Merge branch 'sf/deepseek-ocr' into sf/deepseek-ocr

src/llama-vocab.cpp

@@ -2347,6 +2347,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "_<EOT>"
                     || t.first == "<|end_of_text|>"
                     || t.first == "<end_of_utterance>" // smoldocling
+                    || t.first == "<｜end▁of▁sentence｜>" // deepseek-ocr
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

tools/mtmd/clip.cpp

@@ -803,9 +803,8 @@ struct clip_graph {
 
         cur = sam_layer_norm_2d(ctx0, cur, 256, model.neck_3_w, model.neck_3_b, hparams.eps);
 
-        //TODO : check conv padding
+        cur = ggml_conv_2d(ctx0, model.net_2, cur, 2,2,1,1, 1,1);
-        cur = ggml_conv_2d_s1_ph(ctx0, model.net_2, cur);
+        cur = ggml_conv_2d(ctx0, model.net_3, cur, 2,2,1,1, 1,1);
-        cur = ggml_conv_2d_s1_ph(ctx0, model.net_3, cur);
 
         ggml_build_forward_expand(gf, cur);
         return cur;
@@ -840,22 +839,27 @@ struct clip_graph {
 
         ggml_tensor * global_features_2 = build_dp_ocr_clip(global_features_1);
 
+        // FIXME remove n_patches is hardcoded
+        int clip_n_patches = 256; // FIXME hardcoded for sam 1024x1024 with 16x16 patches
+
         // torch global_features = torch.cat((global_features_2[:, 1:], global_features_1.flatten(2).permute(0, 2, 1)), dim=-1)
         global_features_1 = ggml_cont(ctx0,ggml_permute(ctx0, global_features_1,2,1,0,3));
-        global_features_1 = ggml_reshape_2d(ctx0, global_features_1, n_embd, n_patches);
+        // flatten 2nd and 3rd dims
+        global_features_1 = ggml_reshape_2d(ctx0, global_features_1, global_features_1->ne[0], clip_n_patches);
 
         // remove CLS token
         global_features_2 = ggml_view_2d(ctx0, global_features_2,
-            n_embd, n_patches,
+            n_embd, clip_n_patches,
             ggml_row_size(global_features_2->type, n_embd), 0);
 
         ggml_tensor * global_features = ggml_concat(ctx0, global_features_2, global_features_1, 1);
-        global_features = ggml_reshape_2d(ctx0, global_features, 2* n_embd, n_patches);
+        global_features = ggml_reshape_2d(ctx0, global_features, 2* n_embd,clip_n_patches);
         global_features = ggml_cont(ctx0, global_features);
         global_features = ggml_mul_mat(ctx0, model.fc_w, global_features);
         global_features = ggml_add(ctx0, global_features, model.fc_b);
 
         global_features = build_global_local_features(ctx0,global_features);
+        global_features = ggml_cont(ctx0, ggml_permute(ctx0, global_features, 1, 0, 2, 3));
         ggml_build_forward_expand(gf, global_features);
         return gf;
     }
@@ -870,16 +874,16 @@ struct clip_graph {
         GGML_ASSERT(model.view_seperator != nullptr);
 
         // 1) global_features: [n_dim, h*w] -> [n_dim, w, h] -> [h, w, n_dim]
-        ggml_tensor * t = ggml_reshape_4d(ctx0, global_features, 1280, 64, 64, 1);  // (n_dim, w, h)
+        ggml_tensor * t = ggml_reshape_4d(ctx0, global_features, 1280, 16, 16, 1);  // (n_dim, w, h)
         t               = ggml_cont(ctx0, ggml_permute(ctx0, t, 2, 1, 0, 3)); // (h, w, n_dim)
         ggml_tensor * nl = ggml_cont(ctx0,ggml_permute(ctx0, model.image_newline, 2, 1, 0, 3));
-        nl = ggml_repeat_4d(ctx0, nl, 64, 1, 1280, 1); // n_pos rows
+        nl = ggml_repeat_4d(ctx0, nl, 16, 1, 1280, 1); // n_pos rows
 
 
         // 2) image_newline: [n_dim] -> [1, 1, n_dim] -> repeat to [h, 1, n_dim]
         t = ggml_concat(ctx0, t, nl, 1);  // (h, w+1, n_dim)
 
-        t = ggml_reshape_2d(ctx0, t, 1280, 64 * (64 + 1));  // (n_dim, h*(w+1))
+        t = ggml_reshape_2d(ctx0, t, 1280, 16 * (16 + 1));  // (n_dim, h*(w+1))
 
