feat: add linear and attention scale overrides (#1964)

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -6,7 +6,9 @@ body:
   - type: markdown
     attributes:
       value: |
-        Please use this template and include as many details as possible to help us reproduce and fix the issue.
+        Before submitting a bug report, please read the [Troubleshooting guide](https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/troubleshooting.md) and try the steps relevant to your problem.
+
+        If the problem persists, complete this form and include what you tried and the results, along with enough details to help us reproduce and fix the issue.
   - type: textarea
     id: commit
     attributes:

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,4 @@
+contact_links:
+  - name: Troubleshooting
+    url: https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/troubleshooting.md
+    about: Read the troubleshooting guide first. If the problem persists, submit a bug report.

README.md

@@ -51,6 +51,7 @@ API and command-line option may change frequently.***
     - [LongCat Image](./docs/longcat_image.md)
     - [Z-Image](./docs/z_image.md)
     - [MiniT2I](./docs/minit2i.md)
+    - [SenseNova U1.5](./docs/sensenova_u1.md)
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
@@ -147,6 +148,7 @@ For runtime and parameter backend placement, see the [backend selection guide](.
 
 ## More Guides
 
+- [Troubleshooting](./docs/troubleshooting.md)
 - [Backend selection](./docs/backend.md)
 - [RPC](./docs/rpc.md)
 - [LoRA](./docs/lora.md)

docs/sensenova_u1.md

@@ -0,0 +1,46 @@
+# How to Use
+
+SenseNova U1.5 is an 8B MoT model that performs diffusion directly in RGB pixel
+space. It does not require a separate text encoder or VAE.
+
+## Download weights
+
+- Download SenseNova U1.5 8B MoT
+    - safetensors: https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT
+
+Pass the complete downloaded repository directory to `--model`. The directory
+must contain `model.safetensors.index.json`, every referenced Safetensors shard,
+and the tokenizer files.
+
+## Examples
+
+### CUDA
+
+```bash
+./bin/sd-cli \
+  --model /path/to/SenseNova-U1.5-8B-MoT \
+  --prompt "a red cube on a white background" \
+  --width 2048 \
+  --height 2048 \
+  --steps 50 \
+  --cfg-scale 4 \
+  --flow-shift 3 \
+  --seed 42 \
+  --sampling-method euler \
+  --rng cuda \
+  --fa \
+  --output output.png
+```
+
+## Notes
+
+- To match the official non-thinking text-to-image pipeline, use 50 Euler
+  steps, CFG 4, flow shift 3, seed 42, CUDA RNG, and an empty negative prompt.
+- Width and height must be multiples of 32. The trained 1:1 resolution is
+  2048x2048; lower resolutions are useful for smoke tests but are outside the
+  training buckets.
+- The SenseNova prompt template and unconditional prompt are built
+  automatically.
+- This implementation supports non-thinking text-to-image generation. Image
+  editing, visual understanding, interleaved generation, and thinking-mode
+  prompt expansion are not implemented.

docs/troubleshooting.md

@@ -0,0 +1,45 @@
+# Troubleshooting
+
+## Completely black or white images or videos / NaNs
+
+Some ggml backends can encounter numerical overflow during inference, producing
+NaN (not-a-number) values. This can result in completely black or white images or videos.
+Whether it happens can depend on the backend, device, model, and weight format.
+
+Known overflow issues have been addressed as far as possible, but the maintainer
+has limited hardware and cannot test every combination. Some cases may therefore
+still need a manual workaround.
+
+These options are supported by both `sd-cli` and `sd-server`. If you encounter
+this problem, add them to your CLI generation command or server startup command:
+
+```sh
+--linear-scale 0.0078125 --attn-scale 0.0078125
+```
+
+For `sd-server`, restart the server after changing these startup options. Run the
+same prompt and seed again to see whether the output recovers. If the problem
+persists, try smaller positive values, for example:
+
+```sh
+--linear-scale 0.00390625 --attn-scale 0.00390625
+```
+
+These options reduce intermediate values and compensate afterwards to preserve
+the intended output scale:
+
+- `--linear-scale` scales Linear inputs before matrix multiplication and rescales
+  the result.
+- `--attn-scale` scales attention keys and values (K/V). It takes effect only in
+  the Flash Attention path, where `--fa` or `--diffusion-fa` is enabled and the
+  backend supports it.
+
+The two values can be set independently and apply across model components. The
+default `0` preserves each model's built-in settings; `1` explicitly disables the
+corresponding scaling. Overrides must be finite positive values. C API users can
+set `linear_scale` and `attn_scale` in `sd_ctx_params_t`.
+
+If the problem persists after trying the relevant steps above,
+[submit a bug report](https://github.com/leejet/stable-diffusion.cpp/issues/new?template=bug_report.yml).
+Include your full command, backend and hardware, model and weight format, logs,
+and the scale values you tried with their results.

examples/cli/README.md

@@ -22,3 +22,6 @@ Metadata mode inspects PNG/JPEG container metadata without loading any model:
 ./bin/sd-cli -M metadata --image ./output.png --metadata-raw
 ./bin/sd-cli -M metadata --image ./output.png --metadata-all
 ```
+
+For completely black or white images or videos, NaNs, and the `--linear-scale` /
+`--attn-scale` workaround, see [Troubleshooting](../../docs/troubleshooting.md).

examples/common/common.cpp

@@ -302,8 +302,12 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
                         invalid_arg = true;
                         return;
                     }
-                    *option.target = std::stoi(argv[i]);
+                    try {
-                    found_arg      = true;
+                        *option.target = std::stoi(argv[i]);
+                    } catch (const std::invalid_argument&) {
+                        invalid_arg = true;
+                    }
+                    found_arg = true;
                 }))
                 break;
 
@@ -312,8 +316,12 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
                         invalid_arg = true;
                         return;
                     }
-                    *option.target = std::stof(argv[i]);
+                    try {
-                    found_arg      = true;
+                        *option.target = std::stof(argv[i]);
+                    } catch (const std::invalid_argument&) {
+                        invalid_arg = true;
+                    }
+                    found_arg = true;
                 }))
                 break;
 
@@ -337,7 +345,8 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
 
         if (invalid_arg) {
             if (!valid) {
-                LOG_ERROR("error: invalid parameter for argument: %s", arg.c_str());
+                LOG_ERROR("error: invalid parameter for argument \"%s\": \"%s\"",
+                          arg.c_str(), (i >= argc) ? "" : argv[i]);
             }
             return false;
         }
@@ -350,6 +359,25 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
     return true;
 }
 
+static int parse_scale_override(int argc, const char** argv, int index, float& scale) {
+    if (++index >= argc) {
+        return -1;
+    }
+    try {
+        size_t end              = 0;
+        const std::string value = argv[index];
+        float parsed            = std::stof(value, &end);
+        if (end != value.size() || !std::isfinite(parsed) || parsed < 0.f ||
+            (parsed > 0.f && !std::isfinite(1.f / parsed))) {
+            return -1;
+        }
+        scale = parsed;
+    } catch (const std::exception&) {
+        return -1;
+    }
+    return 1;
+}
+
 ArgOptions SDContextParams::get_options() {
     ArgOptions options;
     options.string_options = {
@@ -678,6 +706,18 @@ ArgOptions SDContextParams::get_options() {
     };
 
     options.manual_options = {
+        {"",
+         "--linear-scale",
+         "linear input scale override (float, default: 0 = model default, 1 = no scaling)",
+         [this](int argc, const char** argv, int index) {
+             return parse_scale_override(argc, argv, index, linear_scale);
+         }},
+        {"",
+         "--attn-scale",
+         "flash-attention K/V scale override (float, default: 0 = model default, 1 = no scaling); requires --fa or --diffusion-fa",
+         [this](int argc, const char** argv, int index) {
+             return parse_scale_override(argc, argv, index, attn_scale);
+         }},
         {"",
          "--auto-fit",
          "on|off (default: on). Use one GPU for diffusion/te/vae computation and place weights on that GPU, "
@@ -886,6 +926,8 @@ std::string SDContextParams::to_string() const {
         << "  vae_on_cpu: " << (vae_on_cpu ? "true" : "false") << ",\n"
         << "  flash_attn: " << (flash_attn ? "true" : "false") << ",\n"
         << "  diffusion_flash_attn: " << (diffusion_flash_attn ? "true" : "false") << ",\n"
+        << "  linear_scale: " << linear_scale << ",\n"
+        << "  attn_scale: " << attn_scale << ",\n"
         << "  diffusion_conv_direct: " << (diffusion_conv_direct ? "true" : "false") << ",\n"
         << "  vae_conv_direct: " << (vae_conv_direct ? "true" : "false") << ",\n"
         << "  prediction: " << sd_prediction_name(prediction) << ",\n"
@@ -939,6 +981,8 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.enable_mmap                     = enable_mmap;
     sd_ctx_params.flash_attn                      = flash_attn;
     sd_ctx_params.diffusion_flash_attn            = diffusion_flash_attn;
+    sd_ctx_params.linear_scale                    = linear_scale;
+    sd_ctx_params.attn_scale                      = attn_scale;
     sd_ctx_params.tae_preview_only                = taesd_preview;
     sd_ctx_params.diffusion_conv_direct           = diffusion_conv_direct;
     sd_ctx_params.vae_conv_direct                 = vae_conv_direct;

examples/common/common.h

@@ -175,6 +175,8 @@ struct SDContextParams {
     lora_apply_mode_t lora_apply_mode = LORA_APPLY_AUTO;
 
     bool force_sdxl_vae_conv_scale = false;
+    float linear_scale             = 0.f;
+    float attn_scale               = 0.f;
 
     float flow_shift = INFINITY;
     ArgOptions get_options();

examples/server/README.md

@@ -129,3 +129,6 @@ For detailed command-line arguments, run:
 ```bash
 ./bin/sd-server -h
 ```
+
+For completely black or white images or videos, NaNs, and the `--linear-scale` /
+`--attn-scale` startup options, see [Troubleshooting](../../docs/troubleshooting.md).

include/stable-diffusion.h

@@ -92,6 +92,7 @@ enum prediction_t {
     FLUX_FLOW_PRED,
     SEFI_FLOW_PRED,
     MINIT2I_FLOW_PRED,
+    SENSENOVA_U1_FLOW_PRED,
     PREDICTION_COUNT
 };
 
@@ -240,6 +241,8 @@ 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
+    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
 } sd_ctx_params_t;
 
 typedef struct {
@@ -493,6 +496,9 @@ SD_API void free_sd_audio(sd_audio_t* audio);
 SD_API void sd_sample_params_init(sd_sample_params_t* sample_params);
 SD_API char* sd_sample_params_to_str(const sd_sample_params_t* sample_params);
 
+// Requires a loaded context; returns a static string owned by the library, or "Unknown".
