refactor: add some sd vesion helper functions

* repair flash attention in _ext
this does not fix the currently broken fa behind the define, which is only used by VAE

Co-authored-by: FSSRepo <FSSRepo@users.noreply.github.com>

* make flash attention in the diffusion model a runtime flag
no support for sd3 or video

* remove old flash attention option and switch vae over to attn_ext

* update docs

* format code

---------

Co-authored-by: FSSRepo <FSSRepo@users.noreply.github.com>
Co-authored-by: leejet <leejet714@gmail.com>

CMakeLists.txt

@@ -29,7 +29,6 @@ option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
 option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
-option(SD_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
 option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
@@ -61,11 +60,6 @@ if (SD_HIPBLAS)
     endif()
 endif ()
 
-if(SD_FLASH_ATTN)
-    message("-- Use Flash Attention for memory optimization")
-    add_definitions(-DSD_USE_FLASH_ATTENTION)
-endif()
-
 set(SD_LIB stable-diffusion)
 
 file(GLOB SD_LIB_SOURCES 

README.md

@@ -24,7 +24,7 @@ Inference of Stable Diffusion and Flux in pure C/C++
 - Full CUDA, Metal, Vulkan and SYCL backend for GPU acceleration.
 - Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models
     - No need to convert to `.ggml` or `.gguf` anymore!
-- Flash Attention for memory usage optimization (only cpu for now)
+- Flash Attention for memory usage optimization
 - Original `txt2img` and `img2img` mode
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
@@ -182,11 +182,21 @@ Example of text2img by using SYCL backend:
 
 ##### Using Flash Attention
 
-Enabling flash attention reduces memory usage by at least 400 MB. At the moment, it is not supported when CUBLAS is enabled because the kernel implementation is missing.
+Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
+eg.:
+ - flux 768x768 ~600mb
+ - SD2 768x768 ~1400mb
 
+For most backends, it slows things down, but for cuda it generally speeds it up too.
+At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
+
+Run by adding `--diffusion-fa` to the arguments and watch for:
 ```
-cmake .. -DSD_FLASH_ATTN=ON
-cmake --build . --config Release
+[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
+```
+and the compute buffer shrink in the debug log:
+```
+[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
 ```
 
 ### Run
@@ -240,6 +250,9 @@ arguments:
   --vae-tiling                       process vae in tiles to reduce memory usage
   --vae-on-cpu                       keep vae in cpu (for low vram)
   --clip-on-cpu                      keep clip in cpu (for low vram)
+  --diffusion-fa                     use flash attention in the diffusion model (for low vram)
+                                     Might lower quality, since it implies converting k and v to f16.
+                                     This might crash if it is not supported by the backend.
   --control-net-cpu                  keep controlnet in cpu (for low vram)
   --canny                            apply canny preprocessor (edge detection)
   --color                            Colors the logging tags according to level
@@ -292,12 +305,15 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra
 
 * Golang: [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
 * C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
+* Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
+* Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)
 
 ## UIs
 
 These projects use `stable-diffusion.cpp` as a backend for their image generation.
 
 - [Jellybox](https://jellybox.com)
+- [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
 
 ## Contributors
 

clip.hpp

@@ -343,6 +343,13 @@ public:
         }
     }
 
+    std::string clean_up_tokenization(std::string& text) {
+        std::regex pattern(R"( ,)");
+        // Replace " ," with ","
+        std::string result = std::regex_replace(text, pattern, ",");
+        return result;
+    }
+
     std::string decode(const std::vector<int>& tokens) {
         std::string text = "";
         for (int t : tokens) {
@@ -351,8 +358,12 @@ public:
             std::u32string ts = decoder[t];
             // printf("%d, %s \n", t,  utf32_to_utf8(ts).c_str());
             std::string s = utf32_to_utf8(ts);
-            if (s.length() >= 4 && ends_with(s, "</w>")) {
-                text += " " + s.replace(s.length() - 4, s.length() - 1, "");
+            if (s.length() >= 4) {
+                if (ends_with(s, "</w>")) {
+                    text += s.replace(s.length() - 4, s.length() - 1, "") + " ";
+                } else {
+                    text += s;
+                }
             } else {
                 text += " " + s;
             }
@@ -364,6 +375,7 @@ public:
 
         // std::string s((char *)bytes.data());
         // std::string s = "";
+        text = clean_up_tokenization(text);
         return trim(text);
     }
 
@@ -711,8 +723,12 @@ public:
         if (return_pooled) {
             auto text_projection = params["text_projection"];
             ggml_tensor* pooled  = ggml_view_1d(ctx, x, hidden_size, x->nb[1] * max_token_idx);
-            pooled               = ggml_mul_mat(ctx, ggml_cont(ctx, ggml_transpose(ctx, text_projection)), pooled);
-            return pooled;
+            if (text_projection != NULL) {
+                pooled = ggml_nn_linear(ctx, pooled, text_projection, NULL);
+            } else {
+                LOG_DEBUG("Missing text_projection matrix, assuming identity...");
+            }
+            return pooled;  // [hidden_size, 1, 1]
         }
 
         return x;  // [N, n_token, hidden_size]
@@ -761,14 +777,17 @@ public:
         auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
         x      = pre_layernorm->forward(ctx, x);
         x      = encoder->forward(ctx, x, -1, false);
-        x      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
+        // print_ggml_tensor(x, true, "ClipVisionModel x: ");
+        auto last_hidden_state = x;
+        x                      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
 
         GGML_ASSERT(x->ne[3] == 1);
         if (return_pooled) {
             ggml_tensor* pooled = ggml_cont(ctx, ggml_view_2d(ctx, x, x->ne[0], x->ne[2], x->nb[2], 0));
             return pooled;  // [N, hidden_size]
         } else {
-            return x;  // [N, n_token, hidden_size]
+            // return x;  // [N, n_token, hidden_size]
+            return last_hidden_state;  // [N, n_token, hidden_size]
         }
     }
 };

common.hpp

@@ -245,16 +245,19 @@ protected:
     int64_t context_dim;
     int64_t n_head;
     int64_t d_head;
+    bool flash_attn;
 
 public:
     CrossAttention(int64_t query_dim,
                    int64_t context_dim,
                    int64_t n_head,
-                   int64_t d_head)
+                   int64_t d_head,
+                   bool flash_attn = false)
         : n_head(n_head),
           d_head(d_head),
           query_dim(query_dim),
-          context_dim(context_dim) {
+          context_dim(context_dim),
+          flash_attn(flash_attn) {
         int64_t inner_dim = d_head * n_head;
 
         blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, false));
@@ -283,7 +286,7 @@ public:
         auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
         auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]
 
-        x = ggml_nn_attention_ext(ctx, q, k, v, n_head, NULL, false);  // [N, n_token, inner_dim]
+        x = ggml_nn_attention_ext(ctx, q, k, v, n_head, NULL, false, false, flash_attn);  // [N, n_token, inner_dim]
 
         x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
         return x;
@@ -301,15 +304,16 @@ public:
                           int64_t n_head,
                           int64_t d_head,
                           int64_t context_dim,
-                          bool ff_in = false)
+                          bool ff_in      = false,
+                          bool flash_attn = false)
         : n_head(n_head), d_head(d_head), ff_in(ff_in) {
         // disable_self_attn is always False
         // disable_temporal_crossattention is always False
         // switch_temporal_ca_to_sa is always False
         // inner_dim is always None or equal to dim
         // gated_ff is always True
-        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head));
-        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head));
+        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head, flash_attn));
+        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head, flash_attn));
         blocks["ff"]    = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
         blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
         blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
@@ -374,7 +378,8 @@ public:
                        int64_t n_head,
                        int64_t d_head,
                        int64_t depth,
-                       int64_t context_dim)
+                       int64_t context_dim,
+                       bool flash_attn = false)
         : in_channels(in_channels),
           n_head(n_head),
           d_head(d_head),
@@ -388,7 +393,7 @@ public:
 
         for (int i = 0; i < depth; i++) {
             std::string name = "transformer_blocks." + std::to_string(i);
-            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim));
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim, false, flash_attn));
         }
 
         blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
@@ -511,4 +516,4 @@ public:
     }
 };
 
-#endif  // __COMMON_HPP__
+#endif  // __COMMON_HPP__

conditioner.hpp

@@ -43,7 +43,8 @@ struct Conditioner {
 // ldm.modules.encoders.modules.FrozenCLIPEmbedder
 // Ref: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cad87bf4e3e0b0a759afa94e933527c3123d59bc/modules/sd_hijack_clip.py#L283
 struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
-    SDVersion version = VERSION_SD1;
+    SDVersion version    = VERSION_SD1;
+    PMVersion pm_version = PM_VERSION_1;
     CLIPTokenizer tokenizer;
     ggml_type wtype;
     std::shared_ptr<CLIPTextModelRunner> text_model;
@@ -59,8 +60,9 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                       ggml_type wtype,
                                       const std::string& embd_dir,
                                       SDVersion version = VERSION_SD1,
+                                      PMVersion pv      = PM_VERSION_1,
                                       int clip_skip     = -1)
-        : version(version), tokenizer(version == VERSION_SD2 ? 0 : 49407), embd_dir(embd_dir), wtype(wtype) {
+        : version(version), pm_version(pv), tokenizer(version == VERSION_SD2 ? 0 : 49407), embd_dir(embd_dir), wtype(wtype) {
         if (clip_skip <= 0) {
             clip_skip = 1;
             if (version == VERSION_SD2 || version == VERSION_SDXL) {
@@ -268,7 +270,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                 std::vector<int> clean_input_ids_tmp;
                 for (uint32_t i = 0; i < class_token_index[0]; i++)
                     clean_input_ids_tmp.push_back(clean_input_ids[i]);
-                for (uint32_t i = 0; i < num_input_imgs; i++)
+                for (uint32_t i = 0; i < (pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs); i++)
                     clean_input_ids_tmp.push_back(class_token);
                 for (uint32_t i = class_token_index[0] + 1; i < clean_input_ids.size(); i++)
                     clean_input_ids_tmp.push_back(clean_input_ids[i]);
@@ -279,13 +281,16 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             tokens.insert(tokens.end(), clean_input_ids.begin(), clean_input_ids.end());
             weights.insert(weights.end(), clean_input_ids.size(), curr_weight);
         }
-        tokens.insert(tokens.begin(), tokenizer.BOS_TOKEN_ID);
-        weights.insert(weights.begin(), 1.0);
+        // BUG!! double couting, pad_tokens will add BOS at the beginning
+        // tokens.insert(tokens.begin(), tokenizer.BOS_TOKEN_ID);
+        // weights.insert(weights.begin(), 1.0);
 
         tokenizer.pad_tokens(tokens, weights, max_length, padding);
-
+        int offset = pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs;
         for (uint32_t i = 0; i < tokens.size(); i++) {
-            if (class_idx + 1 <= i && i < class_idx + 1 + num_input_imgs)
+            // if (class_idx + 1 <= i && i < class_idx + 1 + 2*num_input_imgs) // photomaker V2 has num_tokens(=2)*num_input_imgs
+            if (class_idx + 1 <= i && i < class_idx + 1 + offset)  // photomaker V2 has num_tokens(=2)*num_input_imgs
+                                                                   // hardcode for now
                 class_token_mask.push_back(true);
             else
                 class_token_mask.push_back(false);
@@ -798,21 +803,16 @@ struct SD3CLIPEmbedder : public Conditioner {
                 }
 
                 if (chunk_idx == 0) {
-                    // auto it = std::find(chunk_tokens.begin(), chunk_tokens.end(), clip_l_tokenizer.EOS_TOKEN_ID);
-                    // max_token_idx = std::min<size_t>(std::distance(chunk_tokens.begin(), it), chunk_tokens.size() - 1);
-                    // clip_l->compute(n_threads,
-                    //                 input_ids,
-                    //                 0,
-                    //                 NULL,
-                    //                 max_token_idx,
-                    //                 true,
-                    //                 &pooled_l,
-                    //                 work_ctx);
-
-                    // clip_l.transformer.text_model.text_projection no in file, ignore
-                    // TODO: use torch.eye(embed_dim) as default clip_l.transformer.text_model.text_projection
-                    pooled_l = ggml_new_tensor_1d(work_ctx, GGML_TYPE_F32, 768);
-                    ggml_set_f32(pooled_l, 0.f);
+                    auto it       = std::find(chunk_tokens.begin(), chunk_tokens.end(), clip_l_tokenizer.EOS_TOKEN_ID);
+                    max_token_idx = std::min<size_t>(std::distance(chunk_tokens.begin(), it), chunk_tokens.size() - 1);
+                    clip_l->compute(n_threads,
+                                    input_ids,
+                                    0,
+                                    NULL,
+                                    max_token_idx,
+                                    true,
+                                    &pooled_l,
+                                    work_ctx);
                 }
             }
 
@@ -852,21 +852,16 @@ struct SD3CLIPEmbedder : public Conditioner {
                 }
 
                 if (chunk_idx == 0) {
-                    // auto it = std::find(chunk_tokens.begin(), chunk_tokens.end(), clip_g_tokenizer.EOS_TOKEN_ID);
-                    // max_token_idx = std::min<size_t>(std::distance(chunk_tokens.begin(), it), chunk_tokens.size() - 1);
-                    // clip_g->compute(n_threads,
-                    //                 input_ids,
-                    //                 0,
-                    //                 NULL,
-                    //                 max_token_idx,
-                    //                 true,
-                    //                 &pooled_g,
-                    //                 work_ctx);
-                    // clip_l.transformer.text_model.text_projection no in file, ignore pooled_g too
-
-                    // TODO: fix pooled_g
-                    pooled_g = ggml_new_tensor_1d(work_ctx, GGML_TYPE_F32, 1280);
-                    ggml_set_f32(pooled_g, 0.f);
+                    auto it       = std::find(chunk_tokens.begin(), chunk_tokens.end(), clip_g_tokenizer.EOS_TOKEN_ID);
+                    max_token_idx = std::min<size_t>(std::distance(chunk_tokens.begin(), it), chunk_tokens.size() - 1);
+                    clip_g->compute(n_threads,
+                                    input_ids,
+                                    0,
+                                    NULL,
+                                    max_token_idx,
+                                    true,
+                                    &pooled_g,
+                                    work_ctx);
                 }
             }
 
@@ -1104,21 +1099,17 @@ struct FluxCLIPEmbedder : public Conditioner {
                 auto input_ids       = vector_to_ggml_tensor_i32(work_ctx, chunk_tokens);
                 size_t max_token_idx = 0;
 
-                // auto it = std::find(chunk_tokens.begin(), chunk_tokens.end(), clip_l_tokenizer.EOS_TOKEN_ID);
-                // max_token_idx = std::min<size_t>(std::distance(chunk_tokens.begin(), it), chunk_tokens.size() - 1);
-                // clip_l->compute(n_threads,
-                //                 input_ids,
-                //                 0,
-                //                 NULL,
-                //                 max_token_idx,
-                //                 true,
-                //                 &pooled,
-                //                 work_ctx);
+                auto it       = std::find(chunk_tokens.begin(), chunk_tokens.end(), clip_l_tokenizer.EOS_TOKEN_ID);
+                max_token_idx = std::min<size_t>(std::distance(chunk_tokens.begin(), it), chunk_tokens.size() - 1);
 
-                // clip_l.transformer.text_model.text_projection no in file, ignore
-                // TODO: use torch.eye(embed_dim) as default clip_l.transformer.text_model.text_projection
-                pooled = ggml_new_tensor_1d(work_ctx, GGML_TYPE_F32, 768);
-                ggml_set_f32(pooled, 0.f);
+                clip_l->compute(n_threads,
+                                input_ids,
+                                0,
+                                NULL,
+                                max_token_idx,
+                                true,
+                                &pooled,
+                                work_ctx);
             }
 
