feat: do not convert tensor names by default in convert mode (#1122 )

feat: add seamless texture generation support (#914 )
* global bool * reworked circular to global flag * cleaner implementation of tiling support in sd cpp * cleaned rope * working simplified but still need wraps * Further clean of rope * resolve flux conflict * switch to pad op circular only * Set ggml to most recent * Revert ggml temp * Update ggml to most recent * Revert unneded flux change * move circular flag to the GGMLRunnerContext * Pass through circular param in all places where conv is called * fix of constant and minor cleanup * Added back --circular option * Conv2d circular in vae and various models * Fix temporal padding for qwen image and other vaes * Z Image circular tiling * x and y axis seamless only * First attempt at chroma seamless x and y * refactor into pure x and y, almost there * Fix crash on chroma * Refactor into cleaner variable choices * Removed redundant set_circular_enabled * Sync ggml * simplify circular parameter * format code * no need to perform circular pad on the clip * simplify circular_axes setting * unify function naming * remove unnecessary member variables * simplify rope --------- Co-authored-by: Phylliida <phylliidadev@gmail.com> Co-authored-by: leejet <leejet714@gmail.com>
2026-10-15 05:19:28 +08:00 · 2025-12-21 18:40:10 +08:00 · 2025-12-21 18:06:47 +08:00 · 2025-12-21 15:40:21 +08:00 · 2025-12-21 15:35:38 +08:00 · 2025-12-20 00:55:57 +08:00
21 changed files with 553 additions and 102 deletions
@@ -28,7 +28,7 @@ public:
        if (vae_downsample) {
            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);

-            x = ggml_pad(ctx->ggml_ctx, x, 1, 1, 0, 0);
+            x = ggml_ext_pad(ctx->ggml_ctx, x, 1, 1, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
            x = conv->forward(ctx, x);
        } else {
            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["op"]);
@@ -366,18 +366,18 @@ struct KLOptimalScheduler : SigmaScheduler {

        for (uint32_t i = 0; i < n; ++i) {
            // t goes from 0.0 to 1.0
-            float t = static_cast<float>(i) / static_cast<float>(n-1);
+            float t = static_cast<float>(i) / static_cast<float>(n - 1);

            // Interpolate in the angle domain
            float angle = t * alpha_min + (1.0f - t) * alpha_max;

            // Convert back to sigma
            sigmas.push_back(std::tan(angle));
-            }
+        }

        // Append the final zero to sigma
        sigmas.push_back(0.0f);
-    
+
        return sigmas;
    }
 };
@@ -37,8 +37,9 @@ struct DiffusionModel {
    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) = 0;
    virtual size_t get_params_buffer_size()                                             = 0;
    virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
-    virtual int64_t get_adm_in_channels()             = 0;
-    virtual void set_flash_attn_enabled(bool enabled) = 0;
+    virtual int64_t get_adm_in_channels()                            = 0;
+    virtual void set_flash_attn_enabled(bool enabled)                = 0;
+    virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
 };

 struct UNetModel : public DiffusionModel {
@@ -87,6 +88,10 @@ struct UNetModel : public DiffusionModel {
        unet.set_flash_attention_enabled(enabled);
    }

+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        unet.set_circular_axes(circular_x, circular_y);
+    }
+
    bool compute(int n_threads,
                 DiffusionParams diffusion_params,
                 struct ggml_tensor** output     = nullptr,
@@ -148,6 +153,10 @@ struct MMDiTModel : public DiffusionModel {
        mmdit.set_flash_attention_enabled(enabled);
    }

+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        mmdit.set_circular_axes(circular_x, circular_y);
+    }
+
    bool compute(int n_threads,
                 DiffusionParams diffusion_params,
                 struct ggml_tensor** output     = nullptr,
@@ -210,6 +219,10 @@ struct FluxModel : public DiffusionModel {
        flux.set_flash_attention_enabled(enabled);
    }

+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        flux.set_circular_axes(circular_x, circular_y);
+    }
+
    bool compute(int n_threads,
                 DiffusionParams diffusion_params,
                 struct ggml_tensor** output     = nullptr,
@@ -277,6 +290,10 @@ struct WanModel : public DiffusionModel {
        wan.set_flash_attention_enabled(enabled);
    }

+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        wan.set_circular_axes(circular_x, circular_y);
+    }
+
    bool compute(int n_threads,
                 DiffusionParams diffusion_params,
                 struct ggml_tensor** output     = nullptr,
@@ -343,6 +360,10 @@ struct QwenImageModel : public DiffusionModel {
        qwen_image.set_flash_attention_enabled(enabled);
    }

+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        qwen_image.set_circular_axes(circular_x, circular_y);
+    }
+
    bool compute(int n_threads,
                 DiffusionParams diffusion_params,
                 struct ggml_tensor** output     = nullptr,
@@ -406,6 +427,10 @@ struct ZImageModel : public DiffusionModel {
        z_image.set_flash_attention_enabled(enabled);
    }

+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        z_image.set_circular_axes(circular_x, circular_y);
+    }
+
    bool compute(int n_threads,
                 DiffusionParams diffusion_params,
                 struct ggml_tensor** output     = nullptr,
@@ -14,4 +14,26 @@ curl -L -O https://huggingface.co/madebyollin/taesd/resolve/main/diffusion_pytor

 ```bash
 sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
-```
+```
+
+### Qwen-Image and wan (TAEHV)
+
+sd.cpp also supports [TAEHV](https://github.com/madebyollin/taehv) (#937), which can be used for Qwen-Image and wan.
+
+- For **Qwen-Image and wan2.1 and wan2.2-A14B**, download the wan2.1 tae [safetensors weights](https://github.com/madebyollin/taehv/blob/main/safetensors/taew2_1.safetensors)
+  
+  Or curl
+  
+  ```bash
+  curl -L -O https://github.com/madebyollin/taehv/raw/refs/heads/main/safetensors/taew2_1.safetensors
+  ```
+
+- For **wan2.2-TI2V-5B**, use the wan2.2 tae [safetensors weights](https://github.com/madebyollin/taehv/blob/main/safetensors/taew2_2.safetensors)
+  
+  Or curl
+  
+  ```bash
+  curl -L -O https://github.com/madebyollin/taehv/raw/refs/heads/main/safetensors/taew2_2.safetensors
+  ```
+
+Then simply replace the `--vae xxx.safetensors` with `--tae xxx.safetensors` in the commands. If it still out of VRAM, add `--vae-conv-direct` to your command though might be slower.
