64_t out_dim,
int mult,
bool temperal_downsample = false,
bool down_flag = false)
: mult(mult), down_flag(down_flag) {
blocks["avg_shortcut"] = std::shared_ptr<GGMLBlock>(new AvgDown3D(in_dim, out_dim, temperal_downsample ? 2 : 1, down_flag ? 2 : 1));
int i = 0;
for (; i < mult; i++) {
blocks["downsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new ResidualBlock(in_dim, out_dim));
in_dim = out_dim;
}
if (down_flag) {
std::string mode = temperal_downsample ? "downsample3d" : "downsample2d";
blocks["downsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode, true));
i++;
}
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b,
std::vector<struct ggml_tensor*>& feat_cache,
int& feat_idx,
int chunk_idx) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
struct ggml_tensor* x_copy = x;
auto avg_shortcut = std::dynamic_pointer_cast<AvgDown3D>(blocks["avg_shortcut"]);
int i = 0;
for (; i < mult; i++) {
std::string block_name = "downsamples." + std::to_string(i);
auto block = std::dynamic_pointer_cast<ResidualBlock>(blocks[block_name]);
x = block->forward(ctx, x, b, feat_cache, feat_idx);
}
if (down_flag) {
std::string block_name = "downsamples." + std::to_string(i);
auto block = std::dynamic_pointer_cast<Resample>(blocks[block_name]);
x = block->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
}
auto shortcut = avg_shortcut->forward(ctx, x_copy, b);
x = ggml_add(ctx, x, shortcut);
return x;
}
};
class Up_ResidualBlock : public GGMLBlock {
protected:
int mult;
bool up_flag;
public:
Up_ResidualBlock(int64_t in_dim,
int64_t out_dim,
int mult,
bool temperal_upsample = false,
bool up_flag = false)
: mult(mult), up_flag(up_flag) {
if (up_flag) {
blocks["avg_shortcut"] = std::shared_ptr<GGMLBlock>(new DupUp3D(in_dim, out_dim, temperal_upsample ? 2 : 1, up_flag ? 2 : 1));
}
int i = 0;
for (; i < mult; i++) {
blocks["upsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new ResidualBlock(in_dim, out_dim));
in_dim = out_dim;
}
if (up_flag) {
std::string mode = temperal_upsample ? "upsample3d" : "upsample2d";
blocks["upsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode, true));
i++;
}
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b,
std::vector<struct ggml_tensor*>& feat_cache,
int& feat_idx,
int chunk_idx) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
struct ggml_tensor* x_copy = x;
int i = 0;
for (; i < mult; i++) {
std::string block_name = "upsamples." + std::to_string(i);
auto block = std::dynamic_pointer_cast<ResidualBlock>(blocks[block_name]);
x = block->forward(ctx, x, b, feat_cache, feat_idx);
}
if (up_flag) {
std::string block_name = "upsamples." + std::to_string(i);
auto block = std::dynamic_pointer_cast<Resample>(blocks[block_name]);
x = block->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
auto avg_shortcut = std::dynamic_pointer_cast<DupUp3D>(blocks["avg_shortcut"]);
auto shortcut = avg_shortcut->forward(ctx, x_copy, chunk_idx == 0, b);
x = ggml_add(ctx, x, shortcut);
}
return x;
}
};
class AttentionBlock : public GGMLBlock {
protected:
int64_t dim;
public:
AttentionBlock(int64_t dim)
: dim(dim) {
blocks["norm"] = std::shared_ptr<GGMLBlock>(new RMS_norm(dim));
blocks["to_qkv"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim * 3, {1, 1}));
blocks["proj"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {1, 1}));
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
auto norm = std::dynamic_pointer_cast<RMS_norm>(blocks["norm"]);
auto to_qkv = std::dynamic_pointer_cast<Conv2d>(blocks["to_qkv"]);
auto proj = std::dynamic_pointer_cast<Conv2d>(blocks["proj"]);
auto identity = x;
x = norm->forward(ctx, x);
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2)); // (t, c, h, w)
const int64_t n = x->ne[3];
const int64_t c = x->ne[2];
const int64_t h = x->ne[1];
const int64_t w = x->ne[0];
auto qkv = to_qkv->forward(ctx, x);
auto qkv_vec = split_image_qkv(ctx, qkv);
auto q = qkv_vec[0];
q = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, q, 2, 0, 1, 3)); // [t, h, w, c]
q = ggml_reshape_3d(ctx, q, c, h * w, n); // [t, h * w, c]
auto k = qkv_vec[1];
k = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, k, 2, 0, 1, 3)); // [t, h, w, c]
k = ggml_reshape_3d(ctx, k, c, h * w, n); // [t, h * w, c]
auto v = qkv_vec[2];
v = ggml_reshape_3d(ctx, v, h * w, c, n); // [t, c, h * w]
x = ggml_ext_attention(ctx, q, k, v, false); // [t, h * w, c]
// v = ggml_cont(ctx, ggml_ext_torch_permute(ctx, v, 1, 0, 2, 3)); // [t, h * w, c]
// x = ggml_ext_attention_ext(ctx, q, k, v, q->ne[2], nullptr, false, false, true);
x = ggml_ext_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3)); // [t, c, h * w]
x = ggml_reshape_4d(ctx, x, w, h, c, n); // [t, c, h, w]
x = proj->forward(ctx, x);
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2)); // (c, t, h, w)
x = ggml_add(ctx, x, identity);
return x;
}
};
class Encoder3d : public GGMLBlock {
protected:
bool wan2_2;
int64_t dim;
int64_t z_dim;
std::vector<int> dim_mult;
int num_res_blocks;
std::vector<bool> temperal_downsample;
public:
Encoder3d(int64_t dim = 128,
int64_t z_dim = 4,
std::vector<int> dim_mult = {1, 2, 4, 4},
int num_res_blocks = 2,
std::vector<bool> temperal_downsample = {false, true, true},
bool wan2_2 = false)
: dim(dim),
z_dim(z_dim),
dim_mult(dim_mult),
num_res_blocks(num_res_blocks),
temperal_downsample(temperal_downsample),
wan2_2(wan2_2) {
// attn_scales is always []
std::vector<int64_t> dims = {dim};
for (int u : dim_mult) {
dims.push_back(dim * u);
}
if (wan2_2) {
blocks["conv1"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(12, dims[0], {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
} else {
blocks["conv1"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(3, dims[0], {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
}
int index = 0;
int64_t in_dim;
int64_t out_dim;
for (int i = 0; i < dims.size() - 1; i++) {
in_dim = dims[i];
out_dim = dims[i + 1];
if (wan2_2) {
bool t_down_flag = i < temperal_downsample.size() ? temperal_downsample[i] : false;
auto block = std::shared_ptr<GGMLBlock>(new Down_ResidualBlock(in_dim,
out_dim,
num_res_blocks,
t_down_flag,
i != dim_mult.size() - 1));
blocks["downsamples." + std::to_string(index++)] = block;
} else {
for (int j = 0; j < num_res_blocks; j++) {
auto block = std::shared_ptr<GGMLBlock>(new ResidualBlock(in_dim, out_dim));
blocks["downsamples." + std::to_string(index++)] = block;
in_dim = out_dim;
}
if (i != dim_mult.size() - 1) {
std::string mode = temperal_downsample[i] ? "downsample3d" : "downsample2d";
auto block = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode));
blocks["downsamples." + std::to_string(index++)] = block;
}
}
}
blocks["middle.0"] = std::shared_ptr<GGMLBlock>(new ResidualBlock(out_dim, out_dim));
blocks["middle.1"] = std::shared_ptr<GGMLBlock>(new AttentionBlock(out_dim));
blocks["middle.2"] = std::shared_ptr<GGMLBlock>(new ResidualBlock(out_dim, out_dim));
blocks["head.0"] = std::shared_ptr<GGMLBlock>(new RMS_norm(out_dim));
// head.1 is nn.SiLU()
blocks["head.2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(out_dim, z_dim, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b,
std::vector<struct ggml_tensor*>& feat_cache,
int& feat_idx,
int chunk_idx) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
auto conv1 = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
auto middle_0 = std::dynamic_pointer_cast<ResidualBlock>(blocks["middle.0"]);
auto middle_1 = std::dynamic_pointer_cast<AttentionBlock>(blocks["middle.1"]);
auto middle_2 = std::dynamic_pointer_cast<ResidualBlock>(blocks["middle.2"]);