             // t5

diffusion_model.hpp

@@ -17,7 +17,8 @@ struct DiffusionModel {
                          std::vector<struct ggml_tensor*> controls = {},
                          float control_strength                    = 0.f,
                          struct ggml_tensor** output               = NULL,
-                         struct ggml_context* output_ctx           = NULL)                        = 0;
+                         struct ggml_context* output_ctx           = NULL,
+                         std::vector<int> skip_layers              = std::vector<int>())             = 0;
     virtual void alloc_params_buffer()                                                  = 0;
     virtual void free_params_buffer()                                                   = 0;
     virtual void free_compute_buffer()                                                  = 0;
@@ -31,8 +32,9 @@ struct UNetModel : public DiffusionModel {
 
     UNetModel(ggml_backend_t backend,
               ggml_type wtype,
-              SDVersion version = VERSION_SD1)
-        : unet(backend, wtype, version) {
+              SDVersion version = VERSION_SD1,
+              bool flash_attn   = false)
+        : unet(backend, wtype, version, flash_attn) {
     }
 
     void alloc_params_buffer() {
@@ -70,7 +72,9 @@ struct UNetModel : public DiffusionModel {
                  std::vector<struct ggml_tensor*> controls = {},
                  float control_strength                    = 0.f,
                  struct ggml_tensor** output               = NULL,
-                 struct ggml_context* output_ctx           = NULL) {
+                 struct ggml_context* output_ctx           = NULL,
+                 std::vector<int> skip_layers              = std::vector<int>()) {
+        (void)skip_layers;  // SLG doesn't work with UNet models
         return unet.compute(n_threads, x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength, output, output_ctx);
     }
 };
@@ -119,8 +123,9 @@ struct MMDiTModel : public DiffusionModel {
                  std::vector<struct ggml_tensor*> controls = {},
                  float control_strength                    = 0.f,
                  struct ggml_tensor** output               = NULL,
-                 struct ggml_context* output_ctx           = NULL) {
-        return mmdit.compute(n_threads, x, timesteps, context, y, output, output_ctx);
+                 struct ggml_context* output_ctx           = NULL,
+                 std::vector<int> skip_layers              = std::vector<int>()) {
+        return mmdit.compute(n_threads, x, timesteps, context, y, output, output_ctx, skip_layers);
     }
 };
 
@@ -129,8 +134,9 @@ struct FluxModel : public DiffusionModel {
 
     FluxModel(ggml_backend_t backend,
               ggml_type wtype,
-              SDVersion version = VERSION_FLUX_DEV)
-        : flux(backend, wtype, version) {
+              SDVersion version = VERSION_FLUX_DEV,
+              bool flash_attn   = false)
+        : flux(backend, wtype, version, flash_attn) {
     }
 
     void alloc_params_buffer() {
@@ -168,9 +174,10 @@ struct FluxModel : public DiffusionModel {
                  std::vector<struct ggml_tensor*> controls = {},
                  float control_strength                    = 0.f,
                  struct ggml_tensor** output               = NULL,
-                 struct ggml_context* output_ctx           = NULL) {
-        return flux.compute(n_threads, x, timesteps, context, y, guidance, output, output_ctx);
+                 struct ggml_context* output_ctx           = NULL,
+                 std::vector<int> skip_layers              = std::vector<int>()) {
+        return flux.compute(n_threads, x, timesteps, context, y, guidance, output, output_ctx, skip_layers);
     }
 };
 