@@ -39,6 +39,9 @@
        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
    - wan_2.2_vae (for Wan2.2 TI2V 5B only)
        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/vae/wan2.2_vae.safetensors
+
+    > Wan models vae requires really much VRAM! If you do not have enough VRAM, please try tae instead, though the results may be poorer. For tae usage, please refer to [taesd](taesd.md)
+
 - Download umt5_xxl
    - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp16.safetensors
    - gguf: https://huggingface.co/city96/umt5-xxl-encoder-gguf/tree/main
@@ -9,6 +9,7 @@ CLI Options:
  --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
                              every step)
  --canny                     apply canny preprocessor (edge detection)
+  --convert-name              convert tensor name (for convert mode)
  -v, --verbose               print extra info
  --color                     colors the logging tags according to level
  --taesd-preview-only        prevents usage of taesd for decoding the final image. (for use with --preview tae)
@@ -32,6 +32,7 @@ struct SDCliParams {

    bool verbose          = false;
    bool canny_preprocess = false;
+    bool convert_name     = false;

    preview_t preview_method = PREVIEW_NONE;
    int preview_interval     = 1;
@@ -69,6 +70,10 @@ struct SDCliParams {
             "--canny",
             "apply canny preprocessor (edge detection)",
             true, &canny_preprocess},
+            {"",
+             "--convert-name",
+             "convert tensor name (for convert mode)",
+             true, &canny_preprocess},
            {"-v",
             "--verbose",
             "print extra info",
@@ -174,6 +179,7 @@ struct SDCliParams {
            << "  verbose: " << (verbose ? "true" : "false") << ",\n"
            << "  color: " << (color ? "true" : "false") << ",\n"
            << "  canny_preprocess: " << (canny_preprocess ? "true" : "false") << ",\n"
+            << "  convert_name: " << (convert_name ? "true" : "false") << ",\n"
            << "  preview_method: " << previews_str[preview_method] << ",\n"
            << "  preview_interval: " << preview_interval << ",\n"
            << "  preview_path: \"" << preview_path << "\",\n"
@@ -387,7 +393,8 @@ int main(int argc, const char* argv[]) {
                               ctx_params.vae_path.c_str(),
                               cli_params.output_path.c_str(),
                               ctx_params.wtype,
-                               ctx_params.tensor_type_rules.c_str());
+                               ctx_params.tensor_type_rules.c_str(),
+                               cli_params.convert_name);
        if (!success) {
            LOG_ERROR("convert '%s'/'%s' to '%s' failed",
                      ctx_params.model_path.c_str(),
@@ -579,7 +586,7 @@ int main(int argc, const char* argv[]) {
        }

        if (gen_params.sample_params.scheduler == SCHEDULER_COUNT) {
-            gen_params.sample_params.scheduler = sd_get_default_scheduler(sd_ctx);
+            gen_params.sample_params.scheduler = sd_get_default_scheduler(sd_ctx, gen_params.sample_params.sample_method);
        }

        if (cli_params.mode == IMG_GEN) {
@@ -752,4 +759,4 @@ int main(int argc, const char* argv[]) {
    release_all_resources();

    return 0;
-}
+}
@@ -449,6 +449,10 @@ struct SDContextParams {
    bool diffusion_conv_direct  = false;
    bool vae_conv_direct        = false;

+    bool circular   = false;
+    bool circular_x = false;
+    bool circular_y = false;
+
    bool chroma_use_dit_mask = true;
    bool chroma_use_t5_mask  = false;
    int chroma_t5_mask_pad   = 1;
@@ -605,6 +609,18 @@ struct SDContextParams {
             "--vae-conv-direct",
             "use ggml_conv2d_direct in the vae model",
             true, &vae_conv_direct},
+            {"",
+             "--circular",
+             "enable circular padding for convolutions",
+             true, &circular},
+            {"",
+             "--circularx",
+             "enable circular RoPE wrapping on x-axis (width) only",
+             true, &circular_x},
+            {"",
+             "--circulary",
+             "enable circular RoPE wrapping on y-axis (height) only",
+             true, &circular_y},
            {"",
             "--chroma-disable-dit-mask",
             "disable dit mask for chroma",
@@ -868,6 +884,9 @@ struct SDContextParams {
            << "  diffusion_flash_attn: " << (diffusion_flash_attn ? "true" : "false") << ",\n"
            << "  diffusion_conv_direct: " << (diffusion_conv_direct ? "true" : "false") << ",\n"
            << "  vae_conv_direct: " << (vae_conv_direct ? "true" : "false") << ",\n"
+            << "  circular: " << (circular ? "true" : "false") << ",\n"
+            << "  circular_x: " << (circular_x ? "true" : "false") << ",\n"
+            << "  circular_y: " << (circular_y ? "true" : "false") << ",\n"
            << "  chroma_use_dit_mask: " << (chroma_use_dit_mask ? "true" : "false") << ",\n"
            << "  chroma_use_t5_mask: " << (chroma_use_t5_mask ? "true" : "false") << ",\n"
            << "  chroma_t5_mask_pad: " << chroma_t5_mask_pad << ",\n"
@@ -928,6 +947,8 @@ struct SDContextParams {
            taesd_preview,
            diffusion_conv_direct,
            vae_conv_direct,
+            circular || circular_x,
+            circular || circular_y,
            force_sdxl_vae_conv_scale,
            chroma_use_dit_mask,
            chroma_use_t5_mask,
@@ -744,6 +744,8 @@ namespace Flux {
        int64_t nerf_mlp_ratio   = 4;
        int64_t nerf_depth       = 4;
        int64_t nerf_max_freqs   = 8;
+        bool use_x0              = false;
+        bool use_patch_size_32   = false;
    };

    struct FluxParams {
@@ -781,7 +783,7 @@ namespace Flux {
        Flux(FluxParams params)
            : params(params) {
            if (params.version == VERSION_CHROMA_RADIANCE) {
-                std::pair<int, int> kernel_size = {(int)params.patch_size, (int)params.patch_size};
+                std::pair<int, int> kernel_size = {16, 16};
                std::pair<int, int> stride      = kernel_size;

                blocks["img_in_patch"] = std::make_shared<Conv2d>(params.in_channels,
@@ -858,14 +860,14 @@ namespace Flux {
            }
        }

-        struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
+        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
                                              struct ggml_tensor* x) {
            int64_t W = x->ne[0];
            int64_t H = x->ne[1];

            int pad_h = (params.patch_size - H % params.patch_size) % params.patch_size;
            int pad_w = (params.patch_size - W % params.patch_size) % params.patch_size;
-            x         = ggml_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
+            x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
            return x;
        }

@@ -891,11 +893,11 @@ namespace Flux {
            return x;
        }

-        struct ggml_tensor* process_img(struct ggml_context* ctx,
+        struct ggml_tensor* process_img(GGMLRunnerContext* ctx,
                                        struct ggml_tensor* x) {
            // img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
            x = pad_to_patch_size(ctx, x);
-            x = patchify(ctx, x);
+            x = patchify(ctx->ggml_ctx, x);
            return x;
        }

@@ -1044,6 +1046,15 @@ namespace Flux {
            return img;
        }

+        struct ggml_tensor* _apply_x0_residual(GGMLRunnerContext* ctx,
+                                               struct ggml_tensor* predicted,
+                                               struct ggml_tensor* noisy,
+                                               struct ggml_tensor* timesteps) {
+            auto x = ggml_sub(ctx->ggml_ctx, noisy, predicted);
+            x      = ggml_div(ctx->ggml_ctx, x, timesteps);
+            return x;
+        }
+
        struct ggml_tensor* forward_chroma_radiance(GGMLRunnerContext* ctx,
                                                    struct ggml_tensor* x,
                                                    struct ggml_tensor* timestep,
@@ -1065,9 +1076,16 @@ namespace Flux {
            int pad_h          = (patch_size - H % patch_size) % patch_size;
            int pad_w          = (patch_size - W % patch_size) % patch_size;

-            auto img      = pad_to_patch_size(ctx->ggml_ctx, x);
+            auto img      = pad_to_patch_size(ctx, x);
            auto orig_img = img;

+            if (params.chroma_radiance_params.use_patch_size_32) {
+                // It's supposed to be using GGML_SCALE_MODE_NEAREST, but this seems more stable
+                // Maybe the implementation of nearest-neighbor interpolation in ggml behaves differently than the one in PyTorch?
+                // img = F.interpolate(img, size=(H//2, W//2), mode="nearest")
+                img = ggml_interpolate(ctx->ggml_ctx, img, W / 2, H / 2, C, x->ne[3], GGML_SCALE_MODE_BILINEAR);
+            }
+
            auto img_in_patch = std::dynamic_pointer_cast<Conv2d>(blocks["img_in_patch"]);

            img = img_in_patch->forward(ctx, img);                                                       // [N, hidden_size, H/patch_size, W/patch_size]
@@ -1104,6 +1122,10 @@ namespace Flux {

            out = nerf_final_layer_conv->forward(ctx, img_dct);  // [N, C, H, W]

+            if (params.chroma_radiance_params.use_x0) {
+                out = _apply_x0_residual(ctx, out, orig_img, timestep);
+            }
+
            return out;
        }

@@ -1128,7 +1150,7 @@ namespace Flux {
            int pad_h          = (patch_size - H % patch_size) % patch_size;
            int pad_w          = (patch_size - W % patch_size) % patch_size;

-            auto img            = process_img(ctx->ggml_ctx, x);
+            auto img            = process_img(ctx, x);
            uint64_t img_tokens = img->ne[1];

            if (params.version == VERSION_FLUX_FILL) {
@@ -1136,8 +1158,8 @@ namespace Flux {
                ggml_tensor* masked = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], 0);
                ggml_tensor* mask   = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], 8 * 8, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);

-                masked = process_img(ctx->ggml_ctx, masked);
-                mask   = process_img(ctx->ggml_ctx, mask);
+                masked = process_img(ctx, masked);
+                mask   = process_img(ctx, mask);

                img = ggml_concat(ctx->ggml_ctx, img, ggml_concat(ctx->ggml_ctx, masked, mask, 0), 0);
            } else if (params.version == VERSION_FLEX_2) {
@@ -1146,21 +1168,21 @@ namespace Flux {
                ggml_tensor* mask    = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], 1, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);
                ggml_tensor* control = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * (C + 1));

-                masked  = process_img(ctx->ggml_ctx, masked);
-                mask    = process_img(ctx->ggml_ctx, mask);
-                control = process_img(ctx->ggml_ctx, control);
+                masked  = process_img(ctx, masked);
+                mask    = process_img(ctx, mask);
+                control = process_img(ctx, control);

                img = ggml_concat(ctx->ggml_ctx, img, ggml_concat(ctx->ggml_ctx, ggml_concat(ctx->ggml_ctx, masked, mask, 0), control, 0), 0);
            } else if (params.version == VERSION_FLUX_CONTROLS) {
                GGML_ASSERT(c_concat != nullptr);

-                auto control = process_img(ctx->ggml_ctx, c_concat);
+                auto control = process_img(ctx, c_concat);
                img          = ggml_concat(ctx->ggml_ctx, img, control, 0);
            }

            if (ref_latents.size() > 0) {
                for (ggml_tensor* ref : ref_latents) {
-                    ref = process_img(ctx->ggml_ctx, ref);
+                    ref = process_img(ctx, ref);
                    img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                }
            }
@@ -1290,6 +1312,15 @@ namespace Flux {
                    // not schnell
                    flux_params.guidance_embed = true;
                }
+                if (tensor_name.find("__x0__") != std::string::npos) {
+                    LOG_DEBUG("using x0 prediction");
+                    flux_params.chroma_radiance_params.use_x0 = true;
+                }
+                if (tensor_name.find("__32x32__") != std::string::npos) {
+                    LOG_DEBUG("using patch size 32 prediction");
+                    flux_params.chroma_radiance_params.use_patch_size_32 = true;
+                    flux_params.patch_size                               = 32;
+                }
                if (tensor_name.find("distilled_guidance_layer.in_proj.weight") != std::string::npos) {
                    // Chroma
                    flux_params.is_chroma = true;
@@ -1441,6 +1472,8 @@ namespace Flux {
                                            increase_ref_index,
                                            flux_params.ref_index_scale,
                                            flux_params.theta,
+                                            circular_y_enabled,
+                                            circular_x_enabled,
                                            flux_params.axes_dim);
            int pos_len = pe_vec.size() / flux_params.axes_dim_sum / 2;
            // LOG_DEBUG("pos_len %d", pos_len);
@@ -5,6 +5,7 @@
 #include <inttypes.h>
 #include <stdarg.h>
 #include <algorithm>
+#include <atomic>
 #include <cstring>
 #include <fstream>
 #include <functional>
@@ -993,6 +994,48 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_linear(struct ggml_context* ctx,
    return x;
 }

+__STATIC_INLINE__ struct ggml_tensor* ggml_ext_pad_ext(struct ggml_context* ctx,
+                                                       struct ggml_tensor* x,
+                                                       int lp0,
+                                                       int rp0,
+                                                       int lp1,
+                                                       int rp1,
+                                                       int lp2,
+                                                       int rp2,
+                                                       int lp3,
+                                                       int rp3,
+                                                       bool circular_x = false,
+                                                       bool circular_y = false) {
+    if (circular_x && circular_y) {
+        return ggml_pad_ext_circular(ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
+    }
+
+    if (circular_x && (lp0 != 0 || rp0 != 0)) {
+        x   = ggml_pad_ext_circular(ctx, x, lp0, rp0, 0, 0, 0, 0, 0, 0);
+        lp0 = rp0 = 0;
+    }
+    if (circular_y && (lp1 != 0 || rp1 != 0)) {
+        x   = ggml_pad_ext_circular(ctx, x, 0, 0, lp1, rp1, 0, 0, 0, 0);
+        lp1 = rp1 = 0;
+    }
+
+    if (lp0 != 0 || rp0 != 0 || lp1 != 0 || rp1 != 0 || lp2 != 0 || rp2 != 0 || lp3 != 0 || rp3 != 0) {
+        x = ggml_pad_ext(ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3);
+    }
+    return x;
+}
+
+__STATIC_INLINE__ struct ggml_tensor* ggml_ext_pad(struct ggml_context* ctx,
+                                                   struct ggml_tensor* x,
+                                                   int p0,
+                                                   int p1,
+                                                   int p2          = 0,
+                                                   int p3          = 0,
+                                                   bool circular_x = false,
+                                                   bool circular_y = false) {
+    return ggml_ext_pad_ext(ctx, x, p0, p0, p1, p1, p2, p2, p3, p3, circular_x, circular_y);
+}
+
 // w: [OC，IC, KH, KW]
 // x: [N, IC, IH, IW]
 // b: [OC,]
@@ -1001,20 +1044,29 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_conv_2d(struct ggml_context* ctx,
                                                       struct ggml_tensor* x,
                                                       struct ggml_tensor* w,
                                                       struct ggml_tensor* b,
-                                                       int s0      = 1,
-                                                       int s1      = 1,
-                                                       int p0      = 0,
-                                                       int p1      = 0,
-                                                       int d0      = 1,
-                                                       int d1      = 1,
-                                                       bool direct = false,
-                                                       float scale = 1.f) {
+                                                       int s0          = 1,
+                                                       int s1          = 1,
+                                                       int p0          = 0,
+                                                       int p1          = 0,
+                                                       int d0          = 1,
+                                                       int d1          = 1,
+                                                       bool direct     = false,
+                                                       bool circular_x = false,
+                                                       bool circular_y = false,
+                                                       float scale     = 1.f) {
    if (scale != 1.f) {
        x = ggml_scale(ctx, x, scale);
    }
    if (w->ne[2] != x->ne[2] && ggml_n_dims(w) == 2) {
        w = ggml_reshape_4d(ctx, w, 1, 1, w->ne[0], w->ne[1]);
    }
+
+    if ((p0 != 0 || p1 != 0) && (circular_x || circular_y)) {
+        x  = ggml_ext_pad(ctx, x, p0, p1, 0, 0, circular_x, circular_y);
+        p0 = 0;
+        p1 = 0;
+    }
+
    if (direct) {
        x = ggml_conv_2d_direct(ctx, w, x, s0, s1, p0, p1, d0, d1);
    } else {
@@ -1521,14 +1573,16 @@ struct WeightAdapter {
            float scale         = 1.f;
        } linear;
        struct {
-            int s0      = 1;
-            int s1      = 1;
-            int p0      = 0;
-            int p1      = 0;
-            int d0      = 1;
-            int d1      = 1;
-            bool direct = false;
-            float scale = 1.f;
+            int s0          = 1;
+            int s1          = 1;
+            int p0          = 0;
+            int p1          = 0;
+            int d0          = 1;
+            int d1          = 1;
+            bool direct     = false;
+            bool circular_x = false;
+            bool circular_y = false;
+            float scale     = 1.f;
        } conv2d;
    };
    virtual ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name) = 0;
@@ -1546,6 +1600,8 @@ struct GGMLRunnerContext {
    ggml_context* ggml_ctx                        = nullptr;
    bool flash_attn_enabled                       = false;
    bool conv2d_direct_enabled                    = false;
+    bool circular_x_enabled                       = false;
+    bool circular_y_enabled                       = false;
    std::shared_ptr<WeightAdapter> weight_adapter = nullptr;
 };

@@ -1582,6 +1638,8 @@ protected:

    bool flash_attn_enabled    = false;
    bool conv2d_direct_enabled = false;
+    bool circular_x_enabled    = false;
+    bool circular_y_enabled    = false;

    void alloc_params_ctx() {
        struct ggml_init_params params;
@@ -1859,6 +1917,8 @@ public:
        runner_ctx.backend               = runtime_backend;
        runner_ctx.flash_attn_enabled    = flash_attn_enabled;
        runner_ctx.conv2d_direct_enabled = conv2d_direct_enabled;
+        runner_ctx.circular_x_enabled    = circular_x_enabled;
+        runner_ctx.circular_y_enabled    = circular_y_enabled;
        runner_ctx.weight_adapter        = weight_adapter;
        return runner_ctx;
    }
@@ -2003,6 +2063,11 @@ public:
        conv2d_direct_enabled = enabled;
    }

+    void set_circular_axes(bool circular_x, bool circular_y) {
+        circular_x_enabled = circular_x;
+        circular_y_enabled = circular_y;
+    }
+
    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {
        weight_adapter = adapter;
    }
@@ -2266,15 +2331,17 @@ public:
        }
        if (ctx->weight_adapter) {
            WeightAdapter::ForwardParams forward_params;
-            forward_params.op_type       = WeightAdapter::ForwardParams::op_type_t::OP_CONV2D;
-            forward_params.conv2d.s0     = stride.second;
-            forward_params.conv2d.s1     = stride.first;
-            forward_params.conv2d.p0     = padding.second;
-            forward_params.conv2d.p1     = padding.first;
-            forward_params.conv2d.d0     = dilation.second;
-            forward_params.conv2d.d1     = dilation.first;
-            forward_params.conv2d.direct = ctx->conv2d_direct_enabled;
-            forward_params.conv2d.scale  = scale;
+            forward_params.op_type           = WeightAdapter::ForwardParams::op_type_t::OP_CONV2D;
+            forward_params.conv2d.s0         = stride.second;
+            forward_params.conv2d.s1         = stride.first;
+            forward_params.conv2d.p0         = padding.second;
+            forward_params.conv2d.p1         = padding.first;
+            forward_params.conv2d.d0         = dilation.second;
+            forward_params.conv2d.d1         = dilation.first;
+            forward_params.conv2d.direct     = ctx->conv2d_direct_enabled;
+            forward_params.conv2d.circular_x = ctx->circular_x_enabled;
+            forward_params.conv2d.circular_y = ctx->circular_y_enabled;
+            forward_params.conv2d.scale      = scale;
            return ctx->weight_adapter->forward_with_lora(ctx->ggml_ctx, x, w, b, prefix, forward_params);
        }
        return ggml_ext_conv_2d(ctx->ggml_ctx,
@@ -2288,6 +2355,8 @@ public:
                                dilation.second,
                                dilation.first,
                                ctx->conv2d_direct_enabled,
+                                ctx->circular_x_enabled,
+                                ctx->circular_y_enabled,
                                scale);
    }
 };
@@ -599,6 +599,8 @@ struct LoraModel : public GGMLRunner {
                                      forward_params.conv2d.d0,
                                      forward_params.conv2d.d1,
                                      forward_params.conv2d.direct,
+                                      forward_params.conv2d.circular_x,
+                                      forward_params.conv2d.circular_y,
                                      forward_params.conv2d.scale);
                if (lora_mid) {
                    lx = ggml_ext_conv_2d(ctx,
@@ -612,6 +614,8 @@ struct LoraModel : public GGMLRunner {
                                          1,
                                          1,
                                          forward_params.conv2d.direct,
+                                          forward_params.conv2d.circular_x,
+                                          forward_params.conv2d.circular_y,
                                          forward_params.conv2d.scale);
                }
                lx = ggml_ext_conv_2d(ctx,
@@ -625,6 +629,8 @@ struct LoraModel : public GGMLRunner {
                                      1,
                                      1,
                                      forward_params.conv2d.direct,
+                                      forward_params.conv2d.circular_x,
+                                      forward_params.conv2d.circular_y,
                                      forward_params.conv2d.scale);
            }

@@ -779,6 +785,8 @@ public:
                                   forward_params.conv2d.d0,
                                   forward_params.conv2d.d1,
                                   forward_params.conv2d.direct,
+                                   forward_params.conv2d.circular_x,
+                                   forward_params.conv2d.circular_y,
                                   forward_params.conv2d.scale);
        }
        for (auto& lora_model : lora_models) {
@@ -983,4 +983,4 @@ struct MMDiTRunner : public GGMLRunner {
    }
 };

-#endif
+#endif
@@ -1737,6 +1737,13 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
        // tensor_storage.ne[0], tensor_storage.ne[1], tensor_storage.ne[2], tensor_storage.ne[3],
        // tensor->n_dims, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);

+        if (!tensor->data) {
+            GGML_ASSERT(ggml_nelements(tensor) == 0);
+            // avoid crashing the gguf writer by setting a dummy pointer for zero-sized tensors
+            LOG_DEBUG("setting dummy pointer for zero-sized tensor %s", name.c_str());
+            tensor->data = ggml_get_mem_buffer(ggml_ctx);
+        }
+
        *dst_tensor = tensor;

        gguf_add_tensor(gguf_ctx, tensor);
@@ -1776,7 +1783,12 @@ int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type)
    return mem_size;
 }

-bool convert(const char* input_path, const char* vae_path, const char* output_path, sd_type_t output_type, const char* tensor_type_rules) {
+bool convert(const char* input_path,
+             const char* vae_path,
+             const char* output_path,
+             sd_type_t output_type,
+             const char* tensor_type_rules,
+             bool convert_name) {
    ModelLoader model_loader;

    if (!model_loader.init_from_file(input_path)) {
@@ -1790,7 +1802,9 @@ bool convert(const char* input_path, const char* vae_path, const char* output_pa
            return false;
        }
    }
-    model_loader.convert_tensors_name();
+    if (convert_name) {
+        model_loader.convert_tensors_name();
+    }
    bool success = model_loader.save_to_gguf_file(output_path, (ggml_type)output_type, tensor_type_rules);
    return success;
 }
@@ -835,6 +835,7 @@ std::string convert_sep_to_dot(std::string name) {
        "proj_out",
        "transformer_blocks",
        "single_transformer_blocks",
+        "single_blocks",
        "diffusion_model",
        "cond_stage_model",
        "first_stage_model",
@@ -876,7 +877,18 @@ std::string convert_sep_to_dot(std::string name) {
        "ff_context",
        "norm_added_q",
        "norm_added_v",
-        "to_add_out"};
+        "to_add_out",
+        "txt_mod",
+        "img_mod",
+        "txt_mlp",
+        "img_mlp",
+        "proj_mlp",
+        "wi_0",
+        "wi_1",
+        "norm1_context",
+        "ff_context",
+        "x_embedder",
+    };

    // record the positions of underscores that should NOT be replaced
    std::unordered_set<size_t> protected_positions;
@@ -1020,12 +1032,14 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
            }
        }

-        if (sd_version_is_unet(version) || is_lycoris_underline) {
+        // LOG_DEBUG("name %s %d", name.c_str(), version);
+
+        if (sd_version_is_unet(version) || sd_version_is_flux(version) || is_lycoris_underline) {
            name = convert_sep_to_dot(name);
        }
    }

-    std::vector<std::pair<std::string, std::string>> prefix_map = {
+    std::unordered_map<std::string, std::string> prefix_map = {
        {"diffusion_model.", "model.diffusion_model."},
        {"unet.", "model.diffusion_model."},
        {"transformer.", "model.diffusion_model."},  // dit
@@ -1040,8 +1054,13 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
        // {"te2.text_model.encoder.layers.", "cond_stage_model.1.model.transformer.resblocks."},
        {"te2.", "cond_stage_model.1.transformer."},
        {"te1.", "cond_stage_model.transformer."},
+        {"te3.", "text_encoders.t5xxl.transformer."},
    };

+    if (sd_version_is_flux(version)) {
+        prefix_map["te1."] = "text_encoders.clip_l.transformer.";
+    }
+
    replace_with_prefix_map(name, prefix_map);

    // diffusion model
@@ -354,14 +354,14 @@ namespace Qwen {
            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, params.patch_size * params.patch_size * params.out_channels));
        }

-        struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
+        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
                                              struct ggml_tensor* x) {
            int64_t W = x->ne[0];
            int64_t H = x->ne[1];

            int pad_h = (params.patch_size - H % params.patch_size) % params.patch_size;
            int pad_w = (params.patch_size - W % params.patch_size) % params.patch_size;
-            x         = ggml_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
+            x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
            return x;
        }

@@ -387,10 +387,10 @@ namespace Qwen {
            return x;
        }

-        struct ggml_tensor* process_img(struct ggml_context* ctx,
+        struct ggml_tensor* process_img(GGMLRunnerContext* ctx,
                                        struct ggml_tensor* x) {
            x = pad_to_patch_size(ctx, x);
-            x = patchify(ctx, x);
+            x = patchify(ctx->ggml_ctx, x);
            return x;
        }

@@ -466,12 +466,12 @@ namespace Qwen {
            int64_t C = x->ne[2];
            int64_t N = x->ne[3];

-            auto img            = process_img(ctx->ggml_ctx, x);
+            auto img            = process_img(ctx, x);
            uint64_t img_tokens = img->ne[1];

            if (ref_latents.size() > 0) {
                for (ggml_tensor* ref : ref_latents) {
-                    ref = process_img(ctx->ggml_ctx, ref);
+                    ref = process_img(ctx, ref);
                    img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                }
            }
@@ -565,6 +565,8 @@ namespace Qwen {
                                                  ref_latents,
                                                  increase_ref_index,
                                                  qwen_image_params.theta,
+                                                  circular_y_enabled,
+                                                  circular_x_enabled,
                                                  qwen_image_params.axes_dim);
            int pos_len = pe_vec.size() / qwen_image_params.axes_dim_sum / 2;
            // LOG_DEBUG("pos_len %d", pos_len);
@@ -684,4 +686,4 @@ namespace Qwen {

 }  // namespace name

-#endif  // __QWEN_IMAGE_HPP__
+#endif  // __QWEN_IMAGE_HPP__
@@ -1,6 +1,8 @@
 #ifndef __ROPE_HPP__
 #define __ROPE_HPP__

+#include <algorithm>
+#include <cmath>
 #include <vector>
 #include "ggml_extend.hpp"

@@ -39,7 +41,10 @@ namespace Rope {
        return flat_vec;
    }

-    __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos, int dim, int theta) {
+    __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos,
+                                                           int dim,
+                                                           int theta,
+                                                           const std::vector<int>& axis_wrap_dims = {}) {
        assert(dim % 2 == 0);
        int half_dim = dim / 2;

@@ -47,14 +52,31 @@ namespace Rope {

        std::vector<float> omega(half_dim);
        for (int i = 0; i < half_dim; ++i) {
-            omega[i] = 1.0 / std::pow(theta, scale[i]);
+            omega[i] = 1.0f / std::pow(theta, scale[i]);
        }

        int pos_size = pos.size();
        std::vector<std::vector<float>> out(pos_size, std::vector<float>(half_dim));
        for (int i = 0; i < pos_size; ++i) {
            for (int j = 0; j < half_dim; ++j) {
-                out[i][j] = pos[i] * omega[j];
+                float angle = pos[i] * omega[j];
+                if (!axis_wrap_dims.empty()) {
+                    size_t wrap_size = axis_wrap_dims.size();
+                    // mod batch size since we only store this for one item in the batch
+                    size_t wrap_idx = wrap_size > 0 ? (i % wrap_size) : 0;
+                    int wrap_dim    = axis_wrap_dims[wrap_idx];
+                    if (wrap_dim > 0) {
+                        constexpr float TWO_PI = 6.28318530717958647692f;
+                        float cycles           = omega[j] * wrap_dim / TWO_PI;
+                        // closest periodic harmonic, necessary to ensure things neatly tile
+                        // without this round, things don't tile at the boundaries and you end up
+                        // with the model knowing what is "center"
+                        float rounded = std::round(cycles);
+                        angle         = pos[i] * TWO_PI * rounded / wrap_dim;
+                    }
+                }
+
+                out[i][j] = angle;
            }
        }

@@ -89,16 +111,25 @@ namespace Rope {
                                                                       int patch_size,
                                                                       int bs,
                                                                       int axes_dim_num,
-                                                                       int index    = 0,
-                                                                       int h_offset = 0,
-                                                                       int w_offset = 0) {
+                                                                       int index       = 0,
+                                                                       int h_offset    = 0,
+                                                                       int w_offset    = 0,
+                                                                       bool scale_rope = false) {
        int h_len = (h + (patch_size / 2)) / patch_size;
        int w_len = (w + (patch_size / 2)) / patch_size;

        std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(axes_dim_num, 0.0));

-        std::vector<float> row_ids = linspace<float>(h_offset, h_len - 1 + h_offset, h_len);
-        std::vector<float> col_ids = linspace<float>(w_offset, w_len - 1 + w_offset, w_len);
+        int h_start = h_offset;
+        int w_start = w_offset;
+
+        if (scale_rope) {
+            h_start -= h_len / 2;
+            w_start -= w_len / 2;
+        }
+
+        std::vector<float> row_ids = linspace<float>(h_start, h_start + h_len - 1, h_len);
+        std::vector<float> col_ids = linspace<float>(w_start, w_start + w_len - 1, w_len);

        for (int i = 0; i < h_len; ++i) {
            for (int j = 0; j < w_len; ++j) {
@@ -137,7 +168,8 @@ namespace Rope {
    __STATIC_INLINE__ std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
                                                  int bs,
                                                  int theta,
-                                                  const std::vector<int>& axes_dim) {
+                                                  const std::vector<int>& axes_dim,
+                                                  const std::vector<std::vector<int>>& wrap_dims = {}) {
        std::vector<std::vector<float>> trans_ids = transpose(ids);
        size_t pos_len                            = ids.size() / bs;
        int num_axes                              = axes_dim.size();
@@ -152,7 +184,12 @@ namespace Rope {
        std::vector<std::vector<float>> emb(bs * pos_len, std::vector<float>(emb_dim * 2 * 2, 0.0));
        int offset = 0;
        for (int i = 0; i < num_axes; ++i) {
-            std::vector<std::vector<float>> rope_emb = rope(trans_ids[i], axes_dim[i], theta);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