+SD_API const char* sd_get_model_version_name(const sd_ctx_t* sd_ctx);
+
 SD_API enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx);
 SD_API enum scheduler_t sd_get_default_scheduler(const sd_ctx_t* sd_ctx, enum sample_method_t sample_method);
 

src/conditioning/conditioner.hpp

@@ -16,6 +16,7 @@
 #include "model/te/llm.hpp"
 #include "model/te/t5.hpp"
 #include "model_loader.h"
+#include "tokenizers/sensenova_u1_tokenizer.h"
 
 struct SDCondition {
     sd::Tensor<float> c_crossattn;
@@ -149,6 +150,7 @@ public:
     virtual void set_graph_cut_layer_split_backend_vram_limits(const std::vector<size_t>& limits) {}
     virtual void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {}
     virtual void set_flash_attention_enabled(bool enabled) = 0;
+    virtual void set_scale_overrides(float linear_scale, float attn_scale) {}
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual void runner_end() {}
 };
@@ -231,6 +233,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         }
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        text_model->set_scale_overrides(linear_scale, attn_scale);
+        if (sd_version_is_sdxl(version)) {
+            text_model2->set_scale_overrides(linear_scale, attn_scale);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         text_model->set_weight_adapter(adapter);
         if (sd_version_is_sdxl(version)) {
@@ -736,6 +745,18 @@ struct SD3CLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        if (clip_l) {
+            clip_l->set_scale_overrides(linear_scale, attn_scale);
+        }
+        if (clip_g) {
+            clip_g->set_scale_overrides(linear_scale, attn_scale);
+        }
+        if (t5) {
+            t5->set_scale_overrides(linear_scale, attn_scale);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (clip_l) {
             clip_l->set_weight_adapter(adapter);
@@ -1106,6 +1127,15 @@ struct FluxCLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        if (clip_l) {
+            clip_l->set_scale_overrides(linear_scale, attn_scale);
+        }
+        if (t5) {
+            t5->set_scale_overrides(linear_scale, attn_scale);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (clip_l) {
             clip_l->set_weight_adapter(adapter);
@@ -1368,6 +1398,12 @@ struct T5CLIPEmbedder : public Conditioner {
         }
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        if (t5) {
+            t5->set_scale_overrides(linear_scale, attn_scale);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (t5) {
             t5->set_weight_adapter(adapter);
@@ -1576,6 +1612,12 @@ struct MiniT2IConditioner : public Conditioner {
         }
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        if (t5) {
+            t5->set_scale_overrides(linear_scale, attn_scale);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (t5) {
             t5->set_weight_adapter(adapter);
@@ -1623,6 +1665,71 @@ struct MiniT2IConditioner : public Conditioner {
     }
 };
 
+struct SenseNovaU1Conditioner : public Conditioner {
+    static constexpr size_t kMaxPromptTokens = 12288;
+    SenseNovaU1Tokenizer tokenizer;
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        SD_UNUSED(tensors);
+    }
+
+    void set_flash_attention_enabled(bool enabled) override {
+        SD_UNUSED(enabled);
+    }
+
+    static std::string build_query(const std::string& text, bool is_negative) {
+        static const std::string kSystemMessage =
+            "You are an image generation and editing assistant that accurately understands and executes user intent.\n\n"
+            "You support two modes:\n\n1. Think Mode:\nIf the task requires reasoning, you MUST start with a "
+            "<think></think> block. Put all reasoning inside the block using plain text. DO NOT include any image tags. "
+            "Keep it reasonable and directly useful for producing the final image.\n\n2. Non-Think Mode:\nIf no reasoning "
+            "is needed, directly produce the final image.\n\nTask Types:\n\nA. Text-to-Image Generation:\n- Generate a "
+            "high-quality image based on the user's description.\n- Ensure visual clarity, semantic consistency, and "
+            "completeness.\n- DO NOT introduce elements that contradict or override the user's intent.\n\nB. Image Editing:\n"
+            "- Use the provided image(s) as input or reference for modification or transformation.\n- The result can be an "
+            "edited image or a new image based on the reference(s).\n- Preserve all unspecified attributes unless explicitly "
+            "changed.\n\nGeneral Rules:\n- For any visible text in the image, follow the language specified for the rendered "
+            "text in the user's description, not the language of the prompt. If no language is specified, use the user's input "
+            "language.";
+
+        std::string query;
+        if (!is_negative) {
+            query += "<|im_start|>system\n";
+            query += kSystemMessage;
+            query += "<|im_end|>\n";
+        }
+        query += "<|im_start|>user\n";
+        query += text;
+        query += "<|im_end|>\n<|im_start|>assistant\n";
+        query += is_negative ? "<img>" : "<think>\n\n</think>\n\n<img>";
+        return query;
+    }
+
+    SDCondition tokenize_condition(const std::string& text, bool is_negative) {
+        auto tokens = tokenizer.encode(build_query(text, is_negative));
+        if (tokens.empty() || tokens.size() > kMaxPromptTokens) {
+            LOG_ERROR("SenseNova U1.5 prompt token count %zu is outside [1, %zu]",
+                      tokens.size(),
+                      kMaxPromptTokens);
+            return {};
+        }
+
+        SDCondition result;
+        result.c_input_ids = sd::Tensor<int32_t>({static_cast<int64_t>(tokens.size())}, tokens);
+        return result;
+    }
+
+    SDCondition get_learned_condition(int n_threads,
+                                      const ConditionerParams& conditioner_params) override {
+        SD_UNUSED(n_threads);
+        return tokenize_condition(conditioner_params.text, false);
+    }
+
+    SDCondition get_unconditional_condition(const std::string& text) {
+        return tokenize_condition(text, true);
+    }
+};
+
 struct AnimaConditioner : public Conditioner {
     std::shared_ptr<BPETokenizer> qwen_tokenizer;
     T5UniGramTokenizer t5_tokenizer;
@@ -1672,6 +1779,10 @@ struct AnimaConditioner : public Conditioner {
         llm->set_flash_attention_enabled(enabled);
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        llm->set_scale_overrides(linear_scale, attn_scale);
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         llm->set_weight_adapter(adapter);
     }
@@ -1876,6 +1987,13 @@ struct LLMEmbedder : public Conditioner {
         }
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        llm->set_scale_overrides(linear_scale, attn_scale);
+        if (byt5) {
+            byt5->set_scale_overrides(linear_scale, attn_scale);
+        }
+    }
+
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (llm) {
             llm->set_weight_adapter(adapter);
@@ -2965,6 +3083,11 @@ struct LTXAVEmbedder : public Conditioner {
         projector->set_flash_attention_enabled(enabled);
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) override {
+        llm->set_scale_overrides(linear_scale, attn_scale);
+        projector->set_scale_overrides(linear_scale, attn_scale);
+    }
+
     void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
         llm->set_max_graph_vram_bytes(max_vram_bytes);
         projector->set_max_graph_vram_bytes(max_vram_bytes);

src/core/backend_fit.cpp

@@ -389,7 +389,7 @@ namespace sd::backend_fit {
             tiling_params.temporal_tiling = true;
             retry_mode                    = tiling_params.enabled ? "spatial+temporal" : "temporal";
         } else if (!tiling_params.enabled) {
-            tiling_params.enabled = true;
+            tiling_params.enabled    = true;
             tiling_params.rel_size_x = 0.5f;
             tiling_params.rel_size_y = 0.5f;
             if (tiling_params.tile_size_x <= 0) {

src/core/ggml_graph_cut.cpp

@@ -482,6 +482,16 @@ namespace sd::ggml_graph_cut {
         return ggml_nbytes(cache_src);
     }
 
+    static bool can_ignore_op_params(ggml_op op) {
+        // Exempt only parameters that cannot affect graph layout or backend allocation size.
+        switch (op) {
+            case GGML_OP_SCALE:
+                return true;
+            default:
+                return false;
+        }
+    }
+
     std::vector<uint64_t> graph_layout(ggml_cgraph* graph, bool include_bindings) {
         std::vector<const ggml_tensor*> tensors;
         std::unordered_map<const ggml_tensor*, size_t> indices;
@@ -530,8 +540,10 @@ namespace sd::ggml_graph_cut {
             for (auto source : tensor->src) {
                 signature.push_back(source == nullptr ? 0 : indices.at(source));
             }
-            for (int value : tensor->op_params) {
+            if (!can_ignore_op_params(tensor->op)) {
-                signature.push_back(static_cast<uint32_t>(value));
+                for (int value : tensor->op_params) {
+                    signature.push_back(static_cast<uint32_t>(value));
+                }
             }
         }
         return signature;

src/core/ggml_runner.cpp

@@ -2,6 +2,7 @@
 #include <map>
 #include <utility>
 
+#include "core/ggml_extend.h"
 #include "core/ggml_extend_backend.h"
 #include "core/ggml_runner.h"
 #include "core/ggml_tensor_utils.h"
@@ -11,6 +12,21 @@
 
 using namespace sd;
 
+ggml_tensor* ggml_ext_attention_ext(GGMLRunnerContext* ctx,
+                                    ggml_tensor* q,
+                                    ggml_tensor* k,
+                                    ggml_tensor* v,
+                                    int64_t n_head,
+                                    ggml_tensor* mask,
+                                    bool skip_reshape,
+                                    bool flash_attn,
+                                    float kv_scale) {
+    if (ctx->attn_scale > 0.f) {
+        kv_scale = ctx->attn_scale;
+    }
+    return ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, skip_reshape, flash_attn, kv_scale);
+}
+
 void GGMLRunner::alloc_params_ctx() {
     ggml_init_params params;
     params.mem_size   = static_cast<size_t>(MAX_PARAMS_TENSOR_NUM * ggml_tensor_overhead());
@@ -510,6 +526,8 @@ GGMLRunnerContext GGMLRunner::get_context() {
     runner_ctx.ggml_ctx              = compute_ctx;
     runner_ctx.backend               = runtime_backend;
     runner_ctx.flash_attn_enabled    = flash_attn_enabled;
+    runner_ctx.linear_scale          = linear_scale;
+    runner_ctx.attn_scale            = attn_scale;
     runner_ctx.conv2d_direct_enabled = conv2d_direct_enabled;
     runner_ctx.circular_x_enabled    = circular_x_enabled;
     runner_ctx.circular_y_enabled    = circular_y_enabled;

src/core/ggml_runner.h

@@ -68,6 +68,8 @@ struct GGMLRunnerContext {
     ggml_backend_t backend                                           = nullptr;
     ggml_context* ggml_ctx                                           = nullptr;
     bool flash_attn_enabled                                          = false;
+    float linear_scale                                               = 0.f;
+    float attn_scale                                                 = 0.f;
     bool conv2d_direct_enabled                                       = false;
     bool circular_x_enabled                                          = false;
     bool circular_y_enabled                                          = false;
@@ -113,6 +115,16 @@ struct GGMLRunnerContext {
     }
 };
 
+ggml_tensor* ggml_ext_attention_ext(GGMLRunnerContext* ctx,
+                                    ggml_tensor* q,
+                                    ggml_tensor* k,
+                                    ggml_tensor* v,
+                                    int64_t n_head,
+                                    ggml_tensor* mask = nullptr,
+                                    bool skip_reshape = false,
+                                    bool flash_attn   = false,
+                                    float kv_scale    = 1.f);
+
 struct GGMLRunner {
 private:
     std::map<ggml_backend_t, size_t> logged_compute_bytes_;
@@ -163,6 +175,8 @@ protected:
     const std::string final_result_name = "ggml_runner_final_result_tensor";
 
     bool flash_attn_enabled    = false;
+    float linear_scale         = 0.f;
+    float attn_scale           = 0.f;
     bool conv2d_direct_enabled = false;
     bool circular_x_enabled    = false;
     bool circular_y_enabled    = false;
@@ -323,6 +337,11 @@ public:
         flash_attn_enabled = enabled;
     }
 
+    void set_scale_overrides(float linear_scale, float attn_scale) {
+        this->linear_scale = linear_scale;
+        this->attn_scale   = attn_scale;
+    }
+
     void set_conv2d_direct_enabled(bool enabled) {
         conv2d_direct_enabled = enabled;
     }

src/extensions/photomaker_extension.cpp

@@ -135,6 +135,7 @@ struct PhotoMakerExtension : public GenerationExtension {
                                                            pm_version,
                                                            20.f,
                                                            ctx.model_manager);
+        pmid_model->set_scale_overrides(ctx.params->linear_scale, ctx.params->attn_scale);
         if (pm_version == PM_VERSION_2) {
             LOG_INFO("using PhotoMaker Version 2");
         }

src/model.h

@@ -57,6 +57,7 @@ enum SDVersion {
     VERSION_SEFI_IMAGE,
     VERSION_KREA2,
     VERSION_MAGE_FLOW,
+    VERSION_SENSENOVA_U1_5,
     VERSION_ESRGAN,
     VERSION_COUNT,
 };
@@ -237,6 +238,10 @@ static inline bool sd_version_is_mage_flow(SDVersion version) {
     return version == VERSION_MAGE_FLOW;
 }
 
+static inline bool sd_version_is_sensenova_u1(SDVersion version) {
+    return version == VERSION_SENSENOVA_U1_5;
+}
+
 static inline bool sd_version_uses_flux_vae(SDVersion version) {
     if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_boogu_image(version) || sd_version_is_longcat(version)) {
         return true;
@@ -295,7 +300,8 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_ideogram4(version) ||