auto head_0 = std::dynamic_pointer_cast<RMS_norm>(blocks["head.0"]);
auto head_2 = std::dynamic_pointer_cast<CausalConv3d>(blocks["head.2"]);
// conv1
if (feat_cache.size() > 0) {
int idx = feat_idx;
auto cache_x = ggml_ext_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
// cache last frame of last two chunk
cache_x = ggml_concat(ctx,
ggml_ext_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
cache_x,
2);
}
x = conv1->forward(ctx, x, feat_cache[idx]);
feat_cache[idx] = cache_x;
feat_idx += 1;
} else {
x = conv1->forward(ctx, x);
}
// downsamples
std::vector<int64_t> dims = {dim};
for (int u : dim_mult) {
dims.push_back(dim * u);
}
int index = 0;
for (int i = 0; i < dims.size() - 1; i++) {
if (wan2_2) {
auto layer = std::dynamic_pointer_cast<Down_ResidualBlock>(blocks["downsamples." + std::to_string(index++)]);
x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
} else {
for (int j = 0; j < num_res_blocks; j++) {
auto layer = std::dynamic_pointer_cast<ResidualBlock>(blocks["downsamples." + std::to_string(index++)]);
x = layer->forward(ctx, x, b, feat_cache, feat_idx);
}
if (i != dim_mult.size() - 1) {
auto layer = std::dynamic_pointer_cast<Resample>(blocks["downsamples." + std::to_string(index++)]);
x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
}
}
}
// middle
x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
x = middle_1->forward(ctx, x, b);
x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
// head
x = head_0->forward(ctx, x);
x = ggml_silu(ctx, x);
if (feat_cache.size() > 0) {
int idx = feat_idx;
auto cache_x = ggml_ext_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
// cache last frame of last two chunk
cache_x = ggml_concat(ctx,
ggml_ext_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
cache_x,
2);
}
x = head_2->forward(ctx, x, feat_cache[idx]);
feat_cache[idx] = cache_x;
feat_idx += 1;
} else {
x = head_2->forward(ctx, x);
}
return x;
}
};
class Decoder3d : public GGMLBlock {
protected:
bool wan2_2;
int64_t dim;
int64_t z_dim;
std::vector<int> dim_mult;
int num_res_blocks;
std::vector<bool> temperal_upsample;
public:
Decoder3d(int64_t dim = 128,
int64_t z_dim = 4,
std::vector<int> dim_mult = {1, 2, 4, 4},
int num_res_blocks = 2,
std::vector<bool> temperal_upsample = {true, true, false},
bool wan2_2 = false)
: dim(dim),
z_dim(z_dim),
dim_mult(dim_mult),
num_res_blocks(num_res_blocks),
temperal_upsample(temperal_upsample),
wan2_2(wan2_2) {
// attn_scales is always []
std::vector<int64_t> dims = {dim_mult[dim_mult.size() - 1] * dim};
for (int i = static_cast<int>(dim_mult.size()) - 1; i >= 0; i--) {
dims.push_back(dim * dim_mult[i]);
}
// init block
blocks["conv1"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim, dims[0], {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
// middle blocks
blocks["middle.0"] = std::shared_ptr<GGMLBlock>(new ResidualBlock(dims[0], dims[0]));
blocks["middle.1"] = std::shared_ptr<GGMLBlock>(new AttentionBlock(dims[0]));
blocks["middle.2"] = std::shared_ptr<GGMLBlock>(new ResidualBlock(dims[0], dims[0]));
// upsample blocks
int index = 0;
int64_t in_dim;
int64_t out_dim;
for (int i = 0; i < dims.size() - 1; i++) {
in_dim = dims[i];
out_dim = dims[i + 1];
if (wan2_2) {
bool t_up_flag = i < temperal_upsample.size() ? temperal_upsample[i] : false;
auto block = std::shared_ptr<GGMLBlock>(new Up_ResidualBlock(in_dim,
out_dim,
num_res_blocks + 1,
t_up_flag,
i != dim_mult.size() - 1));
blocks["upsamples." + std::to_string(index++)] = block;
} else {
if (i == 1 || i == 2 || i == 3) {
in_dim = in_dim / 2;
}
for (int j = 0; j < num_res_blocks + 1; j++) {
auto block = std::shared_ptr<GGMLBlock>(new ResidualBlock(in_dim, out_dim));
blocks["upsamples." + std::to_string(index++)] = block;
in_dim = out_dim;
}
if (i != dim_mult.size() - 1) {
std::string mode = temperal_upsample[i] ? "upsample3d" : "upsample2d";
auto block = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode));
blocks["upsamples." + std::to_string(index++)] = block;
}
}
}
// output blocks
blocks["head.0"] = std::shared_ptr<GGMLBlock>(new RMS_norm(out_dim));
// head.1 is nn.SiLU()
if (wan2_2) {
blocks["head.2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(out_dim, 12, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
} else {
blocks["head.2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(out_dim, 3, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
}
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b,
std::vector<struct ggml_tensor*>& feat_cache,
int& feat_idx,
int chunk_idx) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
auto conv1 = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
auto middle_0 = std::dynamic_pointer_cast<ResidualBlock>(blocks["middle.0"]);
auto middle_1 = std::dynamic_pointer_cast<AttentionBlock>(blocks["middle.1"]);
auto middle_2 = std::dynamic_pointer_cast<ResidualBlock>(blocks["middle.2"]);
auto head_0 = std::dynamic_pointer_cast<RMS_norm>(blocks["head.0"]);
auto head_2 = std::dynamic_pointer_cast<CausalConv3d>(blocks["head.2"]);
// conv1
if (feat_cache.size() > 0) {
int idx = feat_idx;
auto cache_x = ggml_ext_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
// cache last frame of last two chunk
cache_x = ggml_concat(ctx,
ggml_ext_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
cache_x,
2);
}
x = conv1->forward(ctx, x, feat_cache[idx]);
feat_cache[idx] = cache_x;
feat_idx += 1;
} else {
x = conv1->forward(ctx, x);
}
// middle
x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
x = middle_1->forward(ctx, x, b);
x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
// upsamples
std::vector<int64_t> dims = {dim_mult[dim_mult.size() - 1] * dim};
for (int i = static_cast<int>(dim_mult.size()) - 1; i >= 0; i--) {
dims.push_back(dim * dim_mult[i]);
}
int index = 0;
for (int i = 0; i < dims.size() - 1; i++) {
if (wan2_2) {
auto layer = std::dynamic_pointer_cast<Up_ResidualBlock>(blocks["upsamples." + std::to_string(index++)]);
x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
} else {
for (int j = 0; j < num_res_blocks + 1; j++) {
auto layer = std::dynamic_pointer_cast<ResidualBlock>(blocks["upsamples." + std::to_string(index++)]);
x = layer->forward(ctx, x, b, feat_cache, feat_idx);
}
if (i != dim_mult.size() - 1) {
auto layer = std::dynamic_pointer_cast<Resample>(blocks["upsamples." + std::to_string(index++)]);
x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
}
}
}
// head
x = head_0->forward(ctx, x);
x = ggml_silu(ctx, x);
if (feat_cache.size() > 0) {
int idx = feat_idx;
auto cache_x = ggml_ext_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
// cache last frame of last two chunk
cache_x = ggml_concat(ctx,
ggml_ext_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
cache_x,
2);
}
x = head_2->forward(ctx, x, feat_cache[idx]);
feat_cache[idx] = cache_x;
feat_idx += 1;
} else {
x = head_2->forward(ctx, x);
}
return x;
}
};
class WanVAE : public GGMLBlock {
public:
bool wan2_2 = false;
bool decode_only = true;
int64_t dim = 96;
int64_t dec_dim = 96;
int64_t z_dim = 16;
std::vector<int> dim_mult = {1, 2, 4, 4};
int num_res_blocks = 2;
std::vector<bool> temperal_upsample = {true, true, false};
std::vector<bool> temperal_downsample = {false, true, true};
int _conv_num = 33;
int _conv_idx = 0;
std::vector<struct ggml_tensor*> _feat_map;
int _enc_conv_num = 28;
int _enc_conv_idx = 0;
std::vector<struct ggml_tensor*> _enc_feat_map;
void clear_cache() {
_conv_idx = 0;
_feat_map = std::vector<struct ggml_tensor*>(_conv_num, nullptr);
_enc_conv_idx = 0;
