feat: add MoSR upscaler support (#2128)

.github/workflows/build.yml

@@ -451,8 +451,8 @@ jobs:
     runs-on: windows-2022
 
     env:
-      ROCM_VERSION: "7.14.0"
-      GPU_TARGETS: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+      ROCM_VERSION: "10.1.0"
+      gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
 
     steps:
       - uses: actions/checkout@v3
@@ -498,7 +498,7 @@ jobs:
 
           write-host "Installing ROCm wheels for multi-arch support"
           # Install ROCm wheels for multi-arch support (this may take several minutes)
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{env.ROCM_VERSION}}"
+          python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{env.ROCM_VERSION}}"
 
           # Pre-expand the devel tree so it is included in the cache
           write-host "Initializing ROCm devel tree"
@@ -594,9 +594,9 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "7.14.0"
+          - ROCM_VERSION: "10.1.0"
             gpu_targets: "gfx900;gfx906;gfx908;gfx90a;gfx90c;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
-            build: x64
+            build: 'x64'
 
     steps:
       - name: Clone
@@ -610,6 +610,16 @@ jobs:
         with:
           key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
           evict-old-files: 1d
+          max-size: "2G"
+
+      - name: Tune ccache for reinstalled ROCm toolchain
+        run: |
+          # ROCm is pip-installed fresh each run, so the clang binary's mtime
+          # changes every time. With the default compiler_check=mtime that
+          # invalidates the cache; hash compiler contents instead so warm
+          # builds hit.
+          ccache --set-config=compiler_check=content
+          ccache --set-config=sloppiness=time_macros,include_file_mtime,include_file_ctime
 
       - name: Dependencies
         id: depends
@@ -646,7 +656,7 @@ jobs:
           # libraries = HIP runtime and CMake configs needed for linking
           # devel = compilers, headers, static libs
           python -m pip install --upgrade pip
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{matrix.ROCM_VERSION}}"
+          python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{matrix.ROCM_VERSION}}"
 
           # Get ROCm installation paths using the rocm-sdk CLI tool
           ROCM_PATH=$(rocm-sdk path --root)

CONTRIBUTING.md

@@ -29,6 +29,10 @@ PRs should include:
 - Linked issue/PR context when applicable.
 - Screenshots or sample outputs for UI/visual behavior changes.
 
+## README UI Listings
+
+The [UIs list in the README](README.md#uis) only includes open-source projects that use `stable-diffusion.cpp` as a backend. Do not submit PRs to add closed-source UIs.
+
 ## Code Style
 
 Format code according to the repository style before submitting changes.

README.md

@@ -67,6 +67,7 @@ API and command-line option may change frequently.***
     - [LLaDA-Image](./docs/llada_image.md)
     - [Ming-Image Design](./docs/ming_image.md)
     - [PixArt](./docs/pixart.md)
+    - [Iris-3B](./docs/iris.md)
   - [Image Edit Models](./docs/edit.md)
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -182,7 +183,6 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra
 These projects use `stable-diffusion.cpp` as a backend for their image generation.
 
 - [GIMP Plugins](https://github.com/themanyone/gimp-plugins)
-- [Jellybox](https://jellybox.com)
 - [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
 - [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
 - [Local Diffusion](https://github.com/rmatif/Local-Diffusion)

docs/esrgan.md

@@ -2,7 +2,9 @@
 
 You can use ESRGAN—such as the model [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth)—to upscale the generated images and improve their overall resolution and clarity.
 
-RGBA images, including Qwen Image 2.1 output, keep their alpha channel during model upscaling and hires fix. ESRGAN processes the RGB channels; the alpha channel is resized with bilinear interpolation and recombined with the upscaled image.
+MoSR models with the PixelShuffle upsampler, such as [4xNomos2_hq_mosr.pth](https://github.com/Phhofm/models/releases/download/4xNomos2_hq_mosr/4xNomos2_hq_mosr.pth), are also supported. The architecture is detected from the weights, so they are used the same way as ESRGAN models.
+
+RGBA images, including Qwen Image 2.1 output, keep their alpha channel during model upscaling and hires fix. The upscaler processes the RGB channels; the alpha channel is resized with bilinear interpolation and recombined with the upscaled image.
 
 - Specify the model path using the `--upscale-model PATH` parameter. example:
 

docs/iris.md

@@ -0,0 +1,24 @@
+# How to Use
+
+Iris-3B generates images directly in pixel space and uses Qwen3-VL-4B-Instruct as the text encoder. No VAE is required.
+
+## Download weights
+
+- Download Iris-3B
+    - safetensors: https://huggingface.co/speridlabs/iris-3b/tree/main (`model.safetensors` in the root directory)
+- Download Qwen3-VL-4B-Instruct
+    - safetensors: https://huggingface.co/Comfy-Org/Krea-2/tree/main/text_encoders
+    - gguf: https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/tree/main
+
+## Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\iris-3b.safetensors --llm ..\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf -p "a lovely cat" --cfg-scale 3 -H 1024 -W 1024 --diffusion-fa
+```
+
+<img width="256" alt="iris-3b example" src="../assets/iris/example.png" />
+
+## Notes
+
+- Width and height must be multiples of 16. Do not pass `--vae`.
+- Captions are limited to 300 tokens including the assistant-turn suffix. Positive and negative prompts use the same template.

docs/performance.md

@@ -21,6 +21,16 @@ CPU fallback. It excludes weights and cache buffers. Within a runner lifecycle,
 the summary is printed only on the first graph or when backend capacities or the
 segment count change.
 
+## Run conditional and unconditional CFG in one batched UNet forward.
+
+For UNet models, the conditional and unconditional guidance branches are
+concatenated into a single batch of two and run through one UNet forward per
+step instead of two separate forwards. This is enabled by default whenever the
+run qualifies for it.
+
+Use `--batched-cfg off` to force separate conditional and unconditional
+forwards.
+
 ## Use VAE tiling to reduce encode and decode memory usage.
 
 `--vae-tiling` enables spatial tiling for both VAE encoding and decoding. The

docs/qwen_image_2.1.md

@@ -16,6 +16,8 @@ Qwen Image 2.1 supports text-to-image generation and image editing, using Qwen3-
 
 Use `qwen_image_2.1_vae_bf16.safetensors` with this model. The earlier Qwen Image and Wan 2.2 VAE weights are not interchangeable with the Qwen Image 2.1 VAE weights.
 
+For faster encoding, decoding, or previews, see [TAEQI2.1](taesd.md#qwen-image-21-taeqi21).
+
 ## Examples
 
 Run the following commands from the build directory. Use image dimensions divisible by 32. The resolution-dependent flow schedule is selected automatically.

docs/taesd.md

@@ -37,3 +37,15 @@ sd.cpp also supports [TAEHV](https://github.com/madebyollin/taehv) (#937), which
   ```
 
 Then simply replace the `--vae xxx.safetensors` with `--tae xxx.safetensors` in the commands. If it still out of VRAM, add `--vae-conv-direct` to your command though might be slower.
+
+### Qwen Image 2.1 (TAEQI2.1)
+
+For Qwen Image 2.1, use [taeqi2_1](https://github.com/madebyollin/taesd), which supports 64-channel latents, 16x spatial scaling, and RGBA images.
+
+Download the official [safetensors weights](https://huggingface.co/madebyollin/taeqi2_1/blob/main/taeqi2_1.safetensors) directly; no conversion is needed:
+
+```bash
+curl -L -o taeqi2_1.safetensors https://huggingface.co/madebyollin/taeqi2_1/resolve/main/taeqi2_1.safetensors
+```
+
+Replace `--vae PATH` with `--taesd taeqi2_1.safetensors` in the [Qwen Image 2.1 examples](qwen_image_2.1.md) to use it for encoding and decoding. For previews only, keep `--vae PATH` and add `--taesd taeqi2_1.safetensors --taesd-preview-only --preview tae`.

examples/common/common.cpp

@@ -378,6 +378,23 @@ static int parse_scale_override(int argc, const char** argv, int index, float& s
     return 1;
 }
 
+static int parse_on_off_arg(int argc, const char** argv, int index, const char* option, bool& value) {
+    if (++index >= argc) {
+        LOG_ERROR("%s requires 'on' or 'off'", option);
+        return -1;
+    }
+    const std::string arg = argv[index];
+    if (arg == "on") {
+        value = true;
+    } else if (arg == "off") {
+        value = false;
+    } else {
+        LOG_ERROR("invalid %s value '%s'; expected 'on' or 'off'", option, argv[index]);
+        return -1;
+    }
+    return 1;
+}
+
 ArgOptions SDContextParams::get_options() {
     ArgOptions options;
     options.string_options = {
@@ -637,23 +654,6 @@ ArgOptions SDContextParams::get_options() {
          true, &vae_conv_direct},
     };
 
-    auto on_auto_fit_arg = [&](int argc, const char** argv, int index) {
-        if (++index >= argc) {
-            LOG_ERROR("--auto-fit requires 'on' or 'off'");
-            return -1;
-        }
-        const std::string arg = argv[index];
-        if (arg == "on") {
-            auto_fit = true;
-        } else if (arg == "off") {
-            auto_fit = false;
-        } else {
-            LOG_ERROR("invalid --auto-fit value '%s'; expected 'on' or 'off'", argv[index]);
-            return -1;
-        }
-        return 1;
-    };
-
     auto on_type_arg = [&](int argc, const char** argv, int index) {
         if (++index >= argc) {
             return -1;
@@ -742,7 +742,15 @@ ArgOptions SDContextParams::get_options() {
          "on|off (default: on). Preserve --backend (otherwise select one GPU) and place weights on the compute GPU, "
          "RAM, another GPU, or disk in that order, according to available memory (--max-vram limits GPU budgets). "
          "Disabled by explicit --params-backend; uses automatic graph segmentation when needed",
-         on_auto_fit_arg},
+         [this](int argc, const char** argv, int index) {
+             return parse_on_off_arg(argc, argv, index, "--auto-fit", auto_fit);
+         }},
+        {"",
+         "--batched-cfg",
+         "on|off (default: on). Run the conditional and unconditional CFG branches in one batched UNet forward when supported",
+         [this](int argc, const char** argv, int index) {
+             return parse_on_off_arg(argc, argv, index, "--batched-cfg", batched_cfg);
+         }},
         {"",
          "--type",
          "weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). "
@@ -940,6 +948,7 @@ std::string SDContextParams::to_string() const {
         << "  max_vram: \"" << max_vram << "\",\n"
         << "  disable_prefetch: " << (disable_prefetch ? "true" : "false") << ",\n"
         << "  disable_segmented_compute: " << (disable_segmented_compute ? "true" : "false") << ",\n"
+        << "  batched_cfg: " << (batched_cfg ? "true" : "false") << ",\n"
         << "  eager_load: " << (eager_load ? "true" : "false") << ",\n"
         << "  backend: \"" << backend << "\",\n"
         << "  params_backend: \"" << params_backend << "\",\n"
@@ -1022,6 +1031,7 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.max_vram                        = max_vram.c_str();
     sd_ctx_params.disable_prefetch                = disable_prefetch;
     sd_ctx_params.disable_segmented_compute       = disable_segmented_compute;
+    sd_ctx_params.batched_cfg                     = batched_cfg;
     sd_ctx_params.eager_load                      = eager_load;
     sd_ctx_params.backend                         = effective_backend.c_str();
     sd_ctx_params.params_backend                  = effective_params_backend.c_str();

examples/common/common.h

@@ -157,6 +157,7 @@ struct SDContextParams {
     bool disable_prefetch          = false;
     bool disable_segmented_compute = false;
     bool eager_load                = false;
+    bool batched_cfg               = true;
     std::string backend;
     std::string params_backend;
     std::string split_mode;

include/stable-diffusion.h

@@ -245,6 +245,7 @@ typedef struct {
     const char* rpc_servers;
     const char* model_args;
     bool disable_segmented_compute;  // Force monolithic graph execution even when automatic graph cutting would fit memory better
+    bool batched_cfg;                // Run the conditional and unconditional CFG branches in one batched UNet forward when supported
     float linear_scale;              // Override linear input scaling; 0 keeps the model default
     float attn_scale;                // Override flash-attention K/V scaling; 0 keeps the model default
     const char* tokenizer;           // tokenizer.json path or main=FILE,clip-l=FILE,clip-g=FILE assignments; required for PiD and Lens

src/conditioning/conditioner.hpp

@@ -2012,7 +2012,8 @@ struct LLMEmbedder : public Conditioner {
                    sd_version_is_sefi_image(version) ||
                    sd_version_is_krea2(version) ||
                    sd_version_is_minimax_h3(version) ||
-                   sd_version_is_mage_flow(version)) {
+                   sd_version_is_mage_flow(version) ||
+                   version == VERSION_IRIS) {
             arch = LLM::LLMArch::QWEN3_VL;
         } else if (sd_version_is_z_image(version) || sd_version_is_z_image_l2p(version) || version == VERSION_OVIS_IMAGE || version == VERSION_FLUX2_KLEIN) {
             arch = LLM::LLMArch::QWEN3;
@@ -2138,9 +2139,10 @@ struct LLMEmbedder : public Conditioner {
 
     std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
                                                                                   const std::pair<int, int>& attn_range,
-                                                                                  size_t min_length = 0,
-                                                                                  size_t max_length = 100000000,
-                                                                                  bool spell_quotes = false) {
+                                                                                  size_t min_length         = 0,
+                                                                                  size_t max_length         = 100000000,
+                                                                                  bool spell_quotes         = false,
+                                                                                  const std::string& suffix = "") {
         std::vector<std::pair<std::string, float>> parsed_attention;
         if (attn_range.first >= 0 && attn_range.second > 0) {
             if (attn_range.first > 0) {
@@ -2185,6 +2187,20 @@ struct LLMEmbedder : public Conditioner {
             weights.insert(weights.end(), curr_tokens.size(), curr_weight);
         }
 
+        if (!suffix.empty()) {
+            std::vector<int> suffix_tokens;
+            if (!tokenizer->encode(suffix, suffix_tokens) || suffix_tokens.size() >= max_length) {
+                return {};
+            }
+            // Reserve the assistant-turn marker before truncating the caption.
+            if (tokens.size() > max_length - suffix_tokens.size()) {
+                tokens.resize(max_length - suffix_tokens.size());
+                weights.resize(tokens.size());
+            }
+            tokens.insert(tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+            weights.insert(weights.end(), suffix_tokens.size(), 1.f);
+        }
+
         std::vector<float> mask;
         tokenizer->pad_tokens(tokens, &weights, &mask, min_length, max_length);
 
@@ -2207,8 +2223,10 @@ struct LLMEmbedder : public Conditioner {
                                     bool spell_quotes                                       = false,
                                     int max_length                                          = 100000000,
                                     const LLM::DeepStackImageEmbeds& deepstack_image_embeds = {},
-                                    const std::vector<LLM::ImageGrid>& image_grids          = {}) {
-        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range, min_length, max_length, spell_quotes);
+                                    const std::vector<LLM::ImageGrid>& image_grids          = {},
+                                    const std::string& prompt_suffix                        = "",
+                                    sd::Tensor<float>* output_mask                          = nullptr) {
+        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range, min_length, max_length, spell_quotes, prompt_suffix);
         auto& tokens             = std::get<0>(tokens_weights_mask);
         auto& weights            = std::get<1>(tokens_weights_mask);
         auto& mask               = std::get<2>(tokens_weights_mask);
@@ -2271,6 +2289,12 @@ struct LLMEmbedder : public Conditioner {
                                                 1);
         }
 
+        if (output_mask != nullptr) {
+            *output_mask = sd::Tensor<float>::zeros({new_hidden_states.shape()[1]});
+            for (size_t i = prompt_template_encode_start_idx; i < mask.size(); ++i) {
+                output_mask->data()[i - prompt_template_encode_start_idx] = mask[i];
+            }
+        }
         return new_hidden_states;
     }
 
@@ -2337,6 +2361,8 @@ struct LLMEmbedder : public Conditioner {
         int hidden_states_min_length         = 0;  // zero pad hidden_states
         bool spell_quotes                    = false;
         std::set<int> out_layers;
+        std::string prompt_suffix;
+        sd::Tensor<float> text_mask;
 
         int64_t t0                     = ggml_time_ms();
         RefImageResizeMode resize_mode = conditioner_params.ref_image_params.vlm_resize_mode;
@@ -3086,6 +3112,23 @@ struct LLMEmbedder : public Conditioner {
             SDCondition result;
             result.c_crossattn = std::move(hidden_states);
             return result;
+        } else if (version == VERSION_IRIS) {
+            prompt =
+                "<|im_start|>system\n"
+                "Describe the image by detailing the color, shape, size, texture, quantity, text, spatial "
+                "relationships of the objects and background:<|im_end|>\n<|im_start|>user\n";
+            auto prefix_tokens = std::get<0>(tokenize(prompt, {0, 0}));
+            if (prefix_tokens.empty()) {
+                return {};
+            }
+            prompt_template_encode_start_idx = static_cast<int>(prefix_tokens.size());
+            hidden_states_min_length         = 300;
+            max_length                       = prompt_template_encode_start_idx + hidden_states_min_length;
+            out_layers                       = {2, 5, 8, 11, 14, 17, 20, 23, 26, 29, 32, 35};
+            prompt_attn_range.first          = static_cast<int>(prompt.size());
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+            prompt_suffix            = "<|im_end|>\n<|im_start|>assistant\n";
         } else {
             GGML_ABORT("unknown version %d", version);
         }
@@ -3101,7 +3144,9 @@ struct LLMEmbedder : public Conditioner {
                                            spell_quotes,
                                            max_length,
                                            deepstack_image_embeds,
-                                           image_grids);
+                                           image_grids,
+                                           prompt_suffix,
+                                           version == VERSION_IRIS ? &text_mask : nullptr);
         if (hidden_states.empty()) {
             return {};
         }
@@ -3166,6 +3211,7 @@ struct LLMEmbedder : public Conditioner {
         SDCondition result;
         result.c_crossattn        = std::move(hidden_states);
         result.extra_c_crossattns = std::move(extra_hidden_states_vec);
+        result.c_vector           = std::move(text_mask);
         if (version == VERSION_QWEN_IMAGE_2_1) {
             auto slots = sd::Tensor<int32_t>::zeros({result.c_crossattn.shape()[1]});
             for (size_t i = 0; i < image_embeds.size(); ++i) {

src/core/backend_fit.cpp

@@ -497,11 +497,7 @@ namespace sd::backend_fit {
                 return false;
             }
             auto smaller_tile = [&](int size) {
-                int next_size = size / 2;
-                if (!tiling_params.enabled) {
-                    next_size = std::min(next_size, 256 / scale_factor);
-                }
-                return std::min(size, std::max(4, next_size));
+                return std::min(size, std::max(4, size / 2));
             };