 
         // 5) append view_separator as an extra "token":
@@ -1540,9 +1544,12 @@ struct clip_graph {
         GGML_ASSERT(model.class_embedding != nullptr);
         GGML_ASSERT(model.position_embeddings != nullptr);
 
-        const int     n_pos = n_patches + 1;
+        ggml_tensor * inp = ggml_cpy(ctx0, patch_embeds, ggml_dup_tensor(ctx0, patch_embeds));
-        ggml_tensor * inp = ggml_cont(ctx0,ggml_permute(ctx0, patch_embeds,2,1,0,3));
+
-        inp = ggml_reshape_2d(ctx0, inp, n_embd, n_patches);
+
+        const int n_pos = 257; // +1 for [CLS]
+        inp = ggml_cont(ctx0,ggml_permute(ctx0, inp,2,1,0,3));
+        inp = ggml_reshape_2d(ctx0, inp, n_embd, inp->ne[1]*inp->ne[2]*inp->ne[3]);
 
 
 
@@ -1554,7 +1561,9 @@ struct clip_graph {
 
         // for selecting learned pos embd, used by ViT
         ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
-        cb(positions, "positions", -1);
+        ggml_set_name(positions, "positions");
+        ggml_set_input(positions);
+
         ggml_tensor * learned_pos_embd = ggml_get_rows(ctx0, model.position_embeddings, positions);
 
 
@@ -5676,10 +5685,20 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_JANUS_PRO:
         case PROJECTOR_TYPE_COGVLM:
-        case PROJECTOR_TYPE_DEEPSEEKOCR:
             {
                 // do nothing
             } break;
+        case PROJECTOR_TYPE_DEEPSEEKOCR:
+        {
+            //FIXME we need correct this when all model configs are set correctly
+            //n_patch is not correct right now
+            int32_t n_pos = 16 * 16 + 1; //hardcode for now
+            std::vector<int32_t> positions(n_pos);
+            for (int i = 0; i < n_pos; i++) {
+                positions[i] = i;
+            }
+            set_input_i32("positions", positions);
+        } break;
         case PROJECTOR_TYPE_LLAMA4:
             {
                 // set the 2D positions

tools/mtmd/mtmd-cli.cpp

@@ -222,14 +222,18 @@ static std::string chat_add_and_format(mtmd_cli_context & ctx, common_chat_msg &
 
 static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg) {
     bool add_bos = ctx.chat_history.empty();
-    auto formatted_chat = chat_add_and_format(ctx, msg);
-    LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.c_str());
 
     mtmd_input_text text;
-    text.text          = formatted_chat.c_str();
+    text.text          = msg.content.c_str();
     text.add_special   = add_bos;
     text.parse_special = true;
 
+    if (!mtmd_is_deepseekocr(ctx.ctx_vision.get())) {
+        auto formatted_chat = chat_add_and_format(ctx, msg);
+        LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.c_str());
+        text.text = formatted_chat.c_str();        
+    }
+
     if (g_is_interrupted) return 0;
 
     mtmd::input_chunks chunks(mtmd_input_chunks_init());
@@ -332,6 +336,11 @@ int main(int argc, char ** argv) {
         }
 
     } else {
+        if (mtmd_is_deepseekocr(ctx.ctx_vision.get())) {
+            LOG_ERR("\n DeepSeek-OCR doesn't support chat mode.");
+            return 1;
+        }
+        
         LOG("\n Running in chat mode, available commands:");
         if (mtmd_support_vision(ctx.ctx_vision.get())) {
             LOG("\n   /image <path>    load an image");

tools/mtmd/mtmd.cpp

@@ -864,6 +864,10 @@ int mtmd_get_audio_bitrate(mtmd_context * ctx) {
     return 16000; // 16kHz
 }
 
+bool mtmd_is_deepseekocr(mtmd_context * ctx) {
+    return ctx->ctx_v && clip_is_deepseekocr(ctx->ctx_v);
+}
+
 //
 // public API functions
 //

tools/mtmd/mtmd.h

@@ -117,6 +117,9 @@ MTMD_API bool mtmd_support_audio(mtmd_context * ctx);
 // return -1 if audio is not supported
 MTMD_API int mtmd_get_audio_bitrate(mtmd_context * ctx);
 
+// whether the current model is DeepSeek-OCR
+MTMD_API bool mtmd_is_deepseekocr(mtmd_context * ctx);
+
 // mtmd_bitmap
 //
 // if bitmap is image: