refactor: reorganize src model layout (#1615)

src/model/vae/tae.hpp

@@ -0,0 +1,753 @@
+#ifndef __SD_MODEL_VAE_TAE_HPP__
+#define __SD_MODEL_VAE_TAE_HPP__
+
+#include "core/ggml_extend.hpp"
+#include "model.h"
+
+/*
+    ===================================    TinyAutoEncoder  ===================================
+    References:
+    https://github.com/huggingface/diffusers/blob/main/src/diffusers/model/autoencoders/vae.py
+    https://github.com/madebyollin/taesd/blob/main/taesd.py
+
+*/
+
+class TAEBlock : public UnaryBlock {
+protected:
+    int n_in;
+    int n_out;
+    bool use_midblock_gn;
+
+public:
+    TAEBlock(int n_in, int n_out, bool use_midblock_gn = false)
+        : n_in(n_in), n_out(n_out), use_midblock_gn(use_midblock_gn) {
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
+        if (n_in != n_out) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {1, 1}, {1, 1}, {1, 1}, {1, 1}, false));
+        }
+        if (use_midblock_gn) {
+            int n_gn         = n_in * 4;
+            blocks["pool.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_gn, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+            blocks["pool.1"] = std::shared_ptr<GGMLBlock>(new GroupNorm(4, n_gn));
+            // pool.2 is ReLU, handled in forward
+            blocks["pool.3"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_gn, n_in, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        // x: [n, n_in, h, w]
+        // return: [n, n_out, h, w]
+
+        if (use_midblock_gn) {
+            auto pool_0 = std::dynamic_pointer_cast<Conv2d>(blocks["pool.0"]);
+            auto pool_1 = std::dynamic_pointer_cast<GroupNorm>(blocks["pool.1"]);
+            auto pool_3 = std::dynamic_pointer_cast<Conv2d>(blocks["pool.3"]);
+
+            auto p = pool_0->forward(ctx, x);
+            p      = pool_1->forward(ctx, p);
+            p      = ggml_relu_inplace(ctx->ggml_ctx, p);
+            p      = pool_3->forward(ctx, p);
+
+            x = ggml_add(ctx->ggml_ctx, x, p);
+        }
+
+        auto conv_0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv_2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
+        auto conv_4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+
+        auto h = conv_0->forward(ctx, x);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv_2->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv_4->forward(ctx, h);
+
+        if (n_in != n_out) {
+            auto skip = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            LOG_DEBUG("skip");
+            x = skip->forward(ctx, x);
+        }
+
+        h = ggml_add(ctx->ggml_ctx, h, x);
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+        return h;
+    }
+};
+
+class TinyEncoder : public UnaryBlock {
+    int in_channels = 3;
+    int channels    = 64;
+    int z_channels  = 4;
+    int num_blocks  = 3;
+
+public:
+    TinyEncoder(int z_channels = 4, bool use_midblock_gn = false)
+        : z_channels(z_channels) {
+        int index                       = 0;
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        // x: [n, in_channels, h, w]
+        // return: [n, z_channels, h/8, w/8]
+
+        for (int i = 0; i < num_blocks * 3 + 6; i++) {
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
+
+            x = block->forward(ctx, x);
+        }
+
+        return x;
+    }
+};
+
+class TinyDecoder : public UnaryBlock {
+    int z_channels   = 4;
+    int channels     = 64;
+    int out_channels = 3;
+    int num_blocks   = 3;
+
+public:
+    TinyDecoder(int z_channels = 4, bool use_midblock_gn = false)
+        : z_channels(z_channels) {
+        int index = 0;
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        index++;  // nn.ReLU()
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
+        // z: [n, z_channels, h, w]
+        // return: [n, out_channels, h*8, w*8]
+
+        auto h = ggml_ext_scale(ctx->ggml_ctx, z, 1.0f / 3.0f);
+        h      = ggml_tanh_inplace(ctx->ggml_ctx, h);
+        h      = ggml_ext_scale(ctx->ggml_ctx, h, 3.0f);
+
+        for (int i = 0; i < num_blocks * 3 + 10; i++) {
+            if (blocks.find(std::to_string(i)) == blocks.end()) {
+                if (i == 1) {
+                    h = ggml_relu_inplace(ctx->ggml_ctx, h);
+                } else {
+                    h = ggml_upscale(ctx->ggml_ctx, h, 2, GGML_SCALE_MODE_NEAREST);
+                }
+                continue;
+            }
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
+
+            h = block->forward(ctx, h);
+        }
+
+        return h;
+    }
+};
+
+class TPool : public UnaryBlock {
+    int stride;
+
+public:
+    TPool(int channels, int stride)
+        : stride(stride) {
+        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * stride, channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
+        auto h    = x;
+        if (stride != 1) {
