implement tiling vae encode support

Author: stduhpf

ggml_extend.hpp

@@ -734,21 +734,31 @@ __STATIC_INLINE__ std::vector<struct ggml_tensor*> ggml_chunk(struct ggml_contex
 typedef std::function<void(ggml_tensor*, ggml_tensor*, bool)> on_tile_process;
 
 // Tiling
-__STATIC_INLINE__ void sd_tiling(ggml_tensor* input, ggml_tensor* output, const int scale, const int tile_size, const float tile_overlap_factor, on_tile_process on_processing) {
+__STATIC_INLINE__ void sd_tiling(ggml_tensor* input, ggml_tensor* output, const int scale, const int tile_size, const float tile_overlap_factor, on_tile_process on_processing, bool scaled_out = true) {
     output = ggml_set_f32(output, 0);
 
     int input_width   = (int)input->ne[0];
     int input_height  = (int)input->ne[1];
     int output_width  = (int)output->ne[0];
     int output_height = (int)output->ne[1];
+
+    int input_tile_size, output_tile_size;
+    if (scaled_out) {
+        input_tile_size  = tile_size;
+        output_tile_size = tile_size * scale;
+    } else {
+        input_tile_size  = tile_size * scale;
+        output_tile_size = tile_size;
+    }
+
     GGML_ASSERT(input_width % 2 == 0 && input_height % 2 == 0 && output_width % 2 == 0 && output_height % 2 == 0);  // should be multiple of 2
 
-    int tile_overlap     = (int32_t)(tile_size * tile_overlap_factor);
-    int non_tile_overlap = tile_size - tile_overlap;
+    int tile_overlap     = (int32_t)(input_tile_size * tile_overlap_factor);
+    int non_tile_overlap = input_tile_size - tile_overlap;
 
     struct ggml_init_params params = {};
-    params.mem_size += tile_size * tile_size * input->ne[2] * sizeof(float);                       // input chunk
-    params.mem_size += (tile_size * scale) * (tile_size * scale) * output->ne[2] * sizeof(float);  // output chunk
+    params.mem_size += input_tile_size * input_tile_size * input->ne[2] * sizeof(float);     // input chunk
+    params.mem_size += output_tile_size * output_tile_size * output->ne[2] * sizeof(float);  // output chunk
     params.mem_size += 3 * ggml_tensor_overhead();
     params.mem_buffer = NULL;
     params.no_alloc   = false;
@@ -763,8 +773,8 @@ __STATIC_INLINE__ void sd_tiling(ggml_tensor* input, ggml_tensor* output, const
     }
 
     // tiling
-    ggml_tensor* input_tile  = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, tile_size, tile_size, input->ne[2], 1);
-    ggml_tensor* output_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, tile_size * scale, tile_size * scale, output->ne[2], 1);
+    ggml_tensor* input_tile  = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, input_tile_size, input_tile_size, input->ne[2], 1);
+    ggml_tensor* output_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, output_tile_size, output_tile_size, output->ne[2], 1);
     on_processing(input_tile, NULL, true);
     int num_tiles = ceil((float)input_width / non_tile_overlap) * ceil((float)input_height / non_tile_overlap);
     LOG_INFO("processing %i tiles", num_tiles);
@@ -773,19 +783,23 @@ __STATIC_INLINE__ void sd_tiling(ggml_tensor* input, ggml_tensor* output, const
     bool last_y = false, last_x = false;
     float last_time = 0.0f;
     for (int y = 0; y < input_height && !last_y; y += non_tile_overlap) {
-        if (y + tile_size >= input_height) {
-            y      = input_height - tile_size;
+        if (y + input_tile_size >= input_height) {
+            y      = input_height - input_tile_size;
             last_y = true;
         }
         for (int x = 0; x < input_width && !last_x; x += non_tile_overlap) {
-            if (x + tile_size >= input_width) {
-                x      = input_width - tile_size;
+            if (x + input_tile_size >= input_width) {
+                x      = input_width - input_tile_size;
                 last_x = true;
             }
             int64_t t1 = ggml_time_ms();
             ggml_split_tensor_2d(input, input_tile, x, y);
             on_processing(input_tile, output_tile, false);
-            ggml_merge_tensor_2d(output_tile, output, x * scale, y * scale, tile_overlap * scale);
+            if (scaled_out) {
+                ggml_merge_tensor_2d(output_tile, output, x * scale, y * scale, tile_overlap * scale);
+            } else {
+                ggml_merge_tensor_2d(output_tile, output, x / scale, y / scale, tile_overlap / scale);
+            }
             int64_t t2 = ggml_time_ms();
             last_time  = (t2 - t1) / 1000.0f;
             pretty_progress(tile_count, num_tiles, last_time);

