mtmd: use RAII for setting and resetting non-causal attention (#25723)

* mtmd: use RAII for setting and resetting non-causal attention

* mtmd: drop dependency on <optional>

* mtmd: shorten class and variable names

diff --git a/tools/mtmd/mtmd-helper.cpp b/tools/mtmd/mtmd-helper.cpp
index 3c73db4431e7fb1bb0a11b1c90f34243f8266264..84422c89f3abee37a550690b9389f5eecf612cfd 100644 (file)
--- a/tools/mtmd/mtmd-helper.cpp
+++ b/tools/mtmd/mtmd-helper.cpp
@@ -238,6 +238,29 @@ struct decode_embd_batch {
     }
 };
 
+// Helper class to set non-causal attention via RAII
+class scope_non_causal {
+public:
+    scope_non_causal(llama_context * context, bool enabled) : context_(context), enabled_(enabled) {
+        if (enabled_) {
+            // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
+            llama_set_causal_attn(context_, false);
+        }
+    }
+    ~scope_non_causal() {
+        if (enabled_) {
+            llama_set_causal_attn(context_, true);
+        }
+    }
+
+    scope_non_causal(const scope_non_causal &) = delete;
+    scope_non_causal & operator=(const scope_non_causal &) = delete;
+
+private:
+    llama_context * context_;
+    bool enabled_;
+};
+
 // Helper function for decoding an image whose embeddings have already been calculated
 int32_t mtmd_helper_decode_image_chunk(
         mtmd_context * ctx,
@@ -288,10 +311,7 @@ int32_t mtmd_helper_decode_image_chunk(
     }
 
     const bool use_non_causal = mtmd_decode_use_non_causal(ctx, chunk);
-    if (use_non_causal) {
-        llama_set_causal_attn(lctx, false);
-        // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
-    }
+    const scope_non_causal non_causal(lctx, use_non_causal);
 
     while (i_batch < n_img_batches) { // split into batches
         int pos_offset = i_batch*n_batch;
@@ -304,9 +324,6 @@ int32_t mtmd_helper_decode_image_chunk(
         int32_t ret = llama_decode(lctx, batch_embd_view);
         if (ret != 0) {
             LOG_ERR("failed to decode %s\n", name);
-            if (use_non_causal) {
-                llama_set_causal_attn(lctx, true);
-            }
             return ret;
         }
 
@@ -314,9 +331,6 @@ int32_t mtmd_helper_decode_image_chunk(
             ret = callback(batch_embd_view, user_data);
             if (ret != 0) {
                 LOG_ERR("post-decode callback failed\n");
-                if (use_non_causal) {
-                    llama_set_causal_attn(lctx, true);
-                }
                 return ret;
             }
         }
@@ -329,9 +343,6 @@ int32_t mtmd_helper_decode_image_chunk(
     n_past += mtmd_input_chunk_get_n_pos(chunk);
     *new_n_past = n_past;
 
-    if (use_non_causal) {
-        llama_set_causal_attn(lctx, true);
-    }
     return 0;
 }