llava : Add Granite Vision Support (#11794)

* Add super wip scripts for multimodal granite gguf

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Add example for converting mmgranite to gguf

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* remove hardcoded path

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Add vision feature layer to gguf params

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Clean up llava surgery and remove name substitution hacks

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Add transformers llava next tensor name mapping

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Make siglip / openclip mutuall exclusive

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix projector linear substitution

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix linear 2 substitution index

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Increase max flattened gridpoints to 64

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix hardcoded concat for multiple feature layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Pull vision feature layers out of gguf keys

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* fix num gridpoints and use all layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Avoid dropping last image encoder layer in llava models

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Use 10 for max number of patches

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Standardize vision feature layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Cleanup logs

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Update comment for vision feature layer init

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Update notes for alternative to legacy llm conversion script

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix notes rendering

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Add v prefix to vision feature layer log

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Use current defaults for feature layer

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Use constant for max gridpoints / feat layers, style fixes

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* clarify non-negative feature layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Remove CLIP_API from func signature

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* USE MAX_IMAGE_FEATURE_LAYERS const in layer calc

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Clarify feature layers are non negative ints and not uint

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix condition for reading feature layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* pop last llava layer when feature layers are unset

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix unset vision layer 0

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Update examples/llava/clip.cpp

Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>

* Reenable assertion for out of bounds get_rows

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Use std vector for gridpoints and feature layers

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Caculate max feature layer at load time

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Include base patch for granite vision allocation

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Fix trailing whitespace

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Add max num patches = 10 back for minicpmv

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Use unordered set to store feature layers

Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Use max feature layer for postnorm

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

* Apply suggestions from code review

---------

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>

examples/llava/llava.cpp

@@ -353,9 +353,10 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         const int32_t * image_grid = clip_image_grid(ctx_clip);
+        const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);
 
         std::vector<std::pair<int, int>> grid_pinpoints;
-        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+        for (size_t i = 0; i < num_gridpoints; i += 2) {
             grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
         }
 
@@ -405,7 +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) {
-    int num_max_patches = 6;
+    // Granite vision uses up to 10 patches + base patch
+    int num_max_patches = 11;
     if (clip_is_minicpmv(ctx_clip)) {
         num_max_patches = 10;
     }