             const int tile_size_w = smaller_tile(latent_tile_size_w);
             const int tile_size_h = smaller_tile(latent_tile_size_h);

src/model.h

@@ -65,6 +65,7 @@ enum SDVersion {
     VERSION_PIXART,
     VERSION_MING_IMAGE,
     VERSION_Z_IMAGE_L2P,
+    VERSION_IRIS,
     VERSION_COUNT,
 };
 
@@ -334,7 +335,8 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_krea2(version) ||
         sd_version_is_mage_flow(version) ||
         sd_version_is_sensenova_u1(version) ||
-        sd_version_is_pixart(version)) {
+        sd_version_is_pixart(version) ||
+        version == VERSION_IRIS) {
         return true;
     }
     return false;

src/model/adapter/lora_ops.cpp

@@ -11,7 +11,9 @@ ggml_tensor* ggml_ext_merge_lora(ggml_context* ctx,
                                  ggml_tensor* lora_mid) {
     ggml_tensor* updown;
     // flat lora tensors to multiply it
-    int64_t lora_up_rows  = lora_up->ne[ggml_n_dims(lora_up) - 1];
+    // ggml_n_dims drops the output dimension when a linear layer has one output.
+    int lora_up_n_dims    = ggml_n_dims(lora_up);
+    int64_t lora_up_rows  = lora_up_n_dims == 1 ? 1 : lora_up->ne[lora_up_n_dims - 1];
     lora_up               = ggml_reshape_2d(ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
     auto lora_down_n_dims = ggml_n_dims(lora_down);
     // assume n_dims should always be a multiple of 2 (otherwise rank 1 doesn't work)

src/model/common/rope.hpp

@@ -1067,10 +1067,22 @@ namespace Rope {
         return result;
     }
 
-    __STATIC_INLINE__ ggml_tensor* apply_rope(ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* apply_rope(GGMLRunnerContext* runner_ctx,
                                               ggml_tensor* x,
                                               ggml_tensor* pe,
                                               bool rope_interleaved = true) {
+        auto ctx = runner_ctx->ggml_ctx;
+#ifndef SD_USE_UPSTREAM_GGML
+        auto backend = runner_ctx->backend;
+        if (backend != nullptr && x->type == GGML_TYPE_F32 && pe->type == GGML_TYPE_F32 &&
+            x->ne[0] % 2 == 0 && pe->ne[0] == 2 && pe->ne[1] == 2 &&
+            pe->ne[2] == x->ne[0] / 2 && pe->ne[3] == x->ne[2]) {
+            auto out = ggml_rope_apply(ctx, x, pe, rope_interleaved);
+            if (ggml_backend_supports_op(backend, out)) {
+                return out;
+            }
+        }
+#endif
         // x: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2], [[cos, -sin], [sin, cos]]
         int64_t d_head = x->ne[0];
@@ -1126,8 +1138,8 @@ namespace Rope {
         // return: [N, L, n_head*d_head]
         int64_t n_head = q->ne[1];
 
-        q = apply_rope(ctx->ggml_ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
-        k = apply_rope(ctx->ggml_ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]
+        q = apply_rope(ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
+        k = apply_rope(ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]
 
         auto x = ggml_ext_attention_ext(ctx, q, k, v, n_head, mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
         return x;

src/model/diffusion/anima.hpp

@@ -235,8 +235,8 @@ namespace Anima {
                 if (pe_k == nullptr) {
                     pe_k = pe_q;
                 }
-                auto q_rope = Rope::apply_rope(ctx->ggml_ctx, q4, pe_q, false);
-                auto k_rope = Rope::apply_rope(ctx->ggml_ctx, k4, pe_k, false);
+                auto q_rope = Rope::apply_rope(ctx, q4, pe_q, false);
+                auto k_rope = Rope::apply_rope(ctx, k4, pe_k, false);
                 attn_out    = ggml_ext_attention_ext(ctx,
                                                      q_rope,
                                                      k_rope,

src/model/diffusion/iris.hpp

@@ -0,0 +1,450 @@
+#ifndef __SD_MODEL_DIFFUSION_IRIS_HPP__
+#define __SD_MODEL_DIFFUSION_IRIS_HPP__
+
+#include <array>
+#include <stdexcept>
+#include "model/diffusion/pid.hpp"
+
+namespace Iris {
+    struct IrisConfig {
+        int64_t hidden_size    = 2560;
+        int64_t depth          = 24;
+        int64_t dual_depth     = 8;
+        int64_t head_dim       = 128;
+        int64_t num_heads      = 20;
+        int64_t kv_heads       = 5;
+        int64_t mlp_dim        = 6826;
+        int64_t patch_size     = 16;
+        int64_t text_dim       = 2560;
+        int64_t text_len       = 300;
+        int64_t text_layers    = 12;
+        int64_t layer_heads    = 32;
+        int64_t layer_mlp_dim  = 3328;
+        int64_t pixel_dim      = 16;
+        int64_t pixel_attn_dim = 1280;
+        int64_t pixel_heads    = 10;
+        int64_t pixel_depth    = 4;
+
+        static IrisConfig detect_from_weights(const String2TensorStorage& weights, const std::string& prefix) {
+            IrisConfig config;
+            auto find = [&](const std::string& name) -> const TensorStorage& {
+                auto it = weights.find(prefix + "." + name);
+                if (it == weights.end()) {
+                    throw std::runtime_error("Iris-3B: missing weight " + prefix + "." + name);
+                }
+                return it->second;
+            };
+            config.hidden_size    = find("s_embedder.proj.weight").ne[1];
+            config.patch_size     = static_cast<int64_t>(std::sqrt(find("s_embedder.proj.weight").ne[0] / 3));
+            config.head_dim       = find("blocks.0.attn.q_norm_x.weight").ne[0];
+            config.num_heads      = config.hidden_size / config.head_dim;
+            config.kv_heads       = find("blocks.0.attn.k_proj_x.weight").ne[1] / config.head_dim;
+            config.mlp_dim        = find("blocks.0.mlp_x.w1.weight").ne[1];
+            config.text_dim       = find("y_embedder.refiner.proj.weight").ne[0];
+            config.text_len       = find("y_pos_embedding").ne[1];
+            config.text_layers    = find("y_embedder.layer_pool.weight").ne[0];
+            config.layer_mlp_dim  = find("y_embedder.layer_blocks.0.mlp.0.weight").ne[1];
+            config.pixel_dim      = find("pixel_embedder.proj.weight").ne[1];
+            config.pixel_attn_dim = find("pixel_blocks.0.compress.weight").ne[1];
+            config.pixel_heads    = config.pixel_attn_dim / find("pixel_blocks.0.attn.q_norm.weight").ne[0];
+            config.depth = config.dual_depth = config.pixel_depth = 0;
+            for (const auto& [name, tensor] : weights) {
+                if (!starts_with(name, prefix + ".")) {
+                    continue;
+                }
+                auto parts = split_string(name.substr(prefix.size() + 1), '.');
+                if (parts.size() > 2 && parts[0] == "blocks") {
+                    int64_t index = std::stoi(parts[1]);
+                    config.depth  = std::max(config.depth, index + 1);
+                    if (parts[2] == "adaln_img") {
+                        config.dual_depth = std::max(config.dual_depth, index + 1);
+                    }
+                } else if (parts.size() > 2 && parts[0] == "pixel_blocks") {
+                    config.pixel_depth = std::max(config.pixel_depth, int64_t(std::stoi(parts[1]) + 1));
+                }
+            }
+            if (config.text_layers != 12 || config.text_dim != 2560 || config.text_len != 300 ||
+                find("modulation_cores.adaln_img.weight").ne[1] != 6 * config.hidden_size ||
+                find("pixel_blocks.0.adaln.weight").ne[1] != 4 * config.pixel_dim * config.patch_size * config.patch_size) {
+                throw std::runtime_error("Iris-3B: unsupported checkpoint configuration");
+            }
+            LOG_VERBOSE("iris: hidden_size = %" PRId64 ", depth = %" PRId64 ", dual_depth = %" PRId64 ", heads = %" PRId64 "/%" PRId64 ", pixel_depth = %" PRId64,
+                        config.hidden_size, config.depth, config.dual_depth, config.num_heads, config.kv_heads, config.pixel_depth);
+            return config;
+        }
+    };
+
+    struct SelfAttention : public GGMLBlock {
+        int64_t dim;
+        int64_t heads;
+        bool qk_norm;
+
+        SelfAttention(int64_t dim, int64_t heads, bool qk_norm = true)
+            : dim(dim), heads(heads), qk_norm(qk_norm) {
+            blocks["qkv"]  = std::make_shared<Linear>(dim, 3 * dim, false);
+            blocks["proj"] = std::make_shared<Linear>(dim, dim, true);
+            if (qk_norm) {
+                blocks["q_norm"] = std::make_shared<RMSNorm>(dim / heads, 1e-6f);
+                blocks["k_norm"] = std::make_shared<RMSNorm>(dim / heads, 1e-6f);
+            }
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* pe = nullptr, ggml_tensor* mask = nullptr) {
+            auto c   = ctx->ggml_ctx;
+            auto qkv = split_qkv(c, std::dynamic_pointer_cast<Linear>(blocks["qkv"])->forward(ctx, x));
+            auto q   = ggml_reshape_4d(c, qkv[0], dim / heads, heads, x->ne[1], x->ne[2]);
+            auto k   = ggml_reshape_4d(c, qkv[1], dim / heads, heads, x->ne[1], x->ne[2]);
+            auto v   = ggml_reshape_4d(c, qkv[2], dim / heads, heads, x->ne[1], x->ne[2]);
+            if (qk_norm) {
+                q = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"])->forward(ctx, q);
+                k = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"])->forward(ctx, k);
+            }
+            if (pe) {
+                x = Rope::attention(ctx, q, k, v, pe, mask);
+            } else {
+                q = ggml_reshape_3d(c, q, dim, x->ne[1], x->ne[2]);
+                k = ggml_reshape_3d(c, k, dim, x->ne[1], x->ne[2]);
+                v = ggml_reshape_3d(c, v, dim, x->ne[1], x->ne[2]);
+                x = ggml_ext_attention_ext(ctx, q, k, v, heads, mask, false, ctx->flash_attn_enabled);
+            }
+            return std::dynamic_pointer_cast<Linear>(blocks["proj"])->forward(ctx, x);
+        }
+    };
+
+    struct TextBlock : public GGMLBlock {
+        bool layerwise;
+
+        TextBlock(int64_t dim, int64_t heads, int64_t mlp_dim, bool layerwise)
+            : layerwise(layerwise) {
+            blocks["norm1"] = std::make_shared<RMSNorm>(dim, 1e-6f);
+            blocks["norm2"] = std::make_shared<RMSNorm>(dim, 1e-6f);
+            blocks["attn"]  = std::make_shared<SelfAttention>(dim, heads, !layerwise);
+            if (layerwise) {
+                blocks["mlp.0"] = std::make_shared<Linear>(dim, mlp_dim, true);
+                blocks["mlp.2"] = std::make_shared<Linear>(mlp_dim, dim, true);
+            } else {
+                blocks["mlp"] = std::make_shared<Pid::FeedForward>(dim, mlp_dim);
+            }
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mask = nullptr) {
+            auto h = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"])->forward(ctx, x);
+            h      = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"])->forward(ctx, h, nullptr, mask);
+            x      = ggml_add(ctx->ggml_ctx, x, h);
+            h      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"])->forward(ctx, x);
+            if (layerwise) {
+                h = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"])->forward(ctx, h);
+                h = ggml_silu(ctx->ggml_ctx, h);
+                h = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"])->forward(ctx, h);
+            } else {
+                h = std::dynamic_pointer_cast<Pid::FeedForward>(blocks["mlp"])->forward(ctx, h);
+            }
+            return ggml_add(ctx->ggml_ctx, x, h);
+        }
+    };
+
+    struct TextEmbedder : public GGMLBlock {
+        IrisConfig config;
+
+        TextEmbedder(const IrisConfig& config)
+            : config(config) {
+            for (int i = 0; i < 2; ++i) {
+                blocks["layer_blocks." + std::to_string(i)]   = std::make_shared<TextBlock>(config.text_dim, config.layer_heads, config.layer_mlp_dim, true);
+                blocks["refiner.blocks." + std::to_string(i)] = std::make_shared<TextBlock>(config.hidden_size, config.num_heads, config.mlp_dim, false);
+            }
+            blocks["layer_pool"]   = std::make_shared<Linear>(config.text_layers, 1, true);
+            blocks["refiner.proj"] = std::make_shared<Linear>(config.text_dim, config.hidden_size, true);
+            blocks["refiner.norm"] = std::make_shared<RMSNorm>(config.hidden_size, 1e-6f);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mask) {
+            auto c         = ctx->ggml_ctx;
+            int64_t tokens = x->ne[1];
+            int64_t batch  = x->ne[2];
+            // The encoder concatenates selected layers within each token's feature dimension.
+            x = ggml_reshape_3d(c, x, config.text_dim, config.text_layers, tokens * batch);
+            for (int i = 0; i < 2; ++i) {
+                x = std::dynamic_pointer_cast<TextBlock>(blocks["layer_blocks." + std::to_string(i)])->forward(ctx, x);
+            }
+            x = ggml_cont(c, ggml_permute(c, x, 1, 0, 2, 3));
+            x = std::dynamic_pointer_cast<Linear>(blocks["layer_pool"])->forward(ctx, x);
+            x = ggml_reshape_3d(c, x, config.text_dim, tokens, batch);
+            x = std::dynamic_pointer_cast<Linear>(blocks["refiner.proj"])->forward(ctx, x);
+            for (int i = 0; i < 2; ++i) {
+                x = std::dynamic_pointer_cast<TextBlock>(blocks["refiner.blocks." + std::to_string(i)])->forward(ctx, x, mask);
+            }
+            return std::dynamic_pointer_cast<RMSNorm>(blocks["refiner.norm"])->forward(ctx, x);
+        }
+    };
+
+    struct TrunkBlock : public GGMLBlock {
+        IrisConfig config;
+        bool dual;
+
+        TrunkBlock(const IrisConfig& config, bool dual)
+            : config(config), dual(dual) {
+            int64_t d = config.hidden_size;