         sd_version_is_sefi_image(version) ||
         sd_version_is_krea2(version) ||
-        sd_version_is_mage_flow(version)) {
+        sd_version_is_mage_flow(version) ||
+        sd_version_is_sensenova_u1(version)) {
         return true;
     }
     return false;

src/model/adapter/ip_adapter.hpp

@@ -95,7 +95,7 @@ namespace IPAdapter {
                 int64_t L          = kv->ne[1];
                 ggml_tensor* k     = ggml_cont(ctx->ggml_ctx, ggml_view_3d(ctx->ggml_ctx, kv, dim, L, N, kv->nb[1], kv->nb[2], 0));
                 ggml_tensor* v     = ggml_cont(ctx->ggml_ctx, ggml_view_3d(ctx->ggml_ctx, kv, dim, L, N, kv->nb[1], kv->nb[2], dim * kv->nb[0]));
-                ggml_tensor* attn  = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, heads, nullptr, false, false);
+                ggml_tensor* attn  = ggml_ext_attention_ext(ctx, q, k, v, heads, nullptr, false, false);
                 attn               = to_out->forward(ctx, attn);
                 latents            = ggml_add(ctx->ggml_ctx, latents, attn);
 

src/model/adapter/pulid.hpp

@@ -63,12 +63,11 @@ public:
         k              = ggml_cont(ctx->ggml_ctx, k);
         v              = ggml_cont(ctx->ggml_ctx, v);
 
-        ggml_tensor* attn_out = ggml_ext_attention_ext(
+        ggml_tensor* attn_out = ggml_ext_attention_ext(ctx,
-            ctx->ggml_ctx, ctx->backend,
+                                                       q, k, v,
-            q, k, v,
+                                                       heads,
-            heads,
+                                                       /*mask=*/nullptr,
-            /*mask=*/nullptr,
+                                                       /*diag_mask_inf=*/false);
-            /*diag_mask_inf=*/false);
 
         ggml_tensor* out = to_out->forward(ctx, attn_out);
         return out;

src/model/common/block.hpp

@@ -380,14 +380,14 @@ public:
         if (xtra_dim) {
             context->ne[0] = 320;  // reset dim to orig
         }
-        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, inner_dim]
+        x = ggml_ext_attention_ext(ctx, q, k, v, n_head, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, inner_dim]
 
         if (has_ip && ctx->ip_context != nullptr && ctx->ip_scale != 0.0f) {
             auto to_k_ip = std::dynamic_pointer_cast<Linear>(blocks["to_k_ip"]);
             auto to_v_ip = std::dynamic_pointer_cast<Linear>(blocks["to_v_ip"]);
             auto k_ip    = to_k_ip->forward(ctx, ctx->ip_context);
             auto v_ip    = to_v_ip->forward(ctx, ctx->ip_context);
-            auto x_ip    = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k_ip, v_ip, n_head, nullptr, false, ctx->flash_attn_enabled);
+            auto x_ip    = ggml_ext_attention_ext(ctx, q, k_ip, v_ip, n_head, nullptr, false, ctx->flash_attn_enabled);
             x            = ggml_add(ctx->ggml_ctx, x, ggml_scale(ctx->ggml_ctx, x_ip, ctx->ip_scale));
         }
 

src/model/common/ggml_block.hpp

@@ -206,6 +206,7 @@ public:
 
     ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         ggml_tensor* w            = params["weight"];
+        const float scale         = ctx->linear_scale > 0.f ? ctx->linear_scale : this->scale;
         ggml_tensor* weight_scale = has_weight_scale ? params["weight_scale"] : nullptr;
         if (w->type == GGML_TYPE_F8_E4M3 || w->type == GGML_TYPE_F8_E5M2) {
             bool supports_fp8_matmul = false;
@@ -309,6 +310,7 @@ public:
 __STATIC_INLINE__ bool support_get_rows(ggml_type wtype) {
     switch (wtype) {
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q5_0:
@@ -869,7 +871,7 @@ public:
             v = v_proj->forward(ctx, x);
         }
 
-        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, false);  // [N, n_token, embed_dim]
+        x = ggml_ext_attention_ext(ctx, q, k, v, n_head, mask, false);  // [N, n_token, embed_dim]
 
         x = out_proj->forward(ctx, x);  // [N, n_token, embed_dim]
         return x;

src/model/common/rope.hpp

@@ -1024,7 +1024,7 @@ namespace Rope {
         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]
 
-        auto x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*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;
     }
 };  // namespace Rope

src/model/diffusion/anima.hpp

@@ -237,8 +237,7 @@ namespace Anima {
                 }
                 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);
-                attn_out    = ggml_ext_attention_ext(ctx->ggml_ctx,
+                attn_out    = ggml_ext_attention_ext(ctx,
-                                                     ctx->backend,
                                                      q_rope,
                                                      k_rope,
                                                      v4,
@@ -249,8 +248,7 @@ namespace Anima {
             } else {
                 auto q_flat = ggml_reshape_3d(ctx->ggml_ctx, q4, head_dim * num_heads, L_q, N);
                 auto k_flat = ggml_reshape_3d(ctx->ggml_ctx, k4, head_dim * num_heads, L_k, N);
-                attn_out    = ggml_ext_attention_ext(ctx->ggml_ctx,
+                attn_out    = ggml_ext_attention_ext(ctx,
-                                                     ctx->backend,
                                                      q_flat,
                                                      k_flat,
                                                      v,

src/model/diffusion/animatediff.hpp

@@ -61,7 +61,7 @@ namespace AnimateDiff {
             auto k = to_k->forward(ctx, x_pe);
             auto v = to_v->forward(ctx, x_pe);
 
-            auto a = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, (int)num_heads, nullptr, false);
+            auto a = ggml_ext_attention_ext(ctx, q, k, v, (int)num_heads, nullptr, false);
             return to_out->forward(ctx, a);
         }
     };

src/model/diffusion/ernie_image.hpp

@@ -183,7 +183,7 @@ namespace ErnieImage {
             k = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, k, 0, 2, 1, 3));  // [N, heads, S, head_dim]
             k = ggml_reshape_3d(ctx->ggml_ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]);
 
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, true, ctx->flash_attn_enabled);  // [N, S, hidden_size]
+            x = ggml_ext_attention_ext(ctx, q, k, v, num_heads, attention_mask, true, ctx->flash_attn_enabled);  // [N, S, hidden_size]
             x = to_out_0->forward(ctx, x);
             return x;
         }

src/model/diffusion/hidream_o1.hpp

@@ -504,6 +504,10 @@ namespace HiDreamO1 {
             vision_runner->set_flash_attention_enabled(enabled);
         }
 
+        void set_scale_overrides(float linear_scale, float attn_scale) override {
+            vision_runner->set_scale_overrides(linear_scale, attn_scale);
+        }
+
         void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
             vision_runner->set_weight_adapter(adapter);
         }

src/model/diffusion/hunyuan.hpp

@@ -54,7 +54,7 @@ namespace Hunyuan {
             auto k       = qkv_vec[1];
             auto v       = qkv_vec[2];
 
-            auto attn_out = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, mask, false, ctx->flash_attn_enabled);
+            auto attn_out = ggml_ext_attention_ext(ctx, q, k, v, num_heads, mask, false, ctx->flash_attn_enabled);
             attn_out      = self_attn_proj->forward(ctx, attn_out);
 
             // adaLN_modulation

src/model/diffusion/krea2.hpp

@@ -232,8 +232,7 @@ namespace Krea2 {
             q          = ggml_reshape_3d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, q), head_dim_ * heads, Lq, N);
             k          = ggml_reshape_3d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, k), head_dim_ * kv_heads, Lk, N);
             v          = ggml_reshape_3d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, v), head_dim_ * kv_heads, Lk, N);
-            return ggml_ext_attention_ext(ctx->ggml_ctx,
+            return ggml_ext_attention_ext(ctx,
-                                          ctx->backend,
                                           q,
                                           k,
                                           v,

src/model/diffusion/ltxv.hpp

@@ -709,8 +709,7 @@ namespace LTXV {
                 k = apply_hidden_rope(ctx->ggml_ctx, k, k_pe, heads, dim_head, rope_interleaved);
             }
 
-            auto out = ggml_ext_attention_ext(ctx->ggml_ctx,
+            auto out = ggml_ext_attention_ext(ctx,
-                                              ctx->backend,
                                               q,
                                               k,
                                               v,

src/model/diffusion/minimax_h3.hpp

@@ -215,8 +215,7 @@ namespace MiniMaxH3 {
                 q = attention_layout(ctx->ggml_ctx, q);
                 k = attention_layout(ctx->ggml_ctx, k);
             }
-            auto out = ggml_ext_attention_ext(ctx->ggml_ctx,
+            auto out = ggml_ext_attention_ext(ctx,
-                                              ctx->backend,
                                               q,
                                               k,
                                               v,

src/model/diffusion/mmdit.hpp

@@ -365,8 +365,8 @@ public:
     ggml_tensor* forward(GGMLRunnerContext* ctx,
                          ggml_tensor* x) {
         auto qkv = pre_attention(ctx, x);
-        x        = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+        x        = ggml_ext_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
-        x        = post_attention(ctx, x);                                                                                                           // [N, n_token, dim]
+        x        = post_attention(ctx, x);                                                                                   // [N, n_token, dim]
         return x;
     }
 };
@@ -587,8 +587,8 @@ public:
             auto qkv2          = std::get<1>(qkv_intermediates);
             auto intermediates = std::get<2>(qkv_intermediates);
 
-            auto attn_out  = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);     // [N, n_token, dim]
+            auto attn_out  = ggml_ext_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);     // [N, n_token, dim]
-            auto attn2_out = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv2[0], qkv2[1], qkv2[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+            auto attn2_out = ggml_ext_attention_ext(ctx, qkv2[0], qkv2[1], qkv2[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
             x              = post_attention_x(ctx,
                                               attn_out,
                                               attn2_out,
@@ -604,7 +604,7 @@ public:
             auto qkv               = qkv_intermediates.first;
             auto intermediates     = qkv_intermediates.second;
 
-            auto attn_out = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+            auto attn_out = ggml_ext_attention_ext(ctx, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
             x             = post_attention(ctx,
                                            attn_out,
                                            intermediates[0],
@@ -648,7 +648,7 @@ block_mixing(GGMLRunnerContext* ctx,
         qkv.push_back(ggml_concat(ctx->ggml_ctx, context_qkv[i], x_qkv[i], 1));
     }
 
-    auto attn = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], x_block->num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_context + n_token, hidden_size]
+    auto attn = ggml_ext_attention_ext(ctx, qkv[0], qkv[1], qkv[2], x_block->num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_context + n_token, hidden_size]
 
     auto context_attn = ggml_view_3d(ctx->ggml_ctx,
                                      attn,
@@ -680,7 +680,7 @@ block_mixing(GGMLRunnerContext* ctx,
     }
 
     if (x_block->self_attn) {
-        auto attn2 = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, x_qkv2[0], x_qkv2[1], x_qkv2[2], x_block->num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, hidden_size]
+        auto attn2 = ggml_ext_attention_ext(ctx, x_qkv2[0], x_qkv2[1], x_qkv2[2], x_block->num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, hidden_size]
 
         x = x_block->post_attention_x(ctx,
                                       x_attn,

src/model/diffusion/model.hpp

@@ -114,6 +114,10 @@ struct MiniT2IDiffusionExtra {
     const sd::Tensor<float>* mask = nullptr;
 };
 
+struct SenseNovaU1DiffusionExtra {
+    const sd::Tensor<int32_t>* input_ids = nullptr;
+};
+
 struct HunyuanVideoDiffusionExtra {
     const sd::Tensor<float>* guidance   = nullptr;
     const sd::Tensor<float>* byt5       = nullptr;
@@ -131,6 +135,7 @@ using DiffusionExtraParams = std::variant<std::monostate,
                                           LTXAVDiffusionExtra,
                                           MiniMaxH3DiffusionExtra,
                                           MiniT2IDiffusionExtra,
+                                          SenseNovaU1DiffusionExtra,
                                           HunyuanVideoDiffusionExtra>;
 
 struct DiffusionParams {

src/model/diffusion/sensenova_u1.h

@@ -0,0 +1,846 @@
+#ifndef __SD_MODEL_DIFFUSION_SENSENOVA_U1_H__
+#define __SD_MODEL_DIFFUSION_SENSENOVA_U1_H__
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "core/ggml_extend.h"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model/te/llm.hpp"
+#include "model_loader.h"
+
+namespace SenseNovaU1 {
+    constexpr int SENSENOVA_U1_GRAPH_SIZE = 327680;
+
+    struct SenseNovaU1Config {
+        int64_t hidden_size                  = 4096;
+        int64_t intermediate_size            = 12288;
+        int64_t num_layers                   = 42;
+        int64_t num_heads                    = 32;
+        int64_t num_kv_heads                 = 8;
+        int64_t head_dim                     = 128;
+        int64_t vocab_size                   = 151936;
+        int64_t max_position_embeddings      = 262144;
+        int64_t max_position_embeddings_hw   = 10000;
+        int64_t vision_hidden_size           = 1024;