_enc_feat_map = std::vector<struct ggml_tensor*>(_enc_conv_num, nullptr);
}
public:
WanVAE(bool decode_only = true, bool wan2_2 = false)
: decode_only(decode_only), wan2_2(wan2_2) {
// attn_scales is always []
if (wan2_2) {
dim = 160;
dec_dim = 256;
z_dim = 48;
_conv_num = 34;
_enc_conv_num = 26;
}
if (!decode_only) {
blocks["encoder"] = std::shared_ptr<GGMLBlock>(new Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks, temperal_downsample, wan2_2));
blocks["conv1"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim * 2, z_dim * 2, {1, 1, 1}));
}
blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder3d(dec_dim, z_dim, dim_mult, num_res_blocks, temperal_upsample, wan2_2));
blocks["conv2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim, z_dim, {1, 1, 1}));
}
struct ggml_tensor* patchify(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t patch_size,
int64_t b = 1) {
// x: [b*c, f, h*q, w*r]
// return: [b*c*r*q, f, h, w]
if (patch_size == 1) {
return x;
}
int64_t r = patch_size;
int64_t q = patch_size;
int64_t c = x->ne[3] / b;
int64_t f = x->ne[2];
int64_t h = x->ne[1] / q;
int64_t w = x->ne[0] / r;
x = ggml_reshape_4d(ctx, x, r * w, q, h, f * c * b); // [b*c*f, h, q, w*r]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [b*c*f, q, h, w*r]
x = ggml_reshape_4d(ctx, x, r, w, h * q, f * c * b); // [b*c*f, q*h, w, r]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3)); // [b*c*f, r, q*h, w]
x = ggml_reshape_4d(ctx, x, w * h, q * r, f, c * b); // [b*c, f, r*q, h*w]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [b*c, r*q, f, h*w]
x = ggml_reshape_4d(ctx, x, w, h, f, q * r * c * b); // [b*c*r*q, f, h, w]
return x;
}
struct ggml_tensor* unpatchify(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t patch_size,
int64_t b = 1) {
// x: [b*c*r*q, f, h, w]
// return: [b*c, f, h*q, w*r]
if (patch_size == 1) {
return x;
}
int64_t r = patch_size;
int64_t q = patch_size;
int64_t c = x->ne[3] / b / q / r;
int64_t f = x->ne[2];
int64_t h = x->ne[1];
int64_t w = x->ne[0];
x = ggml_reshape_4d(ctx, x, w * h, f, q * r, c * b); // [b*c, r*q, f, h*w]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [b*c, f, r*q, h*w]
x = ggml_reshape_4d(ctx, x, w, h * q, r, f * c * b); // [b*c*f, r, q*h, w]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3)); // [b*c*f, q*h, w, r]
x = ggml_reshape_4d(ctx, x, r * w, h, q, f * c * b); // [b*c*f, q, h, w*r]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [b*c*f, h, q, w*r]
x = ggml_reshape_4d(ctx, x, r * w, q * h, f, c * b); // [b*c, f, h*q, w*r]
return x;
}
struct ggml_tensor* encode(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t b = 1) {
// x: [b*c, t, h, w]
GGML_ASSERT(b == 1);
GGML_ASSERT(decode_only == false);
clear_cache();
if (wan2_2) {
x = patchify(ctx, x, 2, b);
}
auto encoder = std::dynamic_pointer_cast<Encoder3d>(blocks["encoder"]);
auto conv1 = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
int64_t t = x->ne[2];
int64_t iter_ = 1 + (t - 1) / 4;
struct ggml_tensor* out;
for (int i = 0; i < iter_; i++) {
_enc_conv_idx = 0;
if (i == 0) {
auto in = ggml_ext_slice(ctx, x, 2, 0, 1); // [b*c, 1, h, w]
out = encoder->forward(ctx, in, b, _enc_feat_map, _enc_conv_idx, i);
} else {
auto in = ggml_ext_slice(ctx, x, 2, 1 + 4 * (i - 1), 1 + 4 * i); // [b*c, 4, h, w]
auto out_ = encoder->forward(ctx, in, b, _enc_feat_map, _enc_conv_idx, i);
out = ggml_concat(ctx, out, out_, 2);
}
}
out = conv1->forward(ctx, out);
auto mu = ggml_ext_chunk(ctx, out, 2, 3)[0];
clear_cache();
return mu;
}
struct ggml_tensor* decode(struct ggml_context* ctx,
struct ggml_tensor* z,
int64_t b = 1) {
// z: [b*c, t, h, w]
GGML_ASSERT(b == 1);
clear_cache();
auto decoder = std::dynamic_pointer_cast<Decoder3d>(blocks["decoder"]);
auto conv2 = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv2"]);
int64_t iter_ = z->ne[2];
auto x = conv2->forward(ctx, z);
struct ggml_tensor* out;
for (int64_t i = 0; i < iter_; i++) {
_conv_idx = 0;
if (i == 0) {
auto in = ggml_ext_slice(ctx, x, 2, i, i + 1); // [b*c, 1, h, w]
out = decoder->forward(ctx, in, b, _feat_map, _conv_idx, i);
} else {
auto in = ggml_ext_slice(ctx, x, 2, i, i + 1); // [b*c, 1, h, w]
auto out_ = decoder->forward(ctx, in, b, _feat_map, _conv_idx, i);
out = ggml_concat(ctx, out, out_, 2);
}
}
if (wan2_2) {
out = unpatchify(ctx, out, 2, b);
}
clear_cache();
return out;
}
struct ggml_tensor* decode_partial(struct ggml_context* ctx,
struct ggml_tensor* z,
int64_t i,
int64_t b = 1) {
// z: [b*c, t, h, w]
GGML_ASSERT(b == 1);
auto decoder = std::dynamic_pointer_cast<Decoder3d>(blocks["decoder"]);
auto conv2 = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv2"]);
auto x = conv2->forward(ctx, z);
auto in = ggml_ext_slice(ctx, x, 2, i, i + 1); // [b*c, 1, h, w]
_conv_idx = 0;
auto out = decoder->forward(ctx, in, b, _feat_map, _conv_idx, i);
if (wan2_2) {
out = unpatchify(ctx, out, 2, b);
}
return out;
}
};
struct WanVAERunner : public VAE {
bool decode_only = true;
WanVAE ae;
WanVAERunner(ggml_backend_t backend,
bool offload_params_to_cpu,
const String2GGMLType& tensor_types = {},
const std::string prefix = "",
bool decode_only = false,
SDVersion version = VERSION_WAN2)
: decode_only(decode_only), ae(decode_only, version == VERSION_WAN2_2_TI2V), VAE(backend, offload_params_to_cpu) {
ae.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() override {
return "wan_vae";
}
void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {
ae.get_param_tensors(tensors, prefix);
}
struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, 10240 * z->ne[2], false);
z = to_backend(z);
struct ggml_tensor* out = decode_graph ? ae.decode(compute_ctx, z) : ae.encode(compute_ctx, z);
ggml_build_forward_expand(gf, out);
return gf;
}
struct ggml_cgraph* build_graph_partial(struct ggml_tensor* z, bool decode_graph, int64_t i) {
struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, 20480, false);
ae.clear_cache();
for (int64_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) {
auto feat_cache = get_cache_tensor_by_name("feat_idx:" + std::to_string(feat_idx));
ae._feat_map[feat_idx] = feat_cache;
}
z = to_backend(z);
struct ggml_tensor* out = decode_graph ? ae.decode_partial(compute_ctx, z, i) : ae.encode(compute_ctx, z);
for (int64_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) {
ggml_tensor* feat_cache = ae._feat_map[feat_idx];
if (feat_cache != nullptr) {
cache("feat_idx:" + std::to_string(feat_idx), feat_cache);
ggml_build_forward_expand(gf, feat_cache);
}
}
ggml_build_forward_expand(gf, out);
return gf;
}
void compute(const int n_threads,
struct ggml_tensor* z,
bool decode_graph,
struct ggml_tensor** output,
struct ggml_context* output_ctx = nullptr) override {
if (true) {
auto get_graph = [&]() -> struct ggml_cgraph* {
return build_graph(z, decode_graph);
};
GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
} else { // chunk 1 result is weird
ae.clear_cache();
int64_t t = z->ne[2];
int64_t i = 0;
auto get_graph = [&]() -> struct ggml_cgraph* {
return build_graph_partial(z, decode_graph, i);
};
struct ggml_tensor* out = nullptr;
GGMLRunner::compute(get_graph, n_threads, true, &out, output_ctx);
ae.clear_cache();
if (t == 1) {
*output = out;
return;
}
*output = ggml_new_tensor_4d(output_ctx, GGML_TYPE_F32, out->ne[0], out->ne[1], (t - 1) * 4 + 1, out->ne[3]);
auto copy_to_output = [&]() {
for (int64_t i3 = 0; i3 < out->ne[3]; i3++) {
for (int64_t i2 = 0; i2 < out->ne[2]; i2++) {
for (int64_t i1 = 0; i1 < out->ne[1]; i1++) {
for (int64_t i0 = 0; i0 < out->ne[0]; i0++) {
float value = ggml_ext_tensor_get_f32(out, i0, i1, i2, i3);
int64_t offset = (i == 0) ? 0 : (1 + (i - 1) * 4);
ggml_ext_tensor_set_f32(*output, value, i0, i1, offset + i2, i3);
}
}
}
}
};
copy_to_output();
out = ggml_new_tensor_4d(output_ctx, GGML_TYPE_F32, out->ne[0], out->ne[1], 4, out->ne[3]);
for (i = 1; i < t; i++) {
GGMLRunner::compute(get_graph, n_threads, true, &out);
ae.clear_cache();
copy_to_output();
}
free_cache_ctx_and_buffer();
}
}
void test() {
struct ggml_init_params params;
params.mem_size = static_cast<size_t>(1024 * 1024) * 1024; // 1G
params.mem_buffer = nullptr;
params.no_alloc = false;
struct ggml_context* work_ctx = ggml_init(params);
GGML_ASSERT(work_ctx != nullptr);
if (true) {
// cpu f32, pass
// cpu f16, pass
// cuda f16, pass
// cuda f32, pass
auto z = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 104, 60, 2, 16);
ggml_set_f32(z, 0.5f);
z = load_tensor_from_file(work_ctx, "wan_vae_z.bin");
print_ggml_tensor(z);
struct ggml_tensor* out = nullptr;
int64_t t0 = ggml_time_ms();
compute(8, z, true, &out, work_ctx);
int64_t t1 = ggml_time_ms();
print_ggml_tensor(out);
LOG_DEBUG("decode test done in %ldms", t1 - t0);
}
};
static void load_from_file_and_test(const std::string& file_path) {
// ggml_backend_t backend = ggml_backend_cuda_init(0);
ggml_backend_t backend = ggml_backend_cpu_init();
ggml_type model_data_type = GGML_TYPE_F16;
std::shared_ptr<WanVAERunner> vae = std::make_shared<WanVAERunner>(backend, false, String2GGMLType{}, "", false, VERSION_WAN2_2_TI2V);
{
LOG_INFO("loading from '%s'", file_path.c_str());
vae->alloc_params_buffer();
std::map<std::string, ggml_tensor*> tensors;
vae->get_param_tensors(tensors, "first_stage_model");
ModelLoader model_loader;
if (!model_loader.init_from_file(file_path, "vae.")) {
LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
return;
}
bool success = model_loader.load_tensors(tensors);
if (!success) {
LOG_ERROR("load tensors from model loader failed");
return;
}
LOG_INFO("vae model loaded");
}
vae->test();
}
};
class WanSelfAttention : public GGMLBlock {
public:
int64_t num_heads;
int64_t head_dim;
bool flash_attn;
public:
WanSelfAttention(int64_t dim,
int64_t num_heads,
bool qk_norm = true,
float eps = 1e-6,
bool flash_attn = false)
: num_heads(num_heads), flash_attn(flash_attn) {
head_dim = dim / num_heads;
blocks["q"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
blocks["k"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
blocks["v"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
blocks["o"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
if (qk_norm) {
blocks["norm_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim, eps));
blocks["norm_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim, eps));
} else {
blocks["norm_q"] = std::shared_ptr<GGMLBlock>(new Identity());
blocks["norm_k"] = std::shared_ptr<GGMLBlock>(new Identity());
}
}
virtual struct ggml_tensor* forward(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* x,
struct ggml_tensor* pe,
struct ggml_tensor* mask = nullptr) {
// x: [N, n_token, dim]
// pe: [n_token, d_head/2, 2, 2]
// return [N, n_token, dim]
int64_t N = x->ne[2];
int64_t n_token = x->ne[1];
auto q_proj = std::dynamic_pointer_cast<Linear>(blocks["q"]);
auto k_proj = std::dynamic_pointer_cast<Linear>(blocks["k"]);
auto v_proj = std::dynamic_pointer_cast<Linear>(blocks["v"]);
auto o_proj = std::dynamic_pointer_cast<Linear>(blocks["o"]);
auto norm_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_q"]);
auto norm_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_k"]);
auto q = q_proj->forward(ctx, x);
q = norm_q->forward(ctx, q);
auto k = k_proj->forward(ctx, x);
k = norm_k->forward(ctx, k);
auto v = v_proj->forward(ctx, x); // [N, n_token, n_head*d_head]
q = ggml_reshape_4d(ctx, q, head_dim, num_heads, n_token, N); // [N, n_token, n_head, d_head]
k = ggml_reshape_4d(ctx, k, head_dim, num_heads, n_token, N); // [N, n_token, n_head, d_head]
v = ggml_reshape_4d(ctx, v, head_dim, num_heads, n_token, N); // [N, n_token, n_head, d_head]
x = Rope::attention(ctx, backend, q, k, v, pe, mask, flash_attn); // [N, n_token, dim]
x = o_proj->forward(ctx, x); // [N, n_token, dim]
return x;
}
};
class WanCrossAttention : public WanSelfAttention {
public:
WanCrossAttention(int64_t dim,
int64_t num_heads,
bool qk_norm = true,
float eps = 1e-6,
bool flash_attn = false)
: WanSelfAttention(dim, num_heads, qk_norm, eps, flash_attn) {}
virtual struct ggml_tensor* forward(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* x,
struct ggml_tensor* context,
int64_t context_img_len) = 0;
};
class WanT2VCrossAttention : public WanCrossAttention {
public:
WanT2VCrossAttention(int64_t dim,
int64_t num_heads,
bool qk_norm = true,
float eps = 1e-6,
bool flash_attn = false)
: WanCrossAttention(dim, num_heads, qk_norm, eps, flash_attn) {}
struct ggml_tensor* forward(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* x,
struct ggml_tensor* context,
int64_t context_img_len) override {
// x: [N, n_token, dim]
// context: [N, n_context, dim]
// context_img_len: unused
// return [N, n_token, dim]
int64_t N = x->ne[2];
int64_t n_token = x->ne[1];
auto q_proj = std::dynamic_pointer_cast<Linear>(blocks["q"]);
auto k_proj = std::dynamic_pointer_cast<Linear>(blocks["k"]);
auto v_proj = std::dynamic_pointer_cast<Linear>(blocks["v"]);
auto o_proj = std::dynamic_pointer_cast<Linear>(blocks["o"]);
auto norm_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_q"]);
auto norm_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_k"]);
auto q = q_proj->forward(ctx, x);
q = norm_q->forward(ctx, q);
auto k = k_proj->forward(ctx, context); // [N, n_context, dim]
k = norm_k->forward(ctx, k);
auto v = v_proj->forward(ctx, context); // [N, n_context, dim]
x = ggml_ext_attention_ext(ctx, backend, q, k, v, num_heads, nullptr, false, false, flash_attn); // [N, n_token, dim]
x = o_proj->forward(ctx, x); // [N, n_token, dim]
return x;
}
};
class WanI2VCrossAttention : public WanCrossAttention {
public:
WanI2VCrossAttention(int64_t dim,
int64_t num_heads,
bool qk_norm = true,
float eps = 1e-6,
bool flash_attn = false)
: WanCrossAttention(dim, num_heads, qk_norm, eps, flash_attn) {
blocks["k_img"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
blocks["v_img"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
if (qk_norm) {
blocks["norm_k_img"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim, eps));
} else {
blocks["norm_k_img"] = std::shared_ptr<GGMLBlock>(new Identity());
}
}
struct ggml_tensor* forward(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* x,
struct ggml_tensor* context,
int64_t context_img_len) override {
// x: [N, n_token, dim]
// context: [N, context_img_len + context_txt_len, dim]
// return [N, n_token, dim]
auto q_proj = std::dynamic_pointer_cast<Linear>(blocks["q"]);
auto k_proj = std::dynamic_pointer_cast<Linear>(blocks["k"]);
auto v_proj = std::dynamic_pointer_cast<Linear>(blocks["v"]);
auto o_proj = std::dynamic_pointer_cast<Linear>(blocks["o"]);
auto k_img_proj = std::dynamic_pointer_cast<Linear>(blocks["k_img"]);
auto v_img_proj = std::dynamic_pointer_cast<Linear>(blocks["v_img"]);
auto norm_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_q"]);
auto norm_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_k"]);
auto norm_k_img = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_k_img"]);
int64_t N = x->ne[2];
int64_t n_token = x->ne[1];
int64_t dim = x->ne[0];
int64_t context_txt_len = context->ne[1] - context_img_len;
context = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, context, 0, 2, 1, 3)); // [context_img_len + context_txt_len, N, dim]
auto context_img = ggml_view_3d(ctx, context, dim, N, context_img_len, context->nb[1], context->nb[2], 0);
auto context_txt = ggml_view_3d(ctx, context, dim, N, context_txt_len, context->nb[1], context->nb[2], context_img_len * context->nb[2]);
context_img = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, context_img, 0, 2, 1, 3)); // [N, context_img_len, dim]
context_txt = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, context_txt, 0, 2, 1, 3)); // [N, context_txt_len, dim]
auto q = q_proj->forward(ctx, x);
q = norm_q->forward(ctx, q);
auto k = k_proj->forward(ctx, context_txt); // [N, context_txt_len, dim]
k = norm_k->forward(ctx, k);
auto v = v_proj->forward(ctx, context_txt); // [N, context_txt_len, dim]
auto k_img = k_img_proj->forward(ctx, context_img); // [N, context_img_len, dim]
k_img = norm_k_img->forward(ctx, k_img);
auto v_img = v_img_proj->forward(ctx, context_img); // [N, context_img_len, dim]
auto img_x = ggml_ext_attention_ext(ctx, backend, q, k_img, v_img, num_heads, nullptr, false, false, flash_attn); // [N, n_token, dim]
x = ggml_ext_attention_ext(ctx, backend, q, k, v, num_heads, nullptr, false, false, flash_attn); // [N, n_token, dim]
x = ggml_add(ctx, x, img_x);
x = o_proj->forward(ctx, x); // [N, n_token, dim]
return x;
}
};
static struct ggml_tensor* modulate_add(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* e) {
// x: [N, n_token, dim]
// e: [N, 1, dim] or [N, T, 1, dim]
if (ggml_n_dims(e) == 3) {
int64_t T = e->ne[2];
x = ggml_reshape_4d(ctx, x, x->ne[0], x->ne[1] / T, T, x->ne[2]); // [N, T, n_token/T, dim]
x = ggml_add(ctx, x, e);
x = ggml_reshape_3d(ctx, x, x->ne[0], x->ne[1] * x->ne[2], x->ne[3]); // [N, n_token, dim]
} else {
x = ggml_add(ctx, x, e);
}
return x;
}
static struct ggml_tensor* modulate_mul(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* e) {
// x: [N, n_token, dim]
// e: [N, 1, dim] or [N, T, 1, dim]
if (ggml_n_dims(e) == 3) {
int64_t T = e->ne[2];
x = ggml_reshape_4d(ctx, x, x->ne[0], x->ne[1] / T, T, x->ne[2]); // [N, T, n_token/T, dim]
x = ggml_mul(ctx, x, e);
x = ggml_reshape_3d(ctx, x, x->ne[0], x->ne[1] * x->ne[2], x->ne[3]); // [N, n_token, dim]
} else {
x = ggml_mul(ctx, x, e);
}
return x;
}
class WanAttentionBlock : public GGMLBlock {
protected:
int dim;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 6, 1);
}
public:
WanAttentionBlock(bool t2v_cross_attn,
int64_t dim,
int64_t ffn_dim,
int64_t num_heads,
bool qk_norm = true,
bool cross_attn_norm = false,
float eps = 1e-6,
bool flash_attn = false)
: dim(dim) {
blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
blocks["self_attn"] = std::shared_ptr<GGMLBlock>(new WanSelfAttention(dim, num_heads, qk_norm, eps, flash_attn));
if (cross_attn_norm) {
blocks["norm3"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, true));
} else {
blocks["norm3"] = std::shared_ptr<GGMLBlock>(new Identity());
}
if (t2v_cross_attn) {
blocks["cross_attn"] = std::shared_ptr<GGMLBlock>(new WanT2VCrossAttention(dim, num_heads, qk_norm, eps, flash_attn));
} else {
blocks["cross_attn"] = std::shared_ptr<GGMLBlock>(new WanI2VCrossAttention(dim, num_heads, qk_norm, eps, flash_attn));
}
blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
blocks["ffn.0"] = std::shared_ptr<GGMLBlock>(new Linear(dim, ffn_dim));
// ffn.1 is nn.GELU(approximate='tanh')
blocks["ffn.2"] = std::shared_ptr<GGMLBlock>(new Linear(ffn_dim, dim));
}
virtual struct ggml_tensor* forward(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* x,
struct ggml_tensor* e,
struct ggml_tensor* pe,
struct ggml_tensor* context,
int64_t context_img_len = 257) {
// x: [N, n_token, dim]
// e: [N, 6, dim] or [N, T, 6, dim]
// context: [N, context_img_len + context_txt_len, dim]
// return [N, n_token, dim]
auto modulation = params["modulation"];
e = ggml_add(ctx, e, modulation); // [N, 6, dim] or [N, T, 6, dim]
auto es = ggml_ext_chunk(ctx, e, 6, 1); // ([N, 1, dim], ...) or [N, T, 1, dim]
auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
auto self_attn = std::dynamic_pointer_cast<WanSelfAttention>(blocks["self_attn"]);
auto norm3 = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm3"]);
auto cross_attn = std::dynamic_pointer_cast<WanCrossAttention>(blocks["cross_attn"]);
auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
auto ffn_0 = std::dynamic_pointer_cast<Linear>(blocks["ffn.0"]);
auto ffn_2 = std::dynamic_pointer_cast<Linear>(blocks["ffn.2"]);
// self-attention
auto y = norm1->forward(ctx, x);
y = ggml_add(ctx, y, modulate_mul(ctx, y, es[1]));
y = modulate_add(ctx, y, es[0]);
y = self_attn->forward(ctx, backend, y, pe);
x = ggml_add(ctx, x, modulate_mul(ctx, y, es[2]));
// cross-attention
x = ggml_add(ctx,
x,
cross_attn->forward(ctx, backend, norm3->forward(ctx, x), context, context_img_len));
// ffn
y = norm2->forward(ctx, x);
y = ggml_add(ctx, y, modulate_mul(ctx, y, es[4]));
y = modulate_add(ctx, y, es[3]);
y = ffn_0->forward(ctx, y);
y = ggml_gelu_inplace(ctx, y);
y = ffn_2->forward(ctx, y);
x = ggml_add(ctx, x, modulate_mul(ctx, y, es[5]));
return x;
}
};
class VaceWanAttentionBlock : public WanAttentionBlock {
protected:
int block_id;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 6, 1);
}
public:
VaceWanAttentionBlock(bool t2v_cross_attn,
int64_t dim,
int64_t ffn_dim,
int64_t num_heads,
bool qk_norm = true,
bool cross_attn_norm = false,
float eps = 1e-6,
int block_id = 0,
bool flash_attn = false)
: WanAttentionBlock(t2v_cross_attn, dim, ffn_dim, num_heads, qk_norm, cross_attn_norm, eps, flash_attn), block_id(block_id) {
if (block_id == 0) {
blocks["before_proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
}
blocks["after_proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
}
std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* c,
struct ggml_tensor* x,
struct ggml_tensor* e,
struct ggml_tensor* pe,
struct ggml_tensor* context,
int64_t context_img_len = 257) {
// x: [N, n_token, dim]
// e: [N, 6, dim] or [N, T, 6, dim]
// context: [N, context_img_len + context_txt_len, dim]
// return [N, n_token, dim]
if (block_id == 0) {
auto before_proj = std::dynamic_pointer_cast<Linear>(blocks["before_proj"]);
c = before_proj->forward(ctx, c);
c = ggml_add(ctx, c, x);
}
auto after_proj = std::dynamic_pointer_cast<Linear>(blocks["after_proj"]);
c = WanAttentionBlock::forward(ctx, backend, c, e, pe, context, context_img_len);
auto c_skip = after_proj->forward(ctx, c);
return {c_skip, c};
}
};
class Head : public GGMLBlock {
protected:
int dim;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 2, 1);
}
public:
Head(int64_t dim,
int64_t out_dim,
std::tuple<int, int, int> patch_size,
float eps = 1e-6)
: dim(dim) {
out_dim = out_dim * std::get<0>(patch_size) * std::get<1>(patch_size) * std::get<2>(patch_size);
blocks["norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
blocks["head"] = std::shared_ptr<GGMLBlock>(new Linear(dim, out_dim));
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* x,
struct ggml_tensor* e) {
// x: [N, n_token, dim]
// e: [N, dim] or [N, T, dim]
// return [N, n_token, out_dim]
auto modulation = params["modulation"];
e = ggml_reshape_4d(ctx, e, e->ne[0], 1, e->ne[1], e->ne[2]); // [N, 1, dim] or [N, T, 1, dim]
e = ggml_repeat_4d(ctx, e, e->ne[0], 2, e->ne[2], e->ne[3]); // [N, 2, dim] or [N, T, 2, dim]
e = ggml_add(ctx, e, modulation); // [N, 2, dim] or [N, T, 2, dim]
auto es = ggml_ext_chunk(ctx, e, 2, 1); // ([N, 1, dim], ...) or ([N, T, 1, dim], ...)