+            h = ggml_reshape_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2] * stride, h->ne[3] / stride);
+        }
+        h = conv->forward(ctx, h);
+        return h;
+    }
+};
+
+class TGrow : public UnaryBlock {
+    int stride;
+
+public:
+    TGrow(int channels, int stride)
+        : stride(stride) {
+        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels * stride, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
+        auto h    = conv->forward(ctx, x);
+        if (stride != 1) {
+            h = ggml_reshape_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2] / stride, h->ne[3] * stride);
+        }
+        return h;
+    }
+};
+
+class MemBlock : public GGMLBlock {
+    bool has_skip_conv = false;
+
+public:
+    MemBlock(int channels, int out_channels)
+        : has_skip_conv(channels != out_channels) {
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * 2, out_channels, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+        if (has_skip_conv) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* past) {
+        // x: [n, channels, h, w]
+        auto conv0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+
+        auto h = ggml_concat(ctx->ggml_ctx, x, past, 2);
+        h      = conv0->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv1->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv2->forward(ctx, h);
+
+        auto skip = x;
+        if (has_skip_conv) {
+            auto skip_conv = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            skip           = skip_conv->forward(ctx, x);
+        }
+        h = ggml_add_inplace(ctx->ggml_ctx, h, skip);
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+        return h;
+    }
+};
+
+class WideMemBlock : public GGMLBlock {
+    bool has_skip_conv = false;
+
+public:
+    WideMemBlock(int channels, int out_channels)
+        : has_skip_conv(channels != out_channels) {
+        int groups       = std::max(1, out_channels / 64);
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * 2, out_channels, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d_grouped(out_channels, out_channels, groups, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {1, 1}, {1, 1}));
+        blocks["conv.6"] = std::shared_ptr<GGMLBlock>(new Conv2d_grouped(out_channels, out_channels, groups, {3, 3}, {1, 1}, {1, 1}));
+        if (has_skip_conv) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* past) {
+        // x: [n, channels, h, w]
+        auto conv0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d_grouped>(blocks["conv.2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+        auto conv3 = std::dynamic_pointer_cast<Conv2d_grouped>(blocks["conv.6"]);
+
+        auto h = ggml_concat(ctx->ggml_ctx, x, past, 2);
+        h      = conv0->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv1->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv2->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv3->forward(ctx, h);
+
+        auto skip = x;
+        if (has_skip_conv) {
+            auto skip_conv = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            skip           = skip_conv->forward(ctx, x);
+        }
+        h = ggml_add_inplace(ctx->ggml_ctx, h, skip);
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+        return h;
+    }
+};
+
+ggml_tensor*
+patchify(ggml_context* ctx,
+         ggml_tensor* x,
+         int64_t patch_size,
+         int64_t b = 1) {
+    // x: [f, b*c, h*q, w*r]
+    // return: [f, b*c*r*q, h, w]
+    if (patch_size == 1) {
+        return x;
+    }
+    int64_t r = patch_size;
+    int64_t q = patch_size;
+
+    int64_t W = x->ne[0];
+    int64_t H = x->ne[1];
+    int64_t C = x->ne[2];
+    int64_t f = x->ne[3];
+
+    int64_t w = W / r;
+    int64_t h = H / q;
+
+    x = ggml_reshape_4d(ctx, x, W, q, h, C * f);                         // [W, q, h, C*f]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));  // [W, h, q, C*f]
+    x = ggml_reshape_4d(ctx, x, r, w, h, q * C * f);                     // [r, w, h, q*C*f]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3));  // [w, h, r, q*C*f]
+    x = ggml_reshape_4d(ctx, x, w, h, r * q * C, f);                     // [f, b*c*r*q, h, w]
+
+    return x;
+}
+
+ggml_tensor* unpatchify(ggml_context* ctx,
+                        ggml_tensor* x,
+                        int64_t patch_size,
+                        int64_t b = 1) {
+    // x: [f, b*c*r*q, h, w]
+    // return: [f, b*c, 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[2] / b / q / r;
+    int64_t f = x->ne[3];
+    int64_t h = x->ne[1];
+    int64_t w = x->ne[0];
+
+    x = ggml_reshape_4d(ctx, x, w, h, r, q * c * b * f);                 // [q*c*b*f, r, h, w]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3));  // [r, w, h, q*c*b*f]
+    x = ggml_reshape_4d(ctx, x, r * w, h, q, c * b * f);                 // [c*b*f, q, h, r*w]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));  // [r*w, q, h, c*b*f]
+    x = ggml_reshape_4d(ctx, x, r * w, q * h, c * b, f);