+        int64_t patch_size                   = 16;
+        int64_t vision_downsample_factor     = 2;
+        int64_t in_channels                  = 3;
+        int64_t timestep_embedding_size      = 256;
+        float rms_norm_eps                   = 1e-6f;
+        float rope_theta                     = 5000000.f;
+        float rope_theta_hw                  = 10000.f;
+        float noise_scale_base_image_seq_len = 64.f;
+        float noise_scale_max_value          = 16.f;
+        float t_eps                          = 0.02f;
+        bool add_noise_scale_embedding       = true;
+
+        int64_t image_token_stride() const {
+            return patch_size * vision_downsample_factor;
+        }
+
+        static SenseNovaU1Config detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                                     const std::string& prefix) {
+            SenseNovaU1Config config;
+            config.num_layers      = 0;
+            const std::string root = prefix.empty() ? "" : prefix + ".";
+
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, root)) {
+                    continue;
+                }
+                if (ends_with(name, "language_model.model.embed_tokens.weight") && tensor_storage.n_dims == 2) {
+                    config.hidden_size = tensor_storage.ne[0];
+                    config.vocab_size  = tensor_storage.ne[1];
+                } else if (ends_with(name, "language_model.model.layers.0.mlp.gate_proj.weight") && tensor_storage.n_dims == 2) {
+                    config.intermediate_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "language_model.model.layers.0.self_attn.q_proj.weight") && tensor_storage.n_dims == 2) {
+                    config.num_heads = tensor_storage.ne[1] / config.head_dim;
+                } else if (ends_with(name, "language_model.model.layers.0.self_attn.k_proj.weight") && tensor_storage.n_dims == 2) {
+                    config.num_kv_heads = tensor_storage.ne[1] / config.head_dim;
+                } else if (ends_with(name, "fm_modules.vision_model_mot_gen.embeddings.patch_embedding.weight") && tensor_storage.n_dims == 4) {
+                    config.patch_size         = tensor_storage.ne[0];
+                    config.in_channels        = tensor_storage.ne[2];
+                    config.vision_hidden_size = tensor_storage.ne[3];
+                } else if (ends_with(name, "fm_modules.vision_model_mot_gen.embeddings.dense_embedding.weight") && tensor_storage.n_dims == 4) {
+                    config.vision_downsample_factor = tensor_storage.ne[0];
+                }
+
+                const std::string layer_prefix = root + "language_model.model.layers.";
+                if (starts_with(name, layer_prefix)) {
+                    const char* index_begin = name.c_str() + layer_prefix.size();
+                    config.num_layers       = std::max<int64_t>(config.num_layers, std::strtoll(index_begin, nullptr, 10) + 1);
+                }
+            }
+
+            if (config.num_layers == 0) {
+                config.num_layers = 42;
+            }
+            config.add_noise_scale_embedding = tensor_storage_map.find(root + "fm_modules.noise_scale_embedder.mlp.0.weight") != tensor_storage_map.end();
+
+            LOG_DEBUG("sensenova-u1.5: layers=%" PRId64 ", hidden=%" PRId64 ", intermediate=%" PRId64 ", heads=%" PRId64 ", kv_heads=%" PRId64 ", patch=%" PRId64 "x%" PRId64,
+                      config.num_layers,
+                      config.hidden_size,
+                      config.intermediate_size,
+                      config.num_heads,
+                      config.num_kv_heads,
+                      config.patch_size,
+                      config.vision_downsample_factor);
+            return config;
+        }
+    };
+
+    class StorageConv2d : public Conv2d {
+    protected:
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            this->prefix     = prefix;
+            ggml_type wtype  = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F16);
+            params["weight"] = ggml_new_tensor_4d(ctx,
+                                                  wtype,
+                                                  kernel_size.second,
+                                                  kernel_size.first,
+                                                  in_channels,
+                                                  out_channels);
+            if (bias) {
+                params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
+            }
+        }
+
+    public:
+        StorageConv2d(int64_t in_channels,
+                      int64_t out_channels,
+                      std::pair<int, int> kernel_size,
+                      std::pair<int, int> stride  = {1, 1},
+                      std::pair<int, int> padding = {0, 0},
+                      bool bias                   = true)
+            : Conv2d(in_channels,
+                     out_channels,
+                     kernel_size,
+                     stride,
+                     padding,
+                     {1, 1},
+                     bias) {}
+    };
+
+    struct TimestepEmbedder : public GGMLBlock {
+        int64_t frequency_embedding_size;
+
+        TimestepEmbedder(int64_t hidden_size, int64_t frequency_embedding_size = 256)
+            : frequency_embedding_size(frequency_embedding_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* timesteps) {
+            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
+            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
+            auto x     = ggml_ext_timestep_embedding(ctx->ggml_ctx,
+                                                     timesteps,
+                                                     static_cast<int>(frequency_embedding_size),
+                                                     10000.f,
+                                                     1.f);
+            x          = mlp_0->forward(ctx, x);
+            x          = ggml_silu_inplace(ctx->ggml_ctx, x);
+            return mlp_2->forward(ctx, x);
+        }
+    };
+
+    inline ggml_tensor* apply_vision_rope(GGMLRunnerContext* ctx,
+                                          ggml_tensor* x,
+                                          ggml_tensor* position_x,
+                                          ggml_tensor* position_y,
+                                          float theta,
+                                          int max_position) {
+        GGML_ASSERT(x->ne[0] % 2 == 0);
+        // ggml_rope_ext addresses positions through ne[2]. The vision
+        // embeddings arrive as [hidden, tokens, batch], so add the singleton
+        // head axis used by the RoPE kernel: [hidden, 1, tokens, batch].
+        x                  = ggml_reshape_4d(ctx->ggml_ctx, x, x->ne[0], 1, x->ne[1], x->ne[2]);
+        const int64_t half = x->ne[0] / 2;
+        auto x_part        = ggml_ext_slice(ctx->ggml_ctx, x, 0, 0, half);
+        auto y_part        = ggml_ext_slice(ctx->ggml_ctx, x, 0, half, x->ne[0]);
+        x_part             = ggml_rope_ext(ctx->ggml_ctx,
+                                           x_part,
+                                           position_x,
+                                           nullptr,
+                                           static_cast<int>(half),
+                                           GGML_ROPE_TYPE_NORMAL,
+                                           max_position,
+                                           theta,
+                                           1.f,
+                                           0.f,
+                                           1.f,
+                                           32.f,
+                                           1.f);
+        y_part             = ggml_rope_ext(ctx->ggml_ctx,
+                                           y_part,
+                                           position_y,
+                                           nullptr,
+                                           static_cast<int>(half),
+                                           GGML_ROPE_TYPE_NORMAL,
+                                           max_position,
+                                           theta,
+                                           1.f,
+                                           0.f,
+                                           1.f,
+                                           32.f,
+                                           1.f);
+        return ggml_concat(ctx->ggml_ctx, x_part, y_part, 0);
+    }
+
+    struct VisionEmbeddings : public GGMLBlock {
+        SenseNovaU1Config config;
+
+        explicit VisionEmbeddings(const SenseNovaU1Config& config)
+            : config(config) {
+            blocks["patch_embedding"] = std::make_shared<StorageConv2d>(config.in_channels,
+                                                                        config.vision_hidden_size,
+                                                                        std::pair<int, int>{static_cast<int>(config.patch_size), static_cast<int>(config.patch_size)},
+                                                                        std::pair<int, int>{static_cast<int>(config.patch_size), static_cast<int>(config.patch_size)},
+                                                                        std::pair<int, int>{0, 0},
+                                                                        true);
+            blocks["dense_embedding"] = std::make_shared<StorageConv2d>(config.vision_hidden_size,
+                                                                        config.hidden_size,
+                                                                        std::pair<int, int>{static_cast<int>(config.vision_downsample_factor), static_cast<int>(config.vision_downsample_factor)},
+                                                                        std::pair<int, int>{static_cast<int>(config.vision_downsample_factor), static_cast<int>(config.vision_downsample_factor)},
+                                                                        std::pair<int, int>{0, 0},
+                                                                        true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* image,
+                             ggml_tensor* position_x,
+                             ggml_tensor* position_y) {
+            auto patch_embedding = std::dynamic_pointer_cast<StorageConv2d>(blocks["patch_embedding"]);
+            auto dense_embedding = std::dynamic_pointer_cast<StorageConv2d>(blocks["dense_embedding"]);
+
+            auto x = patch_embedding->forward(ctx, image);
+            x      = ggml_gelu_erf(ctx->ggml_ctx, x);
+
+            const int64_t grid_w = x->ne[0];
+            const int64_t grid_h = x->ne[1];
+            const int64_t batch  = x->ne[3];
+            x                    = ggml_reshape_3d(ctx->ggml_ctx, x, grid_w * grid_h, x->ne[2], batch);
+            x                    = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));
+            x                    = apply_vision_rope(ctx,
+                                                     x,
+                                                     position_x,
+                                                     position_y,
+                                                     config.rope_theta_hw,
+                                                     static_cast<int>(config.max_position_embeddings_hw));
+            x                    = ggml_reshape_4d(ctx->ggml_ctx, x, config.vision_hidden_size, grid_w, grid_h, batch);
+            x                    = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));
+            x                    = dense_embedding->forward(ctx, x);
+
+            const int64_t token_w = x->ne[0];
+            const int64_t token_h = x->ne[1];
+            x                     = ggml_reshape_3d(ctx->ggml_ctx, x, token_w * token_h, x->ne[2], x->ne[3]);
+            return ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));
+        }
+    };
+
+    inline ggml_tensor* pixel_shuffle(GGMLRunnerContext* ctx,
+                                      ggml_tensor* x,
+                                      int upscale_factor) {
+        GGML_ASSERT(upscale_factor > 0);
+        const int64_t h = x->ne[1];
+        const int64_t w = x->ne[0];
+        GGML_ASSERT(x->ne[2] % (upscale_factor * upscale_factor) == 0);
+        x = ggml_ext_cont(ctx->ggml_ctx,
+                          ggml_ext_torch_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));
+        x = ggml_reshape_3d(ctx->ggml_ctx, x, x->ne[0], x->ne[1] * x->ne[2], x->ne[3]);
+        return DiT::unpatchify(ctx->ggml_ctx, x, h, w, upscale_factor, upscale_factor, true);
+    }
+
+    struct PixelDecoder : public GGMLBlock {
+        explicit PixelDecoder(const SenseNovaU1Config& config) {
+            blocks["conv1"] = std::make_shared<StorageConv2d>(config.hidden_size / 4,
+                                                              1024,
+                                                              std::pair<int, int>{3, 3},
+                                                              std::pair<int, int>{1, 1},
+                                                              std::pair<int, int>{1, 1},
+                                                              true);
+            blocks["conv2"] = std::make_shared<StorageConv2d>(256,
+                                                              192,
+                                                              std::pair<int, int>{3, 3},
+                                                              std::pair<int, int>{1, 1},
+                                                              std::pair<int, int>{1, 1},
+                                                              true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto conv1 = std::dynamic_pointer_cast<StorageConv2d>(blocks["conv1"]);
+            auto conv2 = std::dynamic_pointer_cast<StorageConv2d>(blocks["conv2"]);
+            x          = pixel_shuffle(ctx, x, 2);
+            x          = conv1->forward(ctx, x);
+            x          = ggml_gelu_erf(ctx->ggml_ctx, x);
+            x          = pixel_shuffle(ctx, x, 2);
+            x          = conv2->forward(ctx, x);
