fix: avoid generating black images when running T5 on the GPU (#882)

ggml_extend.hpp

@@ -483,12 +483,15 @@ __STATIC_INLINE__ void ggml_split_tensor_2d(struct ggml_tensor* input,
     int64_t width    = output->ne[0];
     int64_t height   = output->ne[1];
     int64_t channels = output->ne[2];
+    int64_t ne3      = output->ne[3];
     GGML_ASSERT(input->type == GGML_TYPE_F32 && output->type == GGML_TYPE_F32);
     for (int iy = 0; iy < height; iy++) {
         for (int ix = 0; ix < width; ix++) {
             for (int k = 0; k < channels; k++) {
-                float value = ggml_tensor_get_f32(input, ix + x, iy + y, k);
+                for (int l = 0; l < ne3; l++) {
-                ggml_tensor_set_f32(output, value, ix, iy, k);
+                    float value = ggml_tensor_get_f32(input, ix + x, iy + y, k, l);
+                    ggml_tensor_set_f32(output, value, ix, iy, k, l);
+                }
             }
         }
     }
@@ -511,6 +514,7 @@ __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
     int64_t width    = input->ne[0];
     int64_t height   = input->ne[1];
     int64_t channels = input->ne[2];
+    int64_t ne3      = input->ne[3];
 
     int64_t img_width  = output->ne[0];
     int64_t img_height = output->ne[1];
@@ -519,9 +523,10 @@ __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
     for (int iy = y_skip; iy < height; iy++) {
         for (int ix = x_skip; ix < width; ix++) {
             for (int k = 0; k < channels; k++) {
-                float new_value = ggml_tensor_get_f32(input, ix, iy, k);
+                for (int l = 0; l < ne3; l++) {
+                    float new_value = ggml_tensor_get_f32(input, ix, iy, k, l);
                     if (overlap_x > 0 || overlap_y > 0) {  // blend colors in overlapped area
-                    float old_value = ggml_tensor_get_f32(output, x + ix, y + iy, k);
+                        float old_value = ggml_tensor_get_f32(output, x + ix, y + iy, k, l);
 
                         const float x_f_0 = (overlap_x > 0 && x > 0) ? (ix - x_skip) / float(overlap_x) : 1;
                         const float x_f_1 = (overlap_x > 0 && x < (img_width - width)) ? (width - ix) / float(overlap_x) : 1;
@@ -534,9 +539,10 @@ __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
                         ggml_tensor_set_f32(
                             output,
                             old_value + new_value * ggml_smootherstep_f32(y_f) * ggml_smootherstep_f32(x_f),
-                        x + ix, y + iy, k);
+                            x + ix, y + iy, k, l);
                     } else {
-                    ggml_tensor_set_f32(output, new_value, x + ix, y + iy, k);
+                        ggml_tensor_set_f32(output, new_value, x + ix, y + iy, k, l);
+                    }
                 }
             }
         }
@@ -852,8 +858,8 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
     }
 
     struct ggml_init_params params = {};
-    params.mem_size += input_tile_size_x * input_tile_size_y * input->ne[2] * sizeof(float);     // input chunk
+    params.mem_size += input_tile_size_x * input_tile_size_y * input->ne[2] * input->ne[3] * sizeof(float);      // input chunk
-    params.mem_size += output_tile_size_x * output_tile_size_y * output->ne[2] * sizeof(float);  // output chunk
+    params.mem_size += output_tile_size_x * output_tile_size_y * output->ne[2] * output->ne[3] * sizeof(float);  // output chunk
     params.mem_size += 3 * ggml_tensor_overhead();
     params.mem_buffer = NULL;
     params.no_alloc   = false;
@@ -868,8 +874,8 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
     }
 
     // tiling
-    ggml_tensor* input_tile  = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, input_tile_size_x, input_tile_size_y, input->ne[2], 1);
+    ggml_tensor* input_tile  = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, input_tile_size_x, input_tile_size_y, input->ne[2], input->ne[3]);
-    ggml_tensor* output_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, output_tile_size_x, output_tile_size_y, output->ne[2], 1);
+    ggml_tensor* output_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, output_tile_size_x, output_tile_size_y, output->ne[2], output->ne[3]);
     int num_tiles            = num_tiles_x * num_tiles_y;
     LOG_INFO("processing %i tiles", num_tiles);
     pretty_progress(0, num_tiles, 0.0f);

stable-diffusion.cpp

@@ -339,16 +339,6 @@ public:
 
         {
             clip_backend = backend;
-            bool use_t5xxl = false;
-            if (sd_version_is_dit(version) && !sd_version_is_qwen_image(version)) {
-                use_t5xxl = true;
-            }
-            if (!clip_on_cpu && !ggml_backend_is_cpu(backend) && use_t5xxl) {
-                LOG_WARN(
-                    "!!!It appears that you are using the T5 model. Some backends may encounter issues with it."
-                    "If you notice that the generated images are completely black,"
-                    "try running the T5 model on the CPU using the --clip-on-cpu parameter.");
-            }
             if (clip_on_cpu && !ggml_backend_is_cpu(backend)) {
                 LOG_INFO("CLIP: Using CPU backend");
                 clip_backend = ggml_backend_cpu_init();
@@ -1440,10 +1430,19 @@ public:
         if (vae_tiling_params.enabled && !encode_video) {
             // TODO wan2.2 vae support?
             int C = sd_version_is_dit(version) ? 16 : 4;
+            int ne2;
+            int ne3;
+            if (sd_version_is_qwen_image(version)) {
+                ne2 = 1;
+                ne3 = C*x->ne[3];
+            } else {
                 if (!use_tiny_autoencoder) {
                     C *= 2;
                 }
-            result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, C, x->ne[3]);
+                ne2 = C;
+                ne3 = x->ne[3];
+            }
+            result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, ne2, ne3);
         }
 
         if (sd_version_is_qwen_image(version)) {

t5.hpp

@@ -504,7 +504,9 @@ public:
     T5DenseGatedActDense(int64_t model_dim, int64_t ff_dim) {
         blocks["wi_0"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
         blocks["wi_1"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
-        blocks["wo"]   = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false));
+        float scale    = 1.f / 32.f;
+        // The purpose of the scale here is to prevent NaN issues on some backends(CUDA, ...).
+        blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false, false, false, scale));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {