fix attn weight scaling after rebase

HimariO · HimariO · commit b27a8dcb0ed3 · 2025-04-13T21:06:46.000+08:00
diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
@@ -757,7 +757,7 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
             if (full_attn) {
                 KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f);
             } else {
-                KQ = ggml_soft_max_ext(ctx0, KQ, window_mask, 1.0f, 0.0f);
+                KQ = ggml_soft_max_ext(ctx0, KQ, window_mask, 1.0f / sqrtf((float)d_head), 0.0f);
             }
 
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);

Original file line number	Diff line number	Diff line change
`@@ -757,7 +757,7 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im`
`757`	`757`	`if (full_attn) {`
`758`	`758`	`KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f);`
`759`	`759`	`} else {`
`760`		`- KQ = ggml_soft_max_ext(ctx0, KQ, window_mask, 1.0f, 0.0f);`
	`760`	`+ KQ = ggml_soft_max_ext(ctx0, KQ, window_mask, 1.0f / sqrtf((float)d_head), 0.0f);`
`761`	`761`	`}`
`762`	`762`
`763`	`763`	`struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);`