-#endif
+#endif

docs/photo_maker.md

@@ -29,4 +29,26 @@ Example:
 
 ```bash
 bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --stacked-id-embd-dir ../models/photomaker-v1.safetensors --input-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --style-ratio 10 --vae-on-cpu -o output.png
-```
+```
+
+## PhotoMaker Version 2
+
+[PhotoMaker Version 2 (PMV2)](https://github.com/TencentARC/PhotoMaker/blob/main/README_pmv2.md) has some key improvements. Unfortunately it has a very heavy dependency which makes running it a bit involved in ```SD.cpp```. 
+
+Running PMV2 is now a two-step process:
+
+- Run a python script ```face_detect.py``` to obtain **id_embeds** for the given input images
+```
+python face_detect.py input_image_dir
+```
+An ```id_embeds.safetensors``` file will be generated in ```input_images_dir```
+
+**Note: this step is only needed to run once; the same ```id_embeds``` can be reused**
+
+- Run the same command as in version 1 but replacing ```photomaker-v1.safetensors``` with ```photomaker-v2.safetensors```.
+
+  You can download ```photomaker-v2.safetensors``` from [here](https://huggingface.co/bssrdf/PhotoMakerV2)
+
+- All the command line parameters from Version 1 remain the same for Version 2
+
+

examples/cli/main.cpp

@@ -116,9 +116,15 @@ struct SDParams {
     bool normalize_input          = false;
     bool clip_on_cpu              = false;
     bool vae_on_cpu               = false;
+    bool diffusion_flash_attn     = false;
     bool canny_preprocess         = false;
     bool color                    = false;
     int upscale_repeats           = 1;
+
+    std::vector<int> skip_layers = {7, 8, 9};
+    float slg_scale              = 0.;
+    float skip_layer_start       = 0.01;
+    float skip_layer_end         = 0.2;
 };
 
 void print_params(SDParams params) {
@@ -146,11 +152,13 @@ void print_params(SDParams params) {
     printf("    clip on cpu:       %s\n", params.clip_on_cpu ? "true" : "false");
     printf("    controlnet cpu:    %s\n", params.control_net_cpu ? "true" : "false");
     printf("    vae decoder on cpu:%s\n", params.vae_on_cpu ? "true" : "false");
+    printf("    diffusion flash attention:%s\n", params.diffusion_flash_attn ? "true" : "false");
     printf("    strength(control): %.2f\n", params.control_strength);
     printf("    prompt:            %s\n", params.prompt.c_str());
     printf("    negative_prompt:   %s\n", params.negative_prompt.c_str());
     printf("    min_cfg:           %.2f\n", params.min_cfg);
     printf("    cfg_scale:         %.2f\n", params.cfg_scale);
+    printf("    slg_scale:         %.2f\n", params.slg_scale);
     printf("    guidance:          %.2f\n", params.guidance);
     printf("    clip_skip:         %d\n", params.clip_skip);
     printf("    width:             %d\n", params.width);
@@ -177,7 +185,7 @@ void print_usage(int argc, const char* argv[]) {
     printf("  -m, --model [MODEL]                path to full model\n");
     printf("  --diffusion-model                  path to the standalone diffusion model\n");
     printf("  --clip_l                           path to the clip-l text encoder\n");
-    printf("  --clip_g                           path to the clip-l text encoder\n");
+    printf("  --clip_g                           path to the clip-g text encoder\n");
     printf("  --t5xxl                            path to the the t5xxl text encoder\n");
     printf("  --vae [VAE]                        path to vae\n");
     printf("  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)\n");
@@ -197,6 +205,12 @@ void print_usage(int argc, const char* argv[]) {
     printf("  -p, --prompt [PROMPT]              the prompt to render\n");
     printf("  -n, --negative-prompt PROMPT       the negative prompt (default: \"\")\n");
     printf("  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)\n");
+    printf("  --slg-scale SCALE                  skip layer guidance (SLG) scale, only for DiT models: (default: 0)\n");
+    printf("                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium\n");
+    printf("  --skip_layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])\n");
+    printf("  --skip_layer_start START           SLG enabling point: (default: 0.01)\n");
+    printf("  --skip_layer_end END               SLG disabling point: (default: 0.2)\n");
+    printf("                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])\n");
     printf("  --strength STRENGTH                strength for noising/unnoising (default: 0.75)\n");
     printf("  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20%%)\n");
     printf("  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)\n");
@@ -215,6 +229,9 @@ void print_usage(int argc, const char* argv[]) {
     printf("  --vae-tiling                       process vae in tiles to reduce memory usage\n");
     printf("  --vae-on-cpu                       keep vae in cpu (for low vram)\n");
     printf("  --clip-on-cpu                      keep clip in cpu (for low vram)\n");
+    printf("  --diffusion-fa                     use flash attention in the diffusion model (for low vram)\n");
+    printf("                                     Might lower quality, since it implies converting k and v to f16.\n");
+    printf("                                     This might crash if it is not supported by the backend.\n");
     printf("  --control-net-cpu                  keep controlnet in cpu (for low vram)\n");
     printf("  --canny                            apply canny preprocessor (edge detection)\n");
     printf("  --color                            Colors the logging tags according to level\n");
@@ -465,6 +482,8 @@ void parse_args(int argc, const char** argv, SDParams& params) {
             params.clip_on_cpu = true;  // will slow down get_learned_condiotion but necessary for low MEM GPUs
         } else if (arg == "--vae-on-cpu") {
             params.vae_on_cpu = true;  // will slow down latent decoding but necessary for low MEM GPUs
+        } else if (arg == "--diffusion-fa") {
+            params.diffusion_flash_attn = true;  // can reduce MEM significantly
         } else if (arg == "--canny") {
             params.canny_preprocess = true;
         } else if (arg == "-b" || arg == "--batch-count") {
@@ -534,6 +553,61 @@ void parse_args(int argc, const char** argv, SDParams& params) {
             params.verbose = true;
         } else if (arg == "--color") {
             params.color = true;
+        } else if (arg == "--slg-scale") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.slg_scale = std::stof(argv[i]);
+        } else if (arg == "--skip-layers") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            if (argv[i][0] != '[') {
+                invalid_arg = true;
+                break;
+            }
+            std::string layers_str = argv[i];
+            while (layers_str.back() != ']') {
+                if (++i >= argc) {
+                    invalid_arg = true;
+                    break;
+                }
+                layers_str += " " + std::string(argv[i]);
+            }
+            layers_str = layers_str.substr(1, layers_str.size() - 2);
+
+            std::regex regex("[, ]+");
+            std::sregex_token_iterator iter(layers_str.begin(), layers_str.end(), regex, -1);
+            std::sregex_token_iterator end;
+            std::vector<std::string> tokens(iter, end);
+            std::vector<int> layers;
+            for (const auto& token : tokens) {
+                try {
+                    layers.push_back(std::stoi(token));
+                } catch (const std::invalid_argument& e) {
+                    invalid_arg = true;
+                    break;
+                }
+            }
+            params.skip_layers = layers;
+
+            if (invalid_arg) {
+                break;
+            }
+        } else if (arg == "--skip-layer-start") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.skip_layer_start = std::stof(argv[i]);
+        } else if (arg == "--skip-layer-end") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.skip_layer_end = std::stof(argv[i]);
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             print_usage(argc, argv);
@@ -624,6 +698,16 @@ std::string get_image_params(SDParams params, int64_t seed) {
     }
     parameter_string += "Steps: " + std::to_string(params.sample_steps) + ", ";
     parameter_string += "CFG scale: " + std::to_string(params.cfg_scale) + ", ";
+    if (params.slg_scale != 0 && params.skip_layers.size() != 0) {
+        parameter_string += "SLG scale: " + std::to_string(params.cfg_scale) + ", ";
+        parameter_string += "Skip layers: [";
+        for (const auto& layer : params.skip_layers) {
+            parameter_string += std::to_string(layer) + ", ";
+        }
+        parameter_string += "], ";
+        parameter_string += "Skip layer start: " + std::to_string(params.skip_layer_start) + ", ";
+        parameter_string += "Skip layer end: " + std::to_string(params.skip_layer_end) + ", ";
+    }
     parameter_string += "Guidance: " + std::to_string(params.guidance) + ", ";
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(params.width) + "x" + std::to_string(params.height) + ", ";
@@ -791,7 +875,8 @@ int main(int argc, const char* argv[]) {
                                   params.schedule,
                                   params.clip_on_cpu,
                                   params.control_net_cpu,
-                                  params.vae_on_cpu);
+                                  params.vae_on_cpu,
+                                  params.diffusion_flash_attn);
 
     if (sd_ctx == NULL) {
         printf("new_sd_ctx_t failed\n");
@@ -840,7 +925,11 @@ int main(int argc, const char* argv[]) {
                           params.control_strength,
                           params.style_ratio,
                           params.normalize_input,
-                          params.input_id_images_path.c_str());
+                          params.input_id_images_path.c_str(),
+                          params.skip_layers,
+                          params.slg_scale,
+                          params.skip_layer_start,
+                          params.skip_layer_end);
     } else {
         sd_image_t input_image = {(uint32_t)params.width,
                                   (uint32_t)params.height,

face_detect.py

@@ -0,0 +1,88 @@
+import os
+import sys
+
+import numpy as np
+import torch
+from diffusers.utils import load_image
+# pip install insightface==0.7.3
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+from safetensors.torch import save_file
+
+### 
+# https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/165#issue-2055829543
+###
+class FaceAnalysis2(FaceAnalysis):
+    # NOTE: allows setting det_size for each detection call.
+    # the model allows it but the wrapping code from insightface
+    # doesn't show it, and people end up loading duplicate models
+    # for different sizes where there is absolutely no need to
+    def get(self, img, max_num=0, det_size=(640, 640)):
+        if det_size is not None:
+            self.det_model.input_size = det_size
+
+        return super().get(img, max_num)
+
+def analyze_faces(face_analysis: FaceAnalysis, img_data: np.ndarray, det_size=(640, 640)):
+    # NOTE: try detect faces, if no faces detected, lower det_size until it does
+    detection_sizes = [None] + [(size, size) for size in range(640, 256, -64)] + [(256, 256)]
+
+    for size in detection_sizes:
+        faces = face_analysis.get(img_data, det_size=size)
+        if len(faces) > 0:
+            return faces
+
+    return []
+
+if __name__ == "__main__":
+    #face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+    face_detector = FaceAnalysis2(providers=['CPUExecutionProvider'], allowed_modules=['detection', 'recognition'])
+    face_detector.prepare(ctx_id=0, det_size=(640, 640))
+    #input_folder_name = './scarletthead_woman'
+    input_folder_name = sys.argv[1]
+    image_basename_list = os.listdir(input_folder_name)
+    image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
+
+    input_id_images = []
+    for image_path in image_path_list:
+        input_id_images.append(load_image(image_path))
+    
+    id_embed_list = []
+    
+    for img in input_id_images:
+        img = np.array(img)
+        img = img[:, :, ::-1]
+        faces = analyze_faces(face_detector, img)
+        if len(faces) > 0:
+            id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+    
+    if len(id_embed_list) == 0:
+        raise ValueError(f"No face detected in input image pool")
+    
+    id_embeds = torch.stack(id_embed_list)    
+    
+    # for r in id_embeds:
+    #     print(r)
+    # #torch.save(id_embeds, input_folder_name+'/id_embeds.pt');
+    # weights = dict()
+    # weights["id_embeds"] = id_embeds
+    # save_file(weights, input_folder_name+'/id_embeds.safetensors')
+
+    binary_data = id_embeds.numpy().tobytes()
+    two = 4
+    zero = 0
+    one = 1
+    tensor_name = "id_embeds"
+# Write binary data to a file
+    with open(input_folder_name+'/id_embeds.bin', "wb") as f:
+        f.write(two.to_bytes(4, byteorder='little'))
+        f.write((len(tensor_name)).to_bytes(4, byteorder='little'))
+        f.write(zero.to_bytes(4, byteorder='little'))
+        f.write((id_embeds.shape[1]).to_bytes(4, byteorder='little'))
+        f.write((id_embeds.shape[0]).to_bytes(4, byteorder='little'))
+        f.write(one.to_bytes(4, byteorder='little'))
+        f.write(one.to_bytes(4, byteorder='little'))
+        f.write(tensor_name.encode('ascii'))
+        f.write(binary_data)
+
+    

flux.hpp

@@ -115,25 +115,28 @@ namespace Flux {
                                                     struct ggml_tensor* q,
                                                     struct ggml_tensor* k,
                                                     struct ggml_tensor* v,
-                                                    struct ggml_tensor* pe) {
+                                                    struct ggml_tensor* pe,
+                                                    bool flash_attn) {
         // q,k,v: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2]
         // return: [N, L, n_head*d_head]
         q = apply_rope(ctx, q, pe);  // [N*n_head, L, d_head]
         k = apply_rope(ctx, k, pe);  // [N*n_head, L, d_head]
 
-        auto x = ggml_nn_attention_ext(ctx, q, k, v, v->ne[1], NULL, false, true);  // [N, L, n_head*d_head]
+        auto x = ggml_nn_attention_ext(ctx, q, k, v, v->ne[1], NULL, false, true, flash_attn);  // [N, L, n_head*d_head]
         return x;
     }
 
     struct SelfAttention : public GGMLBlock {
     public:
         int64_t num_heads;
+        bool flash_attn;
 
     public:
         SelfAttention(int64_t dim,
                       int64_t num_heads = 8,
-                      bool qkv_bias     = false)
+                      bool qkv_bias     = false,
+                      bool flash_attn   = false)
             : num_heads(num_heads) {
             int64_t head_dim = dim / num_heads;
             blocks["qkv"]    = std::shared_ptr<GGMLBlock>(new Linear(dim, dim * 3, qkv_bias));
@@ -167,9 +170,9 @@ namespace Flux {
             // x: [N, n_token, dim]
             // pe: [n_token, d_head/2, 2, 2]
             // return [N, n_token, dim]
-            auto qkv = pre_attention(ctx, x);                       // q,k,v: [N, n_token, n_head, d_head]
-            x        = attention(ctx, qkv[0], qkv[1], qkv[2], pe);  // [N, n_token, dim]
-            x        = post_attention(ctx, x);                      // [N, n_token, dim]
+            auto qkv = pre_attention(ctx, x);                                   // q,k,v: [N, n_token, n_head, d_head]
+            x        = attention(ctx, qkv[0], qkv[1], qkv[2], pe, flash_attn);  // [N, n_token, dim]
+            x        = post_attention(ctx, x);                                  // [N, n_token, dim]
             return x;
         }
     };
@@ -237,15 +240,19 @@ namespace Flux {
     }
 
     struct DoubleStreamBlock : public GGMLBlock {
+        bool flash_attn;
+
     public:
         DoubleStreamBlock(int64_t hidden_size,
                           int64_t num_heads,
                           float mlp_ratio,
-                          bool qkv_bias = false) {
+                          bool qkv_bias   = false,
+                          bool flash_attn = false)
+            : flash_attn(flash_attn) {
             int64_t mlp_hidden_dim = hidden_size * mlp_ratio;
             blocks["img_mod"]      = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
             blocks["img_norm1"]    = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
-            blocks["img_attn"]     = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias));
+            blocks["img_attn"]     = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias, flash_attn));
 
             blocks["img_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
             blocks["img_mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, mlp_hidden_dim));
@@ -254,7 +261,7 @@ namespace Flux {
 
             blocks["txt_mod"]   = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
             blocks["txt_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
-            blocks["txt_attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias));
+            blocks["txt_attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias, flash_attn));
 
             blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
             blocks["txt_mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, mlp_hidden_dim));
@@ -316,7 +323,7 @@ namespace Flux {
             auto k = ggml_concat(ctx, txt_k, img_k, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
             auto v = ggml_concat(ctx, txt_v, img_v, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
 
-            auto attn         = attention(ctx, q, k, v, pe);                          // [N, n_txt_token + n_img_token, n_head*d_head]
+            auto attn         = attention(ctx, q, k, v, pe, flash_attn);              // [N, n_txt_token + n_img_token, n_head*d_head]
             attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));  // [n_txt_token + n_img_token, N, hidden_size]
             auto txt_attn_out = ggml_view_3d(ctx,
                                              attn,
@@ -364,13 +371,15 @@ namespace Flux {
         int64_t num_heads;
         int64_t hidden_size;
         int64_t mlp_hidden_dim;
+        bool flash_attn;
 
     public:
         SingleStreamBlock(int64_t hidden_size,
                           int64_t num_heads,
                           float mlp_ratio = 4.0f,
-                          float qk_scale  = 0.f)
-            : hidden_size(hidden_size), num_heads(num_heads) {
+                          float qk_scale  = 0.f,
+                          bool flash_attn = false)
+            : hidden_size(hidden_size), num_heads(num_heads), flash_attn(flash_attn) {
             int64_t head_dim = hidden_size / num_heads;
             float scale      = qk_scale;
             if (scale <= 0.f) {
@@ -433,7 +442,7 @@ namespace Flux {
             auto v           = ggml_reshape_4d(ctx, qkv_vec[2], head_dim, num_heads, qkv_vec[2]->ne[1], qkv_vec[2]->ne[2]);  // [N, n_token, n_head, d_head]
             q                = norm->query_norm(ctx, q);
             k                = norm->key_norm(ctx, k);
-            auto attn        = attention(ctx, q, k, v, pe);  // [N, n_token, hidden_size]
+            auto attn        = attention(ctx, q, k, v, pe, flash_attn);  // [N, n_token, hidden_size]
 
             auto attn_mlp = ggml_concat(ctx, attn, ggml_gelu_inplace(ctx, mlp), 0);  // [N, n_token, hidden_size + mlp_hidden_dim]
             auto output   = linear2->forward(ctx, attn_mlp);                         // [N, n_token, hidden_size]
@@ -492,6 +501,7 @@ namespace Flux {
         int theta                   = 10000;
         bool qkv_bias               = true;
         bool guidance_embed         = true;
+        bool flash_attn             = true;
     };
 
     struct Flux : public GGMLBlock {
@@ -646,13 +656,16 @@ namespace Flux {
                 blocks["double_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new DoubleStreamBlock(params.hidden_size,
                                                                                                                 params.num_heads,
                                                                                                                 params.mlp_ratio,