+            return pixel_shuffle(ctx, x, 8);
+        }
+    };
+
+    enum class Branch {
+        UNDERSTANDING,
+        GENERATION,
+    };
+
+    struct Attention : public GGMLBlock {
+        SenseNovaU1Config config;
+        int layer_index;
+
+        Attention(const SenseNovaU1Config& config, int layer_index)
+            : config(config), layer_index(layer_index) {
+            blocks["q_proj"]         = std::make_shared<Linear>(config.hidden_size, config.num_heads * config.head_dim, false);
+            blocks["k_proj"]         = std::make_shared<Linear>(config.hidden_size, config.num_kv_heads * config.head_dim, false);
+            blocks["v_proj"]         = std::make_shared<Linear>(config.hidden_size, config.num_kv_heads * config.head_dim, false);
+            blocks["o_proj"]         = std::make_shared<Linear>(config.num_heads * config.head_dim, config.hidden_size, false);
+            blocks["q_proj_mot_gen"] = std::make_shared<Linear>(config.hidden_size, config.num_heads * config.head_dim, false);
+            blocks["k_proj_mot_gen"] = std::make_shared<Linear>(config.hidden_size, config.num_kv_heads * config.head_dim, false);
+            blocks["v_proj_mot_gen"] = std::make_shared<Linear>(config.hidden_size, config.num_kv_heads * config.head_dim, false);
+            blocks["o_proj_mot_gen"] = std::make_shared<Linear>(config.num_heads * config.head_dim, config.hidden_size, false);
+
+            const int64_t axis_dim      = config.head_dim / 2;
+            blocks["q_norm"]            = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+            blocks["k_norm"]            = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+            blocks["q_norm_hw"]         = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+            blocks["k_norm_hw"]         = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+            blocks["q_norm_mot_gen"]    = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+            blocks["k_norm_mot_gen"]    = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+            blocks["q_norm_hw_mot_gen"] = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+            blocks["k_norm_hw_mot_gen"] = std::make_shared<LLM::LLMRMSNorm>(axis_dim, config.rms_norm_eps);
+        }
+
+        ggml_tensor* apply_axis_rope(GGMLRunnerContext* ctx,
+                                     ggml_tensor* x,
+                                     ggml_tensor* positions,
+                                     int dimensions,
+                                     float theta,
+                                     int max_position) {
+            return ggml_rope_ext(ctx->ggml_ctx,
+                                 x,
+                                 positions,
+                                 nullptr,
+                                 dimensions,
+                                 GGML_ROPE_TYPE_NEOX,
+                                 max_position,
+                                 theta,
+                                 1.f,
+                                 0.f,
+                                 1.f,
+                                 32.f,
+                                 1.f);
+        }
+
+        ggml_tensor* normalize_and_rotate(GGMLRunnerContext* ctx,
+                                          ggml_tensor* x,
+                                          ggml_tensor* position_t,
+                                          ggml_tensor* position_h,
+                                          ggml_tensor* position_w,
+                                          const std::string& norm_name,
+                                          const std::string& norm_hw_name) {
+            const int64_t temporal_dim = config.head_dim / 2;
+            const int64_t spatial_dim  = config.head_dim - temporal_dim;
+            const int64_t axis_dim     = spatial_dim / 2;
+
+            auto temporal = ggml_ext_slice(ctx->ggml_ctx, x, 0, 0, temporal_dim);
+            auto spatial  = ggml_ext_slice(ctx->ggml_ctx, x, 0, temporal_dim, config.head_dim);
+            temporal      = std::dynamic_pointer_cast<LLM::LLMRMSNorm>(blocks[norm_name])->forward(ctx, temporal);
+            spatial       = std::dynamic_pointer_cast<LLM::LLMRMSNorm>(blocks[norm_hw_name])->forward(ctx, spatial);
+
+            auto height = ggml_ext_slice(ctx->ggml_ctx, spatial, 0, 0, axis_dim);
+            auto width  = ggml_ext_slice(ctx->ggml_ctx, spatial, 0, axis_dim, spatial_dim);
+            temporal    = apply_axis_rope(ctx,
+                                          temporal,
+                                          position_t,
+                                          static_cast<int>(temporal_dim),
+                                          config.rope_theta,
+                                          static_cast<int>(config.max_position_embeddings));
+            height      = apply_axis_rope(ctx,
+                                          height,
+                                          position_h,
+                                          static_cast<int>(axis_dim),
+                                          config.rope_theta_hw,
+                                          static_cast<int>(config.max_position_embeddings_hw));
+            width       = apply_axis_rope(ctx,
+                                          width,
+                                          position_w,
+                                          static_cast<int>(axis_dim),
+                                          config.rope_theta_hw,
+                                          static_cast<int>(config.max_position_embeddings_hw));
+            return ggml_concat(ctx->ggml_ctx,
+                               ggml_concat(ctx->ggml_ctx, temporal, height, 0),
+                               width,
+                               0);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* position_t,
+                             ggml_tensor* position_h,
+                             ggml_tensor* position_w,
+                             ggml_tensor* attention_mask,
+                             Branch branch,
+                             const std::string& cache_prefix) {
+            const bool generation    = branch == Branch::GENERATION;
+            const std::string suffix = generation ? "_mot_gen" : "";
+            auto q_proj              = std::dynamic_pointer_cast<Linear>(blocks["q_proj" + suffix]);
+            auto k_proj              = std::dynamic_pointer_cast<Linear>(blocks["k_proj" + suffix]);
+            auto v_proj              = std::dynamic_pointer_cast<Linear>(blocks["v_proj" + suffix]);
+            auto o_proj              = std::dynamic_pointer_cast<Linear>(blocks["o_proj" + suffix]);
+
+            const int64_t n_tokens = x->ne[1];
+            const int64_t batch    = x->ne[2];
+            auto q                 = ggml_reshape_4d(ctx->ggml_ctx,
+                                                     q_proj->forward(ctx, x),
+                                                     config.head_dim,
+                                                     config.num_heads,
+                                                     n_tokens,
+                                                     batch);
+            auto k                 = ggml_reshape_4d(ctx->ggml_ctx,
+                                                     k_proj->forward(ctx, x),
+                                                     config.head_dim,
+                                                     config.num_kv_heads,
+                                                     n_tokens,
+                                                     batch);
+            auto v                 = ggml_reshape_4d(ctx->ggml_ctx,
+                                                     v_proj->forward(ctx, x),
+                                                     config.head_dim,
+                                                     config.num_kv_heads,
+                                                     n_tokens,
+                                                     batch);
+
+            q = normalize_and_rotate(ctx,
+                                     q,
+                                     position_t,
+                                     position_h,
+                                     position_w,
+                                     "q_norm" + suffix,
+                                     "q_norm_hw" + suffix);
+            k = normalize_and_rotate(ctx,
+                                     k,
+                                     position_t,
+                                     position_h,
+                                     position_w,
+                                     "k_norm" + suffix,
+                                     "k_norm_hw" + suffix);
+
+            const std::string layer_cache = cache_prefix + "." + std::to_string(layer_index);
+            if (generation) {
+                auto prefix_k = ctx->load_cache_tensor(layer_cache + ".k");
+                auto prefix_v = ctx->load_cache_tensor(layer_cache + ".v");
+                GGML_ASSERT(prefix_k != nullptr && prefix_v != nullptr);
+                k = ggml_concat(ctx->ggml_ctx, prefix_k, k, 2);
+                v = ggml_concat(ctx->ggml_ctx, prefix_v, v, 2);
+            } else {
+                // Keep dedicated graph outputs alive until the runner copies them
+                // into its persistent cache buffer after graph execution.
+                auto cache_k = ggml_dup_tensor(ctx->ggml_ctx, k);
+                cache_k      = ggml_cpy(ctx->ggml_ctx, k, cache_k);
+                ggml_set_output(cache_k);
+                auto cache_v = ggml_dup_tensor(ctx->ggml_ctx, v);
+                cache_v      = ggml_cpy(ctx->ggml_ctx, v, cache_v);
+                ggml_set_output(cache_v);
+                ctx->persist_cache_tensor(layer_cache + ".k", cache_k);
+                ctx->persist_cache_tensor(layer_cache + ".v", cache_v);
+            }
+
+            q = ggml_cont(ctx->ggml_ctx,
+                          ggml_ext_torch_permute(ctx->ggml_ctx, q, 0, 2, 1, 3));
+            q = ggml_reshape_3d(ctx->ggml_ctx, q, q->ne[0], q->ne[1], q->ne[2] * q->ne[3]);
+            k = ggml_cont(ctx->ggml_ctx,
+                          ggml_ext_torch_permute(ctx->ggml_ctx, k, 0, 2, 1, 3));
+            k = ggml_reshape_3d(ctx->ggml_ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]);
+
+            auto out = ggml_ext_attention_ext(ctx->ggml_ctx,
+                                              ctx->backend,
+                                              q,
+                                              k,
+                                              v,
+                                              config.num_heads,
+                                              attention_mask,
+                                              true,
+                                              ctx->flash_attn_enabled);
+            return o_proj->forward(ctx, out);
+        }
+    };
+
+    struct TransformerBlock : public GGMLBlock {
+        TransformerBlock(const SenseNovaU1Config& config, int layer_index) {
+            blocks["self_attn"]                        = std::make_shared<Attention>(config, layer_index);
+            blocks["mlp"]                              = std::make_shared<LLM::MLP>(config.hidden_size, config.intermediate_size, false);
+            blocks["mlp_mot_gen"]                      = std::make_shared<LLM::MLP>(config.hidden_size, config.intermediate_size, false);
+            blocks["input_layernorm"]                  = std::make_shared<LLM::LLMRMSNorm>(config.hidden_size, config.rms_norm_eps);
+            blocks["input_layernorm_mot_gen"]          = std::make_shared<LLM::LLMRMSNorm>(config.hidden_size, config.rms_norm_eps);
+            blocks["post_attention_layernorm"]         = std::make_shared<LLM::LLMRMSNorm>(config.hidden_size, config.rms_norm_eps);
+            blocks["post_attention_layernorm_mot_gen"] = std::make_shared<LLM::LLMRMSNorm>(config.hidden_size, config.rms_norm_eps);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* position_t,
+                             ggml_tensor* position_h,
+                             ggml_tensor* position_w,
+                             ggml_tensor* attention_mask,
+                             Branch branch,
+                             const std::string& cache_prefix) {
+            const bool generation = branch == Branch::GENERATION;
+            auto input_norm       = std::dynamic_pointer_cast<LLM::LLMRMSNorm>(
+                blocks[generation ? "input_layernorm_mot_gen" : "input_layernorm"]);
+            auto post_norm = std::dynamic_pointer_cast<LLM::LLMRMSNorm>(
+                blocks[generation ? "post_attention_layernorm_mot_gen" : "post_attention_layernorm"]);
+            auto attention = std::dynamic_pointer_cast<Attention>(blocks["self_attn"]);
+            auto mlp       = std::dynamic_pointer_cast<LLM::MLP>(blocks[generation ? "mlp_mot_gen" : "mlp"]);
+
+            auto residual = x;
+            x             = input_norm->forward(ctx, x);
+            x             = attention->forward(ctx,
+                                               x,
+                                               position_t,
+                                               position_h,
+                                               position_w,
+                                               attention_mask,
+                                               branch,
+                                               cache_prefix);
+            x             = ggml_add_inplace(ctx->ggml_ctx, x, residual);
+
+            residual = x;
+            x        = post_norm->forward(ctx, x);
+            x        = mlp->forward(ctx, x);
+            return ggml_add_inplace(ctx->ggml_ctx, x, residual);
+        }
+    };
+
+    struct TextModel : public GGMLBlock {
+        SenseNovaU1Config config;
+
+        explicit TextModel(const SenseNovaU1Config& config)
+            : config(config) {
+            blocks["embed_tokens"] = std::make_shared<Embedding>(config.vocab_size, config.hidden_size);
+            for (int i = 0; i < config.num_layers; ++i) {
+                blocks["layers." + std::to_string(i)] = std::make_shared<TransformerBlock>(config, i);