+            std::vector<int> axis_wrap_dims;
+            if (!wrap_dims.empty() && i < (int)wrap_dims.size()) {
+                axis_wrap_dims = wrap_dims[i];
+            }
+            std::vector<std::vector<float>> rope_emb =
+                rope(trans_ids[i], axes_dim[i], theta, axis_wrap_dims);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
            for (int b = 0; b < bs; ++b) {
                for (int j = 0; j < pos_len; ++j) {
                    for (int k = 0; k < rope_emb[0].size(); ++k) {
@@ -171,7 +208,8 @@ namespace Rope {
                                                                   int axes_dim_num,
                                                                   const std::vector<ggml_tensor*>& ref_latents,
                                                                   bool increase_ref_index,
-                                                                   float ref_index_scale) {
+                                                                   float ref_index_scale,
+                                                                   bool scale_rope) {
        std::vector<std::vector<float>> ids;
        uint64_t curr_h_offset = 0;
        uint64_t curr_w_offset = 0;
@@ -185,6 +223,7 @@ namespace Rope {
                } else {
                    h_offset = curr_h_offset;
                }
+                scale_rope = false;
            }

            auto ref_ids = gen_flux_img_ids(ref->ne[1],
@@ -194,7 +233,8 @@ namespace Rope {
                                            axes_dim_num,
                                            static_cast<int>(index * ref_index_scale),
                                            h_offset,
-                                            w_offset);
+                                            w_offset,
+                                            scale_rope);
            ids          = concat_ids(ids, ref_ids, bs);

            if (increase_ref_index) {
@@ -222,7 +262,7 @@ namespace Rope {

        auto ids = concat_ids(txt_ids, img_ids, bs);
        if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, ref_index_scale);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, ref_index_scale, false);
            ids           = concat_ids(ids, refs_ids, bs);
        }
        return ids;
@@ -239,6 +279,8 @@ namespace Rope {
                                                     bool increase_ref_index,
                                                     float ref_index_scale,
                                                     int theta,
+                                                     bool circular_h,
+                                                     bool circular_w,
                                                     const std::vector<int>& axes_dim) {
        std::vector<std::vector<float>> ids = gen_flux_ids(h,
                                                           w,
@@ -250,7 +292,47 @@ namespace Rope {
                                                           ref_latents,
                                                           increase_ref_index,
                                                           ref_index_scale);
-        return embed_nd(ids, bs, theta, axes_dim);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int h_len = (h + (patch_size / 2)) / patch_size;
+            int w_len = (w + (patch_size / 2)) / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                size_t pos_len = ids.size() / bs;
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+                size_t cursor           = context_len;  // text first
+                const size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][cursor + token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][cursor + token_i] = w_len;
+                    }
+                }
+                cursor += img_tokens;
+                // reference latents
+                for (ggml_tensor* ref : ref_latents) {
+                    if (ref == nullptr) {
+                        continue;
+                    }
+                    int ref_h         = static_cast<int>(ref->ne[1]);
+                    int ref_w         = static_cast<int>(ref->ne[0]);
+                    int ref_h_l       = (ref_h + (patch_size / 2)) / patch_size;
+                    int ref_w_l       = (ref_w + (patch_size / 2)) / patch_size;
+                    size_t ref_tokens = static_cast<size_t>(ref_h_l) * static_cast<size_t>(ref_w_l);
+                    for (size_t token_i = 0; token_i < ref_tokens; ++token_i) {
+                        if (circular_h) {
+                            wrap_dims[1][cursor + token_i] = ref_h_l;
+                        }
+                        if (circular_w) {
+                            wrap_dims[2][cursor + token_i] = ref_w_l;
+                        }
+                    }
+                    cursor += ref_tokens;
+                }
+            }
+        }
+        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
    }

    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int h,
@@ -271,10 +353,10 @@ namespace Rope {
            }
        }
        int axes_dim_num = 3;
-        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num);
+        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, 0, 0, 0, true);
        auto ids         = concat_ids(txt_ids_repeated, img_ids, bs);
        if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, 1.f);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, 1.f, true);
            ids           = concat_ids(ids, refs_ids, bs);
        }
        return ids;
@@ -289,9 +371,57 @@ namespace Rope {
                                                           const std::vector<ggml_tensor*>& ref_latents,
                                                           bool increase_ref_index,
                                                           int theta,
+                                                           bool circular_h,
+                                                           bool circular_w,
                                                           const std::vector<int>& axes_dim) {
        std::vector<std::vector<float>> ids = gen_qwen_image_ids(h, w, patch_size, bs, context_len, ref_latents, increase_ref_index);
-        return embed_nd(ids, bs, theta, axes_dim);
+        std::vector<std::vector<int>> wrap_dims;
+        // This logic simply stores the (pad and patch_adjusted) sizes of images so we can make sure rope correctly tiles
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int pad_h = (patch_size - (h % patch_size)) % patch_size;
+            int pad_w = (patch_size - (w % patch_size)) % patch_size;
+            int h_len = (h + pad_h) / patch_size;
+            int w_len = (w + pad_w) / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                const size_t total_tokens = ids.size();
+                // Track per-token wrap lengths for the row/column axes so only spatial tokens become periodic.
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(total_tokens / bs, 0));
+                size_t cursor           = context_len;  // ignore text tokens
+                const size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][cursor + token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][cursor + token_i] = w_len;
+                    }
+                }
+                cursor += img_tokens;
+                // For each reference image, store wrap sizes as well
+                for (ggml_tensor* ref : ref_latents) {
+                    if (ref == nullptr) {
+                        continue;
+                    }
+                    int ref_h           = static_cast<int>(ref->ne[1]);
+                    int ref_w           = static_cast<int>(ref->ne[0]);
+                    int ref_pad_h       = (patch_size - (ref_h % patch_size)) % patch_size;
+                    int ref_pad_w       = (patch_size - (ref_w % patch_size)) % patch_size;
+                    int ref_h_len       = (ref_h + ref_pad_h) / patch_size;
+                    int ref_w_len       = (ref_w + ref_pad_w) / patch_size;
+                    size_t ref_n_tokens = static_cast<size_t>(ref_h_len) * static_cast<size_t>(ref_w_len);
+                    for (size_t token_i = 0; token_i < ref_n_tokens; ++token_i) {
+                        if (circular_h) {
+                            wrap_dims[1][cursor + token_i] = ref_h_len;
+                        }
+                        if (circular_w) {
+                            wrap_dims[2][cursor + token_i] = ref_w_len;
+                        }
+                    }
+                    cursor += ref_n_tokens;
+                }
+            }
+        }
+        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
    }

    __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
@@ -428,9 +558,33 @@ namespace Rope {
                                                        const std::vector<ggml_tensor*>& ref_latents,