-                                                                                                                params.qkv_bias));
+                                                                                                                params.qkv_bias,
+                                                                                                                params.flash_attn));
             }
 
             for (int i = 0; i < params.depth_single_blocks; i++) {
                 blocks["single_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new SingleStreamBlock(params.hidden_size,
                                                                                                                 params.num_heads,
-                                                                                                                params.mlp_ratio));
+                                                                                                                params.mlp_ratio,
+                                                                                                                0.f,
+                                                                                                                params.flash_attn));
             }
 
             blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new LastLayer(params.hidden_size, 1, out_channels));
@@ -711,7 +724,8 @@ namespace Flux {
                                          struct ggml_tensor* timesteps,
                                          struct ggml_tensor* y,
                                          struct ggml_tensor* guidance,
-                                         struct ggml_tensor* pe) {
+                                         struct ggml_tensor* pe,
+                                         std::vector<int> skip_layers = std::vector<int>()) {
             auto img_in      = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
             auto time_in     = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]);
             auto vector_in   = std::dynamic_pointer_cast<MLPEmbedder>(blocks["vector_in"]);
@@ -733,6 +747,10 @@ namespace Flux {
             txt = txt_in->forward(ctx, txt);
 
             for (int i = 0; i < params.depth; i++) {
+                if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
+                    continue;
+                }
+
                 auto block = std::dynamic_pointer_cast<DoubleStreamBlock>(blocks["double_blocks." + std::to_string(i)]);
 
                 auto img_txt = block->forward(ctx, img, txt, vec, pe);
@@ -742,6 +760,9 @@ namespace Flux {
 
             auto txt_img = ggml_concat(ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
             for (int i = 0; i < params.depth_single_blocks; i++) {
+                if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i + params.depth) != skip_layers.end()) {
+                    continue;
+                }
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, vec, pe);
@@ -769,7 +790,8 @@ namespace Flux {
                                     struct ggml_tensor* context,
                                     struct ggml_tensor* y,
                                     struct ggml_tensor* guidance,
-                                    struct ggml_tensor* pe) {
+                                    struct ggml_tensor* pe,
+                                    std::vector<int> skip_layers = std::vector<int>()) {
             // Forward pass of DiT.
             // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
             // timestep: (N,) tensor of diffusion timesteps
@@ -791,7 +813,7 @@ namespace Flux {
             // img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
             auto img = patchify(ctx, x, patch_size);  // [N, h*w, C * patch_size * patch_size]
 
-            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe);  // [N, h*w, C * patch_size * patch_size]
+            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, skip_layers);  // [N, h*w, C * patch_size * patch_size]
 
             // rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)
             out = unpatchify(ctx, out, (H + pad_h) / patch_size, (W + pad_w) / patch_size, patch_size);  // [N, C, H + pad_h, W + pad_w]
@@ -808,11 +830,16 @@ namespace Flux {
 
         FluxRunner(ggml_backend_t backend,
                    ggml_type wtype,
-                   SDVersion version = VERSION_FLUX_DEV)
+                   SDVersion version = VERSION_FLUX_DEV,
+                   bool flash_attn   = false)
             : GGMLRunner(backend, wtype) {
+            flux_params.flash_attn = flash_attn;
             if (version == VERSION_FLUX_SCHNELL) {
                 flux_params.guidance_embed = false;
             }
+            if (version == VERSION_FLUX_LITE) {
+                flux_params.depth = 8;
+            }
             flux = Flux(flux_params);
             flux.init(params_ctx, wtype);
         }
@@ -829,7 +856,8 @@ namespace Flux {
                                         struct ggml_tensor* timesteps,
                                         struct ggml_tensor* context,
                                         struct ggml_tensor* y,
-                                        struct ggml_tensor* guidance) {
+                                        struct ggml_tensor* guidance,
+                                        std::vector<int> skip_layers = std::vector<int>()) {
             GGML_ASSERT(x->ne[3] == 1);
             struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, FLUX_GRAPH_SIZE, false);
 
@@ -856,7 +884,8 @@ namespace Flux {
                                                    context,
                                                    y,
                                                    guidance,
-                                                   pe);
+                                                   pe,
+                                                   skip_layers);
 
             ggml_build_forward_expand(gf, out);
 
@@ -870,14 +899,15 @@ namespace Flux {
                      struct ggml_tensor* y,
                      struct ggml_tensor* guidance,
                      struct ggml_tensor** output     = NULL,
-                     struct ggml_context* output_ctx = NULL) {
+                     struct ggml_context* output_ctx = NULL,
+                     std::vector<int> skip_layers    = std::vector<int>()) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             // y: [N, adm_in_channels] or [1, adm_in_channels]
             // guidance: [N, ]
             auto get_graph = [&]() -> struct ggml_cgraph* {
-                return build_graph(x, timesteps, context, y, guidance);
+                return build_graph(x, timesteps, context, y, guidance, skip_layers);
             };
 
             GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
@@ -958,4 +988,4 @@ namespace Flux {
 
 }  // namespace Flux
 
-#endif  // __FLUX_HPP__
+#endif  // __FLUX_HPP__

ggml_extend.hpp

@@ -666,32 +666,6 @@ __STATIC_INLINE__ std::vector<struct ggml_tensor*> split_qkv(struct ggml_context
     return {q, k, v};
 }
 
-// q: [N * n_head, n_token, d_head]
-// k: [N * n_head, n_k, d_head]
-// v: [N * n_head, d_head, n_k]
-// return: [N * n_head, n_token, d_head]
-__STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention(struct ggml_context* ctx,
-                                                        struct ggml_tensor* q,
-                                                        struct ggml_tensor* k,
-                                                        struct ggml_tensor* v,
-                                                        bool mask = false) {
-#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL) && !defined(SD_USE_VULKAN) && !defined(SD_USE_SYCL)
-    struct ggml_tensor* kqv = ggml_flash_attn(ctx, q, k, v, false);  // [N * n_head, n_token, d_head]
-#else
-    float d_head = (float)q->ne[0];
-
-    struct ggml_tensor* kq = ggml_mul_mat(ctx, k, q);  // [N * n_head, n_token, n_k]
-    kq                     = ggml_scale_inplace(ctx, kq, 1.0f / sqrt(d_head));
-    if (mask) {
-        kq = ggml_diag_mask_inf_inplace(ctx, kq, 0);
-    }
-    kq = ggml_soft_max_inplace(ctx, kq);
-
-    struct ggml_tensor* kqv = ggml_mul_mat(ctx, v, kq);  // [N * n_head, n_token, d_head]
-#endif
-    return kqv;
-}
-
 // q: [N, L_q, C] or [N*n_head, L_q, d_head]
 // k: [N, L_k, C] or [N*n_head, L_k, d_head]
 // v: [N, L_k, C] or [N, L_k, n_head, d_head]
@@ -703,7 +677,8 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
                                                             int64_t n_head,
                                                             struct ggml_tensor* mask = NULL,
                                                             bool diag_mask_inf       = false,
-                                                            bool skip_reshape        = false) {
+                                                            bool skip_reshape        = false,
+                                                            bool flash_attn          = false) {
     int64_t L_q;
     int64_t L_k;
     int64_t C;
@@ -734,13 +709,42 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
 
     float scale = (1.0f / sqrt((float)d_head));
 
-    bool use_flash_attn = false;
-    ggml_tensor* kqv    = NULL;
-    if (use_flash_attn) {
+    // if (flash_attn) {
+    //     LOG_DEBUG("attention_ext L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N);
+    // }
+    //  is there anything oddly shaped?? ping Green-Sky if you can trip this assert
+    GGML_ASSERT(((L_k % 256 == 0) && L_q == L_k) || !(L_k % 256 == 0));
+
+    bool can_use_flash_attn = true;
+    can_use_flash_attn      = can_use_flash_attn && L_k % 256 == 0;
+    can_use_flash_attn      = can_use_flash_attn && d_head % 64 == 0;  // double check
+
+    // cuda max d_head seems to be 256, cpu does seem to work with 512
+    can_use_flash_attn = can_use_flash_attn && d_head <= 256;  // double check
+
+    if (mask != nullptr) {
+        // TODO(Green-Sky): figure out if we can bend t5 to work too
+        can_use_flash_attn = can_use_flash_attn && mask->ne[2] == 1;
+        can_use_flash_attn = can_use_flash_attn && mask->ne[3] == 1;
+    }
+
+    // TODO(Green-Sky): more pad or disable for funny tensor shapes
+
+    ggml_tensor* kqv = nullptr;
+    // GGML_ASSERT((flash_attn && can_use_flash_attn) || !flash_attn);
+    if (can_use_flash_attn && flash_attn) {
+        // LOG_DEBUG("using flash attention");
+        k = ggml_cast(ctx, k, GGML_TYPE_F16);
+
         v = ggml_cont(ctx, ggml_permute(ctx, v, 0, 2, 1, 3));  // [N, n_head, L_k, d_head]
         v = ggml_reshape_3d(ctx, v, d_head, L_k, n_head * N);  // [N * n_head, L_k, d_head]
-        LOG_DEBUG("k->ne[1] == %d", k->ne[1]);
+        v = ggml_cast(ctx, v, GGML_TYPE_F16);
+
         kqv = ggml_flash_attn_ext(ctx, q, k, v, mask, scale, 0, 0);
+        ggml_flash_attn_ext_set_prec(kqv, GGML_PREC_F32);
+
+        // kqv = ggml_view_3d(ctx, kqv, d_head, n_head, L_k, kqv->nb[1], kqv->nb[2], 0);
+        kqv = ggml_view_3d(ctx, kqv, d_head, n_head, L_q, kqv->nb[1], kqv->nb[2], 0);
     } else {
         v = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));  // [N, n_head, d_head, L_k]
         v = ggml_reshape_3d(ctx, v, L_k, d_head, n_head * N);  // [N * n_head, d_head, L_k]
@@ -756,10 +760,12 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
         kq = ggml_soft_max_inplace(ctx, kq);
 
         kqv = ggml_mul_mat(ctx, v, kq);  // [N * n_head, L_q, d_head]
+
+        kqv = ggml_reshape_4d(ctx, kqv, d_head, L_q, n_head, N);  // [N, n_head, L_q, d_head]
+        kqv = ggml_permute(ctx, kqv, 0, 2, 1, 3);                 // [N, L_q, n_head, d_head]
     }
 
-    kqv = ggml_reshape_4d(ctx, kqv, d_head, L_q, n_head, N);   // [N, n_head, L_q, d_head]
-    kqv = ggml_cont(ctx, ggml_permute(ctx, kqv, 0, 2, 1, 3));  // [N, L_q, n_head, d_head]
+    kqv = ggml_cont(ctx, kqv);
     kqv = ggml_reshape_3d(ctx, kqv, d_head * n_head, L_q, N);  // [N, L_q, C]
 
     return kqv;
@@ -1047,6 +1053,11 @@ public:
                   params_buffer_size / (1024.0 * 1024.0),
                   ggml_backend_is_cpu(backend) ? "RAM" : "VRAM",
                   num_tensors);
+        // printf("%s params backend buffer size = % 6.2f MB(%s) (%i tensors)\n",
+        //           get_desc().c_str(),
+        //           params_buffer_size / (1024.0 * 1024.0),
+        //           ggml_backend_is_cpu(backend) ? "RAM" : "VRAM",
+        //           num_tensors);
         return true;
     }
 

mmdit.hpp

@@ -252,6 +252,7 @@ struct DismantledBlock : public GGMLBlock {
 public:
     int64_t num_heads;
     bool pre_only;
+    bool self_attn;
 
 public:
     DismantledBlock(int64_t hidden_size,
@@ -259,14 +260,19 @@ public:
                     float mlp_ratio     = 4.0,
                     std::string qk_norm = "",
                     bool qkv_bias       = false,
-                    bool pre_only       = false)
-        : num_heads(num_heads), pre_only(pre_only) {
+                    bool pre_only       = false,
+                    bool self_attn      = false)
+        : num_heads(num_heads), pre_only(pre_only), self_attn(self_attn) {
         // rmsnorm is always Flase
         // scale_mod_only is always Flase
         // swiglu is always Flase
         blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
         blocks["attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, pre_only));
 
+        if (self_attn) {
+            blocks["attn2"] = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, false));
+        }
+
         if (!pre_only) {
             blocks["norm2"]        = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
             int64_t mlp_hidden_dim = (int64_t)(hidden_size * mlp_ratio);
@@ -277,9 +283,52 @@ public:
         if (pre_only) {
             n_mods = 2;
         }
+        if (self_attn) {
+            n_mods = 9;
+        }
         blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, n_mods * hidden_size));
     }
 
+    std::tuple<std::vector<struct ggml_tensor*>, std::vector<struct ggml_tensor*>, std::vector<struct ggml_tensor*>> pre_attention_x(struct ggml_context* ctx,
+                                                                                                                                     struct ggml_tensor* x,
+                                                                                                                                     struct ggml_tensor* c) {
+        GGML_ASSERT(self_attn);
+        // x: [N, n_token, hidden_size]
+        // c: [N, hidden_size]
+        auto norm1              = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
+        auto attn               = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"]);
+        auto attn2              = std::dynamic_pointer_cast<SelfAttention>(blocks["attn2"]);
+        auto adaLN_modulation_1 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);
+
+        int64_t n_mods = 9;
+        auto m         = adaLN_modulation_1->forward(ctx, ggml_silu(ctx, c));  // [N, n_mods * hidden_size]
+        m              = ggml_reshape_3d(ctx, m, c->ne[0], n_mods, c->ne[1]);  // [N, n_mods, hidden_size]
+        m              = ggml_cont(ctx, ggml_permute(ctx, m, 0, 2, 1, 3));     // [n_mods, N, hidden_size]
+
+        int64_t offset = m->nb[1] * m->ne[1];
+        auto shift_msa = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 0);  // [N, hidden_size]
+        auto scale_msa = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 1);  // [N, hidden_size]
+        auto gate_msa  = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 2);  // [N, hidden_size]
+
+        auto shift_mlp = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 3);  // [N, hidden_size]
+        auto scale_mlp = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 4);  // [N, hidden_size]
+        auto gate_mlp  = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 5);  // [N, hidden_size]
+
+        auto shift_msa2 = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 6);  // [N, hidden_size]
+        auto scale_msa2 = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 7);  // [N, hidden_size]
+        auto gate_msa2  = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 8);  // [N, hidden_size]
+
+        auto x_norm = norm1->forward(ctx, x);
+
+        auto attn_in = modulate(ctx, x_norm, shift_msa, scale_msa);
+        auto qkv     = attn->pre_attention(ctx, attn_in);
+
+        auto attn2_in = modulate(ctx, x_norm, shift_msa2, scale_msa2);
+        auto qkv2     = attn2->pre_attention(ctx, attn2_in);
+
+        return {qkv, qkv2, {x, gate_msa, shift_mlp, scale_mlp, gate_mlp, gate_msa2}};
+    }
+
     std::pair<std::vector<struct ggml_tensor*>, std::vector<struct ggml_tensor*>> pre_attention(struct ggml_context* ctx,
                                                                                                 struct ggml_tensor* x,
                                                                                                 struct ggml_tensor* c) {
@@ -319,6 +368,44 @@ public:
         }
     }
 
+    struct ggml_tensor* post_attention_x(struct ggml_context* ctx,
+                                         struct ggml_tensor* attn_out,
+                                         struct ggml_tensor* attn2_out,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* gate_msa,
+                                         struct ggml_tensor* shift_mlp,
+                                         struct ggml_tensor* scale_mlp,
+                                         struct ggml_tensor* gate_mlp,
+                                         struct ggml_tensor* gate_msa2) {
+        // attn_out: [N, n_token, hidden_size]
+        // x: [N, n_token, hidden_size]
+        // gate_msa: [N, hidden_size]
+        // shift_mlp: [N, hidden_size]
+        // scale_mlp: [N, hidden_size]
+        // gate_mlp: [N, hidden_size]
+        // return: [N, n_token, hidden_size]
+        GGML_ASSERT(!pre_only);
+
+        auto attn  = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"]);
+        auto attn2 = std::dynamic_pointer_cast<SelfAttention>(blocks["attn2"]);
+        auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
+        auto mlp   = std::dynamic_pointer_cast<Mlp>(blocks["mlp"]);
+
+        gate_msa  = ggml_reshape_3d(ctx, gate_msa, gate_msa->ne[0], 1, gate_msa->ne[1]);     // [N, 1, hidden_size]
+        gate_mlp  = ggml_reshape_3d(ctx, gate_mlp, gate_mlp->ne[0], 1, gate_mlp->ne[1]);     // [N, 1, hidden_size]
+        gate_msa2 = ggml_reshape_3d(ctx, gate_msa2, gate_msa2->ne[0], 1, gate_msa2->ne[1]);  // [N, 1, hidden_size]
+
+        attn_out  = attn->post_attention(ctx, attn_out);
+        attn2_out = attn2->post_attention(ctx, attn2_out);
+
+        x            = ggml_add(ctx, x, ggml_mul(ctx, attn_out, gate_msa));
+        x            = ggml_add(ctx, x, ggml_mul(ctx, attn2_out, gate_msa2));
+        auto mlp_out = mlp->forward(ctx, modulate(ctx, norm2->forward(ctx, x), shift_mlp, scale_mlp));
+        x            = ggml_add(ctx, x, ggml_mul(ctx, mlp_out, gate_mlp));
+
+        return x;
+    }
+
     struct ggml_tensor* post_attention(struct ggml_context* ctx,
                                        struct ggml_tensor* attn_out,
                                        struct ggml_tensor* x,
@@ -357,29 +444,52 @@ public:
         // return: [N, n_token, hidden_size]
 
         auto attn = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"]);
+        if (self_attn) {
+            auto qkv_intermediates = pre_attention_x(ctx, x, c);
+            // auto qkv               = qkv_intermediates.first;
+            // auto intermediates     = qkv_intermediates.second;
+            // no longer a pair, but a tuple
+            auto qkv           = std::get<0>(qkv_intermediates);
+            auto qkv2          = std::get<1>(qkv_intermediates);
+            auto intermediates = std::get<2>(qkv_intermediates);
 
-        auto qkv_intermediates = pre_attention(ctx, x, c);
-        auto qkv               = qkv_intermediates.first;
-        auto intermediates     = qkv_intermediates.second;
+            auto attn_out  = ggml_nn_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads);     // [N, n_token, dim]
+            auto attn2_out = ggml_nn_attention_ext(ctx, qkv2[0], qkv2[1], qkv2[2], num_heads);  // [N, n_token, dim]
+            x              = post_attention_x(ctx,
+                                              attn_out,
+                                              attn2_out,
+                                              intermediates[0],
