Include base patch for granite vision allocation

alex-jw-brooks · alex-jw-brooks · commit 65575237ea1a · 2025-02-23T18:32:16.000-07:00
Signed-off-by: Alex-Brooks &lt;Alex.Brooks@ibm.com&gt;
diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp
@@ -406,8 +406,8 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }
 
 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    // Minicpmv / granite vision use 10 patches
-    int num_max_patches = 10;
+    // Granite vision uses up to 10 patches + base patch
+    int num_max_patches = 11;
     if (clip_is_glm(ctx_clip)) {
         num_max_patches = 1;
     }

Original file line number	Diff line number	Diff line change
`@@ -406,8 +406,8 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *`
`406`	`406`	`}`
`407`	`407`
`408`	`408`	`bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {`
`409`		`- // Minicpmv / granite vision use 10 patches`
`410`		`- int num_max_patches = 10;`
	`409`	`+ // Granite vision uses up to 10 patches + base patch`
	`410`	`+ int num_max_patches = 11;`
`411`	`411`	`if (clip_is_glm(ctx_clip)) {`
`412`	`412`	`num_max_patches = 1;`
`413`	`413`	`}`