+            for (const std::string stream : dual ? std::vector<std::string>{"x", "y"} : std::vector<std::string>{""}) {
+                std::string suffix                                = stream.empty() ? "" : "_" + stream;
+                std::string attn                                  = dual ? "attn." : "";
+                blocks[attn + "q_proj" + suffix]                  = std::make_shared<Linear>(d, d, false);
+                blocks[attn + "k_proj" + suffix]                  = std::make_shared<Linear>(d, config.kv_heads * config.head_dim, false);
+                blocks[attn + "v_proj" + suffix]                  = std::make_shared<Linear>(d, config.kv_heads * config.head_dim, false);
+                blocks[attn + "q_norm" + suffix]                  = std::make_shared<RMSNorm>(config.head_dim, 1e-6f);
+                blocks[attn + "k_norm" + suffix]                  = std::make_shared<RMSNorm>(config.head_dim, 1e-6f);
+                blocks[dual ? "attn.proj" + suffix : "attn_proj"] = std::make_shared<Linear>(d, d, true);
+                blocks["attn_gate" + suffix]                      = std::make_shared<Linear>(d, d, false);
+                blocks["norm" + suffix + "1"]                     = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["norm" + suffix + "2"]                     = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["attn_post_norm" + suffix]                 = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["mlp_post_norm" + suffix]                  = std::make_shared<RMSNorm>(d, 1e-6f);
+                blocks["mlp" + suffix]                            = std::make_shared<Pid::FeedForward>(d, config.mlp_dim);
+            }
+        }
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& = {}, const std::string = "") override {
+            if (dual) {
+                params["adaln_img.bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 6 * config.hidden_size);
+                params["adaln_txt.bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 6 * config.hidden_size);
+            } else {
+                params["adaln.bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 6 * config.hidden_size);
+            }
+        }
+
+        ggml_tensor* norm(GGMLRunnerContext* ctx, const std::string& name, ggml_tensor* x) {
+            return std::dynamic_pointer_cast<RMSNorm>(blocks[name])->forward(ctx, x);
+        }
+
+        ggml_tensor* linear(GGMLRunnerContext* ctx, const std::string& name, ggml_tensor* x) {
+            return std::dynamic_pointer_cast<Linear>(blocks[name])->forward(ctx, x);
+        }
+
+        std::array<ggml_tensor*, 3> project(GGMLRunnerContext* ctx, ggml_tensor* h, const std::string& suffix) {
+            auto c        = ctx->ggml_ctx;
+            std::string a = dual ? "attn." : "";
+            auto q        = linear(ctx, a + "q_proj" + suffix, h);
+            auto k        = linear(ctx, a + "k_proj" + suffix, h);
+            auto v        = linear(ctx, a + "v_proj" + suffix, h);
+            q             = ggml_reshape_4d(c, q, config.head_dim, config.num_heads, h->ne[1], h->ne[2]);
+            k             = ggml_reshape_4d(c, k, config.head_dim, config.kv_heads, h->ne[1], h->ne[2]);
+            v             = ggml_reshape_4d(c, v, config.head_dim, config.kv_heads, h->ne[1], h->ne[2]);
+            return {norm(ctx, a + "q_norm" + suffix, q), norm(ctx, a + "k_norm" + suffix, k), v};
+        }
+
+        ggml_tensor* finish(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* h, ggml_tensor* out, const std::vector<ggml_tensor*>& mods, const std::string& suffix) {
+            auto c = ctx->ggml_ctx;
+            out    = ggml_mul(c, out, ggml_sigmoid(c, linear(ctx, "attn_gate" + suffix, h)));
+            out    = linear(ctx, dual ? "attn.proj" + suffix : "attn_proj", out);
+            out    = norm(ctx, "attn_post_norm" + suffix, out);
+            x      = ggml_add(c, x, ggml_mul(c, out, mods[2]));
+            h      = Pid::apply_adaln(c, norm(ctx, "norm" + suffix + "2", x), mods[3], mods[4]);
+            h      = std::dynamic_pointer_cast<Pid::FeedForward>(blocks["mlp" + suffix])->forward(ctx, h);
+            h      = norm(ctx, "mlp_post_norm" + suffix, h);
+            return ggml_add(c, x, ggml_mul(c, h, mods[5]));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* y, ggml_tensor* img_mod, ggml_tensor* txt_mod, ggml_tensor* pe) {
+            auto c     = ctx->ggml_ctx;
+            int64_t nt = y->ne[1];
+            auto mx    = ggml_ext_chunk(c, ggml_add(c, img_mod, params[dual ? "adaln_img.bias" : "adaln.bias"]), 6, 0);
+            if (!dual) {
+                auto tokens = ggml_concat(c, y, x, 1);
+                auto h      = Pid::apply_adaln(c, norm(ctx, "norm1", tokens), mx[0], mx[1]);
+                auto qkv    = project(ctx, h, "");
+                auto out    = Rope::attention(ctx, qkv[0], qkv[1], qkv[2], pe, nullptr);
+                tokens      = finish(ctx, tokens, h, out, mx, "");
+                return {ggml_ext_slice(c, tokens, 1, nt, tokens->ne[1]), ggml_ext_slice(c, tokens, 1, 0, nt)};
+            }
+            auto my  = ggml_ext_chunk(c, ggml_add(c, txt_mod, params["adaln_txt.bias"]), 6, 0);
+            auto hx  = Pid::apply_adaln(c, norm(ctx, "norm_x1", x), mx[0], mx[1]);
+            auto hy  = Pid::apply_adaln(c, norm(ctx, "norm_y1", y), my[0], my[1]);
+            auto qx  = project(ctx, hx, "_x");
+            auto qy  = project(ctx, hy, "_y");
+            auto out = Rope::attention(ctx, ggml_concat(c, qy[0], qx[0], 2), ggml_concat(c, qy[1], qx[1], 2),
+                                       ggml_concat(c, qy[2], qx[2], 2), pe, nullptr);
+            x        = finish(ctx, x, hx, ggml_ext_slice(c, out, 1, nt, out->ne[1]), mx, "_x");
+            y        = finish(ctx, y, hy, ggml_ext_slice(c, out, 1, 0, nt), my, "_y");
+            return {x, y};
+        }
+    };
+
+    struct PixelBlock : public GGMLBlock {
+        IrisConfig config;
+
+        PixelBlock(const IrisConfig& config)
+            : config(config) {
+            auto d             = config.pixel_dim;
+            auto flat          = d * config.patch_size * config.patch_size;
+            blocks["norm1"]    = std::make_shared<RMSNorm>(d, 1e-6f);
+            blocks["norm2"]    = std::make_shared<RMSNorm>(d, 1e-6f);
+            blocks["compress"] = std::make_shared<Linear>(flat, config.pixel_attn_dim, true);
+            blocks["expand"]   = std::make_shared<Linear>(config.pixel_attn_dim, flat, true);
+            blocks["adaln"]    = std::make_shared<Linear>(config.hidden_size, 4 * flat, true);
+            blocks["attn"]     = std::make_shared<SelfAttention>(config.pixel_attn_dim, config.pixel_heads);
+            blocks["mlp.fc1"]  = std::make_shared<Linear>(d, d * 4, true);
+            blocks["mlp.fc2"]  = std::make_shared<Linear>(d * 4, d, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* cond, ggml_tensor* pe) {
+            auto c       = ctx->ggml_ctx;
+            auto d       = config.pixel_dim;
+            auto p2      = config.patch_size * config.patch_size;
+            auto patches = cond->ne[1];
+            auto batch   = cond->ne[2];
+            auto mods    = std::dynamic_pointer_cast<Linear>(blocks["adaln"])->forward(ctx, cond);
+            mods         = ggml_reshape_3d(c, mods, 4 * d, p2, patches * batch);
+            auto m       = ggml_ext_chunk(c, mods, 4, 0);
+            auto h       = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"])->forward(ctx, x);
+            h            = ggml_reshape_3d(c, h, d * p2, patches, batch);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["compress"])->forward(ctx, h);
+            h            = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"])->forward(ctx, h, pe);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["expand"])->forward(ctx, h);
+            h            = ggml_reshape_3d(c, h, d, p2, patches * batch);
+            x            = ggml_add(c, x, Pid::apply_adaln(c, h, m[1], m[0]));
+            h            = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"])->forward(ctx, x);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc1"])->forward(ctx, h);
+            h            = ggml_gelu_erf(c, h);
+            h            = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc2"])->forward(ctx, h);
+            return ggml_add(c, x, Pid::apply_adaln(c, h, m[3], m[2]));
+        }
+    };
+
+    struct IrisModel : public GGMLBlock {
+        IrisConfig config;
+
+        IrisModel(const IrisConfig& config)
+            : config(config) {
+            blocks["s_embedder"]                 = std::make_shared<Pid::PatchTokenEmbedder>(3 * config.patch_size * config.patch_size, config.hidden_size);
+            blocks["t_embedder"]                 = std::make_shared<Pid::PixelDiTTimestepEmbedder>(config.hidden_size);
+            blocks["y_embedder"]                 = std::make_shared<TextEmbedder>(config);
+            blocks["modulation_cores.adaln_img"] = std::make_shared<Linear>(config.hidden_size, 6 * config.hidden_size, true);
+            blocks["modulation_cores.adaln_txt"] = std::make_shared<Linear>(config.hidden_size, 6 * config.hidden_size, true);
+            for (int64_t i = 0; i < config.depth; ++i) {
+                blocks["blocks." + std::to_string(i)] = std::make_shared<TrunkBlock>(config, i < config.dual_depth);
+            }
+            blocks["pixel_embedder"] = std::make_shared<Pid::PixelTokenEmbedder>(3, config.pixel_dim);
+            for (int64_t i = 0; i < config.pixel_depth; ++i) {
+                blocks["pixel_blocks." + std::to_string(i)] = std::make_shared<PixelBlock>(config);
+            }
+            blocks["final_layer"] = std::make_shared<Pid::FinalLayer>(config.pixel_dim, 3);
+        }
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& = {}, const std::string = "") override {
+            params["y_pos_embedding"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, config.hidden_size, config.text_len, 1);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* image, ggml_tensor* t, ggml_tensor* text, ggml_tensor* mask, ggml_tensor* joint_pe, ggml_tensor* pixel_pe, ggml_tensor* pixel_pos) {
+            auto c    = ctx->ggml_ctx;
+            int p     = static_cast<int>(config.patch_size);
+            auto x    = DiT::patchify(c, image, p, p, true);
+            x         = std::dynamic_pointer_cast<Pid::PatchTokenEmbedder>(blocks["s_embedder"])->forward(ctx, x);
+            auto te   = std::dynamic_pointer_cast<Pid::PixelDiTTimestepEmbedder>(blocks["t_embedder"])->forward(ctx, t);
+            te        = ggml_reshape_3d(c, te, config.hidden_size, 1, image->ne[3]);
+            auto cond = ggml_silu(c, te);
+            auto im   = std::dynamic_pointer_cast<Linear>(blocks["modulation_cores.adaln_img"])->forward(ctx, cond);
+            auto tm   = std::dynamic_pointer_cast<Linear>(blocks["modulation_cores.adaln_txt"])->forward(ctx, cond);
+            auto y    = std::dynamic_pointer_cast<TextEmbedder>(blocks["y_embedder"])->forward(ctx, text, mask);
+            y         = ggml_add(c, y, params["y_pos_embedding"]);
+            for (int64_t i = 0; i < config.depth; ++i) {
+                auto block     = std::dynamic_pointer_cast<TrunkBlock>(blocks["blocks." + std::to_string(i)]);
+                std::tie(x, y) = block->forward(ctx, x, y, im, tm, joint_pe);
+                sd::ggml_graph_cut::mark_graph_cut(x, "iris.blocks." + std::to_string(i), "x");
+                sd::ggml_graph_cut::mark_graph_cut(y, "iris.blocks." + std::to_string(i), "y");
+            }
+            cond = ggml_silu(c, ggml_add(c, x, te));
+            x    = std::dynamic_pointer_cast<Pid::PixelTokenEmbedder>(blocks["pixel_embedder"])->forward(ctx, image, p, pixel_pos);
+            for (int64_t i = 0; i < config.pixel_depth; ++i) {
+                x = std::dynamic_pointer_cast<PixelBlock>(blocks["pixel_blocks." + std::to_string(i)])->forward(ctx, x, cond, pixel_pe);
+                sd::ggml_graph_cut::mark_graph_cut(x, "iris.pixel_blocks." + std::to_string(i), "x");
+            }
+            x = std::dynamic_pointer_cast<Pid::FinalLayer>(blocks["final_layer"])->forward(ctx, x);
+            x = ggml_reshape_3d(c, x, 3 * p * p, cond->ne[1], image->ne[3]);
+            return DiT::unpatchify(c, x, image->ne[1] / p, image->ne[0] / p, p, p, false);
+        }
+    };
+
+    inline std::vector<float> image_rope(int64_t height, int64_t width, int64_t dim) {
+        std::vector<float> pe;
+        pe.reserve(height * width * dim * 2);
+        float step = 16.f / static_cast<float>(std::max<int64_t>(std::max(height, width) - 1, 1));
+        for (int64_t y = 0; y < height; ++y) {
+            for (int64_t x = 0; x < width; ++x) {
+                for (int64_t i = 0; i < dim / 4; ++i) {
+                    float frequency = std::pow(10000.f, -4.f * static_cast<float>(i) / static_cast<float>(dim));
+                    for (int64_t pos : {x, y}) {
+                        float angle = static_cast<float>(pos) * step * frequency;
+                        pe.insert(pe.end(), {std::cos(angle), -std::sin(angle), std::sin(angle), std::cos(angle)});
+                    }
+                }
+            }
+        }
+        return pe;
+    }
+
+    struct IrisRunner : public DiffusionModelRunner {
+        IrisConfig config;
+        IrisModel model;
+        std::vector<float> joint_pe_data;
+        std::vector<float> pixel_pe_data;
+        std::vector<float> pixel_pos_data;
+        std::vector<float> text_mask_data;
+
+        IrisRunner(ggml_backend_t backend, const String2TensorStorage& weights, const std::string& prefix, std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(IrisConfig::detect_from_weights(weights, prefix)),