+                                              intermediates[1],
+                                              intermediates[2],
+                                              intermediates[3],
+                                              intermediates[4],
+                                              intermediates[5]);
+            return x;  // [N, n_token, dim]
+        } else {
+            auto qkv_intermediates = pre_attention(ctx, x, c);
+            auto qkv               = qkv_intermediates.first;
+            auto intermediates     = qkv_intermediates.second;
 
-        auto attn_out = ggml_nn_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
-        x             = post_attention(ctx,
-                                       attn_out,
-                                       intermediates[0],
-                                       intermediates[1],
-                                       intermediates[2],
-                                       intermediates[3],
-                                       intermediates[4]);
-        return x;  // [N, n_token, dim]
+            auto attn_out = ggml_nn_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
+            x             = post_attention(ctx,
+                                           attn_out,
+                                           intermediates[0],
+                                           intermediates[1],
+                                           intermediates[2],
+                                           intermediates[3],
+                                           intermediates[4]);
+            return x;  // [N, n_token, dim]
+        }
     }
 };
 
-__STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*> block_mixing(struct ggml_context* ctx,
-                                                                                   struct ggml_tensor* context,
-                                                                                   struct ggml_tensor* x,
-                                                                                   struct ggml_tensor* c,
-                                                                                   std::shared_ptr<DismantledBlock> context_block,
-                                                                                   std::shared_ptr<DismantledBlock> x_block) {
+__STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*>
+block_mixing(struct ggml_context* ctx,
+             struct ggml_tensor* context,
+             struct ggml_tensor* x,
+             struct ggml_tensor* c,
+             std::shared_ptr<DismantledBlock> context_block,
+             std::shared_ptr<DismantledBlock> x_block) {
     // context: [N, n_context, hidden_size]
     // x: [N, n_token, hidden_size]
     // c: [N, hidden_size]
@@ -387,10 +497,18 @@ __STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*> block_mixi
     auto context_qkv               = context_qkv_intermediates.first;
     auto context_intermediates     = context_qkv_intermediates.second;
 
-    auto x_qkv_intermediates = x_block->pre_attention(ctx, x, c);
-    auto x_qkv               = x_qkv_intermediates.first;
-    auto x_intermediates     = x_qkv_intermediates.second;
+    std::vector<ggml_tensor*> x_qkv, x_qkv2, x_intermediates;
 
+    if (x_block->self_attn) {
+        auto x_qkv_intermediates = x_block->pre_attention_x(ctx, x, c);
+        x_qkv                    = std::get<0>(x_qkv_intermediates);
+        x_qkv2                   = std::get<1>(x_qkv_intermediates);
+        x_intermediates          = std::get<2>(x_qkv_intermediates);
+    } else {
+        auto x_qkv_intermediates = x_block->pre_attention(ctx, x, c);
+        x_qkv                    = x_qkv_intermediates.first;
+        x_intermediates          = x_qkv_intermediates.second;
+    }
     std::vector<struct ggml_tensor*> qkv;
     for (int i = 0; i < 3; i++) {
         qkv.push_back(ggml_concat(ctx, context_qkv[i], x_qkv[i], 1));
@@ -429,13 +547,27 @@ __STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*> block_mixi
         context = NULL;
     }
 
-    x = x_block->post_attention(ctx,
-                                x_attn,
-                                x_intermediates[0],
-                                x_intermediates[1],
-                                x_intermediates[2],
-                                x_intermediates[3],
-                                x_intermediates[4]);
+    if (x_block->self_attn) {
+        auto attn2 = ggml_nn_attention_ext(ctx, x_qkv2[0], x_qkv2[1], x_qkv2[2], x_block->num_heads);  // [N, n_token, hidden_size]
+
+        x = x_block->post_attention_x(ctx,
+                                      x_attn,
+                                      attn2,
+                                      x_intermediates[0],
+                                      x_intermediates[1],
+                                      x_intermediates[2],
+                                      x_intermediates[3],
+                                      x_intermediates[4],
+                                      x_intermediates[5]);
+    } else {
+        x = x_block->post_attention(ctx,
+                                    x_attn,
+                                    x_intermediates[0],
+                                    x_intermediates[1],
+                                    x_intermediates[2],
+                                    x_intermediates[3],
+                                    x_intermediates[4]);
+    }
 
     return {context, x};
 }
@@ -447,9 +579,10 @@ public:
                float mlp_ratio     = 4.0,
                std::string qk_norm = "",
                bool qkv_bias       = false,
-               bool pre_only       = false) {
+               bool pre_only       = false,
+               bool self_attn_x    = false) {
         blocks["context_block"] = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, pre_only));
-        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false));
+        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x));
     }
 
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
@@ -507,6 +640,7 @@ protected:
     int64_t input_size               = -1;
     int64_t patch_size               = 2;
     int64_t in_channels              = 16;
+    int64_t d_self                   = -1;  // >=0 for MMdiT-X
     int64_t depth                    = 24;
     float mlp_ratio                  = 4.0f;
     int64_t adm_in_channels          = 2048;
@@ -561,6 +695,20 @@ public:
             context_size             = 4096;
             context_embedder_out_dim = 2432;
             qk_norm                  = "rms";
+        } else if (version == VERSION_SD3_5_2B) {
+            input_size               = -1;
+            patch_size               = 2;
+            in_channels              = 16;
+            depth                    = 24;
+            d_self                   = 12;
+            mlp_ratio                = 4.0f;
+            adm_in_channels          = 2048;
+            out_channels             = 16;
+            pos_embed_max_size       = 384;
+            num_patchs               = 147456;
+            context_size             = 4096;
+            context_embedder_out_dim = 1536;
+            qk_norm                  = "rms";
         }
         int64_t default_out_channels = in_channels;
         hidden_size                  = 64 * depth;
@@ -581,15 +729,17 @@ public:
                                                                                                     mlp_ratio,
                                                                                                     qk_norm,
                                                                                                     true,
-                                                                                                    i == depth - 1));
+                                                                                                    i == depth - 1,
+                                                                                                    i <= d_self));
         }
 
         blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(hidden_size, patch_size, out_channels));
     }
 
-    struct ggml_tensor* cropped_pos_embed(struct ggml_context* ctx,
-                                          int64_t h,
-                                          int64_t w) {
+    struct ggml_tensor*
+    cropped_pos_embed(struct ggml_context* ctx,
+                      int64_t h,
+                      int64_t w) {
         auto pos_embed = params["pos_embed"];
 
         h = (h + 1) / patch_size;
@@ -651,7 +801,8 @@ public:
     struct ggml_tensor* forward_core_with_concat(struct ggml_context* ctx,
                                                  struct ggml_tensor* x,
                                                  struct ggml_tensor* c_mod,
-                                                 struct ggml_tensor* context) {
+                                                 struct ggml_tensor* context,
+                                                 std::vector<int> skip_layers = std::vector<int>()) {
         // x: [N, H*W, hidden_size]
         // context: [N, n_context, d_context]
         // c: [N, hidden_size]
@@ -659,6 +810,11 @@ public:
         auto final_layer = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
 
         for (int i = 0; i < depth; i++) {
+            // skip iteration if i is in skip_layers
+            if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
+                continue;
+            }
+
             auto block = std::dynamic_pointer_cast<JointBlock>(blocks["joint_blocks." + std::to_string(i)]);
 
             auto context_x = block->forward(ctx, context, x, c_mod);
@@ -674,8 +830,9 @@ public:
     struct ggml_tensor* forward(struct ggml_context* ctx,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* t,
-                                struct ggml_tensor* y       = NULL,
-                                struct ggml_tensor* context = NULL) {
+                                struct ggml_tensor* y        = NULL,
+                                struct ggml_tensor* context  = NULL,
+                                std::vector<int> skip_layers = std::vector<int>()) {
         // Forward pass of DiT.
         // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
         // t: (N,) tensor of diffusion timesteps
@@ -706,7 +863,7 @@ public:
             context = context_embedder->forward(ctx, context);  // [N, L, D] aka [N, L, 1536]
         }
 
-        x = forward_core_with_concat(ctx, x, c, context);  // (N, H*W, patch_size ** 2 * out_channels)
+        x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 
         x = unpatchify(ctx, x, h, w);  // [N, C, H, W]
 
@@ -735,7 +892,8 @@ struct MMDiTRunner : public GGMLRunner {
     struct ggml_cgraph* build_graph(struct ggml_tensor* x,
                                     struct ggml_tensor* timesteps,
                                     struct ggml_tensor* context,
-                                    struct ggml_tensor* y) {
+                                    struct ggml_tensor* y,
+                                    std::vector<int> skip_layers = std::vector<int>()) {
         struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, MMDIT_GRAPH_SIZE, false);
 
         x         = to_backend(x);
@@ -747,7 +905,8 @@ struct MMDiTRunner : public GGMLRunner {
                                                 x,
                                                 timesteps,
                                                 y,
-                                                context);
+                                                context,
+                                                skip_layers);
 
         ggml_build_forward_expand(gf, out);
 
@@ -760,13 +919,14 @@ struct MMDiTRunner : public GGMLRunner {
                  struct ggml_tensor* context,
                  struct ggml_tensor* y,
                  struct ggml_tensor** output     = NULL,
-                 struct ggml_context* output_ctx = NULL) {
+                 struct ggml_context* output_ctx = NULL,
+                 std::vector<int> skip_layers    = std::vector<int>()) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 154, 4096]) or [1, max_position, hidden_size]
         // y: [N, adm_in_channels] or [1, adm_in_channels]
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(x, timesteps, context, y);
+            return build_graph(x, timesteps, context, y, skip_layers);
         };
 
         GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);

model.cpp

@@ -146,6 +146,33 @@ std::unordered_map<std::string, std::string> vae_decoder_name_map = {
     {"first_stage_model.decoder.mid.attn_1.to_v.weight", "first_stage_model.decoder.mid.attn_1.v.weight"},
 };
 
+std::unordered_map<std::string, std::string> pmid_v2_name_map = {
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.0.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.0.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.0.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.0.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.1.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.1.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.1.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.1.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.2.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.2.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.2.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.2.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.3.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.3.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.3.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.3.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.token_proj.0.bias",
+     "pmid.qformer_perceiver.token_proj.fc1.bias"},
+    {"pmid.qformer_perceiver.token_proj.2.bias",
+     "pmid.qformer_perceiver.token_proj.fc2.bias"},
+    {"pmid.qformer_perceiver.token_proj.0.weight",
+     "pmid.qformer_perceiver.token_proj.fc1.weight"},
+    {"pmid.qformer_perceiver.token_proj.2.weight",
+     "pmid.qformer_perceiver.token_proj.fc2.weight"},
+};
+
 std::string convert_open_clip_to_hf_clip(const std::string& name) {
     std::string new_name = name;
     std::string prefix;
@@ -212,6 +239,13 @@ std::string convert_vae_decoder_name(const std::string& name) {
     return name;
 }
 
+std::string convert_pmid_v2_name(const std::string& name) {
+    if (pmid_v2_name_map.find(name) != pmid_v2_name_map.end()) {
+        return pmid_v2_name_map[name];
+    }
+    return name;
+}
+
 /* If not a SDXL LoRA the unet" prefix will have already been replaced by this
  * point and "te2" and "te1" don't seem to appear in non-SDXL only "te_" */
 std::string convert_sdxl_lora_name(std::string tensor_name) {
@@ -443,6 +477,8 @@ std::string convert_tensor_name(std::string name) {
         new_name = convert_open_clip_to_hf_clip(name);
     } else if (starts_with(name, "first_stage_model.decoder")) {
         new_name = convert_vae_decoder_name(name);
+    } else if (starts_with(name, "pmid.qformer_perceiver")) {
+        new_name = convert_pmid_v2_name(name);
     } else if (starts_with(name, "control_model.")) {  // for controlnet pth models
         size_t pos = name.find('.');
         if (pos != std::string::npos) {
@@ -614,6 +650,47 @@ uint16_t f8_e4m3_to_f16(uint8_t f8) {
     return ggml_fp32_to_fp16(*reinterpret_cast<const float*>(&result));
 }
 
+uint16_t f8_e5m2_to_f16(uint8_t fp8) {
+    uint8_t sign     = (fp8 >> 7) & 0x1;
+    uint8_t exponent = (fp8 >> 2) & 0x1F;
+    uint8_t mantissa = fp8 & 0x3;
+
+    uint16_t fp16_sign = sign << 15;
+    uint16_t fp16_exponent;
+    uint16_t fp16_mantissa;
+
+    if (exponent == 0 && mantissa == 0) {  // zero
+        return fp16_sign;
+    }
+
+    if (exponent == 0x1F) {  // NAN and INF
+        fp16_exponent = 0x1F;
+        fp16_mantissa = mantissa ? (mantissa << 8) : 0;
+        return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
+    }
+
+    if (exponent == 0) {  // subnormal numbers
+        fp16_exponent = 0;
+        fp16_mantissa = (mantissa << 8);
+        return fp16_sign | fp16_mantissa;
+    }
+
+    // normal numbers
+    int16_t true_exponent = (int16_t)exponent - 15 + 15;
+    if (true_exponent <= 0) {
+        fp16_exponent = 0;
+        fp16_mantissa = (mantissa << 8);
+    } else if (true_exponent >= 0x1F) {
+        fp16_exponent = 0x1F;
+        fp16_mantissa = 0;
+    } else {
+        fp16_exponent = (uint16_t)true_exponent;
+        fp16_mantissa = mantissa << 8;
+    }
+
+    return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
+}
+
 void bf16_to_f32_vec(uint16_t* src, float* dst, int64_t n) {
     // support inplace op
     for (int64_t i = n - 1; i >= 0; i--) {
@@ -627,6 +704,12 @@ void f8_e4m3_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
         dst[i] = f8_e4m3_to_f16(src[i]);
     }
 }
+void f8_e5m2_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
+    // support inplace op
+    for (int64_t i = n - 1; i >= 0; i--) {
+        dst[i] = f8_e5m2_to_f16(src[i]);
+    }
+}
 
 void convert_tensor(void* src,
                     ggml_type src_type,
@@ -863,6 +946,8 @@ ggml_type str_to_ggml_type(const std::string& dtype) {
         ttype = GGML_TYPE_F32;
     } else if (dtype == "F8_E4M3") {
         ttype = GGML_TYPE_F16;
+    } else if (dtype == "F8_E5M2") {
+        ttype = GGML_TYPE_F16;
     }
     return ttype;
 }
@@ -976,6 +1061,10 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
             tensor_storage.is_f8_e4m3 = true;
             // f8 -> f16
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F8_E5M2") {
+            tensor_storage.is_f8_e5m2 = true;
+            // f8 -> f16
+            GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
         } else {
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size);
         }
@@ -1308,7 +1397,7 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
                         reader.tensor_storage.reverse_ne();
                         reader.tensor_storage.file_index = file_index;
                         // if(strcmp(prefix.c_str(), "scarlett") == 0)
-                        // printf(" got tensor %s \n ", reader.tensor_storage.name.c_str());
+                        // printf(" ZIP got tensor %s \n ", reader.tensor_storage.name.c_str());
                         reader.tensor_storage.name = prefix + reader.tensor_storage.name;
                         tensor_storages.push_back(reader.tensor_storage);
                         // LOG_DEBUG("%s", reader.tensor_storage.name.c_str());
@@ -1345,7 +1434,8 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
             size_t pos       = name.find("data.pkl");
             if (pos != std::string::npos) {
                 std::string dir = name.substr(0, pos);
-                void* pkl_data  = NULL;
+                printf("ZIP %d, name = %s, dir = %s \n", i, name.c_str(), dir.c_str());
+                void* pkl_data = NULL;
                 size_t pkl_size;
                 zip_entry_read(zip, &pkl_data, &pkl_size);
 
@@ -1364,15 +1454,23 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
 
 SDVersion ModelLoader::get_sd_version() {
     TensorStorage token_embedding_weight;
-    bool is_flux = false;
-    bool is_sd3  = false;
+    bool is_flux    = false;
+    bool is_schnell = true;
+    bool is_lite    = true;
+    bool is_sd3     = false;
     for (auto& tensor_storage : tensor_storages) {
         if (tensor_storage.name.find("model.diffusion_model.guidance_in.in_layer.weight") != std::string::npos) {
-            return VERSION_FLUX_DEV;
+            is_schnell = false;
         }
         if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
             is_flux = true;
         }
+        if (tensor_storage.name.find("model.diffusion_model.double_blocks.8") != std::string::npos) {
+            is_lite = false;
+        }
+        if (tensor_storage.name.find("joint_blocks.0.x_block.attn2.ln_q.weight") != std::string::npos) {
+            return VERSION_SD3_5_2B;
+        }
         if (tensor_storage.name.find("joint_blocks.37.x_block.attn.ln_q.weight") != std::string::npos) {
             return VERSION_SD3_5_8B;
         }
@@ -1400,7 +1498,14 @@ SDVersion ModelLoader::get_sd_version() {
         }
     }
     if (is_flux) {
-        return VERSION_FLUX_SCHNELL;
+        if (is_schnell) {
+            GGML_ASSERT(!is_lite);
+            return VERSION_FLUX_SCHNELL;
+        } else if (is_lite) {
+            return VERSION_FLUX_LITE;
+        } else {
+            return VERSION_FLUX_DEV;
+        }
     }
     if (is_sd3) {
         return VERSION_SD3_2B;
@@ -1629,6 +1734,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     } else if (tensor_storage.is_f8_e4m3) {
                         // inplace op
                         f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
+                    } else if (tensor_storage.is_f8_e5m2) {
+                        // inplace op
+                        f8_e5m2_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
                     }
                 } else {
                     read_buffer.resize(tensor_storage.nbytes());
@@ -1640,6 +1748,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     } else if (tensor_storage.is_f8_e4m3) {
                         // inplace op
                         f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                    } else if (tensor_storage.is_f8_e5m2) {
+                        // inplace op
+                        f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
                     }
 
                     convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data,
@@ -1655,6 +1766,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                 } else if (tensor_storage.is_f8_e4m3) {
                     // inplace op
                     f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                } else if (tensor_storage.is_f8_e5m2) {
+                    // inplace op
+                    f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
                 }
 
                 if (tensor_storage.type == dst_tensor->type) {

model.h

@@ -26,14 +26,43 @@ enum SDVersion {
     VERSION_FLUX_DEV,
     VERSION_FLUX_SCHNELL,
     VERSION_SD3_5_8B,
+    VERSION_SD3_5_2B,
+    VERSION_FLUX_LITE,
     VERSION_COUNT,
 };
 
+static inline bool sd_version_is_flux(SDVersion version) {
+    if (version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL || version == VERSION_FLUX_LITE) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sd3(SDVersion version) {
+    if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B || version == VERSION_SD3_5_2B) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_dit(SDVersion version) {
+    if (sd_version_is_flux(version) || sd_version_is_sd3(version)) {
+        return true;
+    }
+    return false;
+}
+
+enum PMVersion {
+    PM_VERSION_1,
+    PM_VERSION_2,
+};
+
 struct TensorStorage {
     std::string name;
     ggml_type type          = GGML_TYPE_F32;
     bool is_bf16            = false;
     bool is_f8_e4m3         = false;
+    bool is_f8_e5m2         = false;
     int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
     int n_dims              = 0;
 
@@ -63,7 +92,7 @@ struct TensorStorage {
     }
 
     int64_t nbytes_to_read() const {
-        if (is_bf16 || is_f8_e4m3) {
+        if (is_bf16 || is_f8_e4m3 || is_f8_e5m2) {
             return nbytes() / 2;
         } else {
             return nbytes();
@@ -113,6 +142,8 @@ struct TensorStorage {
             type_name = "bf16";
         } else if (is_f8_e4m3) {
             type_name = "f8_e4m3";
+        } else if (is_f8_e5m2) {
+            type_name = "f8_e5m2";
         }
         ss << name << " | " << type_name << " | ";
         ss << n_dims << " [";
@@ -157,6 +188,7 @@ public:
     bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
                       ggml_backend_t backend,
                       std::set<std::string> ignore_tensors = {});
+
     bool save_to_gguf_file(const std::string& file_path, ggml_type type);
     bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
     int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);

pmid.hpp

@@ -42,6 +42,370 @@ public:
     }
 };
 
+/*
+class QFormerPerceiver(nn.Module):
+    def __init__(self, id_embeddings_dim, cross_attention_dim, num_tokens, embedding_dim=1024, use_residual=True, ratio=4):
+        super().__init__()
+
+        self.num_tokens = num_tokens
+        self.cross_attention_dim = cross_attention_dim
+        self.use_residual = use_residual
+        print(cross_attention_dim*num_tokens)
+        self.token_proj = nn.Sequential(
+            nn.Linear(id_embeddings_dim, id_embeddings_dim*ratio),
+            nn.GELU(),
+            nn.Linear(id_embeddings_dim*ratio, cross_attention_dim*num_tokens),
+        )
+        self.token_norm = nn.LayerNorm(cross_attention_dim)
+        self.perceiver_resampler = FacePerceiverResampler(
+            dim=cross_attention_dim,
+            depth=4,
+            dim_head=128,
+            heads=cross_attention_dim // 128,
+            embedding_dim=embedding_dim,
+            output_dim=cross_attention_dim,
+            ff_mult=4,
+        )
+
+    def forward(self, x, last_hidden_state):
+        x = self.token_proj(x)
+        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
+        x = self.token_norm(x) # cls token
+        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
+        if self.use_residual: # TODO: if use_residual is not true
+            out = x + 1.0 * out
+        return out
+*/
+
+struct PMFeedForward : public GGMLBlock {
+    // network hparams
+    int dim;
+
+public:
+    PMFeedForward(int d, int multi = 4)
+        : dim(d) {
+        int inner_dim = dim * multi;
+        blocks["0"]   = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["1"]   = std::shared_ptr<GGMLBlock>(new Mlp(dim, inner_dim, dim, false));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x) {
+        auto norm = std::dynamic_pointer_cast<LayerNorm>(blocks["0"]);
+        auto ff   = std::dynamic_pointer_cast<Mlp>(blocks["1"]);
+
+        x = norm->forward(ctx, x);
+        x = ff->forward(ctx, x);
+        return x;
+    }
+};
+
+struct PerceiverAttention : public GGMLBlock {
+    // network hparams
+    float scale;   // = dim_head**-0.5
+    int dim_head;  // = dim_head
+    int heads;     // = heads
+public:
+    PerceiverAttention(int dim, int dim_h = 64, int h = 8)
+        : scale(powf(dim_h, -0.5)), dim_head(dim_h), heads(h) {
+        int inner_dim    = dim_head * heads;
+        blocks["norm1"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["norm2"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["to_q"]   = std::shared_ptr<GGMLBlock>(new Linear(dim, inner_dim, false));
+        blocks["to_kv"]  = std::shared_ptr<GGMLBlock>(new Linear(dim, inner_dim * 2, false));
+        blocks["to_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim, false));
+    }
+
+    struct ggml_tensor* reshape_tensor(struct ggml_context* ctx,
+                                       struct ggml_tensor* x,
+                                       int heads) {
+        int64_t ne[4];
+        for (int i = 0; i < 4; ++i)
+            ne[i] = x->ne[i];
+        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 0: ");
+        // printf("heads = %d \n", heads);
+        // x = ggml_view_4d(ctx, x, x->ne[0], x->ne[1], heads, x->ne[2]/heads,
+        //                          x->nb[1], x->nb[2], x->nb[3], 0);
+        x = ggml_reshape_4d(ctx, x, x->ne[0] / heads, heads, x->ne[1], x->ne[2]);
+        // x = ggml_view_4d(ctx, x, x->ne[0]/heads, heads, x->ne[1], x->ne[2],
+        //                          x->nb[1], x->nb[2], x->nb[3], 0);
+        // x = ggml_cont(ctx, x);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
+        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 1: ");
+        // x  = ggml_reshape_4d(ctx, x, ne[0], heads, ne[1], ne[2]/heads);
+        return x;
+    }
+
+    std::vector<struct ggml_tensor*> chunk_half(struct ggml_context* ctx,
+                                                struct ggml_tensor* x) {
+        auto tlo = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], 0);
+        auto tli = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], x->nb[0] * x->ne[0] / 2);
+        return {ggml_cont(ctx, tlo),
+                ggml_cont(ctx, tli)};
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* latents) {
+        // x (torch.Tensor): image features
+        //     shape (b, n1, D)
+        // latent (torch.Tensor): latent features
+        //     shape (b, n2, D)
+        int64_t ne[4];
+        for (int i = 0; i < 4; ++i)
+            ne[i] = latents->ne[i];
+
+        auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
+        auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
+        x          = norm1->forward(ctx, x);
+        latents    = norm2->forward(ctx, latents);
+        auto to_q  = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+        auto q     = to_q->forward(ctx, latents);
+
+        auto kv_input = ggml_concat(ctx, x, latents, 1);
+        auto to_kv    = std::dynamic_pointer_cast<Linear>(blocks["to_kv"]);
+        auto kv       = to_kv->forward(ctx, kv_input);
+        auto k        = ggml_view_4d(ctx, kv, kv->ne[0] / 2, kv->ne[1], kv->ne[2], kv->ne[3], kv->nb[1] / 2, kv->nb[2] / 2, kv->nb[3] / 2, 0);
+        auto v        = ggml_view_4d(ctx, kv, kv->ne[0] / 2, kv->ne[1], kv->ne[2], kv->ne[3], kv->nb[1] / 2, kv->nb[2] / 2, kv->nb[3] / 2, kv->nb[0] * (kv->ne[0] / 2));
+        k             = ggml_cont(ctx, k);
+        v             = ggml_cont(ctx, v);
+        q             = reshape_tensor(ctx, q, heads);
+        k             = reshape_tensor(ctx, k, heads);
+        v             = reshape_tensor(ctx, v, heads);
+        scale         = 1.f / sqrt(sqrt((float)dim_head));
+        k             = ggml_scale_inplace(ctx, k, scale);
+        q             = ggml_scale_inplace(ctx, q, scale);
+        // auto weight = ggml_mul_mat(ctx, q, k);
+        auto weight = ggml_mul_mat(ctx, k, q);  // NOTE order of mul is opposite to pytorch
+
+        // GGML's softmax() is equivalent to pytorch's softmax(x, dim=-1)
+        // in this case, dimension along which Softmax will be computed is the last dim
+        // in torch and the first dim in GGML, consistent with the convention that pytorch's
+        // last dimension (varying most rapidly) corresponds to GGML's first (varying most rapidly).
+        // weight = ggml_soft_max(ctx, weight);
+        weight = ggml_soft_max_inplace(ctx, weight);
+        v      = ggml_cont(ctx, ggml_transpose(ctx, v));
+        // auto out = ggml_mul_mat(ctx, weight, v);
+        auto out    = ggml_mul_mat(ctx, v, weight);  // NOTE order of mul is opposite to pytorch
+        out         = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));
+        out         = ggml_reshape_3d(ctx, out, ne[0], ne[1], ggml_nelements(out) / (ne[0] * ne[1]));
+        auto to_out = std::dynamic_pointer_cast<Linear>(blocks["to_out"]);
+        out         = to_out->forward(ctx, out);
+        return out;
+    }
+};
+
+struct FacePerceiverResampler : public GGMLBlock {
+    // network hparams
+    int depth;
+
+public:
+    FacePerceiverResampler(int dim           = 768,
+                           int d             = 4,
+                           int dim_head      = 64,
+                           int heads         = 16,
+                           int embedding_dim = 1280,
+                           int output_dim    = 768,
+                           int ff_mult       = 4)
+        : depth(d) {
+        blocks["proj_in"]  = std::shared_ptr<GGMLBlock>(new Linear(embedding_dim, dim, true));
+        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(dim, output_dim, true));
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new LayerNorm(output_dim));
+
+        for (int i = 0; i < depth; i++) {
+            std::string name = "layers." + std::to_string(i) + ".0";
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new PerceiverAttention(dim, dim_head, heads));
+            name             = "layers." + std::to_string(i) + ".1";
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new PMFeedForward(dim, ff_mult));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* latents,
+                                struct ggml_tensor* x) {
+        // x: [N, channels, h, w]
+        auto proj_in  = std::dynamic_pointer_cast<Linear>(blocks["proj_in"]);
+        auto proj_out = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
+        auto norm_out = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_out"]);
+
+        x = proj_in->forward(ctx, x);
+        for (int i = 0; i < depth; i++) {
+            std::string name = "layers." + std::to_string(i) + ".0";
+            auto attn        = std::dynamic_pointer_cast<PerceiverAttention>(blocks[name]);
+            name             = "layers." + std::to_string(i) + ".1";
+            auto ff          = std::dynamic_pointer_cast<PMFeedForward>(blocks[name]);
+            auto t           = attn->forward(ctx, x, latents);
+            latents          = ggml_add(ctx, t, latents);
+            t                = ff->forward(ctx, latents);
+            latents          = ggml_add(ctx, t, latents);
+        }
+        latents = proj_out->forward(ctx, latents);
+        latents = norm_out->forward(ctx, latents);
+        return latents;
+    }
+};
+
+struct QFormerPerceiver : public GGMLBlock {
+    // network hparams
+    int num_tokens;
+    int cross_attention_dim;
+    bool use_residul;
+
+public:
+    QFormerPerceiver(int id_embeddings_dim, int cross_attention_d, int num_t, int embedding_dim = 1024, bool use_r = true, int ratio = 4)
+        : cross_attention_dim(cross_attention_d), num_tokens(num_t), use_residul(use_r) {
+        blocks["token_proj"]          = std::shared_ptr<GGMLBlock>(new Mlp(id_embeddings_dim,
+                                                                           id_embeddings_dim * ratio,
+                                                                           cross_attention_dim * num_tokens,
+                                                                           true));
+        blocks["token_norm"]          = std::shared_ptr<GGMLBlock>(new LayerNorm(cross_attention_d));
+        blocks["perceiver_resampler"] = std::shared_ptr<GGMLBlock>(new FacePerceiverResampler(
+            cross_attention_dim,
+            4,
+            128,
+            cross_attention_dim / 128,
+            embedding_dim,
+            cross_attention_dim,
+            4));
+    }
+
+    /*
+    def forward(self, x, last_hidden_state):
+        x = self.token_proj(x)
+        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
+        x = self.token_norm(x) # cls token
+        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
+        if self.use_residual: # TODO: if use_residual is not true
+            out = x + 1.0 * out
+        return out
+    */
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* last_hidden_state) {
+        // x: [N, channels, h, w]
+        auto token_proj          = std::dynamic_pointer_cast<Mlp>(blocks["token_proj"]);
+        auto token_norm          = std::dynamic_pointer_cast<LayerNorm>(blocks["token_norm"]);
+        auto perceiver_resampler = std::dynamic_pointer_cast<FacePerceiverResampler>(blocks["perceiver_resampler"]);
+
+        x                       = token_proj->forward(ctx, x);
+        int64_t nel             = ggml_nelements(x);
+        x                       = ggml_reshape_3d(ctx, x, cross_attention_dim, num_tokens, nel / (cross_attention_dim * num_tokens));
+        x                       = token_norm->forward(ctx, x);
+        struct ggml_tensor* out = perceiver_resampler->forward(ctx, x, last_hidden_state);
+        if (use_residul)
+            out = ggml_add(ctx, x, out);
+        return out;
+    }
+};
+
+/*
+class FacePerceiverResampler(torch.nn.Module):
+    def __init__(
+        self,
+        *,
+        dim=768,
+        depth=4,
+        dim_head=64,
+        heads=16,
+        embedding_dim=1280,
+        output_dim=768,
+        ff_mult=4,
+    ):
+        super().__init__()
+
+        self.proj_in = torch.nn.Linear(embedding_dim, dim)
+        self.proj_out = torch.nn.Linear(dim, output_dim)
+        self.norm_out = torch.nn.LayerNorm(output_dim)
+        self.layers = torch.nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                torch.nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, latents, x):
+        x = self.proj_in(x)
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+*/
+
+/*
+
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+*/
+
 struct FuseModule : public GGMLBlock {
     // network hparams
     int embed_dim;
@@ -61,12 +425,19 @@ public:
         auto mlp2       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp2"]);
         auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm"]);
 
-        auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
-        auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
-        // concat is along dim 2
-        auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0, 2);
-        stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
+        // print_ggml_tensor(id_embeds, true, "Fuseblock id_embeds: ");
+        // print_ggml_tensor(prompt_embeds, true, "Fuseblock prompt_embeds: ");
 
+        // auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
+        // auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
+        // print_ggml_tensor(id_embeds0, true, "Fuseblock id_embeds0: ");
+        // print_ggml_tensor(prompt_embeds0, true, "Fuseblock prompt_embeds0: ");
+        // concat is along dim 2
+        // auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0, 2);
+        auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds, id_embeds, 0);
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 0: ");
+        // stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
         // stacked_id_embeds = mlp1.forward(ctx, stacked_id_embeds);
         // stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
         // stacked_id_embeds = mlp2.forward(ctx, stacked_id_embeds);
@@ -77,6 +448,8 @@ public:
         stacked_id_embeds = mlp2->forward(ctx, stacked_id_embeds);
         stacked_id_embeds = layer_norm->forward(ctx, stacked_id_embeds);
 
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
+
         return stacked_id_embeds;
     }
 
@@ -98,23 +471,31 @@ public:
         // print_ggml_tensor(class_tokens_mask_pos, true, "class_tokens_mask_pos");
         struct ggml_tensor* image_token_embeds = ggml_get_rows(ctx, prompt_embeds, class_tokens_mask_pos);
         ggml_set_name(image_token_embeds, "image_token_embeds");
+        valid_id_embeds                       = ggml_reshape_2d(ctx, valid_id_embeds, valid_id_embeds->ne[0],
+                                                                ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
         struct ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
 
-        stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "AA stacked_id_embeds");
+        // print_ggml_tensor(left, true, "AA left");
+        // print_ggml_tensor(right, true, "AA right");
         if (left && right) {
-            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 2);
-            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 2);
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
         } else if (left) {
-            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 2);
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
         } else if (right) {
-            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 2);
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
         }
-        stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "BB stacked_id_embeds");
+        // stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "CC stacked_id_embeds");
         class_tokens_mask                         = ggml_cont(ctx, ggml_transpose(ctx, class_tokens_mask));
         class_tokens_mask                         = ggml_repeat(ctx, class_tokens_mask, prompt_embeds);
         prompt_embeds                             = ggml_mul(ctx, prompt_embeds, class_tokens_mask);
         struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx, prompt_embeds, stacked_id_embeds);
         ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
+        // print_ggml_tensor(updated_prompt_embeds, true, "updated_prompt_embeds: ");
         return updated_prompt_embeds;
     }
 };
@@ -159,10 +540,77 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
     }
 };
 
+struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionModelProjection {
+    int cross_attention_dim;
+    int num_tokens;
+
+    PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock(int id_embeddings_dim = 512)
+        : CLIPVisionModelProjection(OPENAI_CLIP_VIT_L_14),