+
+    return x;
+}
+
+class TinyVideoEncoder : public UnaryBlock {
+    int in_channels = 3;
+    int hidden      = 64;
+    int z_channels  = 4;
+    int num_blocks  = 3;
+    int num_layers  = 3;
+    int patch_size  = 1;
+
+public:
+    int t_downscale = 1;
+    TinyVideoEncoder(int z_channels = 4, int patch_size = 1, std::vector<bool> time_downscale = {true, true, false})
+        : z_channels(z_channels), patch_size(patch_size) {
+        t_downscale = 1;
+        for (bool downscale : time_downscale) {
+            if (downscale) {
+                t_downscale *= 2;
+            }
+        }
+        int index                       = 0;
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels * patch_size * patch_size, hidden, {3, 3}, {1, 1}, {1, 1}));
+        index++;  // nn.ReLU()
+        for (int i = 0; i < num_layers; i++) {
+            int stride                      = time_downscale[i] ? 2 : 1;
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TPool(hidden, stride));
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, hidden, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+            for (int j = 0; j < num_blocks; j++) {
+                blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(hidden, hidden));
+            }
+        }
+        blocks[std::to_string(index)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
+        auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["0"]);
+
+        if (patch_size > 1) {
+            z = patchify(ctx->ggml_ctx, z, patch_size, 1);
+        }
+
+        auto h = first_conv->forward(ctx, z);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+
+        int index = 2;
+        for (int i = 0; i < num_layers; i++) {
+            auto pool = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+            auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+
+            h = pool->forward(ctx, h);
+            h = conv->forward(ctx, h);
+            for (int j = 0; j < num_blocks; j++) {
+                auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
+                auto mem   = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                mem        = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
+                h          = block->forward(ctx, h, mem);
+            }
+        }
+        auto last_conv = std::dynamic_pointer_cast<Conv2d>(blocks[std::to_string(index)]);
+        h              = last_conv->forward(ctx, h);
+        return h;
+    }
+};
+
+class TinyVideoDecoder : public UnaryBlock {
+    int z_channels               = 4;
+    int out_channels             = 3;
+    int num_blocks               = 3;
+    static const int num_layers  = 3;
+    int channels[num_layers + 1] = {256, 128, 64, 64};
+    int patch_size               = 1;
+    int t_upscale                = 1;
+    bool is_wide                 = false;
+
+public:
+    TinyVideoDecoder(int z_channels = 4, int patch_size = 1, std::vector<bool> time_upscale = {false, true, true}, bool is_wide = false)
+        : z_channels(z_channels), patch_size(patch_size), is_wide(is_wide) {
+        t_upscale = 1;
+        if (is_wide) {
+            channels[0] = 1024;
+            channels[1] = 512;
+            channels[2] = 256;
+        }
+
+        for (bool upscale : time_upscale) {
+            if (upscale) {
+                t_upscale *= 2;
+            }
+        }
+        int index                       = 1;  // Clamp()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels[0], {3, 3}, {1, 1}, {1, 1}));
+        index++;  // nn.ReLU()
+        for (int i = 0; i < num_layers; i++) {
+            int stride = time_upscale[i] ? 2 : 1;
+            for (int j = 0; j < num_blocks; j++) {
+                if (is_wide) {
+                    blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new WideMemBlock(channels[i], channels[i]));
+                } else {
+                    blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(channels[i], channels[i]));
+                }
+            }
+            index++;  // nn.Upsample()
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TGrow(channels[i], stride));
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels[i], channels[i + 1], {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        }
+        index++;  // nn.ReLU()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels[num_layers], out_channels * patch_size * patch_size, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
+        auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["1"]);
+
+        // Clamp()
+        auto h = ggml_ext_scale(ctx->ggml_ctx,
+                                ggml_tanh_inplace(ctx->ggml_ctx,
+                                                  ggml_ext_scale(ctx->ggml_ctx, z, 1.0f / 3.0f)),
+                                3.0f,
+                                true);
+
+        h         = first_conv->forward(ctx, h);
+        h         = ggml_relu_inplace(ctx->ggml_ctx, h);
+        int index = 3;
+        for (int i = 0; i < num_layers; i++) {
+            for (int j = 0; j < num_blocks; j++) {
+                auto mem = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                mem      = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
+                if (is_wide) {
+                    auto block = std::dynamic_pointer_cast<WideMemBlock>(blocks[std::to_string(index++)]);
+                    h          = block->forward(ctx, h, mem);
+                } else {
+                    auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
+                    h          = block->forward(ctx, h, mem);
+                }
+            }
+            // upsample
+            index++;
+            h          = ggml_upscale(ctx->ggml_ctx, h, 2, GGML_SCALE_MODE_NEAREST);
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+            h          = block->forward(ctx, h);
+            block      = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+            h          = block->forward(ctx, h);
+        }
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+
+        auto last_conv = std::dynamic_pointer_cast<Conv2d>(blocks[std::to_string(++index)]);
+        h              = last_conv->forward(ctx, h);
+        if (patch_size > 1) {
+            h = unpatchify(ctx->ggml_ctx, h, patch_size, 1);
+        }
+        // shape(W, H, 3, (t_upscale - 1) + T) => shape(W, H, 3, T)
+        h = ggml_view_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2], h->ne[3] - (t_upscale - 1), h->nb[1], h->nb[2], h->nb[3], (t_upscale - 1) * h->nb[3]);
+        return h;
+    }
+};
+
+class TAEHV : public GGMLBlock {
+protected:
+    bool decode_only;
+    SDVersion version;
+    bool is_wide;
+
+public:
+    int z_channels                   = 16;
+    std::vector<bool> time_downscale = {true, true, false};
+    std::vector<bool> time_upscale   = {false, true, true};
+
+public:
+    TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2, bool is_wide = false)
+        : decode_only(decode_only), version(version), is_wide(is_wide) {
+        int patch = 1;
+        if (version == VERSION_WAN2_2_TI2V) {
+            z_channels = 48;
+            patch      = 2;
+        } else if (sd_version_is_ltxav(version)) {
+            z_channels     = 128;
+            patch          = 4;
+            time_downscale = {true, true, true};
+            time_upscale   = {true, true, true};
+        }
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoDecoder(z_channels, patch, time_upscale, is_wide));
+        if (!decode_only) {
+            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoEncoder(z_channels, patch, time_downscale));
+        }
+    }
+
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
+        auto decoder = std::dynamic_pointer_cast<TinyVideoDecoder>(blocks["decoder"]);
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
+            // (W, H, C, T) -> (W, H, T, C)
+            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 0, 1, 3, 2));
+        }
+        auto result = decoder->forward(ctx, z);
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
+            // (W, H, C, T) -> (W, H, T, C)
+            result = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, result, 0, 1, 3, 2));
+        }
+        return result;
+    }
+
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        auto encoder = std::dynamic_pointer_cast<TinyVideoEncoder>(blocks["encoder"]);
+        // (W, H, T, C) -> (W, H, C, T)
+        x                  = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        int64_t num_frames = x->ne[3];
+        if (num_frames % encoder->t_downscale) {
+            // pad to multiple of encoder->t_downscale at the end
+            auto last_frame = ggml_view_4d(ctx->ggml_ctx, x, x->ne[0], x->ne[1], x->ne[2], 1, x->nb[1], x->nb[2], x->nb[3], (num_frames - 1) * x->nb[3]);
+            for (int i = 0; i < encoder->t_downscale - num_frames % encoder->t_downscale; i++) {
+                x = ggml_concat(ctx->ggml_ctx, x, last_frame, 3);
+            }
+        }
+        x = encoder->forward(ctx, x);
+        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        return x;
+    }
+};
+
+class TAESD : public GGMLBlock {
+protected:
+    bool decode_only;
+    bool taef2 = false;
+
+public:
+    int z_channels = 4;
+
+public:
+    TAESD(bool decode_only = true, SDVersion version = VERSION_SD1)
+        : decode_only(decode_only) {
+        bool use_midblock_gn = false;
+        taef2                = sd_version_uses_flux2_vae(version);
+
+        if (sd_version_is_dit(version)) {
+            z_channels = 16;
+        }
+        if (taef2) {
+            z_channels      = 32;
+            use_midblock_gn = true;
+        }
+        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels, use_midblock_gn));
+
+        if (!decode_only) {