+            }
+            blocks["norm"]         = std::make_shared<LLM::LLMRMSNorm>(config.hidden_size, config.rms_norm_eps);
+            blocks["norm_mot_gen"] = std::make_shared<LLM::LLMRMSNorm>(config.hidden_size, config.rms_norm_eps);
+        }
+
+        ggml_tensor* embed(GGMLRunnerContext* ctx, ggml_tensor* input_ids) {
+            return std::dynamic_pointer_cast<Embedding>(blocks["embed_tokens"])->forward(ctx, input_ids);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* position_t,
+                             ggml_tensor* position_h,
+                             ggml_tensor* position_w,
+                             ggml_tensor* attention_mask,
+                             Branch branch,
+                             const std::string& cache_prefix) {
+            for (int i = 0; i < config.num_layers; ++i) {
+                auto layer = std::dynamic_pointer_cast<TransformerBlock>(blocks["layers." + std::to_string(i)]);
+                x          = layer->forward(ctx,
+                                            x,
+                                            position_t,
+                                            position_h,
+                                            position_w,
+                                            attention_mask,
+                                            branch,
+                                            cache_prefix);
+            }
+            auto norm = std::dynamic_pointer_cast<LLM::LLMRMSNorm>(
+                blocks[branch == Branch::GENERATION ? "norm_mot_gen" : "norm"]);
+            return norm->forward(ctx, x);
+        }
+    };
+
+    struct SenseNovaU1Model : public GGMLBlock {
+        SenseNovaU1Config config;
+
+        explicit SenseNovaU1Model(const SenseNovaU1Config& config)
+            : config(config) {
+            blocks["language_model.model"]                       = std::make_shared<TextModel>(config);
+            blocks["fm_modules.vision_model_mot_gen.embeddings"] = std::make_shared<VisionEmbeddings>(config);
+            blocks["fm_modules.timestep_embedder"]               = std::make_shared<TimestepEmbedder>(config.hidden_size,
+                                                                                        config.timestep_embedding_size);
+            if (config.add_noise_scale_embedding) {
+                blocks["fm_modules.noise_scale_embedder"] = std::make_shared<TimestepEmbedder>(config.hidden_size,
+                                                                                               config.timestep_embedding_size);
+            }
+            blocks["fm_modules.fm_head"] = std::make_shared<PixelDecoder>(config);
+        }
+
+        std::shared_ptr<TextModel> text_model() {
+            return std::dynamic_pointer_cast<TextModel>(blocks["language_model.model"]);
+        }
+
+        std::shared_ptr<VisionEmbeddings> vision_embeddings() {
+            return std::dynamic_pointer_cast<VisionEmbeddings>(blocks["fm_modules.vision_model_mot_gen.embeddings"]);
+        }
+
+        std::shared_ptr<TimestepEmbedder> timestep_embedder() {
+            return std::dynamic_pointer_cast<TimestepEmbedder>(blocks["fm_modules.timestep_embedder"]);
+        }
+
+        std::shared_ptr<TimestepEmbedder> noise_scale_embedder() {
+            if (!config.add_noise_scale_embedding) {
+                return nullptr;
+            }
+            return std::dynamic_pointer_cast<TimestepEmbedder>(blocks["fm_modules.noise_scale_embedder"]);
+        }
+
+        std::shared_ptr<PixelDecoder> pixel_decoder() {
+            return std::dynamic_pointer_cast<PixelDecoder>(blocks["fm_modules.fm_head"]);
+        }
+    };
+
+    struct SenseNovaU1Runner : public DiffusionModelRunner {
+        SenseNovaU1Config config;
+        SenseNovaU1Model model;
+        std::unordered_set<uint64_t> cached_prefix_hashes;
+        std::vector<int32_t> position_t_vec;
+        std::vector<int32_t> position_h_vec;
+        std::vector<int32_t> position_w_vec;
+        std::vector<float> attention_mask_vec;
+        std::vector<float> noise_scale_vec;
+
+        SenseNovaU1Runner(ggml_backend_t backend,
+                          const String2TensorStorage& tensor_storage_map      = {},
+                          const std::string& prefix                           = "",
+                          std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(SenseNovaU1Config::detect_from_weights(tensor_storage_map, prefix)),
+              model(config) {
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "SenseNova U1.5";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                               const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        static uint64_t hash_input_ids(const sd::Tensor<int32_t>& input_ids) {
+            uint64_t hash = 1469598103934665603ULL;
+            for (int32_t token : input_ids.values()) {
+                uint32_t value = static_cast<uint32_t>(token);
+                for (int byte = 0; byte < 4; ++byte) {
+                    hash ^= static_cast<uint8_t>(value & 0xffU);
+                    hash *= 1099511628211ULL;
+                    value >>= 8;
+                }
+            }
+            hash ^= static_cast<uint64_t>(input_ids.numel());
+            hash *= 1099511628211ULL;
+            return hash;
+        }
+
+        static std::string cache_prefix(uint64_t hash) {
+            return "snu15." + std::to_string(hash);
+        }
+
+        ggml_tensor* make_position_tensor(const std::vector<int32_t>& values,
+                                          const std::string& name) {
+            auto tensor = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_I32, values.size());
+            ggml_set_name(tensor, name.c_str());
+            set_backend_tensor_data(tensor, values.data());
+            return tensor;
+        }
+
+        ggml_cgraph* build_prefix_graph(const sd::Tensor<int32_t>& input_ids_tensor,
+                                        const std::string& prefix_cache) {
+            ggml_cgraph* graph   = new_graph_custom(SENSENOVA_U1_GRAPH_SIZE);
+            ggml_tensor* ids     = make_input(input_ids_tensor);
+            const int64_t length = input_ids_tensor.numel();
+
+            position_t_vec.resize(length);
+            position_h_vec.assign(length, 0);
+            position_w_vec.assign(length, 0);
+            for (int64_t i = 0; i < length; ++i) {
+                position_t_vec[i] = static_cast<int32_t>(i);
+            }
+            auto position_t = make_position_tensor(position_t_vec, "snu15.prefix.position_t");
+            auto position_h = make_position_tensor(position_h_vec, "snu15.prefix.position_h");
+            auto position_w = make_position_tensor(position_w_vec, "snu15.prefix.position_w");
+
+            attention_mask_vec.assign(static_cast<size_t>(length * length), 0.f);
+            for (int64_t query = 0; query < length; ++query) {
+                for (int64_t key = query + 1; key < length; ++key) {
+                    attention_mask_vec[static_cast<size_t>(query * length + key)] = -INFINITY;
+                }
+            }
+            auto attention_mask = ggml_new_tensor_2d(compute_ctx,
+                                                     GGML_TYPE_F32,
+                                                     length,
+                                                     length);
+            ggml_set_name(attention_mask, "snu15.prefix.attention_mask");
+            set_backend_tensor_data(attention_mask, attention_mask_vec.data());
+
+            auto runner_ctx = get_context();
+            auto text_model = model.text_model();
+            auto hidden     = text_model->embed(&runner_ctx, ids);
+            hidden          = text_model->forward(&runner_ctx,
+                                                  hidden,
+                                                  position_t,
+                                                  position_h,
+                                                  position_w,
+                                                  attention_mask,
+                                                  Branch::UNDERSTANDING,
+                                                  prefix_cache);
+            ggml_build_forward_expand(graph, hidden);
+            return graph;
+        }
+
+        bool ensure_prefix_cache(int n_threads,
+                                 const sd::Tensor<int32_t>& input_ids,
+                                 std::string* prefix_cache) {
+            const uint64_t hash = hash_input_ids(input_ids);
+            *prefix_cache       = cache_prefix(hash);
+            if (cached_prefix_hashes.find(hash) != cached_prefix_hashes.end() &&
+                get_cache_tensor_by_name(*prefix_cache + ".0.k") != nullptr) {
+                return true;
+            }
+
+            if (cached_prefix_hashes.size() >= 2) {
+                free_cache_ctx_and_buffer();
+                cached_prefix_hashes.clear();
+            }
+            auto get_graph = [&]() {
+                return build_prefix_graph(input_ids, *prefix_cache);
+            };
+            auto result = GGMLRunner::compute(get_graph, n_threads, false, true);
+            if (!result.has_value()) {
+                LOG_ERROR("SenseNova U1.5 prefix cache computation failed");
+                return false;
+            }
+            cached_prefix_hashes.insert(hash);
+            return true;
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timestep_tensor,
+                                 const std::string& prefix_cache,
+                                 int64_t prefix_length) {
+            ggml_cgraph* graph = new_graph_custom(SENSENOVA_U1_GRAPH_SIZE);
+            ggml_tensor* x     = make_input(x_tensor);
+            ggml_tensor* t     = make_input(timestep_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(x->ne[0] % config.image_token_stride() == 0);
+            GGML_ASSERT(x->ne[1] % config.image_token_stride() == 0);
+
+            const int64_t grid_w  = x->ne[0] / config.patch_size;
+            const int64_t grid_h  = x->ne[1] / config.patch_size;
+            const int64_t token_w = grid_w / config.vision_downsample_factor;
+            const int64_t token_h = grid_h / config.vision_downsample_factor;
+            const int64_t tokens  = token_w * token_h;
+
+            position_h_vec.resize(grid_w * grid_h);
+            position_w_vec.resize(grid_w * grid_h);
+            for (int64_t index = 0; index < grid_w * grid_h; ++index) {
+                position_h_vec[index] = static_cast<int32_t>(index / grid_w);
+                position_w_vec[index] = static_cast<int32_t>(index % grid_w);
+            }
+            auto vision_position_x = make_position_tensor(position_w_vec, "snu15.vision.position_x");
+            auto vision_position_y = make_position_tensor(position_h_vec, "snu15.vision.position_y");
+
+            auto runner_ctx     = get_context();
+            auto hidden         = model.vision_embeddings()->forward(&runner_ctx,
+                                                                     x,
+                                                                     vision_position_x,
+                                                                     vision_position_y);
+            auto time_embedding = model.timestep_embedder()->forward(&runner_ctx, t);
+            time_embedding      = ggml_reshape_3d(compute_ctx, time_embedding, config.hidden_size, 1, 1);
+            hidden              = ggml_add(compute_ctx, hidden, time_embedding);
+
+            if (config.add_noise_scale_embedding) {
+                const float image_tokens = static_cast<float>(tokens);
+                const float noise_scale  = std::min(config.noise_scale_max_value,
+                                                    std::sqrt(image_tokens / config.noise_scale_base_image_seq_len));
+                noise_scale_vec          = {noise_scale / config.noise_scale_max_value};
+                auto noise_scale_tensor  = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_F32, 1);
+                ggml_set_name(noise_scale_tensor, "snu15.noise_scale");
+                set_backend_tensor_data(noise_scale_tensor, noise_scale_vec.data());
+                auto noise_embedding = model.noise_scale_embedder()->forward(&runner_ctx, noise_scale_tensor);
+                noise_embedding      = ggml_reshape_3d(compute_ctx, noise_embedding, config.hidden_size, 1, 1);
+                hidden               = ggml_add(compute_ctx, hidden, noise_embedding);
+            }
+
+            position_t_vec.assign(tokens, static_cast<int32_t>(prefix_length));
+            position_h_vec.resize(tokens);
+            position_w_vec.resize(tokens);
+            for (int64_t index = 0; index < tokens; ++index) {
+                position_h_vec[index] = static_cast<int32_t>(index / token_w);
+                position_w_vec[index] = static_cast<int32_t>(index % token_w);
+            }
+            auto position_t = make_position_tensor(position_t_vec, "snu15.image.position_t");
+            auto position_h = make_position_tensor(position_h_vec, "snu15.image.position_h");
+            auto position_w = make_position_tensor(position_w_vec, "snu15.image.position_w");
+
+            hidden            = model.text_model()->forward(&runner_ctx,
+                                                            hidden,
+                                                            position_t,
+                                                            position_h,
+                                                            position_w,
+                                                            nullptr,
+                                                            Branch::GENERATION,
+                                                            prefix_cache);
+            hidden            = ggml_reshape_4d(compute_ctx,
+                                                hidden,
+                                                config.hidden_size,
+                                                token_w,
+                                                token_h,
+                                                x->ne[3]);
