Use ggml_pad instead of concat with empty tensor

Author: stduhpf

flux.hpp

@@ -884,7 +884,7 @@ namespace Flux {
                 vec = approx->forward(ctx, vec);                           // [344, N, hidden_size]
 
                 if (y != NULL) {
-                    txt_img_mask = ggml_concat(ctx, y, ggml_scale_inplace(ctx, ggml_new_tensor_1d(ctx, GGML_TYPE_F32, img->ne[1]), 0), 0);
+                    txt_img_mask = ggml_pad(ctx, y, img->ne[1], 0, 0, 0);
                 }
             } else {
                 auto time_in   = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]);

t5.hpp

@@ -434,6 +434,7 @@ class T5UniGramTokenizer {
                 tokens.insert(tokens.end(), length - tokens.size(), pad_token_id);
                 weights.insert(weights.end(), length - weights.size(), 1.0);
                 if (attention_mask != nullptr) {
+                    // maybe keep some padding tokens unmasked? 
                     attention_mask->insert(attention_mask->end(), length - attention_mask->size(), -HUGE_VALF);
                 }
             }
@@ -594,7 +595,7 @@ class T5Attention : public GGMLBlock {
         }
         if (past_bias != NULL) {
             if (mask != NULL) {
-                mask = ggml_repeat(ctx,mask,past_bias);
+                mask = ggml_repeat(ctx, mask, past_bias);
                 mask = ggml_add(ctx, mask, past_bias);
             } else {
                 mask = past_bias;