+          cross_attention_dim(2048),
+          num_tokens(2) {
+        blocks["visual_projection_2"] = std::shared_ptr<GGMLBlock>(new Linear(1024, 1280, false));
+        blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
+        /*
+        cross_attention_dim = 2048
+        # projection
+        self.num_tokens = 2
+        self.cross_attention_dim = cross_attention_dim
+        self.qformer_perceiver = QFormerPerceiver(
+                                    id_embeddings_dim,
+                                    cross_attention_dim,
+                                    self.num_tokens,
+                                )*/
+        blocks["qformer_perceiver"] = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
+                                                                                      cross_attention_dim,
+                                                                                      num_tokens));
+    }
+
+    /*
+    def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds):
+        b, num_inputs, c, h, w = id_pixel_values.shape
+        id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
+
+        last_hidden_state = self.vision_model(id_pixel_values)[0]
+        id_embeds = id_embeds.view(b * num_inputs, -1)
+
+        id_embeds = self.qformer_perceiver(id_embeds, last_hidden_state)
+        id_embeds = id_embeds.view(b, num_inputs, self.num_tokens, -1)
+        updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
+    */
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* id_pixel_values,
+                                struct ggml_tensor* prompt_embeds,
+                                struct ggml_tensor* class_tokens_mask,
+                                struct ggml_tensor* class_tokens_mask_pos,
+                                struct ggml_tensor* id_embeds,
+                                struct ggml_tensor* left,
+                                struct ggml_tensor* right) {
+        // x: [N, channels, h, w]
+        auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
+        auto fuse_module       = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
+        auto qformer_perceiver = std::dynamic_pointer_cast<QFormerPerceiver>(blocks["qformer_perceiver"]);
+
+        // struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
+        id_embeds                             = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);
+
+        struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
+                                                                         prompt_embeds,
+                                                                         id_embeds,
+                                                                         class_tokens_mask,
+                                                                         class_tokens_mask_pos,
+                                                                         left, right);
+        return updated_prompt_embeds;
+    }
+};
+
 struct PhotoMakerIDEncoder : public GGMLRunner {
 public:
-    SDVersion version = VERSION_SDXL;
+    SDVersion version    = VERSION_SDXL;
+    PMVersion pm_version = PM_VERSION_1;
     PhotoMakerIDEncoderBlock id_encoder;
+    PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock id_encoder2;
     float style_strength;
 
     std::vector<float> ctm;
@@ -175,25 +623,38 @@ public:
     std::vector<float> zeros_right;
 
 public:
-    PhotoMakerIDEncoder(ggml_backend_t backend, ggml_type wtype, SDVersion version = VERSION_SDXL, float sty = 20.f)
+    PhotoMakerIDEncoder(ggml_backend_t backend, ggml_type wtype, SDVersion version = VERSION_SDXL, PMVersion pm_v = PM_VERSION_1, float sty = 20.f)
         : GGMLRunner(backend, wtype),
           version(version),
+          pm_version(pm_v),
           style_strength(sty) {
-        id_encoder.init(params_ctx, wtype);
+        if (pm_version == PM_VERSION_1) {
+            id_encoder.init(params_ctx, wtype);
+        } else if (pm_version == PM_VERSION_2) {
+            id_encoder2.init(params_ctx, wtype);
+        }
     }
 
     std::string get_desc() {
         return "pmid";
     }
 
+    PMVersion get_version() const {
+        return pm_version;
+    }
+
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
-        id_encoder.get_param_tensors(tensors, prefix);
+        if (pm_version == PM_VERSION_1)
+            id_encoder.get_param_tensors(tensors, prefix);
+        else if (pm_version == PM_VERSION_2)
+            id_encoder2.get_param_tensors(tensors, prefix);
     }
 
     struct ggml_cgraph* build_graph(  // struct ggml_allocr* allocr,
         struct ggml_tensor* id_pixel_values,
         struct ggml_tensor* prompt_embeds,
-        std::vector<bool>& class_tokens_mask) {
+        std::vector<bool>& class_tokens_mask,
+        struct ggml_tensor* id_embeds) {
         ctm.clear();
         ctmf16.clear();
         ctmpos.clear();
@@ -214,25 +675,32 @@ public:
 
         struct ggml_tensor* id_pixel_values_d = to_backend(id_pixel_values);
         struct ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
+        struct ggml_tensor* id_embeds_d       = to_backend(id_embeds);
 
         struct ggml_tensor* left  = NULL;
         struct ggml_tensor* right = NULL;
         for (int i = 0; i < class_tokens_mask.size(); i++) {
             if (class_tokens_mask[i]) {
+                // printf(" 1,");
                 ctm.push_back(0.f);                        // here use 0.f instead of 1.f to make a scale mask
                 ctmf16.push_back(ggml_fp32_to_fp16(0.f));  // here use 0.f instead of 1.f to make a scale mask
                 ctmpos.push_back(i);
             } else {
+                // printf(" 0,");
                 ctm.push_back(1.f);                        // here use 1.f instead of 0.f to make a scale mask
                 ctmf16.push_back(ggml_fp32_to_fp16(1.f));  // here use 0.f instead of 1.f to make a scale mask
             }
         }
+        // printf("\n");
         if (ctmpos[0] > 0) {
-            left = ggml_new_tensor_3d(ctx0, type, hidden_size, 1, ctmpos[0]);
+            // left = ggml_new_tensor_3d(ctx0, type, hidden_size, 1, ctmpos[0]);
+            left = ggml_new_tensor_3d(ctx0, type, hidden_size, ctmpos[0], 1);
         }
         if (ctmpos[ctmpos.size() - 1] < seq_length - 1) {
+            // right = ggml_new_tensor_3d(ctx0, type,
+            //                            hidden_size, 1, seq_length - ctmpos[ctmpos.size() - 1] - 1);
             right = ggml_new_tensor_3d(ctx0, type,
-                                       hidden_size, 1, seq_length - ctmpos[ctmpos.size() - 1] - 1);
+                                       hidden_size, seq_length - ctmpos[ctmpos.size() - 1] - 1, 1);
         }
         struct ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ctmpos.size());
 
@@ -265,12 +733,23 @@ public:
                 }
             }
         }
-        struct ggml_tensor* updated_prompt_embeds = id_encoder.forward(ctx0,
-                                                                       id_pixel_values_d,
-                                                                       prompt_embeds_d,
-                                                                       class_tokens_mask_d,
-                                                                       class_tokens_mask_pos,
-                                                                       left, right);
+        struct ggml_tensor* updated_prompt_embeds = NULL;
+        if (pm_version == PM_VERSION_1)
+            updated_prompt_embeds = id_encoder.forward(ctx0,
+                                                       id_pixel_values_d,
+                                                       prompt_embeds_d,
+                                                       class_tokens_mask_d,
+                                                       class_tokens_mask_pos,
+                                                       left, right);
+        else if (pm_version == PM_VERSION_2)
+            updated_prompt_embeds = id_encoder2.forward(ctx0,
+                                                        id_pixel_values_d,
+                                                        prompt_embeds_d,
+                                                        class_tokens_mask_d,
+                                                        class_tokens_mask_pos,
+                                                        id_embeds_d,
+                                                        left, right);
+
         ggml_build_forward_expand(gf, updated_prompt_embeds);
 
         return gf;
@@ -279,12 +758,13 @@ public:
     void compute(const int n_threads,
                  struct ggml_tensor* id_pixel_values,
                  struct ggml_tensor* prompt_embeds,
+                 struct ggml_tensor* id_embeds,
                  std::vector<bool>& class_tokens_mask,
                  struct ggml_tensor** updated_prompt_embeds,
                  ggml_context* output_ctx) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
             // return build_graph(compute_allocr, id_pixel_values, prompt_embeds, class_tokens_mask);
-            return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask);
+            return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds);
         };
 
         // GGMLRunner::compute(get_graph, n_threads, updated_prompt_embeds);
@@ -292,4 +772,75 @@ public:
     }
 };
 
+struct PhotoMakerIDEmbed : public GGMLRunner {
+    std::map<std::string, struct ggml_tensor*> tensors;
+    std::string file_path;
+    ModelLoader* model_loader;
+    bool load_failed = false;
+    bool applied     = false;
+
+    PhotoMakerIDEmbed(ggml_backend_t backend,
+                      ggml_type wtype,
+                      ModelLoader* ml,
+                      const std::string& file_path = "",
+                      const std::string& prefix    = "")
+        : file_path(file_path), GGMLRunner(backend, wtype), model_loader(ml) {
+        if (!model_loader->init_from_file(file_path, prefix)) {
+            load_failed = true;
+        }
+    }
+
+    std::string get_desc() {
+        return "id_embeds";
+    }
+
+    bool load_from_file(bool filter_tensor = false) {
+        LOG_INFO("loading PhotoMaker ID Embeds from '%s'", file_path.c_str());
+
+        if (load_failed) {
+            LOG_ERROR("init photomaker id embed from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool dry_run          = true;
+        auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+            const std::string& name = tensor_storage.name;
+
+            if (filter_tensor && !contains(name, "pmid.id_embeds")) {
+                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
+                return true;
+            }
+            if (dry_run) {
+                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+                                                           tensor_storage.type,
+                                                           tensor_storage.n_dims,
+                                                           tensor_storage.ne);
+                tensors[name]            = real;
+            } else {
+                auto real   = tensors[name];
+                *dst_tensor = real;
+            }
+
+            return true;
+        };
+
+        model_loader->load_tensors(on_new_tensor_cb, backend);
+        alloc_params_buffer();
+
+        dry_run = false;
+        model_loader->load_tensors(on_new_tensor_cb, backend);
+
+        LOG_DEBUG("finished loading PhotoMaker ID Embeds ");
+        return true;
+    }
+
+    struct ggml_tensor* get() {
+        std::map<std::string, struct ggml_tensor*>::iterator pos;
+        pos = tensors.find("pmid.id_embeds");
+        if (pos != tensors.end())
+            return pos->second;
+        return NULL;
+    }
+};
+
 #endif  // __PMI_HPP__

stable-diffusion.cpp

@@ -32,7 +32,9 @@ const char* model_version_to_str[] = {
     "SD3 2B",
     "Flux Dev",
     "Flux Schnell",
-    "SD3.5 8B"};
+    "SD3.5 8B",
+    "SD3.5 2B",
+    "Flux Lite 8B"};
 
 const char* sampling_methods_str[] = {
     "Euler A",
@@ -93,6 +95,7 @@ public:
     std::shared_ptr<ControlNet> control_net;
     std::shared_ptr<PhotoMakerIDEncoder> pmid_model;
     std::shared_ptr<LoraModel> pmid_lora;
+    std::shared_ptr<PhotoMakerIDEmbed> pmid_id_embeds;
 
     std::string taesd_path;
     bool use_tiny_autoencoder = false;
@@ -153,7 +156,8 @@ public:
                         schedule_t schedule,
                         bool clip_on_cpu,
                         bool control_net_cpu,
-                        bool vae_on_cpu) {
+                        bool vae_on_cpu,
+                        bool diffusion_flash_attn) {
         use_tiny_autoencoder = taesd_path.size() > 0;
 #ifdef SD_USE_CUBLAS
         LOG_DEBUG("Using CUDA backend");
@@ -182,13 +186,7 @@ public:
             LOG_DEBUG("Using CPU backend");
             backend = ggml_backend_cpu_init();
         }
-#ifdef SD_USE_FLASH_ATTENTION
-#if defined(SD_USE_CUBLAS) || defined(SD_USE_METAL) || defined(SD_USE_SYCL) || defined(SD_USE_VULKAN)
-        LOG_WARN("Flash Attention not supported with GPU Backend");
-#else
-        LOG_INFO("Flash Attention enabled");
-#endif
-#endif
+
         ModelLoader model_loader;
 
         vae_tiling = vae_tiling_;
@@ -288,9 +286,9 @@ public:
                     "try specifying SDXL VAE FP16 Fix with the --vae parameter. "
                     "You can find it here: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors");
             }
-        } else if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B) {
+        } else if (sd_version_is_sd3(version)) {
             scale_factor = 1.5305f;
-        } else if (version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
+        } else if (sd_version_is_flux(version)) {
             scale_factor = 0.3611;
             // TODO: shift_factor
         }
@@ -311,7 +309,7 @@ public:
         } else {
             clip_backend   = backend;
             bool use_t5xxl = false;
-            if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B || version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
+            if (sd_version_is_dit(version)) {
                 use_t5xxl = true;
             }
             if (!ggml_backend_is_cpu(backend) && use_t5xxl && conditioner_wtype != GGML_TYPE_F32) {
@@ -322,15 +320,25 @@ public:
                 LOG_INFO("CLIP: Using CPU backend");
                 clip_backend = ggml_backend_cpu_init();
             }
-            if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B) {
+            if (diffusion_flash_attn) {
+                LOG_INFO("Using flash attention in the diffusion model");
+            }
+            if (sd_version_is_sd3(version)) {
+                if (diffusion_flash_attn) {
+                    LOG_WARN("flash attention in this diffusion model is currently unsupported!");
+                }
                 cond_stage_model = std::make_shared<SD3CLIPEmbedder>(clip_backend, conditioner_wtype);
                 diffusion_model  = std::make_shared<MMDiTModel>(backend, diffusion_model_wtype, version);
-            } else if (version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
+            } else if (sd_version_is_flux(version)) {
                 cond_stage_model = std::make_shared<FluxCLIPEmbedder>(clip_backend, conditioner_wtype);
-                diffusion_model  = std::make_shared<FluxModel>(backend, diffusion_model_wtype, version);
+                diffusion_model  = std::make_shared<FluxModel>(backend, diffusion_model_wtype, version, diffusion_flash_attn);
             } else {
-                cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, conditioner_wtype, embeddings_path, version);
-                diffusion_model  = std::make_shared<UNetModel>(backend, diffusion_model_wtype, version);
+                if (id_embeddings_path.find("v2") != std::string::npos) {
+                    cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, conditioner_wtype, embeddings_path, version, PM_VERSION_2);
+                } else {
+                    cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, conditioner_wtype, embeddings_path, version);
+                }
+                diffusion_model = std::make_shared<UNetModel>(backend, diffusion_model_wtype, version, diffusion_flash_attn);
             }
             cond_stage_model->alloc_params_buffer();
             cond_stage_model->get_param_tensors(tensors);
@@ -364,7 +372,12 @@ public:
                 control_net = std::make_shared<ControlNet>(controlnet_backend, diffusion_model_wtype, version);
             }
 
-            pmid_model = std::make_shared<PhotoMakerIDEncoder>(clip_backend, model_wtype, version);
+            if (id_embeddings_path.find("v2") != std::string::npos) {
+                pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_wtype, version, PM_VERSION_2);
+                LOG_INFO("using PhotoMaker Version 2");
+            } else {
+                pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_wtype, version);
+            }
             if (id_embeddings_path.size() > 0) {
                 pmid_lora = std::make_shared<LoraModel>(backend, model_wtype, id_embeddings_path, "");
                 if (!pmid_lora->load_from_file(true)) {
@@ -383,14 +396,8 @@ public:
                     LOG_ERROR(" pmid model params buffer allocation failed");
                     return false;
                 }
-                // LOG_INFO("pmid param memory buffer size = %.2fMB ",
-                //     pmid_model->params_buffer_size / 1024.0 / 1024.0);
                 pmid_model->get_param_tensors(tensors, "pmid");
             }
-            // if(stacked_id){
-            //    pmid_model.init_params(GGML_TYPE_F32);
-            //    pmid_model.map_by_name(tensors, "pmid.");
-            // }
         }
 
         struct ggml_init_params params;
@@ -520,10 +527,10 @@ public:
             is_using_v_parameterization = true;
         }
 
-        if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B) {
+        if (sd_version_is_sd3(version)) {
             LOG_INFO("running in FLOW mode");
             denoiser = std::make_shared<DiscreteFlowDenoiser>();
-        } else if (version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
+        } else if (sd_version_is_flux(version)) {
             LOG_INFO("running in Flux FLOW mode");
             float shift = 1.15f;
             if (version == VERSION_FLUX_SCHNELL) {
@@ -673,10 +680,10 @@ public:
     ggml_tensor* id_encoder(ggml_context* work_ctx,
                             ggml_tensor* init_img,
                             ggml_tensor* prompts_embeds,
+                            ggml_tensor* id_embeds,
                             std::vector<bool>& class_tokens_mask) {
         ggml_tensor* res = NULL;
-        pmid_model->compute(n_threads, init_img, prompts_embeds, class_tokens_mask, &res, work_ctx);
-
+        pmid_model->compute(n_threads, init_img, prompts_embeds, id_embeds, class_tokens_mask, &res, work_ctx);
         return res;
     }
 
@@ -771,7 +778,11 @@ public:
                         sample_method_t method,
                         const std::vector<float>& sigmas,
                         int start_merge_step,
-                        SDCondition id_cond) {
+                        SDCondition id_cond,
+                        std::vector<int> skip_layers = {},
+                        float slg_scale              = 2.5,
+                        float skip_layer_start       = 0.01,
+                        float skip_layer_end         = 0.2) {
         size_t steps = sigmas.size() - 1;
         // noise = load_tensor_from_file(work_ctx, "./rand0.bin");
         // print_ggml_tensor(noise);
@@ -782,13 +793,24 @@ public:
         struct ggml_tensor* noised_input = ggml_dup_tensor(work_ctx, noise);
 
         bool has_unconditioned = cfg_scale != 1.0 && uncond.c_crossattn != NULL;
+        bool has_skiplayer     = slg_scale != 0.0 && skip_layers.size() > 0;
 
         // denoise wrapper
         struct ggml_tensor* out_cond   = ggml_dup_tensor(work_ctx, x);
         struct ggml_tensor* out_uncond = NULL;
+        struct ggml_tensor* out_skip   = NULL;
+
         if (has_unconditioned) {
             out_uncond = ggml_dup_tensor(work_ctx, x);
         }
+        if (has_skiplayer) {
+            if (sd_version_is_dit(version)) {
+                out_skip = ggml_dup_tensor(work_ctx, x);
+            } else {
+                has_skiplayer = false;