+              model(config) {
+            model.init(params_ctx, weights, prefix);
+        }
+
+        std::string get_desc() override { return "Iris-3B"; }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        sd::Tensor<float> compute(int n_threads, const DiffusionParams& inputs) override {
+            if (!inputs.x || !inputs.timesteps || !inputs.context || !inputs.y || inputs.y->empty()) {
+                LOG_ERROR("Iris-3B requires text features and their padding mask");
+                return {};
+            }
+            auto get_graph = [&]() {
+                auto gf   = new_graph_custom(196608);
+                auto x    = make_input(*inputs.x);
+                auto t    = make_input(*inputs.timesteps);
+                auto text = make_input(*inputs.context);
+                int h     = static_cast<int>(x->ne[1]);
+                int w     = static_cast<int>(x->ne[0]);
+                int nt    = static_cast<int>(config.text_len);
+                GGML_ASSERT(h % config.patch_size == 0 && w % config.patch_size == 0);
+                GGML_ASSERT(text->ne[0] == config.text_dim * config.text_layers && text->ne[1] == nt);
+                GGML_ASSERT(inputs.y->numel() == nt && x->ne[3] == 1);
+                joint_pe_data = Pid::make_rope_1d(nt, static_cast<int>(config.head_dim), 10000.f);
+                auto img      = image_rope(h / config.patch_size, w / config.patch_size, config.head_dim);
+                joint_pe_data.insert(joint_pe_data.end(), img.begin(), img.end());
+                pixel_pe_data  = image_rope(h / config.patch_size, w / config.patch_size, config.pixel_attn_dim / config.pixel_heads);
+                pixel_pos_data = Pid::make_pixel_abs_pos(h, w, static_cast<int>(config.pixel_dim));
+                text_mask_data.resize(nt * nt);
+                for (int q = 0; q < nt; ++q) {
+                    for (int k = 0; k < nt; ++k) {
+                        text_mask_data[q * nt + k] = inputs.y->values()[k] != 0.f || q == k ? 0.f : -INFINITY;
+                    }
+                }
+                auto pe   = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, nt + img.size() / (2 * config.head_dim));
+                auto pp   = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.pixel_attn_dim / config.pixel_heads / 2, (h / config.patch_size) * (w / config.patch_size));
+                auto pos  = ggml_new_tensor_3d(compute_ctx, GGML_TYPE_F32, config.pixel_dim, h * w, 1);
+                auto mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, nt, nt);
+                set_backend_tensor_data(pe, joint_pe_data.data());
+                set_backend_tensor_data(pp, pixel_pe_data.data());
+                set_backend_tensor_data(pos, pixel_pos_data.data());
+                set_backend_tensor_data(mask, text_mask_data.data());
+                auto ctx = get_context();
+                auto out = model.forward(&ctx, x, t, text, mask, pe, pp, pos);
+                ggml_build_forward_expand(gf, out);
+                return gf;
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, false), inputs.x->dim());
+        }
+    };
+}  // namespace Iris
+
+#endif  // __SD_MODEL_DIFFUSION_IRIS_HPP__

src/model/diffusion/ltxv.hpp

@@ -548,21 +548,22 @@ namespace LTXV {
         return build_rope_matrix_from_frequencies(freqs, dim);
     }
 
-    __STATIC_INLINE__ ggml_tensor* apply_hidden_rope(ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* apply_hidden_rope(GGMLRunnerContext* runner_ctx,
                                                      ggml_tensor* x,
                                                      ggml_tensor* pe,
                                                      int64_t heads,
                                                      int64_t dim_head,
                                                      bool rope_interleaved) {
+        auto ctx = runner_ctx->ggml_ctx;
         GGML_ASSERT(x->ne[0] == heads * dim_head);
         auto x4 = ggml_reshape_4d(ctx, x, dim_head, heads, x->ne[1], x->ne[2]);
         if (pe != nullptr && pe->ne[3] == x->ne[1] * heads) {
             auto x_flat   = ggml_reshape_4d(ctx, x4, dim_head, 1, x->ne[1] * heads, x->ne[2]);
-            auto out_flat = Rope::apply_rope(ctx, x_flat, pe, rope_interleaved);
+            auto out_flat = Rope::apply_rope(runner_ctx, x_flat, pe, rope_interleaved);
             auto out4     = ggml_reshape_4d(ctx, out_flat, dim_head, heads, x->ne[1], x->ne[2]);
             return ggml_reshape_3d(ctx, out4, heads * dim_head, x->ne[1], x->ne[2]);
         }
-        return Rope::apply_rope(ctx, x4, pe, rope_interleaved);
+        return Rope::apply_rope(runner_ctx, x4, pe, rope_interleaved);
     }
 
     struct TimestepEmbedder : public GGMLBlock {
@@ -705,8 +706,8 @@ namespace LTXV {
                 if (k_pe == nullptr) {
                     k_pe = pe;
                 }
-                q = apply_hidden_rope(ctx->ggml_ctx, q, pe, heads, dim_head, rope_interleaved);
-                k = apply_hidden_rope(ctx->ggml_ctx, k, k_pe, heads, dim_head, rope_interleaved);
+                q = apply_hidden_rope(ctx, q, pe, heads, dim_head, rope_interleaved);
+                k = apply_hidden_rope(ctx, k, k_pe, heads, dim_head, rope_interleaved);
             }
 
             auto out = ggml_ext_attention_ext(ctx,

src/model/diffusion/minimax_h3.hpp

@@ -160,12 +160,13 @@ namespace MiniMaxH3 {
         return ggml_reshape_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2] * x->ne[3]);
     }
 
-    static ggml_tensor* apply_partial_rope(ggml_context* ctx,
+    static ggml_tensor* apply_partial_rope(GGMLRunnerContext* runner_ctx,
                                            ggml_tensor* x,
                                            ggml_tensor* pe) {
+        auto ctx        = runner_ctx->ggml_ctx;
         int64_t rot_dim = pe->ne[2] * 2;
         GGML_ASSERT(rot_dim <= x->ne[0]);
-        auto rotated = Rope::apply_rope(ctx,
+        auto rotated = Rope::apply_rope(runner_ctx,
                                         ggml_ext_slice(ctx, x, 0, 0, rot_dim),
                                         pe,
                                         false);
@@ -209,8 +210,8 @@ namespace MiniMaxH3 {
             q                = q_norm->forward(ctx, q);
             k                = k_norm->forward(ctx, k);
             if (pe != nullptr) {
-                q = apply_partial_rope(ctx->ggml_ctx, q, pe);
-                k = apply_partial_rope(ctx->ggml_ctx, k, pe);
+                q = apply_partial_rope(ctx, q, pe);
+                k = apply_partial_rope(ctx, k, pe);
             } else {
                 q = attention_layout(ctx->ggml_ctx, q);
                 k = attention_layout(ctx->ggml_ctx, k);

src/model/diffusion/qwen_image_2_1.hpp

@@ -188,8 +188,8 @@ namespace Qwen {
             auto v = project("to_v");
             q      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"])->forward(ctx, q);
             k      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"])->forward(ctx, k);
-            q      = Rope::apply_rope(ctx->ggml_ctx, q, pe);
-            k      = Rope::apply_rope(ctx->ggml_ctx, k, pe);
+            q      = Rope::apply_rope(ctx, q, pe);
+            k      = Rope::apply_rope(ctx, k, pe);
             if (cache.mode == QwenImage21PrefixCache::Mode::STORE) {
                 // Preserve query-first attention evaluation while writing each layer's
                 // prefix before its full-sequence K/V can accumulate across layers.

src/model/diffusion/unet.hpp

@@ -587,7 +587,9 @@ public:
             label_emb      = ggml_silu_inplace(ctx->ggml_ctx, label_emb);
             label_emb      = label_embed_2->forward(ctx, label_emb);  // [N, time_embed_dim]
 
-            emb = ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
+            emb = label_emb->ne[1] > emb->ne[1]
+                      ? ggml_add(ctx->ggml_ctx, label_emb, emb)
+                      : ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
         }
         // sd::ggml_graph_cut::mark_graph_cut(emb, "unet.prelude", "emb");
 

src/model/upscaler/mosr.hpp

@@ -0,0 +1,337 @@
+#ifndef __SD_MODEL_UPSCALER_MOSR_HPP__
+#define __SD_MODEL_UPSCALER_MOSR_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "core/ggml_extend.h"
+#include "core/ggml_runner.h"
+#include "core/util.h"
+#include "model/common/ggml_block.hpp"
+
+/*
+    ===================================    MoSR  ===================================
+    References:
+    https://github.com/umzi2/MoSR
+    https://github.com/chaiNNer-org/spandrel/blob/main/libs/spandrel/spandrel/architectures/MoSR/__arch/mosr_arch.py
+
+    Only the PixelShuffle ("ps") upsampler is supported.
+*/
+
+struct MoSRConfig {
+    int scale       = 4;
+    int n_block     = 24;
+    int in_ch       = 3;
+    int dim         = 64;
+    int hidden      = 96;
+    int conv_ch     = 64;
+    int kernel_size = 7;
+
+    bool is_valid() const {
+        // The upscaler feeds RGB tiles, so only 3-channel models are usable.
+        return in_ch == 3 && scale > 0 && n_block > 0 && conv_ch <= hidden;
+    }
+
+    static MoSRConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                          const std::string& prefix = "") {
+        MoSRConfig config;
+        auto find_weight = [&](const std::string& suffix) -> const TensorStorage* {
+            std::string name = prefix.empty() ? suffix : prefix + "." + suffix;
+            auto iter        = tensor_storage_map.find(name);
+            if (iter == tensor_storage_map.end() || iter->second.n_dims != 4) {
+                return nullptr;
+            }
+            return &iter->second;
+        };
+
+        const TensorStorage* first = find_weight("gblocks.0.weight");
+        const TensorStorage* fc1   = find_weight("gblocks.1.fc1.weight");
+        const TensorStorage* conv  = find_weight("gblocks.1.conv.weight");
+        const TensorStorage* up    = find_weight("upsampler.0.weight");
+        if (first == nullptr || fc1 == nullptr || conv == nullptr || up == nullptr) {
+            // Missing tensors, or the unsupported "gps"/"dys" upsamplers.
+            config.n_block = 0;
+            return config;
+        }
+
+        int seq_len                      = 0;
+        const std::string gblocks_prefix = prefix.empty() ? "gblocks." : prefix + ".gblocks.";
+        for (const auto& [name, _] : tensor_storage_map) {
+            if (!starts_with(name, gblocks_prefix)) {
+                continue;
+            }
+            size_t pos = name.find('.', gblocks_prefix.size());
+            if (pos == std::string::npos) {
+                continue;
+            }
+            try {
+                int idx = std::stoi(name.substr(gblocks_prefix.size(), pos - gblocks_prefix.size()));
+                seq_len = std::max(seq_len, idx + 1);
+            } catch (...) {
+            }
+        }
+
+        // gblocks = [conv_in, n_block x GatedCNNBlock, conv, Mish, conv, Mish, conv_1x1]
+        config.n_block     = seq_len - 6;
+        config.in_ch       = (int)first->ne[2];
+        config.dim         = (int)first->ne[3];
+        config.hidden      = (int)fc1->ne[3] / 2;
+        config.conv_ch     = (int)conv->ne[3];
+        config.kernel_size = (int)conv->ne[0];
+        config.scale       = (int)std::lround(std::sqrt((double)up->ne[3] / config.in_ch));
+        if (config.scale * config.scale * config.in_ch != up->ne[3]) {
+            config.scale = 0;
+        }
+
+        LOG_VERBOSE("mosr: scale = %d, n_block = %d, in_ch = %d, dim = %d, hidden = %d, conv_ch = %d, kernel_size = %d",
+                    config.scale,
+                    config.n_block,
+                    config.in_ch,
+                    config.dim,
+                    config.hidden,
+                    config.conv_ch,
+                    config.kernel_size);
+        return config;
+    }
+};
+
+static inline ggml_tensor* mosr_mish(ggml_context* ctx, ggml_tensor* x) {
+    return ggml_mul(ctx, x, ggml_tanh(ctx, ggml_softplus(ctx, x)));
+}
+
+class MoSRLayerNorm2d : public UnaryBlock {
+protected:
+    int64_t channels;
+
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
+        params["bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
+    }
+
+public:
+    explicit MoSRLayerNorm2d(int64_t channels)
+        : channels(channels) {}
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        // [N, C, H, W] -> [N, H, W, C] so the norm reduces over channels.
+        x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));
+        x = ggml_ext_layer_norm(ctx->ggml_ctx, x, params["weight"], params["bias"], 1e-6f);
+        return ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));
+    }
+};
+
+class MoSRDepthwiseConv2d : public UnaryBlock {
+protected:
+    int64_t channels;
+    int kernel_size;
+
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        // The direct depthwise kernels of the CUDA and Vulkan backends need an F32 kernel.
+        params["weight"] = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, kernel_size, kernel_size, 1, channels);
+        params["bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, channels);
+    }
+
+public:
+    MoSRDepthwiseConv2d(int64_t channels, int kernel_size)
+        : channels(channels), kernel_size(kernel_size) {}
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        const int pad = kernel_size / 2;
+        x             = ggml_conv_2d_dw_direct(ctx->ggml_ctx, params["weight"], ggml_ext_cont(ctx->ggml_ctx, x), 1, 1, pad, pad, 1, 1);
