feat: increase work_ctx memory buffer size (#814)

Dockerfile.sycl

@@ -0,0 +1,19 @@
+ARG SYCL_VERSION=2025.1.0-0
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build
+
+RUN apt-get update && apt-get install -y cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release -j$(nproc)
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

conditioner.hpp

@@ -115,7 +115,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             return true;
         }
         struct ggml_init_params params;
-        params.mem_size               = 10 * 1024 * 1024;  // max for custom embeddings 10 MB
+        params.mem_size               = 100 * 1024 * 1024;  // max for custom embeddings 100 MB
         params.mem_buffer             = NULL;
         params.no_alloc               = false;
         struct ggml_context* embd_ctx = ggml_init(params);

model.cpp

@@ -107,7 +107,7 @@ const char* unused_tensors[] = {
 };
 
 bool is_unused_tensor(std::string name) {
-    for (int i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
+    for (size_t i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
         if (starts_with(name, unused_tensors[i])) {
             return true;
         }

model.h

@@ -119,7 +119,7 @@ struct TensorStorage {
 
     size_t file_index = 0;
     int index_in_zip  = -1;  // >= means stored in a zip file
-    size_t offset     = 0;   // offset in file
+    uint64_t offset   = 0;   // offset in file
 
     TensorStorage() = default;
 
@@ -164,10 +164,10 @@ struct TensorStorage {
 
     std::vector<TensorStorage> chunk(size_t n) {
         std::vector<TensorStorage> chunks;
-        size_t chunk_size = nbytes_to_read() / n;
+        uint64_t chunk_size = nbytes_to_read() / n;
         // printf("%d/%d\n", chunk_size, nbytes_to_read());
         reverse_ne();
-        for (int i = 0; i < n; i++) {
+        for (size_t i = 0; i < n; i++) {
             TensorStorage chunk_i = *this;
             chunk_i.ne[0]         = ne[0] / n;
             chunk_i.offset        = offset + i * chunk_size;

preprocessing.hpp

@@ -164,7 +164,7 @@ void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float lo
 
 uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
     struct ggml_init_params params;
-    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10
+    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10MB
     params.mem_buffer             = NULL;
     params.no_alloc               = false;
     struct ggml_context* work_ctx = ggml_init(params);

stable-diffusion.cpp

@@ -2196,19 +2196,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
     }
 
     struct ggml_init_params params;
-    params.mem_size = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-    if (sd_version_is_sd3(sd_ctx->sd->version)) {
-        params.mem_size *= 3;
-    }
-    if (sd_version_is_flux(sd_ctx->sd->version)) {
-        params.mem_size *= 4;
-    }
-    if (sd_ctx->sd->stacked_id) {
-        params.mem_size += static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-    }
-    params.mem_size += width * height * 3 * sizeof(float) * 3;
-    params.mem_size += width * height * 3 * sizeof(float) * 3 * sd_img_gen_params->ref_images_count;
-    params.mem_size *= sd_img_gen_params->batch_count;
+    params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
     params.mem_buffer = NULL;
     params.no_alloc   = false;
     // LOG_DEBUG("mem_size %u ", params.mem_size);
@@ -2448,8 +2436,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
     }
 
     struct ggml_init_params params;
-    params.mem_size = static_cast<size_t>(200 * 1024) * 1024;  // 200 MB
-    params.mem_size += width * height * frames * 3 * sizeof(float) * 2;
+    params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
     params.mem_buffer = NULL;
     params.no_alloc   = false;
     // LOG_DEBUG("mem_size %u ", params.mem_size);

upscaler.cpp

@@ -69,8 +69,7 @@ struct UpscalerGGML {
                  input_image.width, input_image.height, output_width, output_height);
 
         struct ggml_init_params params;
-        params.mem_size = output_width * output_height * 3 * sizeof(float) * 2;
-        params.mem_size += 2 * ggml_tensor_overhead();
+        params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
         params.mem_buffer = NULL;
         params.no_alloc   = false;
 
@@ -80,7 +79,7 @@ struct UpscalerGGML {
             LOG_ERROR("ggml_init() failed");
             return upscaled_image;
         }
-        LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
+        // LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
         ggml_tensor* input_image_tensor = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, input_image.width, input_image.height, 3, 1);
         sd_image_to_tensor(input_image.data, input_image_tensor);
 

wan.hpp

@@ -1219,7 +1219,7 @@ namespace WAN {
 
         void test() {
             struct ggml_init_params params;
-            params.mem_size   = static_cast<size_t>(1000 * 1024 * 1024);  // 10 MB
+            params.mem_size = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
             params.mem_buffer = NULL;
             params.no_alloc   = false;