+                LOG_WARN("SLG is incompatible with %s models", model_version_to_str[version]);
+            }
+        }
         struct ggml_tensor* denoised = ggml_dup_tensor(work_ctx, x);
 
         auto denoise = [&](ggml_tensor* input, float sigma, int step) -> ggml_tensor* {
@@ -869,6 +891,28 @@ public:
                                          &out_uncond);
                 negative_data = (float*)out_uncond->data;
             }
+
+            int step_count         = sigmas.size();
+            bool is_skiplayer_step = has_skiplayer && step > (int)(skip_layer_start * step_count) && step < (int)(skip_layer_end * step_count);
+            float* skip_layer_data = NULL;
+            if (is_skiplayer_step) {
+                LOG_DEBUG("Skipping layers at step %d\n", step);
+                // skip layer (same as conditionned)
+                diffusion_model->compute(n_threads,
+                                         noised_input,
+                                         timesteps,
+                                         cond.c_crossattn,
+                                         cond.c_concat,
+                                         cond.c_vector,
+                                         guidance_tensor,
+                                         -1,
+                                         controls,
+                                         control_strength,
+                                         &out_skip,
+                                         NULL,
+                                         skip_layers);
+                skip_layer_data = (float*)out_skip->data;
+            }
             float* vec_denoised  = (float*)denoised->data;
             float* vec_input     = (float*)input->data;
             float* positive_data = (float*)out_cond->data;
@@ -885,6 +929,9 @@ public:
                         latent_result = negative_data[i] + cfg_scale * (positive_data[i] - negative_data[i]);
                     }
                 }
+                if (is_skiplayer_step) {
+                    latent_result = latent_result + (positive_data[i] - skip_layer_data[i]) * slg_scale;
+                }
                 // v = latent_result, eps = latent_result
                 // denoised = (v * c_out + input * c_skip) or (input + eps * c_out)
                 vec_denoised[i] = latent_result * c_out + vec_input[i] * c_skip;
@@ -948,9 +995,9 @@ public:
         if (use_tiny_autoencoder) {
             C = 4;
         } else {
-            if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B) {
+            if (sd_version_is_sd3(version)) {
                 C = 32;
-            } else if (version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
+            } else if (sd_version_is_flux(version)) {
                 C = 32;
             }
         }
@@ -1035,7 +1082,8 @@ sd_ctx_t* new_sd_ctx(const char* model_path_c_str,
                      enum schedule_t s,
                      bool keep_clip_on_cpu,
                      bool keep_control_net_cpu,
-                     bool keep_vae_on_cpu) {
+                     bool keep_vae_on_cpu,
+                     bool diffusion_flash_attn) {
     sd_ctx_t* sd_ctx = (sd_ctx_t*)malloc(sizeof(sd_ctx_t));
     if (sd_ctx == NULL) {
         return NULL;
@@ -1076,7 +1124,8 @@ sd_ctx_t* new_sd_ctx(const char* model_path_c_str,
                                     s,
                                     keep_clip_on_cpu,
                                     keep_control_net_cpu,
-                                    keep_vae_on_cpu)) {
+                                    keep_vae_on_cpu,
+                                    diffusion_flash_attn)) {
         delete sd_ctx->sd;
         sd_ctx->sd = NULL;
         free(sd_ctx);
@@ -1111,7 +1160,11 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx,
                            float control_strength,
                            float style_ratio,
                            bool normalize_input,
-                           std::string input_id_images_path) {
+                           std::string input_id_images_path,
+                           std::vector<int> skip_layers = {},
+                           float slg_scale              = 2.5,
+                           float skip_layer_start       = 0.01,
+                           float skip_layer_end         = 0.2) {
     if (seed < 0) {
         // Generally, when using the provided command line, the seed is always >0.
         // However, to prevent potential issues if 'stable-diffusion.cpp' is invoked as a library
@@ -1161,11 +1214,15 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx,
         }
         // preprocess input id images
         std::vector<sd_image_t*> input_id_images;
+        bool pmv2 = sd_ctx->sd->pmid_model->get_version() == PM_VERSION_2;
         if (sd_ctx->sd->pmid_model && input_id_images_path.size() > 0) {
             std::vector<std::string> img_files = get_files_from_dir(input_id_images_path);
             for (std::string img_file : img_files) {
                 int c = 0;
                 int width, height;
+                if (ends_with(img_file, "safetensors")) {
+                    continue;
+                }
                 uint8_t* input_image_buffer = stbi_load(img_file.c_str(), &width, &height, &c, 3);
                 if (input_image_buffer == NULL) {
                     LOG_ERROR("PhotoMaker load image from '%s' failed", img_file.c_str());
@@ -1203,18 +1260,23 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx,
                 else
                     sd_mul_images_to_tensor(init_image->data, init_img, i, NULL, NULL);
             }
-            t0                = ggml_time_ms();
-            auto cond_tup     = sd_ctx->sd->cond_stage_model->get_learned_condition_with_trigger(work_ctx,
-                                                                                                 sd_ctx->sd->n_threads, prompt,
-                                                                                                 clip_skip,
-                                                                                                 width,
-                                                                                                 height,
-                                                                                                 num_input_images,
-                                                                                                 sd_ctx->sd->diffusion_model->get_adm_in_channels());
-            id_cond           = std::get<0>(cond_tup);
-            class_tokens_mask = std::get<1>(cond_tup);  //
-
-            id_cond.c_crossattn = sd_ctx->sd->id_encoder(work_ctx, init_img, id_cond.c_crossattn, class_tokens_mask);
+            t0                            = ggml_time_ms();
+            auto cond_tup                 = sd_ctx->sd->cond_stage_model->get_learned_condition_with_trigger(work_ctx,
+                                                                                                             sd_ctx->sd->n_threads, prompt,
+                                                                                                             clip_skip,
+                                                                                                             width,
+                                                                                                             height,
+                                                                                                             num_input_images,
+                                                                                                             sd_ctx->sd->diffusion_model->get_adm_in_channels());
+            id_cond                       = std::get<0>(cond_tup);
+            class_tokens_mask             = std::get<1>(cond_tup);  //
+            struct ggml_tensor* id_embeds = NULL;
+            if (pmv2) {
+                // id_embeds = sd_ctx->sd->pmid_id_embeds->get();
+                id_embeds = load_tensor_from_file(work_ctx, path_join(input_id_images_path, "id_embeds.bin"));
+                // print_ggml_tensor(id_embeds, true, "id_embeds:");
+            }
+            id_cond.c_crossattn = sd_ctx->sd->id_encoder(work_ctx, init_img, id_cond.c_crossattn, id_embeds, class_tokens_mask);
             t1                  = ggml_time_ms();
             LOG_INFO("Photomaker ID Stacking, taking %" PRId64 " ms", t1 - t0);
             if (sd_ctx->sd->free_params_immediately) {
@@ -1281,9 +1343,9 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx,
     // Sample
     std::vector<struct ggml_tensor*> final_latents;  // collect latents to decode
     int C = 4;
-    if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B) {
+    if (sd_version_is_sd3(sd_ctx->sd->version)) {
         C = 16;
-    } else if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
+    } else if (sd_version_is_flux(sd_ctx->sd->version)) {
         C = 16;
     }
     int W = width / 8;
@@ -1320,7 +1382,11 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx,
                                                      sample_method,
                                                      sigmas,
                                                      start_merge_step,
-                                                     id_cond);
+                                                     id_cond,
+                                                     skip_layers,
+                                                     slg_scale,
+                                                     skip_layer_start,
+                                                     skip_layer_end);
         // struct ggml_tensor* x_0 = load_tensor_from_file(ctx, "samples_ddim.bin");
         // print_ggml_tensor(x_0);
         int64_t sampling_end = ggml_time_ms();
@@ -1386,7 +1452,11 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,
                     float control_strength,
                     float style_ratio,
                     bool normalize_input,
-                    const char* input_id_images_path_c_str) {
+                    const char* input_id_images_path_c_str,
+                    std::vector<int> skip_layers,
+                    float slg_scale,
+                    float skip_layer_start,
+                    float skip_layer_end) {
     LOG_DEBUG("txt2img %dx%d", width, height);
     if (sd_ctx == NULL) {
         return NULL;
@@ -1394,10 +1464,10 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,
 
     struct ggml_init_params params;
     params.mem_size = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-    if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B) {
+    if (sd_version_is_sd3(sd_ctx->sd->version)) {
         params.mem_size *= 3;
     }
-    if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
+    if (sd_version_is_flux(sd_ctx->sd->version)) {
         params.mem_size *= 4;
     }
     if (sd_ctx->sd->stacked_id) {
@@ -1420,17 +1490,17 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,
     std::vector<float> sigmas = sd_ctx->sd->denoiser->get_sigmas(sample_steps);
 
     int C = 4;
-    if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B) {
+    if (sd_version_is_sd3(sd_ctx->sd->version)) {
         C = 16;
-    } else if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
+    } else if (sd_version_is_flux(sd_ctx->sd->version)) {
         C = 16;
     }
     int W                    = width / 8;
     int H                    = height / 8;
     ggml_tensor* init_latent = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, C, 1);
-    if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B) {
+    if (sd_version_is_sd3(sd_ctx->sd->version)) {
         ggml_set_f32(init_latent, 0.0609f);
-    } else if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
+    } else if (sd_version_is_flux(sd_ctx->sd->version)) {
         ggml_set_f32(init_latent, 0.1159f);
     } else {
         ggml_set_f32(init_latent, 0.f);
@@ -1454,7 +1524,11 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,
                                                control_strength,
                                                style_ratio,
                                                normalize_input,
-                                               input_id_images_path_c_str);
+                                               input_id_images_path_c_str,
+                                               skip_layers,
+                                               slg_scale,
+                                               skip_layer_start,
+                                               skip_layer_end);
 
     size_t t1 = ggml_time_ms();
 
@@ -1481,7 +1555,11 @@ sd_image_t* img2img(sd_ctx_t* sd_ctx,
                     float control_strength,
                     float style_ratio,
                     bool normalize_input,
-                    const char* input_id_images_path_c_str) {
+                    const char* input_id_images_path_c_str,
+                    std::vector<int> skip_layers,
+                    float slg_scale,
+                    float skip_layer_start,
+                    float skip_layer_end) {
     LOG_DEBUG("img2img %dx%d", width, height);
     if (sd_ctx == NULL) {
         return NULL;
@@ -1489,10 +1567,10 @@ sd_image_t* img2img(sd_ctx_t* sd_ctx,
 
     struct ggml_init_params params;
     params.mem_size = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-    if (sd_ctx->sd->version == VERSION_SD3_2B || sd_ctx->sd->version == VERSION_SD3_5_8B) {
+    if (sd_version_is_sd3(sd_ctx->sd->version)) {
         params.mem_size *= 2;
     }
-    if (sd_ctx->sd->version == VERSION_FLUX_DEV || sd_ctx->sd->version == VERSION_FLUX_SCHNELL) {
+    if (sd_version_is_flux(sd_ctx->sd->version)) {
         params.mem_size *= 3;
     }
     if (sd_ctx->sd->stacked_id) {
@@ -1555,7 +1633,11 @@ sd_image_t* img2img(sd_ctx_t* sd_ctx,
                                                control_strength,
                                                style_ratio,
                                                normalize_input,
-                                               input_id_images_path_c_str);
+                                               input_id_images_path_c_str,
+                                               skip_layers,
+                                               slg_scale,
+                                               skip_layer_start,
+                                               skip_layer_end);
 
     size_t t2 = ggml_time_ms();
 

stable-diffusion.h

@@ -142,7 +142,8 @@ SD_API sd_ctx_t* new_sd_ctx(const char* model_path,
                             enum schedule_t s,
                             bool keep_clip_on_cpu,
                             bool keep_control_net_cpu,
-                            bool keep_vae_on_cpu);
+                            bool keep_vae_on_cpu,
+                            bool diffusion_flash_attn);
 
 SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
 
@@ -162,7 +163,11 @@ SD_API sd_image_t* txt2img(sd_ctx_t* sd_ctx,
                            float control_strength,
                            float style_strength,
                            bool normalize_input,
-                           const char* input_id_images_path);
+                           const char* input_id_images_path,
+                           std::vector<int> skip_layers = {},
+                           float slg_scale              = 2.5,
+                           float skip_layer_start       = 0.01,
+                           float skip_layer_end         = 0.2);
 
 SD_API sd_image_t* img2img(sd_ctx_t* sd_ctx,
                            sd_image_t init_image,
@@ -182,7 +187,11 @@ SD_API sd_image_t* img2img(sd_ctx_t* sd_ctx,
                            float control_strength,
                            float style_strength,
                            bool normalize_input,
-                           const char* input_id_images_path);
+                           const char* input_id_images_path,
+                           std::vector<int> skip_layers = {},
+                           float slg_scale              = 2.5,
+                           float skip_layer_start       = 0.01,
+                           float skip_layer_end         = 0.2);
 
 SD_API sd_image_t* img2vid(sd_ctx_t* sd_ctx,
                            sd_image_t init_image,

unet.hpp

@@ -183,7 +183,7 @@ public:
     int model_channels  = 320;
     int adm_in_channels = 2816;  // only for VERSION_SDXL/SVD
 
-    UnetModelBlock(SDVersion version = VERSION_SD1)
+    UnetModelBlock(SDVersion version = VERSION_SD1, bool flash_attn = false)
         : version(version) {
         if (version == VERSION_SD2) {
             context_dim       = 1024;
@@ -242,7 +242,7 @@ public:
             if (version == VERSION_SVD) {
                 return new SpatialVideoTransformer(in_channels, n_head, d_head, depth, context_dim);
             } else {
-                return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim);
+                return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, flash_attn);
             }
         };
 
@@ -533,8 +533,9 @@ struct UNetModelRunner : public GGMLRunner {
 
     UNetModelRunner(ggml_backend_t backend,
                     ggml_type wtype,
-                    SDVersion version = VERSION_SD1)
-        : GGMLRunner(backend, wtype), unet(version) {
+                    SDVersion version = VERSION_SD1,
+                    bool flash_attn   = false)
+        : GGMLRunner(backend, wtype), unet(version, flash_attn) {
         unet.init(params_ctx, wtype);
     }
 
@@ -649,4 +650,4 @@ struct UNetModelRunner : public GGMLRunner {
     }
 };
 
-#endif  // __UNET_HPP__
+#endif  // __UNET_HPP__

util.cpp

@@ -276,6 +276,23 @@ std::string path_join(const std::string& p1, const std::string& p2) {
     return p1 + "/" + p2;
 }
 
+std::vector<std::string> splitString(const std::string& str, char delimiter) {
+    std::vector<std::string> result;
+    size_t start = 0;
+    size_t end   = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        result.push_back(str.substr(start, end - start));
+        start = end + 1;
+        end   = str.find(delimiter, start);
+    }
+
+    // Add the last segment after the last delimiter
+    result.push_back(str.substr(start));
+
+    return result;
+}
+
 sd_image_t* preprocess_id_image(sd_image_t* img) {
     int shortest_edge   = 224;
     int size            = shortest_edge;

util.h

@@ -45,7 +45,7 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
 sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size);
 
 std::string path_join(const std::string& p1, const std::string& p2);
-
+std::vector<std::string> splitString(const std::string& str, char delimiter);
 void pretty_progress(int step, int steps, float time);
 
 void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);

vae.hpp

@@ -99,10 +99,12 @@ public:
         k      = ggml_cont(ctx, ggml_permute(ctx, k, 1, 2, 0, 3));  // [N, h, w, in_channels]
         k      = ggml_reshape_3d(ctx, k, c, h * w, n);              // [N, h * w, in_channels]
 
-        auto v = v_proj->forward(ctx, h_);              // [N, in_channels, h, w]
-        v      = ggml_reshape_3d(ctx, v, h * w, c, n);  // [N, in_channels, h * w]
+        auto v = v_proj->forward(ctx, h_);                          // [N, in_channels, h, w]
+        v      = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));  // [N, h, w, in_channels]
+        v      = ggml_reshape_3d(ctx, v, c, h * w, n);              // [N, h * w, in_channels]
 
-        h_ = ggml_nn_attention(ctx, q, k, v, false);  // [N, h * w, in_channels]
+        // h_ = ggml_nn_attention(ctx, q, k, v, false);  // [N, h * w, in_channels]
+        h_ = ggml_nn_attention_ext(ctx, q, k, v, 1, nullptr, false, true, false);
 
         h_ = ggml_cont(ctx, ggml_permute(ctx, h_, 1, 0, 2, 3));  // [N, in_channels, h * w]
         h_ = ggml_reshape_4d(ctx, h_, w, h, c, n);               // [N, in_channels, h, w]
@@ -457,7 +459,7 @@ public:
                        bool use_video_decoder = false,
                        SDVersion version      = VERSION_SD1)
         : decode_only(decode_only), use_video_decoder(use_video_decoder) {
-        if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B || version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL) {
+        if (sd_version_is_dit(version)) {
             dd_config.z_channels = 16;
             use_quant            = false;
         }
@@ -612,4 +614,4 @@ struct AutoEncoderKL : public GGMLRunner {
     };
 };
 
-#endif
+#endif