+            hidden            = ggml_cont(compute_ctx, ggml_permute(compute_ctx, hidden, 2, 0, 1, 3));
+            auto x_prediction = model.pixel_decoder()->forward(&runner_ctx, hidden);
+
+            const float timestep = timestep_tensor.values()[0];
+            const float denom    = std::max(1.f - timestep, config.t_eps);
+            auto velocity        = ggml_scale(compute_ctx,
+                                              ggml_sub(compute_ctx, x_prediction, x),
+                                              1.f / denom);
+            ggml_build_forward_expand(graph, velocity);
+            return graph;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timestep,
+                                  const sd::Tensor<int32_t>& input_ids) {
+            std::string prefix_cache;
+            if (!ensure_prefix_cache(n_threads, input_ids, &prefix_cache)) {
+                return {};
+            }
+            auto get_graph = [&]() {
+                return build_graph(x, timestep, prefix_cache, input_ids.numel());
+            };
+            return restore_trailing_singleton_dims(
+                GGMLRunner::compute(get_graph, n_threads, false),
+                x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            const auto* extra = diffusion_extra_as<SenseNovaU1DiffusionExtra>(diffusion_params);
+            GGML_ASSERT(extra->input_ids != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           *extra->input_ids);
+        }
+    };
+}  // namespace SenseNovaU1
+
+#endif  // __SD_MODEL_DIFFUSION_SENSENOVA_U1_H__

src/model/diffusion/wan.hpp

@@ -193,7 +193,7 @@ namespace WAN {
             k      = norm_k->forward(ctx, k);
             auto v = v_proj->forward(ctx, context);  // [N, n_context, dim]
 
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+            x = ggml_ext_attention_ext(ctx, q, k, v, num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
 
             x = o_proj->forward(ctx, x);  // [N, n_token, dim]
             return x;
@@ -255,8 +255,8 @@ namespace WAN {
             k_img      = norm_k_img->forward(ctx, k_img);
             auto v_img = v_img_proj->forward(ctx, context_img);  // [N, context_img_len, dim]
 
-            auto img_x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k_img, v_img, num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+            auto img_x = ggml_ext_attention_ext(ctx, q, k_img, v_img, num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
-            x          = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, nullptr, false, ctx->flash_attn_enabled);          // [N, n_token, dim]
+            x          = ggml_ext_attention_ext(ctx, q, k, v, num_heads, nullptr, false, ctx->flash_attn_enabled);          // [N, n_token, dim]
 
             x = ggml_add(ctx->ggml_ctx, x, img_x);
 

src/model/te/clip.hpp

@@ -145,6 +145,10 @@ protected:
         params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
     }
 
+    enum ggml_op param_usage_op(const std::string& name) const override {
+        return name == "token_embedding.weight" ? GGML_OP_GET_ROWS : GGML_OP_NONE;
+    }
+
 public:
     CLIPEmbeddings(int64_t embed_dim,
                    int64_t vocab_size    = 49408,

src/model/te/llm.hpp

@@ -1359,7 +1359,7 @@ namespace LLM {
                 x        = ggml_ext_cont(ctx->ggml_ctx, kqv);
                 x        = ggml_reshape_3d(ctx->ggml_ctx, x, head_dim * num_heads, n_token, N);
             } else {
-                x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, true, false);  // [N, n_token, hidden_size]
+                x = ggml_ext_attention_ext(ctx, q, k, v, num_heads, attention_mask, true, false);  // [N, n_token, hidden_size]
             }
 
             x = out_proj->forward(ctx, x);  // [N, n_token, hidden_size]

src/model/te/t5.hpp

@@ -251,7 +251,7 @@ public:
 
         k = ggml_ext_scale(ctx->ggml_ctx, k, ::sqrtf(static_cast<float>(d_head)), true);
 
-        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, mask);  // [N, n_token, d_head * n_head]
+        x = ggml_ext_attention_ext(ctx, q, k, v, num_heads, mask);  // [N, n_token, d_head * n_head]
 
         x = out_proj->forward(ctx, x);  // [N, n_token, model_dim]
         return {x, past_bias};

src/model/vae/auto_encoder_kl.hpp

@@ -142,7 +142,7 @@ public:
             v = ggml_reshape_3d(ctx->ggml_ctx, v, c, h * w, n);                        // [N, h * w, in_channels]
         }
 
-        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);
+        h_ = ggml_ext_attention_ext(ctx, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);
 
         if (use_linear) {
             h_ = proj_out->forward(ctx, h_);  // [N, h * w, in_channels]

src/model/vae/hunyuan_vae.hpp

@@ -193,7 +193,7 @@ namespace Hunyuan {
             v = ggml_reshape_3d(ctx->ggml_ctx, v, w * h * t, c, b);                                  // [b, c, t*h*w]
             v = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, v, 1, 0, 2, 3));  // [b, t*h*w, c]
 
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);  // [b, t*h*w, c]
+            x = ggml_ext_attention_ext(ctx, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);  // [b, t*h*w, c]
 
             x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [b, c, t*h*w]
             x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, t, c * b);                         // [b*c, t, h, w]

src/model/vae/mage_vae.hpp

@@ -253,7 +253,7 @@ namespace MageVAE {
             q      = to_patches(ctx->ggml_ctx, q);
             k      = to_patches(ctx->ggml_ctx, k);
             v      = to_patches(ctx->ggml_ctx, v);
-            h      = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);
+            h      = ggml_ext_attention_ext(ctx, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);
             h      = from_patches(ctx->ggml_ctx, h, np, batch, hp, wp);
             if (pad_h > 0) {
                 h = ggml_ext_slice(ctx->ggml_ctx, h, 1, 0, height);

src/model/vae/minimax_h3_audio_vae.hpp

@@ -174,8 +174,7 @@ namespace MiniMaxH3 {
             auto mask        = ggml_diag_mask_inf(ctx->ggml_ctx,
                                                   ggml_ext_zeros(ctx->ggml_ctx, sequence, sequence, 1, 1),
                                                   0);
-            auto attn_out    = ggml_ext_attention_ext(ctx->ggml_ctx,
+            auto attn_out    = ggml_ext_attention_ext(ctx,
-                                                      ctx->backend,
                                                       q,
                                                       k,
                                                       v,

src/model/vae/minimax_h3_vae.hpp

@@ -291,8 +291,7 @@ namespace MiniMaxH3VAE {
             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);
-            auto out            = ggml_ext_attention_ext(ctx->ggml_ctx,
+            auto out            = ggml_ext_attention_ext(ctx,
-                                                         ctx->backend,
                                                          q,
                                                          k,
                                                          v,

src/model/vae/vae.hpp

@@ -166,7 +166,7 @@ public:
             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)) {
+        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version) || sd_version_is_sensenova_u1(version)) {
             scale_factor = 1;
         }
         return scale_factor;

src/model/vae/wan_vae.hpp

@@ -615,8 +615,8 @@ namespace WAN {
             auto v = qkv_vec[2];
             v      = ggml_reshape_3d(ctx->ggml_ctx, v, h * w, c, n);  // [t, c, h * w]
 
-            v = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, v, 1, 0, 2, 3));                            // [t, h * w, c]
+            v = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, v, 1, 0, 2, 3));    // [t, h * w, c]
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);  // [t, h * w, c]
+            x = ggml_ext_attention_ext(ctx, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);  // [t, h * w, c]
 
             x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [t, c, h * w]
             x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, c, n);                             // [t, c, h, w]

src/model_loader.cpp

@@ -69,6 +69,8 @@ const char* unused_tensors[] = {
     // "v_pred", // Used to detect SDXL vpred models
     "text_encoders.llm.output.weight",
     "text_encoders.llm.lm_head.",
+    "language_model.lm_head.",
+    "vision_model.",
 };
 
 bool is_unused_tensor(const std::string& name) {
@@ -185,6 +187,15 @@ bool ModelLoader::parse_file(const std::string& file_path, const std::string& pr
     }
     parsed_dependencies_.push_back(stamp);
     if (is_directory(file_path)) {
+        const std::string diffusers_index_path = path_join(file_path, "model_index.json");
+        const std::string diffusers_unet_path  = path_join(file_path, "unet/diffusion_pytorch_model.safetensors");
+        const bool has_diffusers_layout        = file_exists(diffusers_index_path) || file_exists(diffusers_unet_path);
+
+        const std::string safetensors_index_path = path_join(file_path, "model.safetensors.index.json");
+        if (!has_diffusers_layout && file_exists(safetensors_index_path)) {
+            LOG_INFO("load %s using root safetensors index", file_path.c_str());
+            return parse_file(safetensors_index_path, prefix);
+        }
         LOG_INFO("load %s using diffusers format", file_path.c_str());
         return init_from_diffusers_file(file_path, prefix);
     } else if (is_gguf_file(file_path)) {
@@ -463,6 +474,9 @@ SDVersion ModelLoader::get_sd_version() const {
         if (tensor_storage.name.find("net.img_embedder.proj1.weight") != std::string::npos) {
             return VERSION_MINIT2I;
         }
+        if (tensor_storage.name.find("language_model.model.layers.0.self_attn.q_proj_mot_gen.weight") != std::string::npos) {
+            return VERSION_SENSENOVA_U1_5;
+        }
         if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
             auto img_in = tensor_storage_map.find("model.diffusion_model.img_in.weight");
             if (img_in != tensor_storage_map.end() && img_in->second.ne[0] == 128) {

src/pipeline/diffusion_engine.cpp

@@ -96,9 +96,13 @@ const char* model_version_to_str[] = {
     "SeFi-Image",
     "Krea2",
     "Mage Flow",
+    "SenseNova U1.5",
     "ESRGAN",
 };
 
+static_assert(VERSION_COUNT == sizeof(model_version_to_str) / sizeof(model_version_to_str[0]),
+              "\nnumber of elements in model_version_to_str[] != VERSION_COUNT");
+
 void calculate_alphas_cumprod(float* alphas_cumprod,
                               float linear_start = 0.00085f,
                               float linear_end   = 0.0120f,
@@ -843,6 +847,12 @@ bool StableDiffusionGGML::init_model_loader(ModelLoader& model_loader, ModelConf
 }
 
 bool StableDiffusionGGML::init(const sd_ctx_params_t* sd_ctx_params) {
+    for (float scale : {sd_ctx_params->linear_scale, sd_ctx_params->attn_scale}) {
+        if (!std::isfinite(scale) || scale < 0.f || (scale > 0.f && !std::isfinite(1.f / scale))) {
+            LOG_ERROR("scale overrides must be finite positive values, or 0 to keep model defaults");
+            return false;
+        }
+    }
     auto configuration        = std::make_unique<ModelConfig>(*sd_ctx_params);
     n_threads                 = sd_ctx_params->n_threads;
     enable_mmap               = sd_ctx_params->enable_mmap;
@@ -1306,6 +1316,9 @@ bool StableDiffusionGGML::build_denoiser() {
             pred_type = SEFI_FLOW_PRED;
         } else if (sd_version_is_minit2i(version)) {
             pred_type = MINIT2I_FLOW_PRED;
+        } else if (sd_version_is_sensenova_u1(version)) {
+            pred_type          = SENSENOVA_U1_FLOW_PRED;
+            default_flow_shift = 3.f;
         } else {
             pred_type = EPS_PRED;
         }
@@ -1351,6 +1364,11 @@ bool StableDiffusionGGML::build_denoiser() {
             denoiser = std::make_shared<MiniT2IFlowDenoiser>();
             break;
         }
+        case SENSENOVA_U1_FLOW_PRED: {
+            LOG_INFO("running in SenseNova U1.5 FLOW mode");
+            denoiser = std::make_shared<SenseNovaU1FlowDenoiser>(default_flow_shift);
+            break;
+        }
         default: {
             LOG_ERROR("Unknown predition type %i", pred_type);
             return false;
@@ -2333,6 +2351,9 @@ sd::Tensor<float> StableDiffusionGGML::sample(const std::shared_ptr<DiffusionMod
             } else if (sd_version_is_minit2i(version)) {
                 diffusion_params.extra = MiniT2IDiffusionExtra{
                     condition.c_vector.empty() ? nullptr : &condition.c_vector};
+            } else if (sd_version_is_sensenova_u1(version)) {
+                diffusion_params.extra = SenseNovaU1DiffusionExtra{
+                    condition.c_input_ids.empty() ? nullptr : &condition.c_input_ids};
             } else {
                 diffusion_params.extra = std::monostate{};
             }
@@ -2489,7 +2510,9 @@ int StableDiffusionGGML::get_vae_scale_factor() {
 int StableDiffusionGGML::get_diffusion_model_down_factor() {
     int down_factor = 8;  // unet
     if (sd_version_is_dit(version)) {
-        if (sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version)) {
+        if (sd_version_is_sensenova_u1(version)) {
+            down_factor = 32;
+        } else if (sd_version_is_wan(version) || sd_version_is_lingbot_video(version) || sd_version_is_minimax_h3(version)) {
             down_factor = 2;
         } else {
             down_factor = 1;
@@ -2515,6 +2538,8 @@ int StableDiffusionGGML::get_latent_channel() {
             latent_channel = 3;
         } else if (sd_version_is_minit2i(version)) {
             latent_channel = 3;
+        } else if (sd_version_is_sensenova_u1(version)) {
+            latent_channel = 3;
         } else if (sd_version_is_pid(version)) {
             latent_channel = 3;
         } else if (sd_version_is_sefi_image(version)) {

src/pipeline/image.cpp

@@ -455,6 +455,9 @@ namespace sd::pipeline {
                 // states with a zeroed prompt mask, so no extra text encode is needed.