auto norm = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
auto head = std::dynamic_pointer_cast<Linear>(blocks["head"]);
x = norm->forward(ctx, x);
x = ggml_add(ctx, x, modulate_mul(ctx, x, es[1]));
x = modulate_add(ctx, x, es[0]);
x = head->forward(ctx, x);
return x;
}
};
class MLPProj : public GGMLBlock {
protected:
int in_dim;
int flf_pos_embed_token_number;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
if (flf_pos_embed_token_number > 0) {
params["emb_pos"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, in_dim, flf_pos_embed_token_number, 1);
}
}
public:
MLPProj(int64_t in_dim,
int64_t out_dim,
int64_t flf_pos_embed_token_number = 0)
: in_dim(in_dim), flf_pos_embed_token_number(flf_pos_embed_token_number) {
blocks["proj.0"] = std::shared_ptr<GGMLBlock>(new LayerNorm(in_dim));
blocks["proj.1"] = std::shared_ptr<GGMLBlock>(new Linear(in_dim, in_dim));
// proj.2 is nn.GELU()
blocks["proj.3"] = std::shared_ptr<GGMLBlock>(new Linear(in_dim, out_dim));
blocks["proj.4"] = std::shared_ptr<GGMLBlock>(new LayerNorm(out_dim));
}
struct ggml_tensor* forward(struct ggml_context* ctx,
struct ggml_tensor* image_embeds) {
if (flf_pos_embed_token_number > 0) {
auto emb_pos = params["emb_pos"];
auto a = ggml_ext_slice(ctx, image_embeds, 1, 0, emb_pos->ne[1]);
auto b = ggml_ext_slice(ctx, emb_pos, 1, 0, image_embeds->ne[1]);
image_embeds = ggml_add(ctx, a, b);
}
auto proj_0 = std::dynamic_pointer_cast<LayerNorm>(blocks["proj.0"]);
auto proj_1 = std::dynamic_pointer_cast<Linear>(blocks["proj.1"]);
auto proj_3 = std::dynamic_pointer_cast<Linear>(blocks["proj.3"]);
auto proj_4 = std::dynamic_pointer_cast<LayerNorm>(blocks["proj.4"]);
auto x = proj_0->forward(ctx, image_embeds);
x = proj_1->forward(ctx, x);
x = ggml_gelu_inplace(ctx, x);
x = proj_3->forward(ctx, x);
x = proj_4->forward(ctx, x);
return x; // clip_extra_context_tokens
}
};
struct WanParams {
std::string model_type = "t2v";
std::tuple<int, int, int> patch_size = {1, 2, 2};
int64_t text_len = 512;
int64_t in_dim = 16;
int64_t dim = 2048;
int64_t ffn_dim = 8192;
int64_t freq_dim = 256;
int64_t text_dim = 4096;
int64_t out_dim = 16;
int64_t num_heads = 16;
int64_t num_layers = 32;
int64_t vace_layers = 0;
int64_t vace_in_dim = 96;
std::map<int, int> vace_layers_mapping = {};
bool qk_norm = true;
bool cross_attn_norm = true;
float eps = 1e-6;
int64_t flf_pos_embed_token_number = 0;
int theta = 10000;
// wan2.1 1.3B: 1536/12, wan2.1/2.2 14B: 5120/40, wan2.2 5B: 3074/24
std::vector<int> axes_dim = {44, 42, 42};
int64_t axes_dim_sum = 128;
bool flash_attn = false;
};
class Wan : public GGMLBlock {
protected:
WanParams params;
public:
Wan() {}
Wan(WanParams params)
: params(params) {
// patch_embedding
blocks["patch_embedding"] = std::shared_ptr<GGMLBlock>(new Conv3d(params.in_dim, params.dim, params.patch_size, params.patch_size));
// text_embedding
blocks["text_embedding.0"] = std::shared_ptr<GGMLBlock>(new Linear(params.text_dim, params.dim));
// text_embedding.1 is nn.GELU()
blocks["text_embedding.2"] = std::shared_ptr<GGMLBlock>(new Linear(params.dim, params.dim));
// time_embedding
blocks["time_embedding.0"] = std::shared_ptr<GGMLBlock>(new Linear(params.freq_dim, params.dim));
// time_embedding.1 is nn.SiLU()
blocks["time_embedding.2"] = std::shared_ptr<GGMLBlock>(new Linear(params.dim, params.dim));
// time_projection.0 is nn.SiLU()
blocks["time_projection.1"] = std::shared_ptr<GGMLBlock>(new Linear(params.dim, params.dim * 6));
// blocks
for (int i = 0; i < params.num_layers; i++) {
auto block = std::shared_ptr<GGMLBlock>(new WanAttentionBlock(params.model_type == "t2v",
params.dim,
params.ffn_dim,
params.num_heads,
params.qk_norm,
params.cross_attn_norm,
params.eps,
params.flash_attn));
blocks["blocks." + std::to_string(i)] = block;
}
// head
blocks["head"] = std::shared_ptr<GGMLBlock>(new Head(params.dim, params.out_dim, params.patch_size, params.eps));
// img_emb
if (params.model_type == "i2v") {
blocks["img_emb"] = std::shared_ptr<GGMLBlock>(new MLPProj(1280, params.dim, params.flf_pos_embed_token_number));
}
// vace
if (params.vace_layers > 0) {
for (int i = 0; i < params.vace_layers; i++) {
auto block = std::shared_ptr<GGMLBlock>(new VaceWanAttentionBlock(params.model_type == "t2v",
params.dim,
params.ffn_dim,
params.num_heads,
params.qk_norm,
params.cross_attn_norm,
params.eps,
i,
params.flash_attn));
blocks["vace_blocks." + std::to_string(i)] = block;
}
int step = params.num_layers / params.vace_layers;
int n = 0;
for (int i = 0; i < params.num_layers; i += step) {
this->params.vace_layers_mapping[i] = n;
n++;
}
blocks["vace_patch_embedding"] = std::shared_ptr<GGMLBlock>(new Conv3d(params.vace_in_dim, params.dim, params.patch_size, params.patch_size));
}
}
struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
struct ggml_tensor* x) {
int64_t W = x->ne[0];
int64_t H = x->ne[1];
int64_t T = x->ne[2];
int pad_t = (std::get<0>(params.patch_size) - T % std::get<0>(params.patch_size)) % std::get<0>(params.patch_size);
int pad_h = (std::get<1>(params.patch_size) - H % std::get<1>(params.patch_size)) % std::get<1>(params.patch_size);
int pad_w = (std::get<2>(params.patch_size) - W % std::get<2>(params.patch_size)) % std::get<2>(params.patch_size);
x = ggml_pad(ctx, x, pad_w, pad_h, pad_t, 0); // [N*C, T + pad_t, H + pad_h, W + pad_w]
return x;
}