+        auto b        = ggml_reshape_4d(ctx->ggml_ctx, params["bias"], 1, 1, channels, 1);
+        return ggml_add_inplace(ctx->ggml_ctx, x, b);
+    }
+};
+
+class MoSRGatedCNNBlock : public GGMLBlock {
+protected:
+    int hidden;
+    int conv_ch;
+
+public:
+    MoSRGatedCNNBlock(int dim, int hidden, int conv_ch, int kernel_size)
+        : hidden(hidden), conv_ch(conv_ch) {
+        blocks["norm"] = std::shared_ptr<GGMLBlock>(new MoSRLayerNorm2d(dim));
+        blocks["fc1"]  = std::shared_ptr<GGMLBlock>(new Conv2d(dim, hidden * 2, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv"] = std::shared_ptr<GGMLBlock>(new MoSRDepthwiseConv2d(conv_ch, kernel_size));
+        blocks["fc2"]  = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, dim, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [n, dim, h, w]
+        // return: [n, dim, h, w]
+        auto norm = std::dynamic_pointer_cast<MoSRLayerNorm2d>(blocks["norm"]);
+        auto fc1  = std::dynamic_pointer_cast<Conv2d>(blocks["fc1"]);
+        auto conv = std::dynamic_pointer_cast<MoSRDepthwiseConv2d>(blocks["conv"]);
+        auto fc2  = std::dynamic_pointer_cast<Conv2d>(blocks["fc2"]);
+
+        auto h = fc1->forward(ctx, norm->forward(ctx, x));
+
+        auto channel_view = [&](int64_t offset, int64_t count) {
+            return ggml_view_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], count, h->ne[3],
+                                h->nb[1], h->nb[2], h->nb[3], offset * h->nb[2]);
+        };
+        auto g = channel_view(0, hidden);
+        auto i = channel_view(hidden, hidden - conv_ch);
+        auto c = conv->forward(ctx, channel_view(2 * hidden - conv_ch, conv_ch));
+        if (hidden > conv_ch) {
+            c = ggml_concat(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, i), c, 2);
+        }
+
+        h = ggml_mul(ctx->ggml_ctx, c, mosr_mish(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, g)));
+        h = mosr_mish(ctx->ggml_ctx, fc2->forward(ctx, h));
+        return ggml_add(ctx->ggml_ctx, h, ggml_scale_bias(ctx->ggml_ctx, x, 1.f, -0.5f));
+    }
+};
+
+class MoSRConvBlock : public GGMLBlock {
+public:
+    MoSRConvBlock(int in_ch, int out_ch) {
+        blocks["block.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_ch, out_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["block.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_ch, out_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv11"]  = std::shared_ptr<GGMLBlock>(new Conv2d(in_ch, out_ch, {1, 1}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        auto conv0  = std::dynamic_pointer_cast<Conv2d>(blocks["block.0"]);
+        auto conv2  = std::dynamic_pointer_cast<Conv2d>(blocks["block.2"]);
+        auto conv11 = std::dynamic_pointer_cast<Conv2d>(blocks["conv11"]);
+
+        auto out = mosr_mish(ctx->ggml_ctx, conv0->forward(ctx, x));
+        out      = mosr_mish(ctx->ggml_ctx, conv2->forward(ctx, out));
+        return ggml_add(ctx->ggml_ctx, out, conv11->forward(ctx, x));
+    }
+};
+
+class MoSRNet : public GGMLBlock {
+protected:
+    MoSRConfig config;
+
+    static std::string gblock_name(int idx) {
+        return "gblocks." + std::to_string(idx);
+    }
+
+    // torch.nn.PixelShuffle: [n, c*r*r, h, w] -> [n, c, h*r, w*r]
+    static ggml_tensor* pixel_shuffle(ggml_context* ctx, ggml_tensor* x, int r) {
+        const int64_t w = x->ne[0];
+        const int64_t h = x->ne[1];
+        const int64_t c = x->ne[2] / (r * r);
+        const int64_t n = x->ne[3];
+        x               = ggml_reshape_4d(ctx, x, w, h, r, r * c * n);       // [n*c*r, r, h, w]
+        x               = ggml_cont(ctx, ggml_permute(ctx, x, 1, 2, 0, 3));  // [n*c*r, h, w, r]
+        x               = ggml_reshape_4d(ctx, x, r * w, h, r, c * n);       // [n*c, r, h, w*r]
+        x               = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [n*c, h, r, w*r]
+        return ggml_reshape_4d(ctx, x, r * w, r * h, c, n);
+    }
+
+public:
+    explicit MoSRNet(MoSRConfig config)
+        : config(std::move(config)) {
+        const int dim = this->config.dim;
+        const int nb  = this->config.n_block;
+
+        blocks[gblock_name(0)] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.in_ch, dim, {3, 3}, {1, 1}, {1, 1}));
+        for (int i = 1; i <= nb; i++) {
+            blocks[gblock_name(i)] = std::shared_ptr<GGMLBlock>(new MoSRGatedCNNBlock(dim,
+                                                                                      this->config.hidden,
+                                                                                      this->config.conv_ch,
+                                                                                      this->config.kernel_size));
+        }
+        blocks[gblock_name(nb + 1)] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim * 2, {3, 3}, {1, 1}, {1, 1}));
+        blocks[gblock_name(nb + 3)] = std::shared_ptr<GGMLBlock>(new Conv2d(dim * 2, dim, {3, 3}, {1, 1}, {1, 1}));
+        blocks[gblock_name(nb + 5)] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {1, 1}));
+        blocks["shortcut"]          = std::shared_ptr<GGMLBlock>(new MoSRConvBlock(this->config.in_ch, dim));
+        blocks["upsampler.0"]       = std::shared_ptr<GGMLBlock>(new Conv2d(dim,
+                                                                            this->config.in_ch * this->config.scale * this->config.scale,
+                                                                            {3, 3},
+                                                                            {1, 1},
+                                                                            {1, 1}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [n, in_ch, h, w]
+        // return: [n, in_ch, h*scale, w*scale]
+        const int nb   = config.n_block;
+        auto conv_in   = std::dynamic_pointer_cast<Conv2d>(blocks[gblock_name(0)]);
+        auto conv_mid1 = std::dynamic_pointer_cast<Conv2d>(blocks[gblock_name(nb + 1)]);
+        auto conv_mid2 = std::dynamic_pointer_cast<Conv2d>(blocks[gblock_name(nb + 3)]);
+        auto conv_out  = std::dynamic_pointer_cast<Conv2d>(blocks[gblock_name(nb + 5)]);
+        auto shortcut  = std::dynamic_pointer_cast<MoSRConvBlock>(blocks["shortcut"]);
+        auto upsampler = std::dynamic_pointer_cast<Conv2d>(blocks["upsampler.0"]);
+
+        auto skip = shortcut->forward(ctx, x);
+        sd::ggml_graph_cut::mark_graph_cut(skip, "mosr.shortcut", "skip");
+
+        auto feat = conv_in->forward(ctx, x);
+        sd::ggml_graph_cut::mark_graph_cut(feat, "mosr.prelude", "feat");
+        for (int i = 1; i <= nb; i++) {
+            auto block = std::dynamic_pointer_cast<MoSRGatedCNNBlock>(blocks[gblock_name(i)]);
+            feat       = block->forward(ctx, feat);
+            sd::ggml_graph_cut::mark_graph_cut(feat, "mosr.gblocks." + std::to_string(i), "feat");
+        }
+        feat = mosr_mish(ctx->ggml_ctx, conv_mid1->forward(ctx, feat));
+        feat = mosr_mish(ctx->ggml_ctx, conv_mid2->forward(ctx, feat));
+        feat = conv_out->forward(ctx, feat);
+        feat = ggml_add(ctx->ggml_ctx, feat, ggml_scale_bias(ctx->ggml_ctx, skip, 1.f, -0.5f));
+        sd::ggml_graph_cut::mark_graph_cut(feat, "mosr.body.out", "feat");
+
+        auto out = pixel_shuffle(ctx->ggml_ctx, upsampler->forward(ctx, feat), config.scale);
+        sd::ggml_graph_cut::mark_graph_cut(out, "mosr.final", "out");
+        return out;
+    }
+};
+
+struct MoSR : public GGMLRunner {
+    MoSRConfig config;
+    std::unique_ptr<MoSRNet> mosr_net;
+
+    MoSR(ggml_backend_t backend,
+         const String2TensorStorage& tensor_storage_map      = {},
+         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : GGMLRunner(backend, weight_manager),
+          config(MoSRConfig::detect_from_weights(tensor_storage_map)) {
+        if (!config.is_valid()) {
+            return;
+        }
+        mosr_net = std::make_unique<MoSRNet>(config);
+        mosr_net->init(params_ctx, tensor_storage_map, "");
+    }
+
+    std::string get_desc() override {
+        return "mosr";
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
+        if (!mosr_net) {
+            return;
+        }
+
+        mosr_net->get_param_tensors(tensors);
+    }
+
+    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
+        if (!mosr_net)
+            return nullptr;
+        constexpr int kGraphNodes = 1 << 16;  // 65k
+        ggml_cgraph* gf           = new_graph_custom(kGraphNodes);
+        ggml_tensor* x            = make_input(x_tensor);
+
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = mosr_net->forward(&runner_ctx, x);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    sd::Tensor<float> compute(const int n_threads,
+                              const sd::Tensor<float>& x) {
+        auto get_graph = [&]() -> ggml_cgraph* { return build_graph(x); };
+        auto result    = restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, false), x.dim());
+        return result;
+    }
+};
+
+#endif  // __SD_MODEL_UPSCALER_MOSR_HPP__

src/model/vae/minimax_h3_vae.hpp

@@ -228,11 +228,12 @@ namespace MiniMaxH3VAE {
         return ggml_reshape_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2] * x->ne[3]);
     }
 
-    static ggml_tensor* apply_partial_rope(ggml_context* ctx,
+    static ggml_tensor* apply_partial_rope(GGMLRunnerContext* runner_ctx,
                                            ggml_tensor* x,
                                            ggml_tensor* pe) {
+        auto ctx        = runner_ctx->ggml_ctx;
         int64_t rot_dim = pe->ne[2] * 2;
-        auto rotated    = Rope::apply_rope(ctx,
+        auto rotated    = Rope::apply_rope(runner_ctx,
                                            ggml_ext_slice(ctx, x, 0, 0, rot_dim),
                                            pe,
                                            false);
@@ -289,8 +290,8 @@ namespace MiniMaxH3VAE {
                                                   batch_size);
             q                   = ggml_rms_norm(ctx->ggml_ctx, q, 1e-5f);
             k                   = ggml_rms_norm(ctx->ggml_ctx, k, 1e-5f);
-            q                   = apply_partial_rope(ctx->ggml_ctx, q, pe);
-            k                   = apply_partial_rope(ctx->ggml_ctx, k, pe);
+            q                   = apply_partial_rope(ctx, q, pe);
+            k                   = apply_partial_rope(ctx, k, pe);
             auto out            = ggml_ext_attention_ext(ctx,
                                                          q,
                                                          k,

src/model/vae/tae.hpp

@@ -84,12 +84,17 @@ class TinyEncoder : public UnaryBlock {
     int channels    = 64;
     int z_channels  = 4;
     int num_blocks  = 3;
+    bool f16;
 
 public:
-    TinyEncoder(int z_channels = 4, bool use_midblock_gn = false)
-        : z_channels(z_channels) {
-        int index                       = 0;
+    TinyEncoder(int z_channels = 4, bool use_midblock_gn = false, bool f16 = false)
+        : z_channels(z_channels), f16(f16) {
+        in_channels                     = f16 ? 16 : 3;
+        int index                       = f16 ? 1 : 0;
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        if (f16) {
+            index++;  // nn.ReLU()
+        }
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
 
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
@@ -97,12 +102,14 @@ public:
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
         }
 
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels * 2 : channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        channels *= f16 ? 2 : 1;
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
         }
 
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels * 2 : channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        channels *= f16 ? 2 : 1;
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
         }
@@ -114,7 +121,11 @@ public:
         // x: [n, in_channels, h, w]
         // return: [n, z_channels, h/8, w/8]
 
-        for (int i = 0; i < num_blocks * 3 + 6; i++) {
+        for (int i = f16 ? 1 : 0; i < num_blocks * 3 + 6 + (f16 ? 2 : 0); i++) {
+            if (f16 && i == 2) {
+                x = ggml_relu_inplace(ctx->ggml_ctx, x);
+                continue;
+            }
             auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
 
             x = block->forward(ctx, x);
@@ -131,9 +142,11 @@ class TinyDecoder : public UnaryBlock {
     int num_blocks   = 3;
 
 public:
-    TinyDecoder(int z_channels = 4, bool use_midblock_gn = false)
+    TinyDecoder(int z_channels = 4, bool use_midblock_gn = false, bool f16 = false)
         : z_channels(z_channels) {
-        int index = 0;
+        channels     = f16 ? 256 : 64;
+        out_channels = f16 ? 16 : 3;
+        int index    = 0;
 
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels, {3, 3}, {1, 1}, {1, 1}));
         index++;  // nn.ReLU()
@@ -142,13 +155,15 @@ public:
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
         }
         index++;  // nn.Upsample()
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels / 2 : channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        channels /= f16 ? 2 : 1;
 
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
         }
         index++;  // nn.Upsample()
-        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, f16 ? channels / 2 : channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        channels /= f16 ? 2 : 1;
 
         for (int i = 0; i < num_blocks; i++) {
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
@@ -691,6 +706,7 @@ class TAESD : public GGMLBlock {
