fix attn weight scaling after rebase

HimariO · HimariO · commit b28ad7ecca78 · 2025-04-07T22:07:56.000+08:00
diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
@@ -778,7 +778,7 @@ 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
`@@ -778,7 +778,7 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im`
`778`	`778`	`if (full_attn) {`
`779`	`779`	`KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f);`
`780`	`780`	`} else {`
`781`		`- KQ = ggml_soft_max_ext(ctx0, KQ, window_mask, 1.0f, 0.0f);`
	`781`	`+ KQ = ggml_soft_max_ext(ctx0, KQ, window_mask, 1.0f / sqrtf((float)d_head), 0.0f);`
`782`	`782`	`}`
`783`	`783`
`784`	`784`	`struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);`