Merge branch 'prepare-PR-of-minicpm-v2.6' into master

Author: tc-mb

examples/llava/clip.cpp

@@ -3,6 +3,7 @@
 // I'll gradually clean and extend it
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
+#include "common.h"
 #include "log.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
@@ -81,6 +82,7 @@ static std::string format(const char * fmt, ...) {
 #define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
 #define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
 #define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
 #define KEY_USE_GELU            "clip.use_gelu"
 #define KEY_N_EMBD              "clip.%s.embedding_length"
 #define KEY_N_FF                "clip.%s.feed_forward_length"
@@ -526,6 +528,7 @@ struct clip_ctx {
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
     bool has_minicpmv_projector = false;
+    int minicpmv_version = 2;
 
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
@@ -641,7 +644,12 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     if (ctx->has_minicpmv_projector) {
         int pos_w = image_size_width/patch_size;
         int pos_h = image_size_height/patch_size;
-        pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        if (ctx->minicpmv_version == 2) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        }
+        else if (ctx->minicpmv_version == 3) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
+        }
         ggml_set_name(pos_embed, "pos_embed");
         ggml_set_input(pos_embed);
     }
@@ -769,7 +777,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
 
-        } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        } 
+        else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
             embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
             // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
@@ -987,6 +996,72 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             GGML_ASSERT(false);
         }
     }
+    // minicpmv projector
+    else if (ctx->has_minicpmv_projector) {
+        if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
+            struct ggml_tensor * q = model.mm_model_query;
+            { // layernorm
+                q = ggml_norm(ctx0, q, eps);
+                q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+            }
+            struct ggml_tensor *k, *v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
+            { // layernorm
+                v = ggml_norm(ctx0, v, eps);
+                v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b);
+            }
+            { // position
+                // q = ggml_add(ctx0, q, model.mm_model_pos_embed);
+                k = ggml_add(ctx0, v, pos_embed);
+            }
+
+            { // attention
+                int hidden_size = 4096;
+                const int d_head = 128;
+                int n_head = hidden_size/d_head;
+                int num_query = 96;
+                if (ctx->minicpmv_version == 2) {
+                    hidden_size = 4096;
+                    n_head = hidden_size/d_head;
+                    num_query = 96;
+                }
+                else if (ctx->minicpmv_version == 3) {
+                    hidden_size = 3584;
+                    n_head = hidden_size/d_head;
+                    num_query = 64;
+                }
+                struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
+                Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
+                struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b);
+                struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b);
+                // permute
+                Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size);
+                Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+                Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size);
+                K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
+                K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+                K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
+                V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
+                V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
+                V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+                KQ = ggml_soft_max_inplace(ctx0, KQ);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
+                KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size);
+                KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size);
+
+                embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b);
+            }
+            { // layernorm
+                embeddings = ggml_norm(ctx0, embeddings, eps);
+                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b);
+            }
+            embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
+        }
+        else {
+            GGML_ASSERT(false);
+        }
+    }
 
     // build the graph
     ggml_build_forward_expand(gf, embeddings);
@@ -1149,6 +1224,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx);
         }
 
+        idx = gguf_find_key(ctx, KEY_MINICPMV_VERSION);
+        if (idx != -1) {
+            new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
+        }
+
         // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
 
         GGML_ASSERT(new_clip->has_vision_encoder);
@@ -1910,10 +1990,18 @@ int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
-    if (clip_is_minicpmv(ctx)) {
-        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img);
+
+    if(clip_is_minicpmv(ctx)){
+        int max_slice_nums = 9;
+        if (ctx->minicpmv_version == 2) {
+            max_slice_nums = 9;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            max_slice_nums = 9;
+        }
+        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img, max_slice_nums);
         res_imgs->size = 0;
-        for (size_t i = 0; i < imgs.size(); ++i) {
+        for (size_t i = 0; i < imgs.size(); ++i){
             res_imgs->size += imgs[i].size();
         }
         res_imgs->data = new clip_image_f32[res_imgs->size];
@@ -2146,7 +2234,12 @@ int clip_n_patches(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
     } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
-        n_patches = 96;
+        if (ctx->minicpmv_version == 2) {
+            n_patches = 96;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            n_patches = 64;
+        }
     }
 
     return n_patches;
@@ -2282,6 +2375,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     const int patch_size    = hparams.patch_size;
     const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    if(ctx->load_image_size==nullptr){
+        ctx->load_image_size= clip_image_size_init();
+    }
+    const int pos_w = ctx->load_image_size->width/patch_size;
+    const int pos_h = ctx->load_image_size->height/patch_size;
 