                                                        bool increase_ref_index,
                                                        int theta,
+                                                        bool circular_h,
+                                                        bool circular_w,
                                                        const std::vector<int>& axes_dim) {
        std::vector<std::vector<float>> ids = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, increase_ref_index);
-        return embed_nd(ids, bs, theta, axes_dim);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int pad_h = (patch_size - (h % patch_size)) % patch_size;
+            int pad_w = (patch_size - (w % patch_size)) % patch_size;
+            int h_len = (h + pad_h) / patch_size;
+            int w_len = (w + pad_w) / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                size_t pos_len = ids.size() / bs;
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+                size_t cursor     = context_len + bound_mod(context_len, seq_multi_of);  // skip text (and its padding)
+                size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][cursor + token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][cursor + token_i] = w_len;
+                    }
+                }
+            }
+        }
+
+        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
    }

    __STATIC_INLINE__ struct ggml_tensor* apply_rope(struct ggml_context* ctx,
@@ -165,7 +165,27 @@ public:
 #endif
 #ifdef SD_USE_VULKAN
        LOG_DEBUG("Using Vulkan backend");
-        for (int device = 0; device < ggml_backend_vk_get_device_count(); ++device) {
+        size_t device          = 0;
+        const int device_count = ggml_backend_vk_get_device_count();
+        if (device_count) {
+            const char* SD_VK_DEVICE = getenv("SD_VK_DEVICE");
+            if (SD_VK_DEVICE != nullptr) {
+                std::string sd_vk_device_str = SD_VK_DEVICE;
+                try {
+                    device = std::stoull(sd_vk_device_str);
+                } catch (const std::invalid_argument&) {
+                    LOG_WARN("SD_VK_DEVICE environment variable is not a valid integer (%s). Falling back to device 0.", SD_VK_DEVICE);
+                    device = 0;
+                } catch (const std::out_of_range&) {
+                    LOG_WARN("SD_VK_DEVICE environment variable value is out of range for `unsigned long long` type (%s). Falling back to device 0.", SD_VK_DEVICE);
+                    device = 0;
+                }
+                if (device >= device_count) {
+                    LOG_WARN("Cannot find targeted vulkan device (%llu). Falling back to device 0.", device);
+                    device = 0;
+                }
+            }
+            LOG_INFO("Vulkan: Using device %llu", device);
            backend = ggml_backend_vk_init(device);
        }
        if (!backend) {
@@ -385,6 +405,10 @@ public:
            vae_decode_only = false;
        }

+        if (sd_ctx_params->circular_x || sd_ctx_params->circular_y) {
+            LOG_INFO("Using circular padding for convolutions");
+        }
+
        bool clip_on_cpu = sd_ctx_params->keep_clip_on_cpu;

        {
@@ -685,6 +709,20 @@ public:
                }
                pmid_model->get_param_tensors(tensors, "pmid");
            }
+
+            diffusion_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
+            if (high_noise_diffusion_model) {
+                high_noise_diffusion_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
+            }
+            if (control_net) {
+                control_net->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
+            }
+            if (first_stage_model) {
+                first_stage_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
+            }
+            if (tae_first_stage) {
+                tae_first_stage->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
+            }
        }

        struct ggml_init_params params;
@@ -708,6 +746,8 @@ public:
        if (stacked_id) {
            ignore_tensors.insert("pmid.unet.");
        }
+        ignore_tensors.insert("model.diffusion_model.__x0__");
+        ignore_tensors.insert("model.diffusion_model.__32x32__");

        if (vae_decode_only) {
            ignore_tensors.insert("first_stage_model.encoder");
@@ -842,6 +882,7 @@ public:
                    }
                } else if (sd_version_is_flux(version)) {
                    pred_type = FLUX_FLOW_PRED;
+
                    if (flow_shift == INFINITY) {
                        flow_shift = 1.0f;  // TODO: validate
                        for (const auto& [name, tensor_storage] : tensor_storage_map) {
@@ -1496,7 +1537,18 @@ public:
        }
        std::vector<int> skip_layers(guidance.slg.layers, guidance.slg.layers + guidance.slg.layer_count);

-        float cfg_scale     = guidance.txt_cfg;
+        float cfg_scale = guidance.txt_cfg;
+        if (cfg_scale < 1.f) {
+            if (cfg_scale == 0.f) {
+                // Diffusers follow the convention from the original paper
+                // (https://arxiv.org/abs/2207.12598v1), so many distilled model docs
+                // recommend 0 as guidance; warn the user that it'll disable prompt folowing
+                LOG_WARN("unconditioned mode, images won't follow the prompt (use cfg-scale=1 for distilled models)");
+            } else {
+                LOG_WARN("cfg value out of expected range may produce unexpected results");
+            }
+        }
+
        float img_cfg_scale = std::isfinite(guidance.img_cfg) ? guidance.img_cfg : guidance.txt_cfg;
        float slg_scale     = guidance.slg.scale;

@@ -2525,6 +2577,8 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
    sd_ctx_params->keep_control_net_on_cpu = false;
    sd_ctx_params->keep_vae_on_cpu         = false;
    sd_ctx_params->diffusion_flash_attn    = false;
+    sd_ctx_params->circular_x              = false;
+    sd_ctx_params->circular_y              = false;
    sd_ctx_params->chroma_use_dit_mask     = true;
    sd_ctx_params->chroma_use_t5_mask      = false;
    sd_ctx_params->chroma_t5_mask_pad      = 1;
@@ -2564,6 +2618,8 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
             "keep_control_net_on_cpu: %s\n"
             "keep_vae_on_cpu: %s\n"
             "diffusion_flash_attn: %s\n"
+             "circular_x: %s\n"
+             "circular_y: %s\n"
             "chroma_use_dit_mask: %s\n"
             "chroma_use_t5_mask: %s\n"
             "chroma_t5_mask_pad: %d\n",
@@ -2593,6 +2649,8 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
             BOOL_STR(sd_ctx_params->keep_control_net_on_cpu),
             BOOL_STR(sd_ctx_params->keep_vae_on_cpu),
             BOOL_STR(sd_ctx_params->diffusion_flash_attn),
+             BOOL_STR(sd_ctx_params->circular_x),
+             BOOL_STR(sd_ctx_params->circular_y),
             BOOL_STR(sd_ctx_params->chroma_use_dit_mask),
             BOOL_STR(sd_ctx_params->chroma_use_t5_mask),
             sd_ctx_params->chroma_t5_mask_pad);
@@ -2777,13 +2835,16 @@ enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx) {
    return EULER_A_SAMPLE_METHOD;
 }

-enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx) {
+enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx, enum sample_method_t sample_method) {
    if (sd_ctx != nullptr && sd_ctx->sd != nullptr) {
        auto edm_v_denoiser = std::dynamic_pointer_cast<EDMVDenoiser>(sd_ctx->sd->denoiser);
        if (edm_v_denoiser) {
            return EXPONENTIAL_SCHEDULER;
        }
    }
+    if (sample_method == LCM_SAMPLE_METHOD) {
+        return LCM_SCHEDULER;
+    }
    return DISCRETE_SCHEDULER;
 }

@@ -3218,9 +3279,13 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
            LOG_WARN("sample_steps != custom_sigmas_count - 1, set sample_steps to %d", sample_steps);
        }
    } else {
+        scheduler_t scheduler = sd_img_gen_params->sample_params.scheduler;
+        if (scheduler == SCHEDULER_COUNT) {
+            scheduler = sd_get_default_scheduler(sd_ctx, sample_method);
+        }
        sigmas = sd_ctx->sd->denoiser->get_sigmas(sample_steps,
                                                  sd_ctx->sd->get_image_seq_len(height, width),
-                                                  sd_img_gen_params->sample_params.scheduler,
+                                                  scheduler,
                                                  sd_ctx->sd->version);
    }

@@ -3503,9 +3568,13 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
            }
        }
    } else {
+        scheduler_t scheduler = sd_vid_gen_params->sample_params.scheduler;
+        if (scheduler == SCHEDULER_COUNT) {
+            scheduler = sd_get_default_scheduler(sd_ctx, sample_method);
+        }
        sigmas = sd_ctx->sd->denoiser->get_sigmas(total_steps,
                                                  0,
-                                                  sd_vid_gen_params->sample_params.scheduler,