                 uncond.c_crossattn = cond.c_crossattn;
                 uncond.c_vector    = sd::Tensor<float>::zeros_like(cond.c_vector);
+            } else if (sd_version_is_sensenova_u1(sd->version)) {
+                auto* sensenova_conditioner = static_cast<SenseNovaU1Conditioner*>(sd->cond_stage_model.get());
+                uncond                      = sensenova_conditioner->get_unconditional_condition(request->negative_prompt);
             } else {
                 bool zero_out_masked = false;
                 if (sd_version_is_sdxl(sd->version) &&

src/pipeline/model_builders.cpp

@@ -27,6 +27,7 @@
 #include "model/diffusion/model.hpp"
 #include "model/diffusion/pid.hpp"
 #include "model/diffusion/qwen_image.hpp"
+#include "model/diffusion/sensenova_u1.h"
 #include "model/diffusion/unet.hpp"
 #include "model/diffusion/wan.hpp"
 #include "model/diffusion/z_image.hpp"
@@ -306,6 +307,12 @@ namespace sd::model_builders {
                                                                         tensor_storage_map,
                                                                         "model.diffusion_model.model.net",
                                                                         weight_manager);
+        } else if (sd_version_is_sensenova_u1(version)) {
+            result.conditioner = std::make_shared<SenseNovaU1Conditioner>();
+            result.diffusion   = std::make_shared<SenseNovaU1::SenseNovaU1Runner>(ctx.backends.runtime_backend(SDBackendModule::DIFFUSION),
+                                                                                tensor_storage_map,
+                                                                                "",
+                                                                                weight_manager);
         } else if (sd_version_is_anima(version)) {
             result.conditioner = std::make_shared<AnimaConditioner>(ctx.backends.runtime_backend(SDBackendModule::TE),
                                                                     tensor_storage_map,
@@ -396,6 +403,21 @@ namespace sd::model_builders {
                                                                              "ip_adapter",
                                                                              weight_manager);
         }
+        if (result.conditioner) {
+            result.conditioner->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
+        if (result.diffusion) {
+            result.diffusion->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
+        if (result.high_noise_diffusion) {
+            result.high_noise_diffusion->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
+        if (result.clip_vision) {
+            result.clip_vision->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
+        if (result.ip_adapter) {
+            result.ip_adapter->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
         runners = std::move(result);
         return true;
     }
@@ -493,7 +515,7 @@ namespace sd::model_builders {
             }
         };
 
-        if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version)) {
+        if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1 || sd_version_is_minit2i(version) || sd_version_is_sensenova_u1(version)) {
             LOG_INFO("using FakeVAE");
             result.vae = std::make_shared<FakeVAE>(version,
                                                    ctx.backends.runtime_backend(SDBackendModule::VAE),
@@ -531,6 +553,15 @@ namespace sd::model_builders {
                 result.preview->set_conv2d_direct_enabled(true);
             }
         }
+        if (result.vae) {
+            result.vae->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
+        if (result.preview) {
+            result.preview->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
+        if (result.audio) {
+            result.audio->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
+        }
         runners = std::move(result);
         return true;
     }
@@ -552,6 +583,7 @@ namespace sd::model_builders {
             LOG_INFO("Using Conv2d direct in the control net");
             control_net->set_conv2d_direct_enabled(true);
         }
+        control_net->set_scale_overrides(sd_ctx_params->linear_scale, sd_ctx_params->attn_scale);
         runner = std::move(control_net);
         return true;
     }

src/runtime/denoiser.hpp

@@ -1488,6 +1488,82 @@ struct MiniT2IFlowDenoiser : public Denoiser {
     }
 };
 
+// SenseNova U1.5 integrates velocity over t=0..1 while the generic sampler
+// integrates over descending sigma. With sigma=1-t, returning
+// denoised=x+sigma*v makes the generic Euler derivative exactly -v, so the
+// descending-sigma update is identical to the official ascending-time update.
+struct SenseNovaU1FlowDenoiser : public DiscreteFlowDenoiser {
+    explicit SenseNovaU1FlowDenoiser(float shift = 3.f)
+        : DiscreteFlowDenoiser(shift) {}
+
+    float sigma_min() override {
+        return 0.f;
+    }
+
+    float sigma_max() override {
+        return 1.f;
+    }
+
+    float sigma_to_t(float sigma) override {
+        return 1.f - sigma;
+    }
+
+    float t_to_sigma(float t) override {
+        float sigma = 1.f - t;
+        return shift * sigma / (1.f + (shift - 1.f) * sigma);
+    }
+
+    std::vector<float> get_scalings(float sigma) override {
+        return {1.f, sigma, 1.f};
+    }
+
+    sd::Tensor<float> noise_scaling(float sigma,
+                                    const sd::Tensor<float>& noise,
+                                    const sd::Tensor<float>& latent) override {
+        SD_UNUSED(sigma);
+        SD_UNUSED(latent);
+        GGML_ASSERT(noise.dim() >= 2);
+        const float token_w     = static_cast<float>(noise.shape()[0]) / 32.f;
+        const float token_h     = static_cast<float>(noise.shape()[1]) / 32.f;
+        const float noise_scale = std::min(16.f, std::sqrt((token_w * token_h) / 64.f));
+        return noise * noise_scale;
+    }
+
+    sd::Tensor<float> inverse_noise_scaling(float sigma,
+                                            const sd::Tensor<float>& latent) override {
+        SD_UNUSED(sigma);
+        return latent;
+    }
+
+    float noise_level_to_sigma(float noise_level) override {
+        SD_UNUSED(noise_level);
+        return 1.f;
+    }
+
+    std::vector<float> get_sigmas(uint32_t n,
+                                  int image_seq_len,
+                                  scheduler_t scheduler_type,
+                                  SDVersion version,
+                                  const char* extra_sample_args = nullptr) override {
+        SD_UNUSED(image_seq_len);
+        SD_UNUSED(scheduler_type);
+        SD_UNUSED(version);
+        SD_UNUSED(extra_sample_args);
+        std::vector<float> sigmas;
+        sigmas.reserve(n + 1);
+        if (n == 0) {
+            sigmas.push_back(0.f);
+            return sigmas;
+        }
+        for (uint32_t i = 0; i <= n; ++i) {
+            const float t = static_cast<float>(i) / static_cast<float>(n);
+            sigmas.push_back(t_to_sigma(t));
+        }
+        sigmas.back() = 0.f;
+        return sigmas;
+    }
+};
+
 typedef std::function<sd::guidance::GuiderOutput(const sd::Tensor<float>&, float, int)> denoise_cb_t;
 
 static std::pair<float, float> get_ancestral_step(float sigma_from,

src/stable-diffusion.cpp

@@ -156,6 +156,7 @@ const char* prediction_to_str[] = {
     "flux_flow",
     "sefi_flow",
     "minit2i_flow",
+    "sensenova_u1_flow",
 };
 
 const char* sd_prediction_name(enum prediction_t prediction) {
@@ -322,6 +323,8 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->eager_load                = false;
     sd_ctx_params->enable_mmap               = false;
     sd_ctx_params->diffusion_flash_attn      = false;
+    sd_ctx_params->linear_scale              = 0.f;
+    sd_ctx_params->attn_scale                = 0.f;
     sd_ctx_params->vae_format                = SD_VAE_FORMAT_AUTO;
     sd_ctx_params->backend                   = nullptr;
     sd_ctx_params->params_backend            = nullptr;
@@ -373,6 +376,8 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              "auto_fit: %s\n"
              "flash_attn: %s\n"
              "diffusion_flash_attn: %s\n"
+             "linear_scale: %g\n"
+             "attn_scale: %g\n"
              "vae_format: %s\n",
              SAFE_STR(sd_ctx_params->model_path),
              SAFE_STR(sd_ctx_params->clip_l_path),
@@ -408,6 +413,8 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              BOOL_STR(sd_ctx_params->auto_fit),
              BOOL_STR(sd_ctx_params->flash_attn),
              BOOL_STR(sd_ctx_params->diffusion_flash_attn),
+             sd_ctx_params->linear_scale,
+             sd_ctx_params->attn_scale,
              sd_vae_format_name(sd_ctx_params->vae_format));
 
     return buf;
@@ -695,6 +702,13 @@ SD_API bool sd_ctx_has_control_net(const sd_ctx_t* sd_ctx) {
     return sd_ctx->sd->control_net != nullptr;
 }
 
+const char* sd_get_model_version_name(const sd_ctx_t* sd_ctx) {
+    if (sd_ctx == nullptr || sd_ctx->sd == nullptr || sd_ctx->sd->version >= VERSION_COUNT) {
+        return "Unknown";
+    }
+    return model_version_to_str[sd_ctx->sd->version];
+}
+
 enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx) {
     return sd::pipeline::default_sample_method(sd_ctx != nullptr ? sd_ctx->sd : nullptr);
 }

src/tokenizers/qwen2_tokenizer.cpp

@@ -45,16 +45,8 @@ void Qwen2Tokenizer::load_from_merges(const std::string& merges_utf8_str) {
     bpe_len = rank;
 }
 
-Qwen2Tokenizer::Qwen2Tokenizer(const std::string& merges_utf8_str) {
+static const std::vector<std::string>& qwen2_special_tokens() {
-    UNK_TOKEN = "<|endoftext|>";
+    static const std::vector<std::string> tokens = {
-    EOS_TOKEN = "<|endoftext|>";
-    PAD_TOKEN = "<|endoftext|>";
-
-    UNK_TOKEN_ID = 151643;
-    EOS_TOKEN_ID = 151643;
-    PAD_TOKEN_ID = 151643;
-
-    special_tokens = {
         "<|endoftext|>",
         "<|im_start|>",
         "<|im_end|>",
@@ -87,6 +79,24 @@ Qwen2Tokenizer::Qwen2Tokenizer(const std::string& merges_utf8_str) {
         "<|bot_token|>",
         "<|tms_token|>",
     };
+    return tokens;
+}
+
+Qwen2Tokenizer::Qwen2Tokenizer(const std::string& merges_utf8_str)
+    : Qwen2Tokenizer(merges_utf8_str, qwen2_special_tokens()) {
+}
+
+Qwen2Tokenizer::Qwen2Tokenizer(const std::string& merges_utf8_str,
+                               const std::vector<std::string>& special_tokens_override) {
+    UNK_TOKEN = "<|endoftext|>";
+    EOS_TOKEN = "<|endoftext|>";
+    PAD_TOKEN = "<|endoftext|>";
+
+    UNK_TOKEN_ID = 151643;
+    EOS_TOKEN_ID = 151643;
+    PAD_TOKEN_ID = 151643;
+
+    special_tokens = special_tokens_override;
 
     if (merges_utf8_str.size() > 0) {
         load_from_merges(merges_utf8_str);

src/tokenizers/qwen2_tokenizer.h

@@ -2,12 +2,15 @@
 #define __SD_TOKENIZERS_QWEN2_TOKENIZER_H__
 
 #include <string>
+#include <vector>
 
 #include "bpe_tokenizer.h"
 
 class Qwen2Tokenizer : public BPETokenizer {
 protected:
     void load_from_merges(const std::string& merges_utf8_str);
+    Qwen2Tokenizer(const std::string& merges_utf8_str,
+                   const std::vector<std::string>& special_tokens_override);
 
 public:
     explicit Qwen2Tokenizer(const std::string& merges_utf8_str = "");

src/tokenizers/sensenova_u1_tokenizer.cpp

@@ -0,0 +1,44 @@
+#include "sensenova_u1_tokenizer.h"
+
+#include <vector>
+
+static const std::vector<std::string>& sensenova_u1_special_tokens() {
+    static const std::vector<std::string> tokens = {
+        "<|endoftext|>",
+        "<|im_start|>",
+        "<|im_end|>",
+        "<|object_ref_start|>",
+        "<|object_ref_end|>",
+        "<|box_start|>",
+        "<|box_end|>",
+        "<|quad_start|>",
+        "<|quad_end|>",
+        "<|vision_start|>",
+        "<|vision_end|>",
+        "<|vision_pad|>",
+        "<|image_pad|>",
+        "<|video_pad|>",
+        "<tool_call>",
+        "</tool_call>",
+        "<|fim_prefix|>",
+        "<|fim_middle|>",
+        "<|fim_suffix|>",
+        "<|fim_pad|>",
+        "<|repo_name|>",
+        "<|file_sep|>",
+        "<tool_response>",
+        "</tool_response>",
+        "<think>",
+        "</think>",
+        "<IMG_CONTEXT>",
+        "<img>",
+        "</img>",
+    };
+    return tokens;
+}
+
+SenseNovaU1Tokenizer::SenseNovaU1Tokenizer(const std::string& merges_utf8_str)
+    : Qwen2Tokenizer(merges_utf8_str, sensenova_u1_special_tokens()) {
+    EOS_TOKEN    = "<|im_end|>";
+    EOS_TOKEN_ID = 151645;
+}

src/tokenizers/sensenova_u1_tokenizer.h

@@ -0,0 +1,13 @@
+#ifndef __SD_TOKENIZERS_SENSENOVA_U1_TOKENIZER_H__
+#define __SD_TOKENIZERS_SENSENOVA_U1_TOKENIZER_H__
+
+#include <string>
+
+#include "qwen2_tokenizer.h"
+
+class SenseNovaU1Tokenizer : public Qwen2Tokenizer {
+public:
+    explicit SenseNovaU1Tokenizer(const std::string& merges_utf8_str = "");
+};
+
+#endif  // __SD_TOKENIZERS_SENSENOVA_U1_TOKENIZER_H__