 protected:
     bool decode_only;
     bool taef2 = false;
+    bool f16   = false;
 
 public:
     int z_channels = 4;
@@ -708,10 +724,14 @@ public:
             z_channels      = 32;
             use_midblock_gn = true;
         }
-        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels, use_midblock_gn));
+        f16 = version == VERSION_QWEN_IMAGE_2_1;
+        if (f16) {
+            z_channels = 64;
+        }
+        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels, use_midblock_gn, f16));
 
         if (!decode_only) {
-            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels, use_midblock_gn));
+            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels, use_midblock_gn, f16));
         }
     }
 
@@ -720,10 +740,15 @@ public:
         if (taef2) {
             z = unpatchify(ctx->ggml_ctx, z, 2);
         }
-        return decoder->forward(ctx, z);
+        auto x = decoder->forward(ctx, z);
+        return f16 ? unpatchify(ctx->ggml_ctx, x, 2) : x;
     }
 
     ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        if (f16) {
+            GGML_ASSERT(x->ne[0] % 2 == 0 && x->ne[1] % 2 == 0);
+            x = patchify(ctx->ggml_ctx, x, 2);
+        }
         auto encoder = std::dynamic_pointer_cast<TinyEncoder>(blocks["encoder.layers"]);
         auto z       = encoder->forward(ctx, x);
         if (taef2) {

src/model/vae/vae.hpp

@@ -176,11 +176,20 @@ public:
         int scale_factor = 8;
         if (version == VERSION_LTXAV) {
             scale_factor = 32;
-        } else if (version == VERSION_WAN2_2_TI2V || version == VERSION_QWEN_IMAGE_2_1 || sd_version_is_hunyuan_video(version) || sd_version_is_mage_flow(version) || sd_version_is_minimax_h3(version)) {
+        } else if (version == VERSION_WAN2_2_TI2V ||
+                   version == VERSION_QWEN_IMAGE_2_1 ||
+                   sd_version_is_hunyuan_video(version) ||
+                   sd_version_is_mage_flow(version) ||
+                   sd_version_is_minimax_h3(version)) {
             scale_factor = 16;
         } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
-        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version) || sd_version_is_sensenova_u1(version) || sd_version_is_z_image_l2p(version)) {
+        } else if (version == VERSION_CHROMA_RADIANCE ||
+                   version == VERSION_HIDREAM_O1 ||
+                   sd_version_is_minit2i(version) ||
+                   sd_version_is_sensenova_u1(version) ||
+                   sd_version_is_z_image_l2p(version) ||
+                   version == VERSION_IRIS) {
             scale_factor = 1;
         }
         return scale_factor;

src/model_loader.cpp

@@ -556,6 +556,10 @@ SDVersion ModelLoader::get_sd_version() const {
             tensor_storage_map.find("model.diffusion_model.audio_patch_proj.weight") != tensor_storage_map.end()) {
             return VERSION_MINIMAX_H3;
         }
+        if (name == "model.diffusion_model.y_embedder.layer_pool.weight" &&
+            tensor_storage_map.count("model.diffusion_model.modulation_cores.adaln_img.weight")) {
+            return VERSION_IRIS;
+        }
         if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
             is_wan = true;
         }

src/model_manager.cpp

@@ -1904,6 +1904,49 @@ bool ModelManager::ensure_compute_backend_capacity(
         }
     }
 
+    // The loader retains mmap buffers after storage blocks are released, so
+    // demoting mapped states would only force duplicate allocations on reload.
+    std::unordered_set<TensorState*> mapped_states;
+    for (const auto& block : params_storage_blocks_) {
+        if (!block->mmap_tensor_stores.empty()) {
+            mapped_states.insert(block->states.begin(), block->states.end());
+        }
+    }
+    std::unordered_set<TensorState*> required_set(required_states.begin(), required_states.end());
+    std::vector<TensorState*> demoted_states;
+    auto revert_demotions = [&]() {
+        for (TensorState* state : demoted_states) {
+            if (state->loaded_to_params_backend) {
+                state->residency_mode = ResidencyMode::ParamBackend;
+            }
+        }
+    };
+    auto can_demote_to_disk = [&](TensorState* state) {
+        return state != nullptr && state->tensor != nullptr && state->tensor->view_src == nullptr &&
+               state->residency_mode == ResidencyMode::ParamBackend &&
+               state->params_backend != nullptr && state->params_backend == state->compute_backend &&
+               state->compute_backend == compute_backend && !sd_backend_is_cpu(state->params_backend) &&
+               state->pin_count == 0 && !state->staged_to_compute_backend && state->loaded_to_params_backend &&
+               protected_states.find(state) == protected_states.end() &&
+               required_set.find(state) == required_set.end() &&
+               mapped_states.find(state) == mapped_states.end() &&
+               !state->sources.empty() && state->applied_lora_epoch == UINT64_MAX &&
+               state->split_buffer_type == nullptr;
+    };
+    for (TensorState* state : global_candidates) {
+        if (!can_demote_to_disk(state)) {
+            continue;
+        }
+        state->residency_mode = ResidencyMode::Disk;
+        demoted_states.push_back(state);
+        add_evictable_state(state);
+        if (release_eviction_states()) {
+            revert_demotions();
+            return true;
+        }
+    }
+    revert_demotions();
+
     const auto capacity                = check_capacity(request, required_states, true);
     const std::string available_device = capacity.available_device_bytes == SIZE_MAX
                                              ? "unknown"

src/pipeline/diffusion_engine.cpp

@@ -108,6 +108,7 @@ const char* model_version_to_str[] = {
     "PixArt",
     "Ming-Image",
     "Z-Image L2P",
+    "Iris-3B",
 };
 
 static_assert(VERSION_COUNT == sizeof(model_version_to_str) / sizeof(model_version_to_str[0]),
@@ -1382,7 +1383,8 @@ bool StableDiffusionGGML::build_denoiser() {
                    sd_version_is_llada_image(version) ||
                    sd_version_is_boogu_image(version) ||
                    sd_version_is_pid(version) ||
-                   sd_version_is_ideogram4(version)) {
+                   sd_version_is_ideogram4(version) ||
+                   version == VERSION_IRIS) {
             pred_type = FLOW_PRED;
             if (sd_version_is_wan(version)) {
                 default_flow_shift = 5.f;
@@ -1390,7 +1392,7 @@ bool StableDiffusionGGML::build_denoiser() {
                 default_flow_shift = 7.f;
             } else if (sd_version_is_minimax_h3(version)) {
                 default_flow_shift = 12.f;
-            } else if (sd_version_is_ernie_image(version)) {
+            } else if (sd_version_is_ernie_image(version) || version == VERSION_IRIS) {
                 default_flow_shift = 4.f;
             } else if (sd_version_is_pid(version)) {
                 default_flow_shift = 1.5f;
@@ -2225,6 +2227,27 @@ void StableDiffusionGGML::report_sample_progress(int step,
     }
 }
 
+static sd::Tensor<float> batch_two_condition_tensors(const sd::Tensor<float>& a, const sd::Tensor<float>& b) {
+    if (a.empty() || b.empty() || a.dim() != b.dim()) {
+        return {};
+    }
+    if (a.dim() == 1) {
+        if (a.shape() != b.shape()) {
+            return {};
+        }
+        auto batched = sd::ops::concat(a, b, 0);
+        batched.reshape_({a.shape()[0], 2});
+        return batched;
+    }
+    const int64_t batch_dim = a.dim() - 1;
+    for (int64_t d = 0; d < batch_dim; d++) {
+        if (a.shape()[d] != b.shape()[d]) {
+            return {};
+        }
+    }
+    return sd::ops::concat(a, b, static_cast<size_t>(batch_dim));
+}
+
 void StableDiffusionGGML::compute_sample_controls(const sd::Tensor<float>& control_image,
                                                   const sd::Tensor<float>& noised_input,
                                                   const sd::Tensor<float>& timesteps_tensor,
@@ -2602,6 +2625,57 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             return output_opt;
         };
 
+        auto run_batched_condition = [&](const SDCondition& condition,
+                                         const sd::Tensor<float>* c_concat_override) -> sd::Tensor<float> {
+            const sd::Tensor<float>& condition_concat =
+                c_concat_override != nullptr ? *c_concat_override : condition.c_concat;
+
+            sd::Tensor<float> batched_context = batch_two_condition_tensors(condition.c_crossattn, uncond.c_crossattn);
+            sd::Tensor<float> batched_y       = batch_two_condition_tensors(condition.c_vector, uncond.c_vector);
+            sd::Tensor<float> batched_concat  = batch_two_condition_tensors(condition_concat, uncond.c_concat);
+            if (!condition.c_crossattn.empty() && batched_context.empty()) {
+                return {};
+            }
+            if ((!condition.c_vector.empty() || !uncond.c_vector.empty()) && batched_y.empty()) {
+                return {};
+            }
+            if ((!condition_concat.empty() || !uncond.c_concat.empty()) && batched_concat.empty()) {
+                return {};
+            }
+
+            std::vector<sd::Tensor<float>> uncond_controls;
+            compute_sample_controls(control_image, noised_input, timesteps_tensor, uncond, &uncond_controls);
+            if (controls.size() != uncond_controls.size()) {
+                return {};
+            }
+            std::vector<sd::Tensor<float>> batched_controls;
+            batched_controls.reserve(controls.size());
+            for (size_t i = 0; i < controls.size(); i++) {
+                sd::Tensor<float> batched_control = batch_two_condition_tensors(controls[i], uncond_controls[i]);
+                if (batched_control.empty()) {
+                    return {};
+                }
+                batched_controls.push_back(std::move(batched_control));
+            }
+
+            sd::Tensor<float> batched_x =
+                sd::ops::concat(noised_input, noised_input, static_cast<size_t>(noised_input.dim() - 1));
+
+            DiffusionParams batched_params = diffusion_params;
+            batched_params.x               = &batched_x;
+            batched_params.context         = batched_context.empty() ? nullptr : &batched_context;
+            batched_params.c_concat        = batched_concat.empty() ? nullptr : &batched_concat;
+            batched_params.y               = batched_y.empty() ? nullptr : &batched_y;
+            batched_params.ref_latents     = nullptr;
+            batched_params.extra           = UNetDiffusionExtra{1, &batched_controls, control_strength};
+
+            sd::Tensor<float> output = work_diffusion_model->compute(n_threads, batched_params);
+            if (output.empty()) {
+                LOG_ERROR("batched diffusion model compute failed");
+            }
+            return output;
+        };
+
         const SDCondition* positive_condition      = &cond;
         const sd::Tensor<float>* c_concat_override = nullptr;
         for (const auto& extension : generation_extensions) {
@@ -2641,12 +2715,40 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             }
         }
 
-        cond_out = run_condition(*positive_condition, c_concat_override);
-        if (cond_out.empty()) {
-            return {};
+        const bool batch_cfg_ok = config_->params.batched_cfg &&
+                                  sd_version_is_unet(version) &&
+                                  !uncond.empty() &&
+                                  img_uncond.empty() &&
+                                  !skip_uncond &&
+                                  !cache_runtime.ucache_enabled() &&
+                                  !(is_skiplayer_step && slg_uncond) &&
+                                  ip_adapter_tokens.empty() &&
+                                  ip_adapter_uncond_tokens.empty() &&
+                                  !config_->animatediff_loaded &&
+                                  (noised_input.dim() < 4 || noised_input.shape()[3] <= 1) &&
+                                  std::none_of(generation_extensions.begin(),
+                                               generation_extensions.end(),
+                                               [](const std::shared_ptr<GenerationExtension>& extension) {
+                                                   return extension->is_enabled();
+                                               });
+
+        if (batch_cfg_ok) {
+            sd::Tensor<float> batched_out = run_batched_condition(*positive_condition, c_concat_override);
+            if (!batched_out.empty() && batched_out.dim() >= 4 && batched_out.shape()[3] == 2) {
+                auto parts    = sd::ops::chunk(batched_out, 2, 3);
+                cond_out      = std::move(parts[0]);
+                uncond_out    = std::move(parts[1]);
+            }
         }
 
-        if (!uncond.empty()) {
+        if (cond_out.empty()) {
+            cond_out = run_condition(*positive_condition, c_concat_override);
+            if (cond_out.empty()) {
+                return {};
+            }
+        }
+
+        if (uncond_out.empty() && !uncond.empty()) {
             if (!skip_uncond) {
                 const std::vector<int>* uncond_skip_layers = nullptr;
                 if (is_skiplayer_step && slg_uncond) {
@@ -2765,7 +2867,7 @@ int StableDiffusionGGML::get_diffusion_model_down_factor() {
     if (sd_version_is_dit(version)) {
         if (sd_version_is_sensenova_u1(version)) {
             down_factor = 32;
-        } else if (sd_version_is_z_image_l2p(version)) {
+        } else if (sd_version_is_z_image_l2p(version) || version == VERSION_IRIS) {
             down_factor = 16;
         } else if (version == VERSION_QWEN_IMAGE_2_1 || version == VERSION_MING_IMAGE || sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version) || sd_version_is_pixart(version)) {
             down_factor = 2;
@@ -2799,7 +2901,7 @@ int StableDiffusionGGML::get_latent_channel() {
             latent_channel = 3;
         } else if (sd_version_is_z_image_l2p(version)) {
             latent_channel = 3;
-        } else if (sd_version_is_pid(version)) {
+        } else if (sd_version_is_pid(version) || version == VERSION_IRIS) {
             latent_channel = 3;
         } else if (sd_version_is_sefi_image(version)) {
             latent_channel = 144;
@@ -2900,7 +3002,7 @@ sd::Tensor<float> StableDiffusionGGML::encode_first_stage(const sd::Tensor<float
 }
 
 sd::Tensor<float> StableDiffusionGGML::decode_first_stage(const sd::Tensor<float>& x, bool decode_video) {
-    if (sd_version_is_pid(version) || sd_version_is_minit2i(version)) {
+    if (sd_version_is_pid(version) || sd_version_is_minit2i(version) || version == VERSION_IRIS) {
         return sd::ops::clamp((x + 1.f) * 0.5f, 0.0f, 1.0f);
     }
     auto latents                      = first_stage_model->diffusion_to_vae_latents(x);

src/pipeline/model_builders.cpp

@@ -18,6 +18,7 @@
 #include "model/diffusion/hidream_o1.hpp"
 #include "model/diffusion/hunyuan.hpp"
 #include "model/diffusion/ideogram4.hpp"
+#include "model/diffusion/iris.hpp"
 #include "model/diffusion/krea2.hpp"
 #include "model/diffusion/lens.hpp"
 #include "model/diffusion/lingbot_video.hpp"
@@ -456,6 +457,11 @@ namespace sd::model_builders {
                                                                   tensor_storage_map,
                                                                   "model.diffusion_model",
                                                                   weight_manager);
+        } else if (version == VERSION_IRIS) {
+            result.conditioner = std::make_shared<LLMEmbedder>(ctx.backends.runtime_backend(SDBackendModule::TE),
+                                                               tensor_storage_map, version, "", false, weight_manager, tokenizers);
+            result.diffusion   = std::make_shared<Iris::IrisRunner>(ctx.backends.runtime_backend(SDBackendModule::DIFFUSION),
+                                                                  tensor_storage_map, "model.diffusion_model", weight_manager);
         } else {  // SD1.x SD2.x SDXL
             std::map<std::string, std::string> embbeding_map;
             for (uint32_t i = 0; i < sd_ctx_params->embedding_count; i++) {
@@ -527,7 +533,7 @@ namespace sd::model_builders {
         }
 
         auto create_tae = [&](bool decode_only) -> std::shared_ptr<VAE> {
-            if (sd_version_uses_wan_vae(version) || sd_version_is_hunyuan_video(version) || sd_version_is_ltxav(version) || sd_version_is_minimax_h3(version)) {
+            if ((sd_version_uses_wan_vae(version) && version != VERSION_QWEN_IMAGE_2_1) || sd_version_is_hunyuan_video(version) || sd_version_is_ltxav(version) || sd_version_is_minimax_h3(version)) {
                 return std::make_shared<TinyVideoAutoEncoder>(ctx.backends.runtime_backend(SDBackendModule::VAE),
                                                               tensor_storage_map,
                                                               "decoder",
@@ -626,7 +632,12 @@ namespace sd::model_builders {
             }
         };
 
-        if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version) || sd_version_is_sensenova_u1(version) || sd_version_is_z_image_l2p(version)) {
+        if (version == VERSION_CHROMA_RADIANCE ||
+            version == VERSION_HIDREAM_O1 ||
+            sd_version_is_minit2i(version) ||
+            sd_version_is_sensenova_u1(version) ||
+            sd_version_is_z_image_l2p(version) ||
+            version == VERSION_IRIS) {
             LOG_INFO("using FakeVAE");
             result.vae = std::make_shared<FakeVAE>(version,
                                                    ctx.backends.runtime_backend(SDBackendModule::VAE),

src/stable-diffusion.cpp

@@ -335,6 +335,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->max_vram                  = nullptr;
     sd_ctx_params->disable_prefetch          = false;
     sd_ctx_params->disable_segmented_compute = false;
+    sd_ctx_params->batched_cfg               = true;
     sd_ctx_params->eager_load                = false;
     sd_ctx_params->enable_mmap               = false;
     sd_ctx_params->diffusion_flash_attn      = false;

src/upscaler.cpp

@@ -25,15 +25,40 @@ UpscalerGGML::~UpscalerGGML() {
     // ModelManager holds raw ggml tensor pointers owned by the runner context.
     model_manager.reset();
     esrgan_upscaler.reset();
+    mosr_upscaler.reset();
 }
 
 void UpscalerGGML::set_max_graph_vram_bytes(size_t max_vram_bytes) {
     max_graph_vram_bytes = max_vram_bytes;
-    if (esrgan_upscaler) {
-        esrgan_upscaler->set_max_graph_vram_bytes(max_vram_bytes);
+    if (runner() != nullptr) {
+        runner()->set_max_graph_vram_bytes(max_vram_bytes);
     }
 }
 
+GGMLRunner* UpscalerGGML::runner() {
+    if (mosr_upscaler) {
+        return mosr_upscaler.get();
+    }
+    return esrgan_upscaler.get();
+}
+
+int UpscalerGGML::get_scale() {
+    if (mosr_upscaler) {
+        return mosr_upscaler->config.scale;
+    }
+    if (esrgan_upscaler) {
+        return esrgan_upscaler->config.scale;
+    }
+    return 1;
+}
+
+sd::Tensor<float> UpscalerGGML::compute(const sd::Tensor<float>& input_tensor) {
+    if (mosr_upscaler) {
+        return mosr_upscaler->compute(n_threads, input_tensor);
+    }
+    return esrgan_upscaler->compute(n_threads, input_tensor);
+}
+
 bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
                                   int n_threads) {
     sd::ParallelScope tensor_scope(&tensor_executor);
@@ -82,20 +107,33 @@ bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
     }
     model_loader.set_wtype_override(model_data_type);
     LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
-    esrgan_upscaler = std::make_shared<ESRGAN>(backend_for(SDBackendModule::UPSCALER),
-                                               model_loader.get_tensor_storage_map(),
+    const auto& tensor_storage_map = model_loader.get_tensor_storage_map();
+    std::map<std::string, ggml_tensor*> tensors;
+    if (tensor_storage_map.count("gblocks.0.weight") > 0) {
+        mosr_upscaler = std::make_shared<MoSR>(backend_for(SDBackendModule::UPSCALER),
+                                               tensor_storage_map,
                                                model_manager);
-    if (esrgan_upscaler == nullptr || esrgan_upscaler->rrdb_net == nullptr) {
-        LOG_ERROR("init esrgan model from metadata failed: '%s'", esrgan_path.c_str());
-        return false;
+        if (mosr_upscaler->mosr_net == nullptr) {
+            LOG_ERROR("init mosr model from metadata failed: '%s'", esrgan_path.c_str());
+            return false;
+        }
+        mosr_upscaler->get_param_tensors(tensors);
+    } else {
+        esrgan_upscaler = std::make_shared<ESRGAN>(backend_for(SDBackendModule::UPSCALER),
+                                                   tensor_storage_map,
+                                                   model_manager);
+        if (esrgan_upscaler->rrdb_net == nullptr) {
+            LOG_ERROR("init esrgan model from metadata failed: '%s'", esrgan_path.c_str());
+            return false;
+        }
+        esrgan_upscaler->get_param_tensors(tensors);
     }
-    esrgan_upscaler->set_max_graph_vram_bytes(max_graph_vram_bytes);
+    LOG_INFO("Upscaler architecture: %s", runner()->get_desc().c_str());
+    runner()->set_max_graph_vram_bytes(max_graph_vram_bytes);
     if (direct) {
-        esrgan_upscaler->set_conv2d_direct_enabled(true);
+        runner()->set_conv2d_direct_enabled(true);
     }
 
-    std::map<std::string, ggml_tensor*> tensors;
-    esrgan_upscaler->get_param_tensors(tensors);
     if (!model_manager->set_loader(model_loader) ||
         !model_manager->register_param_tensors(ModelComponent::Upscaler,
                                                std::move(tensors),
@@ -103,7 +141,7 @@ bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
                                                backend_for(SDBackendModule::UPSCALER),
                                                params_backend_for(SDBackendModule::UPSCALER)) ||
         !model_manager->validate_registered_tensors()) {
-        LOG_ERROR("register esrgan tensors with model manager failed");
+        LOG_ERROR("register upscaler tensors with model manager failed");
         return false;
     }
     return true;
@@ -113,7 +151,7 @@ sd::Tensor<float> UpscalerGGML::upscale_tensor(const sd::Tensor<float>& input_te
     sd::ParallelScope tensor_scope(&tensor_executor);
     if (input_tensor.empty() || input_tensor.dim() != 4 ||
         (input_tensor.shape()[2] != 3 && input_tensor.shape()[2] != 4)) {
-        LOG_ERROR("esrgan expects a 4D RGB or RGBA image tensor");
+        LOG_ERROR("upscaler expects a 4D RGB or RGBA image tensor");
         return {};
     }
 
@@ -124,14 +162,14 @@ sd::Tensor<float> UpscalerGGML::upscale_tensor(const sd::Tensor<float>& input_te
     }
     const sd::Tensor<float>& model_input = has_alpha ? rgb : input_tensor;
     sd::Tensor<float> upscaled;
-    const int scale = esrgan_upscaler->config.scale;
+    const int scale = get_scale();
     if (tile_size <= 0 || (input_tensor.shape()[0] <= tile_size && input_tensor.shape()[1] <= tile_size)) {
-        upscaled = esrgan_upscaler->compute(n_threads, model_input);
+        upscaled = compute(model_input);
     } else {
         auto on_processing = [&](const sd::Tensor<float>& input_tile) -> sd::Tensor<float> {
-            auto output_tile = esrgan_upscaler->compute(n_threads, input_tile);
+            auto output_tile = compute(input_tile);
             if (output_tile.empty()) {
-                LOG_ERROR("esrgan compute failed while processing a tile");
+                LOG_ERROR("upscaler compute failed while processing a tile");
                 return {};
             }
             return output_tile;
@@ -148,9 +186,9 @@ sd::Tensor<float> UpscalerGGML::upscale_tensor(const sd::Tensor<float>& input_te
                                     false,
                                     on_processing);
     }
-    esrgan_upscaler->runner_end();
+    runner()->runner_end();
     if (upscaled.empty()) {
-        LOG_ERROR("esrgan compute failed");
+        LOG_ERROR("upscaler compute failed");
         return {};
     }
     if (has_alpha) {
@@ -168,7 +206,7 @@ sd_image_t UpscalerGGML::upscale(sd_image_t input_image, uint32_t upscale_factor
     sd::ParallelScope tensor_scope(&tensor_executor);
     // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
     sd_image_t upscaled_image = {0, 0, 0, nullptr};
-    const int scale           = esrgan_upscaler->config.scale;
+    const int scale           = get_scale();
     int output_width          = (int)input_image.width * scale;
     int output_height         = (int)input_image.height * scale;
     LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
@@ -256,10 +294,10 @@ bool upscale(upscaler_ctx_t* upscaler_ctx,
 }
 
 int get_upscale_factor(upscaler_ctx_t* upscaler_ctx) {
-    if (upscaler_ctx == nullptr || upscaler_ctx->upscaler == nullptr || upscaler_ctx->upscaler->esrgan_upscaler == nullptr) {
+    if (upscaler_ctx == nullptr || upscaler_ctx->upscaler == nullptr) {
         return 1;
     }
-    return upscaler_ctx->upscaler->esrgan_upscaler->config.scale;
+    return upscaler_ctx->upscaler->get_scale();
 }
 
 int get_upscaler_model_scale(const char* model_path) {
@@ -272,8 +310,12 @@ int get_upscaler_model_scale(const char* model_path) {
             return 0;
         }
         const auto& tensors = loader.get_tensor_storage_map();
-        auto first          = tensors.find("conv_first.weight");
-        auto last           = tensors.find("conv_last.weight");
+        if (tensors.count("gblocks.0.weight") > 0) {
+            MoSRConfig config = MoSRConfig::detect_from_weights(tensors);
+            return config.is_valid() ? config.scale : 0;
+        }
+        auto first = tensors.find("conv_first.weight");
+        auto last  = tensors.find("conv_last.weight");
         if (first == tensors.end() || last == tensors.end() ||
             tensors.count("body.0.rdb1.conv1.weight") == 0 ||
             first->second.n_dims != 4 || last->second.n_dims != 4 ||

src/upscaler.h

@@ -4,6 +4,7 @@
 #include "core/ggml_extend_backend.h"
 #include "core/tensor.hpp"
 #include "model/upscaler/esrgan.hpp"
+#include "model/upscaler/mosr.hpp"
 #include "model_manager.h"
 #include "stable-diffusion.h"
 
@@ -15,6 +16,7 @@ struct UpscalerGGML {
     std::shared_ptr<ModelManager> model_manager;
     ggml_type model_data_type = GGML_TYPE_F16;
     std::shared_ptr<ESRGAN> esrgan_upscaler;
+    std::shared_ptr<MoSR> mosr_upscaler;
     std::string esrgan_path;
     int n_threads;
     sd::ParallelExecutor tensor_executor;
@@ -34,6 +36,9 @@ struct UpscalerGGML {
     bool load_from_file(const std::string& esrgan_path,
                         int n_threads);
     void set_max_graph_vram_bytes(size_t max_vram_bytes);
+    GGMLRunner* runner();
+    int get_scale();
+    sd::Tensor<float> compute(const sd::Tensor<float>& input_tensor);
     sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor);
     sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor);
 };