+                                                  scheduler,
                                                  sd_ctx->sd->version);
    }

@@ -189,6 +189,8 @@ typedef struct {
    bool tae_preview_only;
    bool diffusion_conv_direct;
    bool vae_conv_direct;
+    bool circular_x;
+    bool circular_y;
    bool force_sdxl_vae_conv_scale;
    bool chroma_use_dit_mask;
    bool chroma_use_t5_mask;
@@ -335,7 +337,7 @@ SD_API void sd_sample_params_init(sd_sample_params_t* sample_params);
 SD_API char* sd_sample_params_to_str(const sd_sample_params_t* sample_params);

 SD_API enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx);
-SD_API enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx);
+SD_API enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx, enum sample_method_t sample_method);

 SD_API void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params);
 SD_API char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params);
@@ -363,7 +365,8 @@ SD_API bool convert(const char* input_path,
                    const char* vae_path,
                    const char* output_path,
                    enum sd_type_t output_type,
-                    const char* tensor_type_rules);
+                    const char* tensor_type_rules,
+                    bool convert_name);

 SD_API bool preprocess_canny(sd_image_t image,
                             float high_threshold,
@@ -75,7 +75,7 @@ namespace WAN {
                lp2 -= (int)cache_x->ne[2];
            }

-            x = ggml_pad_ext(ctx->ggml_ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, 0, 0);
+            x = ggml_ext_pad_ext(ctx->ggml_ctx, x, lp0, rp0, lp1, rp1, lp2, rp2, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
            return ggml_ext_conv_3d(ctx->ggml_ctx, x, w, b, in_channels,
                                    std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
                                    0, 0, 0,
@@ -206,9 +206,9 @@ namespace WAN {
                } else if (mode == "upsample3d") {
                    x = ggml_upscale(ctx->ggml_ctx, x, 2, GGML_SCALE_MODE_NEAREST);
                } else if (mode == "downsample2d") {
-                    x = ggml_pad(ctx->ggml_ctx, x, 1, 1, 0, 0);
+                    x = ggml_ext_pad(ctx->ggml_ctx, x, 1, 1, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
                } else if (mode == "downsample3d") {
-                    x = ggml_pad(ctx->ggml_ctx, x, 1, 1, 0, 0);
+                    x = ggml_ext_pad(ctx->ggml_ctx, x, 1, 1, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
                }
                x = resample_1->forward(ctx, x);
                x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));  // (c, t, h, w)
@@ -1826,7 +1826,7 @@ namespace WAN {
            }
        }

-        struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
+        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
                                              struct ggml_tensor* x) {
            int64_t W = x->ne[0];
            int64_t H = x->ne[1];
@@ -1835,8 +1835,7 @@ namespace WAN {
            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]
-
+            ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, pad_t, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
            return x;
        }

@@ -1986,14 +1985,14 @@ namespace WAN {
            int64_t T = x->ne[2];
            int64_t C = x->ne[3];

-            x = pad_to_patch_size(ctx->ggml_ctx, x);
+            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->ggml_ctx, time_dim_concat);
+                time_dim_concat = pad_to_patch_size(ctx, time_dim_concat);
                x               = ggml_concat(ctx->ggml_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));
            }
@@ -324,14 +324,14 @@ namespace ZImage {
            blocks["final_layer"] = std::make_shared<FinalLayer>(z_image_params.hidden_size, z_image_params.patch_size, z_image_params.out_channels);
        }

-        struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
+        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
                                              struct ggml_tensor* x) {
            int64_t W = x->ne[0];
            int64_t H = x->ne[1];

            int pad_h = (z_image_params.patch_size - H % z_image_params.patch_size) % z_image_params.patch_size;
            int pad_w = (z_image_params.patch_size - W % z_image_params.patch_size) % z_image_params.patch_size;
-            x         = ggml_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
+            x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
            return x;
        }

@@ -357,10 +357,10 @@ namespace ZImage {
            return x;
        }

-        struct ggml_tensor* process_img(struct ggml_context* ctx,
+        struct ggml_tensor* process_img(GGMLRunnerContext* ctx,
                                        struct ggml_tensor* x) {
            x = pad_to_patch_size(ctx, x);
-            x = patchify(ctx, x);
+            x = patchify(ctx->ggml_ctx, x);
            return x;
        }

@@ -473,12 +473,12 @@ namespace ZImage {
            int64_t C = x->ne[2];
            int64_t N = x->ne[3];

-            auto img             = process_img(ctx->ggml_ctx, x);
+            auto img             = process_img(ctx, x);
            uint64_t n_img_token = img->ne[1];

            if (ref_latents.size() > 0) {
                for (ggml_tensor* ref : ref_latents) {
-                    ref = process_img(ctx->ggml_ctx, ref);
+                    ref = process_img(ctx, ref);
                    img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                }
            }
@@ -552,6 +552,8 @@ namespace ZImage {
                                               ref_latents,
                                               increase_ref_index,
                                               z_image_params.theta,
+                                               circular_y_enabled,
+                                               circular_x_enabled,
                                               z_image_params.axes_dim);
            int pos_len = pe_vec.size() / z_image_params.axes_dim_sum / 2;
            // LOG_DEBUG("pos_len %d", pos_len);
Author	SHA1	Message	Date
leejet	ca5b1969a8	feat: do not convert tensor names by default in convert mode (#1122 )	2025-12-21 18:40:10 +08:00
Phylliida Dev Phylliida leejet	50ff966445	feat: add seamless texture generation support (#914 ) * global bool * reworked circular to global flag * cleaner implementation of tiling support in sd cpp * cleaned rope * working simplified but still need wraps * Further clean of rope * resolve flux conflict * switch to pad op circular only * Set ggml to most recent * Revert ggml temp * Update ggml to most recent * Revert unneded flux change * move circular flag to the GGMLRunnerContext * Pass through circular param in all places where conv is called * fix of constant and minor cleanup * Added back --circular option * Conv2d circular in vae and various models * Fix temporal padding for qwen image and other vaes * Z Image circular tiling * x and y axis seamless only * First attempt at chroma seamless x and y * refactor into pure x and y, almost there * Fix crash on chroma * Refactor into cleaner variable choices * Removed redundant set_circular_enabled * Sync ggml * simplify circular parameter * format code * no need to perform circular pad on the clip * simplify circular_axes setting * unify function naming * remove unnecessary member variables * simplify rope --------- Co-authored-by: Phylliida <phylliidadev@gmail.com> Co-authored-by: leejet <leejet714@gmail.com>	2025-12-21 18:06:47 +08:00
leejet	88ec9d30b1	feat: add scale_rope support (#1121 )	2025-12-21 15:40:21 +08:00
stduhpf	60abda56e0	feat: select vulkan device with env variable (#629 )	2025-12-21 15:35:38 +08:00
stduhpfandleejet	23fce0bd84	feat: add support for Chroma Radiance x0 (#1091 ) * Add x0 Flux pred (+prepare for others) * Fix convert models with empty tensors * patch_32 exp support attempt * improve support for patch_32 * follow official pipeline --------- Co-authored-by: leejet <leejet714@gmail.com>	2025-12-20 00:55:57 +08:00
Wagner Bruna	7c88c4765c	chore: give feedback about cfg values smaller than 1 (#1088 )	2025-12-19 23:41:52 +08:00
Weiqi Gao	1f77545cf8	docs: document usage of tae for VRAM reduction using wan (#1108 )	2025-12-19 23:31:09 +08:00
leejet	8e9f3a4d9e	feat: add support for underline style lora of flux (#1103 ) * feat: add support for underline style lora of flux * add support for underline style lora of t5 * add more protected tokens	2025-12-18 21:44:16 +08:00
Wagner Brunaandleejet	78e15bd4af	feat: default to LCM scheduler for LCM sampling (#1109 ) * feat: default to LCM scheduler for LCM sampling * fix bug and attempt to get default scheduler for vid_gen when none is set --------- Co-authored-by: leejet <leejet714@gmail.com>	2025-12-18 21:43:39 +08:00