+            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels, use_midblock_gn));
+        }
+    }
+
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
+        auto decoder = std::dynamic_pointer_cast<TinyDecoder>(blocks["decoder.layers"]);
+        if (taef2) {
+            z = unpatchify(ctx->ggml_ctx, z, 2);
+        }
+        return decoder->forward(ctx, z);
+    }
+
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        auto encoder = std::dynamic_pointer_cast<TinyEncoder>(blocks["encoder.layers"]);
+        auto z       = encoder->forward(ctx, x);
+        if (taef2) {
+            z = patchify(ctx->ggml_ctx, z, 2);
+        }
+        return z;
+    }
+};
+
+struct TinyImageAutoEncoder : public VAE {
+    TAESD taesd;
+    bool decode_only = false;
+
+    TinyImageAutoEncoder(ggml_backend_t backend,
+                         ggml_backend_t params_backend,
+                         const String2TensorStorage& tensor_storage_map,
+                         const std::string prefix,
+                         bool decoder_only = true,
+                         SDVersion version = VERSION_SD1)
+        : decode_only(decoder_only),
+          taesd(decoder_only, version),
+          VAE(version, backend, params_backend) {
+        scale_input = false;
+        taesd.init(params_ctx, tensor_storage_map, prefix);
+    }
+
+    std::string get_desc() override {
+        return "taesd";
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        taesd.get_param_tensors(tensors, prefix);
+    }
+
+    sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
+        SD_UNUSED(rng);
+        return vae_output;
+    }
+
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        return latents;
+    }
+
+    sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+        return latents;
+    }
+
+    int get_encoder_output_channels(int input_channels) {
+        return taesd.z_channels;
+    }
+
+    ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
+        ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        ggml_tensor* z   = make_input(z_tensor);
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = decode_graph ? taesd.decode(&runner_ctx, z) : taesd.encode(&runner_ctx, z);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    sd::Tensor<float> _compute(const int n_threads,
+                               const sd::Tensor<float>& z_tensor,
+                               bool decode_graph) override {
+        auto get_graph = [&]() -> ggml_cgraph* {
+            return build_graph(z_tensor, decode_graph);
+        };
+
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z_tensor.dim());
+    }
+};
+
+struct TinyVideoAutoEncoder : public VAE {
+    TAEHV taehv;
+    bool decode_only = false;
+    bool is_wide     = false;
+
+    TinyVideoAutoEncoder(ggml_backend_t backend,
+                         ggml_backend_t params_backend,
+                         const String2TensorStorage& tensor_storage_map,
+                         const std::string prefix,
+                         bool decoder_only = true,
+                         SDVersion version = VERSION_WAN2)
+        : decode_only(decoder_only),
+          VAE(version, backend, params_backend) {
+        for (auto tensor_storage : tensor_storage_map) {
+            if (tensor_storage.first.find(prefix + ".3.conv.6.weight") != std::string::npos) {
+                is_wide = true;
+                break;
+            }
+        }
+        taehv       = TAEHV(decoder_only, version, is_wide);
+        scale_input = false;
+        taehv.init(params_ctx, tensor_storage_map, prefix);
+    }
+
+    std::string get_desc() override {
+        return "taehv";
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        taehv.get_param_tensors(tensors, prefix);
+    }
+
+    sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
+        SD_UNUSED(rng);
+        return vae_output;
+    }
+
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        return latents;
+    }
+
+    sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+        return latents;
+    }
+
+    int get_encoder_output_channels(int input_channels) {
+        return taehv.z_channels;
+    }
+
+    ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
+        ggml_cgraph* gf  = decode_graph && is_wide ? ggml_new_graph_custom(compute_ctx, 4096, false)
+                                                   : ggml_new_graph(compute_ctx);
+        ggml_tensor* z   = make_input(z_tensor);
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    sd::Tensor<float> _compute(const int n_threads,
+                               const sd::Tensor<float>& z_tensor,
+                               bool decode_graph) override {
+        auto get_graph = [&]() -> ggml_cgraph* {
+            return build_graph(z_tensor, decode_graph);
+        };
+
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z_tensor.dim());
+    }
+};
+
+#endif  // __SD_MODEL_VAE_TAE_HPP__