     {
         struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
@@ -2316,8 +2414,18 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
             int* positions_data = (int*)malloc(ggml_nbytes(positions));
-            for (int i = 0; i < num_positions; i++) {
-                positions_data[i] = std::floor(70.0*i/num_positions);
+            int bucket_coords_h[70];
+            int bucket_coords_w[70];
+            for (int i = 0; i < pos_h; i++){
+                bucket_coords_h[i] = std::floor(70.0*i/pos_h);
+            }
+            for (int i = 0; i < pos_w; i++){
+                bucket_coords_w[i] = std::floor(70.0*i/pos_w);
+            }
+            for (int i = 0, id = 0; i < pos_h; i++){
+                for (int j = 0; j < pos_w; j++){
+                    positions_data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j];
+                }
             }
             ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
             free(positions_data);
@@ -2328,12 +2436,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
             //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
             struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
-            if(ctx->load_image_size==nullptr){
-                ctx->load_image_size= clip_image_size_init();
-            }
-            int pos_w = ctx->load_image_size->width/patch_size;
-            int pos_h = ctx->load_image_size->height/patch_size;
             int embed_dim = 4096;
+            if (ctx->minicpmv_version == 2) {
+                embed_dim = 4096;
+            }
+            else if (ctx->minicpmv_version == 3) {
+                embed_dim = 3584;
+            }
             auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
 
             float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
@@ -2346,7 +2455,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed));
             free(pos_embed_data);
         }
-    } else {
+    }
+    else{
         {
             if (ctx->has_class_embedding) {
                 struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
@@ -2548,13 +2658,21 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         return ctx->vision_model.mm_3_b->ne[0];
     }
     if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
-        return 4096;
+        if (ctx->minicpmv_version == 2) {
+            return 4096;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            return 3584;
+        }
     }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
 }
 
-bool clip_is_minicpmv(const struct clip_ctx * ctx) {
-    return ctx->has_minicpmv_projector;
+int clip_is_minicpmv(const struct clip_ctx * ctx) { 
+    if (ctx->has_minicpmv_projector) {
+        return ctx->minicpmv_version;
+    }
+    return 0;
 }

examples/llava/clip.h

@@ -85,7 +85,7 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
-CLIP_API bool clip_is_minicpmv(const struct clip_ctx * ctx);
+CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
 
 #ifdef __cplusplus
 }

examples/llava/llava.cpp

@@ -254,9 +254,16 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
             image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
             int patch_size=14;
             load_image_size->width = img_res_v.data[i].nx;
-            load_image_size->height = img_res_v.data[i].ny;
+            load_image_size->height = img_res_v.data[i].ny; 
             clip_add_load_image_size(ctx_clip, load_image_size);
-            const bool encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+            bool encoded = false;
+            int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
+            if (has_minicpmv_projector == 2) {
+                encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+            }
+            else if (has_minicpmv_projector == 3) {
+                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+            }            
             if (!encoded) {
                 LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;

examples/llava/minicpmv-cli.cpp

@@ -134,7 +134,13 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
     std::string system_prompt;
     int idx = 0;
     int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
-    system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
+    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
+    if (has_minicpmv_projector == 2) {
+        system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
+    }
+    else if (has_minicpmv_projector == 3) {
+        system_prompt = "<|im_start|>user\n";
+    }     
     LOG_TEE("%s: image token past: %d\n", __func__, n_past);
     eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
     process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
@@ -210,10 +216,24 @@ static struct llava_context * minicpmv_init(gpt_params * params, const std::stri
 
 static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
     std::string user_prompt = prompt;
-    if (!is_first) user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
+    int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
+    if (!is_first) {
+        if (has_minicpmv_projector == 2) {
+            user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
+        }
+        else if (has_minicpmv_projector == 3) {
+            user_prompt = "<|im_start|>user\n" + prompt;
+        }    
+    }
 
     eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
-    eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
+    if (has_minicpmv_projector == 2) {
+        eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
+    }
+    else if (has_minicpmv_projector == 3) {
+        eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
+    } 
+    
     // generate the response
 
     LOG_TEE("\n");