stable-diffusion.cpp

@@ -351,8 +351,8 @@ class StableDiffusionGGML {
                                                                      offload_params_to_cpu,
                                                                      model_loader.tensor_storages_types);
                 diffusion_model  = std::make_shared<MMDiTModel>(backend,
-                                                               offload_params_to_cpu,
-                                                               model_loader.tensor_storages_types);
+                                                                offload_params_to_cpu,
+                                                                model_loader.tensor_storages_types);
             } else if (sd_version_is_flux(version)) {
                 bool is_chroma = false;
                 for (auto pair : model_loader.tensor_storages_types) {
@@ -388,11 +388,11 @@ class StableDiffusionGGML {
                                                                     1,
                                                                     true);
                 diffusion_model  = std::make_shared<WanModel>(backend,
-                                                             offload_params_to_cpu,
-                                                             model_loader.tensor_storages_types,
-                                                             "model.diffusion_model",
-                                                             version,
-                                                             sd_ctx_params->diffusion_flash_attn);
+                                                              offload_params_to_cpu,
+                                                              model_loader.tensor_storages_types,
+                                                              "model.diffusion_model",
+                                                              version,
+                                                              sd_ctx_params->diffusion_flash_attn);
                 if (strlen(SAFE_STR(sd_ctx_params->high_noise_diffusion_model_path)) > 0) {
                     high_noise_diffusion_model = std::make_shared<WanModel>(backend,
                                                                             offload_params_to_cpu,
@@ -1286,7 +1286,15 @@ class StableDiffusionGGML {
         ggml_tensor* result = NULL;
         if (!use_tiny_autoencoder) {
             process_vae_input_tensor(x);
-            first_stage_model->compute(n_threads, x, false, &result, work_ctx);
+            if (vae_tiling && !decode_video) {
+                // split latent in 32x32 tiles and compute in several steps
+                auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
+                    first_stage_model->compute(n_threads, in, true, &out, NULL);
+                };
+                sd_tiling(x, result, 8, 32, 0.5f, on_tiling, false);
+            } else {
+                first_stage_model->compute(n_threads, x, false, &result, work_ctx);
+            }
             first_stage_model->free_compute_buffer();
         } else {
             tae_first_stage->compute(n_threads, x, false, &result, work_ctx);
@@ -1313,12 +1321,12 @@ class StableDiffusionGGML {
                                     -0.0313f, -0.1649f, 0.0117f, 0.0723f, -0.2839f, -0.2083f, -0.0520f, 0.3748f,
                                     0.0152f, 0.1957f, 0.1433f, -0.2944f, 0.3573f, -0.0548f, -0.1681f, -0.0667f};
                 latents_std_vec  = {
-                     0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
-                     0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
-                     0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
-                     0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
-                     0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
-                     0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
+                    0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
+                    0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
+                    0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
+                    0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
+                    0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
+                    0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
             }
             for (int i = 0; i < latent->ne[3]; i++) {
                 float mean = latents_mean_vec[i];
@@ -1353,12 +1361,12 @@ class StableDiffusionGGML {
                                     -0.0313f, -0.1649f, 0.0117f, 0.0723f, -0.2839f, -0.2083f, -0.0520f, 0.3748f,
                                     0.0152f, 0.1957f, 0.1433f, -0.2944f, 0.3573f, -0.0548f, -0.1681f, -0.0667f};
                 latents_std_vec  = {
-                     0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
-                     0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
-                     0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
-                     0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
-                     0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
-                     0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
+                    0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
+                    0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
+                    0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
+                    0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
+                    0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
+                    0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
             }
             for (int i = 0; i < latent->ne[3]; i++) {
                 float mean = latents_mean_vec[i];
@@ -1416,7 +1424,7 @@ class StableDiffusionGGML {
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
                     first_stage_model->compute(n_threads, in, true, &out, NULL);
                 };
-                sd_tiling(x, result, 8, 32, 0.5f, on_tiling);
+                sd_tiling(x, result, 8, 32, 0.5f, on_tiling, true);
             } else {
                 first_stage_model->compute(n_threads, x, true, &result, work_ctx);
             }
@@ -1428,7 +1436,7 @@ class StableDiffusionGGML {
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
                     tae_first_stage->compute(n_threads, in, true, &out);
                 };
-                sd_tiling(x, result, 8, 64, 0.5f, on_tiling);
+                sd_tiling(x, result, 8, 64, 0.5f, on_tiling, true);
             } else {
                 tae_first_stage->compute(n_threads, x, true, &result);
             }