readme : model : mtdm : lfm2 improvements (ggml-org#15476)

* Support untied embeddings

* Increase number of image tokens to 1024

* Add LFM2-VL to readme

* Actually use untied embeddings

Author: tdakhran

README.md

@@ -151,6 +151,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 - [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
 - [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
+- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
 
 </details>
 

gguf-py/gguf/constants.py

@@ -2590,6 +2590,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.ATTN_K,
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.OUTPUT,
     ],
     MODEL_ARCH.SMALLTHINKER: [
         MODEL_TENSOR.TOKEN_EMBD,

src/llama-arch.cpp

@@ -2010,6 +2010,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
             { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT,            "output" },
         }
     },
     {

src/llama-model.cpp

@@ -5474,8 +5474,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_LFM2:
                 {
-                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
                     tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
+                    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,          "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -17787,8 +17792,7 @@ struct llm_build_lfm2 : public llm_graph_context {
         cb(cur, "model.embedding_norm", -1);
         res->t_embd = cur;
 
-        // lm_head is tied with embeddings
-        cur = build_lora_mm(model.tok_embd, cur);
+        cur = build_lora_mm(model.output, cur);
         cb(cur, "lm_head", -1);
 
         res->t_logits = cur;

tools/mtmd/clip.cpp

@@ -3513,7 +3513,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         const int height = img->ny;
         const int total_factor = params.patch_size * params.proj_scale_factor;
         constexpr int min_image_tokens = 64;
-        constexpr int max_image_tokens = 256;
+        constexpr int max_image_tokens = 1024;
         const float min_pixels = min_image_tokens * total_factor * total_factor;
         const float max_pixels = max_image_tokens * total_factor * total_factor;