struct ggml_tensor* unpatchify(struct ggml_context* ctx,
struct ggml_tensor* x,
int64_t t_len,
int64_t h_len,
int64_t w_len) {
// x: [N, t_len*h_len*w_len, pt*ph*pw*C]
// return: [N*C, t_len*pt, h_len*ph, w_len*pw]
int64_t N = x->ne[3];
int64_t pt = std::get<0>(params.patch_size);
int64_t ph = std::get<1>(params.patch_size);
int64_t pw = std::get<2>(params.patch_size);
int64_t C = x->ne[0] / pt / ph / pw;
GGML_ASSERT(C * pt * ph * pw == x->ne[0]);
x = ggml_reshape_4d(ctx, x, C, pw * ph * pt, w_len * h_len * t_len, N); // [N, t_len*h_len*w_len, pt*ph*pw, C]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3)); // [N, C, t_len*h_len*w_len, pt*ph*pw]
x = ggml_reshape_4d(ctx, x, pw, ph * pt, w_len, h_len * t_len * C * N); // [N*C*t_len*h_len, w_len, pt*ph, pw]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [N*C*t_len*h_len, pt*ph, w_len, pw]
x = ggml_reshape_4d(ctx, x, pw * w_len, ph, pt, h_len * t_len * C * N); // [N*C*t_len*h_len, pt, ph, w_len*pw]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [N*C*t_len*h_len, ph, pt, w_len*pw]
x = ggml_reshape_4d(ctx, x, pw * w_len, pt, ph * h_len, t_len * C * N); // [N*C*t_len, h_len*ph, pt, w_len*pw]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3)); // [N*C*t_len, pt, h_len*ph, w_len*pw]
x = ggml_reshape_4d(ctx, x, pw * w_len, ph * h_len, pt * t_len, C * N); // [N*C, t_len*pt, h_len*ph, w_len*pw]
return x;
}
struct ggml_tensor* forward_orig(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* x,
struct ggml_tensor* timestep,
struct ggml_tensor* context,
struct ggml_tensor* pe,
struct ggml_tensor* clip_fea = nullptr,
struct ggml_tensor* vace_context = nullptr,
float vace_strength = 1.f,
int64_t N = 1) {
// x: [N*C, T, H, W], C => in_dim
// vace_context: [N*vace_in_dim, T, H, W]
// timestep: [N,] or [T]
// context: [N, L, text_dim]
// return: [N, t_len*h_len*w_len, out_dim*pt*ph*pw]
GGML_ASSERT(N == 1);
auto patch_embedding = std::dynamic_pointer_cast<Conv3d>(blocks["patch_embedding"]);
auto text_embedding_0 = std::dynamic_pointer_cast<Linear>(blocks["text_embedding.0"]);
auto text_embedding_2 = std::dynamic_pointer_cast<Linear>(blocks["text_embedding.2"]);
auto time_embedding_0 = std::dynamic_pointer_cast<Linear>(blocks["time_embedding.0"]);
auto time_embedding_2 = std::dynamic_pointer_cast<Linear>(blocks["time_embedding.2"]);
auto time_projection_1 = std::dynamic_pointer_cast<Linear>(blocks["time_projection.1"]);
auto head = std::dynamic_pointer_cast<Head>(blocks["head"]);
// patch_embedding
x = patch_embedding->forward(ctx, x); // [N*dim, t_len, h_len, w_len]
x = ggml_reshape_3d(ctx, x, x->ne[0] * x->ne[1] * x->ne[2], x->ne[3] / N, N); // [N, dim, t_len*h_len*w_len]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 0, 2, 3)); // [N, t_len*h_len*w_len, dim]
// time_embedding
auto e = ggml_ext_timestep_embedding(ctx, timestep, params.freq_dim);
e = time_embedding_0->forward(ctx, e);
e = ggml_silu_inplace(ctx, e);
e = time_embedding_2->forward(ctx, e); // [N, dim] or [N, T, dim]
// time_projection
auto e0 = ggml_silu(ctx, e);
e0 = time_projection_1->forward(ctx, e0);
e0 = ggml_reshape_4d(ctx, e0, e0->ne[0] / 6, 6, e0->ne[1], e0->ne[2]); // [N, 6, dim] or [N, T, 6, dim]
context = text_embedding_0->forward(ctx, context);
context = ggml_gelu(ctx, context);
context = text_embedding_2->forward(ctx, context); // [N, context_txt_len, dim]
int64_t context_img_len = 0;
if (clip_fea != nullptr) {
if (params.model_type == "i2v") {
auto img_emb = std::dynamic_pointer_cast<MLPProj>(blocks["img_emb"]);
auto context_img = img_emb->forward(ctx, clip_fea); // [N, context_img_len, dim]
context = ggml_concat(ctx, context_img, context, 1); // [N, context_img_len + context_txt_len, dim]
}
context_img_len = clip_fea->ne[1]; // 257
}
// vace_patch_embedding
ggml_tensor* c = nullptr;
if (params.vace_layers > 0) {
auto vace_patch_embedding = std::dynamic_pointer_cast<Conv3d>(blocks["vace_patch_embedding"]);
c = vace_patch_embedding->forward(ctx, vace_context); // [N*dim, t_len, h_len, w_len]
c = ggml_reshape_3d(ctx, c, c->ne[0] * c->ne[1] * c->ne[2], c->ne[3] / N, N); // [N, dim, t_len*h_len*w_len]
c = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, c, 1, 0, 2, 3)); // [N, t_len*h_len*w_len, dim]
}
auto x_orig = x;
for (int i = 0; i < params.num_layers; i++) {
auto block = std::dynamic_pointer_cast<WanAttentionBlock>(blocks["blocks." + std::to_string(i)]);
x = block->forward(ctx, backend, x, e0, pe, context, context_img_len);
auto iter = params.vace_layers_mapping.find(i);
if (iter != params.vace_layers_mapping.end()) {
int n = iter->second;
auto vace_block = std::dynamic_pointer_cast<VaceWanAttentionBlock>(blocks["vace_blocks." + std::to_string(n)]);
auto result = vace_block->forward(ctx, backend, c, x_orig, e0, pe, context, context_img_len);
auto c_skip = result.first;
c = result.second;
c_skip = ggml_scale(ctx, c_skip, vace_strength);
x = ggml_add(ctx, x, c_skip);
}
}
x = head->forward(ctx, x, e); // [N, t_len*h_len*w_len, pt*ph*pw*out_dim]
return x;
}
struct ggml_tensor* forward(struct ggml_context* ctx,
ggml_backend_t backend,
struct ggml_tensor* x,
struct ggml_tensor* timestep,
struct ggml_tensor* context,
struct ggml_tensor* pe,
struct ggml_tensor* clip_fea = nullptr,
struct ggml_tensor* time_dim_concat = nullptr,
struct ggml_tensor* vace_context = nullptr,
float vace_strength = 1.f,
int64_t N = 1) {
// Forward pass of DiT.
// x: [N*C, T, H, W]
// timestep: [N,]
// context: [N, L, D]
// pe: [L, d_head/2, 2, 2]
// time_dim_concat: [N*C, T2, H, W]
// return: [N*C, T, H, W]
GGML_ASSERT(N == 1);
int64_t W = x->ne[0];
int64_t H = x->ne[1];
int64_t T = x->ne[2];
int64_t C = x->ne[3];
x = pad_to_patch_size(ctx, x);
int64_t t_len = ((T + (std::get<0>(params.patch_size) / 2)) / std::get<0>(params.patch_size));
int64_t h_len = ((H + (std::get<1>(params.patch_size) / 2)) / std::get<1>(params.patch_size));
int64_t w_len = ((W + (std::get<2>(params.patch_size) / 2)) / std::get<2>(params.patch_size));
if (time_dim_concat != nullptr) {
time_dim_concat = pad_to_patch_size(ctx, time_dim_concat);
x = ggml_concat(ctx, x, time_dim_concat, 2); // [N*C, (T+pad_t) + (T2+pad_t2), H + pad_h, W + pad_w]
t_len = ((x->ne[2] + (std::get<0>(params.patch_size) / 2)) / std::get<0>(params.patch_size));
}
auto out = forward_orig(ctx, backend, x, timestep, context, pe, clip_fea, vace_context, vace_strength, N); // [N, t_len*h_len*w_len, pt*ph*pw*C]
out = unpatchify(ctx, out, t_len, h_len, w_len); // [N*C, (T+pad_t) + (T2+pad_t2), H + pad_h, W + pad_w]
// slice
out = ggml_ext_slice(ctx, out, 2, 0, T); // [N*C, T, H + pad_h, W + pad_w]
out = ggml_ext_slice(ctx, out, 1, 0, H); // [N*C, T, H, W + pad_w]
out = ggml_ext_slice(ctx, out, 0, 0, W); // [N*C, T, H, W]
return out;
}
};
struct WanRunner : public GGMLRunner {
public:
std::string desc = "wan";
WanParams wan_params;
Wan wan;
std::vector<float> pe_vec;
SDVersion version;
WanRunner(ggml_backend_t backend,
bool offload_params_to_cpu,
const String2GGMLType& tensor_types = {},
const std::string prefix = "",
SDVersion version = VERSION_WAN2,
bool flash_attn = false)
: GGMLRunner(backend, offload_params_to_cpu) {
wan_params.flash_attn = flash_attn;
wan_params.num_layers = 0;
for (auto pair : tensor_types) {
std::string tensor_name = pair.first;
if (tensor_name.find(prefix) == std::string::npos)
continue;
size_t pos = tensor_name.find("vace_blocks.");
if (pos != std::string::npos) {
tensor_name = tensor_name.substr(pos); // remove prefix
auto items = split_string(tensor_name, '.');
if (items.size() > 1) {
int block_index = atoi(items[1].c_str());
if (block_index + 1 > wan_params.vace_layers) {
wan_params.vace_layers = block_index + 1;
}
}
continue;
}
pos = tensor_name.find("blocks.");
if (pos != std::string::npos) {
tensor_name = tensor_name.substr(pos); // remove prefix
auto items = split_string(tensor_name, '.');
if (items.size() > 1) {
int block_index = atoi(items[1].c_str());
if (block_index + 1 > wan_params.num_layers) {
wan_params.num_layers = block_index + 1;
}
}
continue;
}
if (tensor_name.find("img_emb") != std::string::npos) {
wan_params.model_type = "i2v";
}
if (tensor_name.find("img_emb.emb_pos") != std::string::npos) {
wan_params.flf_pos_embed_token_number = 514;
}
}
if (wan_params.num_layers == 30) {
if (version == VERSION_WAN2_2_TI2V) {
desc = "Wan2.2-TI2V-5B";
wan_params.dim = 3072;
wan_params.eps = 1e-06;
wan_params.ffn_dim = 14336;
wan_params.freq_dim = 256;
wan_params.in_dim = 48;
wan_params.num_heads = 24;
wan_params.out_dim = 48;
wan_params.text_len = 512;
} else {
if (wan_params.vace_layers > 0) {
desc = "Wan2.1-VACE-1.3B";
} else {
desc = "Wan2.1-T2V-1.3B";
}
wan_params.dim = 1536;
wan_params.eps = 1e-06;
wan_params.ffn_dim = 8960;
wan_params.freq_dim = 256;
wan_params.in_dim = 16;
wan_params.num_heads = 12;
wan_params.out_dim = 16;
wan_params.text_len = 512;
}
} else if (wan_params.num_layers == 40) {
if (wan_params.model_type == "t2v") {
if (version == VERSION_WAN2_2_I2V) {
desc = "Wan2.2-I2V-14B";
wan_params.in_dim = 36;
} else {
if (wan_params.vace_layers > 0) {
desc = "Wan2.x-VACE-14B";
} else {
desc = "Wan2.x-T2V-14B";
}
wan_params.in_dim = 16;
}
} else {
wan_params.in_dim = 36;
if (wan_params.flf_pos_embed_token_number > 0) {
desc = "Wan2.1-FLF2V-14B";
} else {
desc = "Wan2.1-I2V-14B";
}
}
wan_params.dim = 5120;
wan_params.eps = 1e-06;
wan_params.ffn_dim = 13824;
wan_params.freq_dim = 256;
wan_params.num_heads = 40;
wan_params.out_dim = 16;
wan_params.text_len = 512;
} else {
GGML_ABORT("invalid num_layers(%ld) of wan", wan_params.num_layers);
}
LOG_INFO("%s", desc.c_str());
wan = Wan(wan_params);
wan.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() override {
return desc;
}
void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
wan.get_param_tensors(tensors, prefix);
}
struct ggml_cgraph* build_graph(struct ggml_tensor* x,
struct ggml_tensor* timesteps,
struct ggml_tensor* context,
struct ggml_tensor* clip_fea = nullptr,
struct ggml_tensor* c_concat = nullptr,
struct ggml_tensor* time_dim_concat = nullptr,
struct ggml_tensor* vace_context = nullptr,
float vace_strength = 1.f) {
struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, WAN_GRAPH_SIZE, false);
x = to_backend(x);
timesteps = to_backend(timesteps);
context = to_backend(context);
clip_fea = to_backend(clip_fea);
c_concat = to_backend(c_concat);
time_dim_concat = to_backend(time_dim_concat);
vace_context = to_backend(vace_context);
pe_vec = Rope::gen_wan_pe(x->ne[2],
x->ne[1],
x->ne[0],
std::get<0>(wan_params.patch_size),
std::get<1>(wan_params.patch_size),
std::get<2>(wan_params.patch_size),
1,
wan_params.theta,
wan_params.axes_dim);
int pos_len = pe_vec.size() / wan_params.axes_dim_sum / 2;
// LOG_DEBUG("pos_len %d", pos_len);
auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, wan_params.axes_dim_sum / 2, pos_len);
// pe->data = pe_vec.data();
// print_ggml_tensor(pe);
// pe->data = nullptr;
set_backend_tensor_data(pe, pe_vec.data());
if (c_concat != nullptr) {
x = ggml_concat(compute_ctx, x, c_concat, 3);
}
struct ggml_tensor* out = wan.forward(compute_ctx,
runtime_backend,
x,
timesteps,
context,
pe,
clip_fea,
time_dim_concat,
vace_context,
vace_strength);
ggml_build_forward_expand(gf, out);
return gf;
}
void compute(int n_threads,
struct ggml_tensor* x,
struct ggml_tensor* timesteps,
struct ggml_tensor* context,
struct ggml_tensor* clip_fea = nullptr,
struct ggml_tensor* c_concat = nullptr,
struct ggml_tensor* time_dim_concat = nullptr,
struct ggml_tensor* vace_context = nullptr,
float vace_strength = 1.f,
struct ggml_tensor** output = nullptr,
struct ggml_context* output_ctx = nullptr) {
auto get_graph = [&]() -> struct ggml_cgraph* {
return build_graph(x, timesteps, context, clip_fea, c_concat, time_dim_concat, vace_context, vace_strength);
};
GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
}
void test() {
struct ggml_init_params params;
params.mem_size = static_cast<size_t>(200 * 1024 * 1024); // 200 MB
params.mem_buffer = nullptr;
params.no_alloc = false;
struct ggml_context* work_ctx = ggml_init(params);
GGML_ASSERT(work_ctx != nullptr);
{
// cpu f16: pass
// cuda f16: pass
// cpu q8_0: pass
// auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 104, 60, 1, 16);
// ggml_set_f32(x, 0.01f);
auto x = load_tensor_from_file(work_ctx, "wan_dit_x.bin");
print_ggml_tensor(x);
std::vector<float> timesteps_vec(3, 1000.f);
timesteps_vec[0] = 0.f;
auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
// auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 4096, 512, 1);
// ggml_set_f32(context, 0.01f);
auto context = load_tensor_from_file(work_ctx, "wan_dit_context.bin");
print_ggml_tensor(context);
// auto clip_fea = load_tensor_from_file(work_ctx, "wan_dit_clip_fea.bin");
// print_ggml_tensor(clip_fea);
struct ggml_tensor* out = nullptr;
int t0 = ggml_time_ms();
compute(8, x, timesteps, context, nullptr, nullptr, nullptr, nullptr, 1.f, &out, work_ctx);
int t1 = ggml_time_ms();
print_ggml_tensor(out);
LOG_DEBUG("wan test done in %dms", t1 - t0);
}
}
static void load_from_file_and_test(const std::string& file_path) {
// ggml_backend_t backend = ggml_backend_cuda_init(0);
ggml_backend_t backend = ggml_backend_cpu_init();
ggml_type model_data_type = GGML_TYPE_F16;
LOG_INFO("loading from '%s'", file_path.c_str());
ModelLoader model_loader;
if (!model_loader.init_from_file(file_path, "model.diffusion_model.")) {
LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
return;
}
auto tensor_types = model_loader.tensor_storages_types;
for (auto& item : tensor_types) {
// LOG_DEBUG("%s %u", item.first.c_str(), item.second);
if (ends_with(item.first, "weight")) {
item.second = model_data_type;
}
}
std::shared_ptr<WanRunner> wan = std::make_shared<WanRunner>(backend,
false,
tensor_types,
"model.diffusion_model",
VERSION_WAN2_2_TI2V,
true);
wan->alloc_params_buffer();
std::map<std::string, ggml_tensor*> tensors;
wan->get_param_tensors(tensors, "model.diffusion_model");
bool success = model_loader.load_tensors(tensors);
if (!success) {
LOG_ERROR("load tensors from model loader failed");
return;
}
LOG_INFO("wan model loaded");
wan->test();
}
};
} // namespace WAN
#endif // __WAN_HPP__
Reference in New Issue View Git Blame Copy Permalink