Support Youtu-VL Model

convert_hf_to_gguf.py

@@ -1230,6 +1230,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "4a2e2abae11ca2b86d570fc5b44be4d5eb5e72cc8f22dd136a94b37da83ab665":
             # ref: https://huggingface.co/KORMo-Team/KORMo-tokenizer
             res = "kormo"
+        if chkhsh == "9d70134b369a70e5735009b6de918f7581b5211f7c074d1f89f753aea8248af1":
+            # ref: ./Youtu-VL
+            res = "utu-vl"
 
         if res is None:
             logger.warning("\n")
@@ -7133,6 +7136,7 @@ def prepare_tensors(self):
     "DeepseekV2ForCausalLM",
     "DeepseekV3ForCausalLM",
     "KimiVLForConditionalGeneration",
+    "UTUVLForCausalLM",
 )
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
@@ -7211,11 +7215,26 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
         self.gguf_writer.add_value_length_mla(hparams["v_head_dim"])
 
-        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
-        self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
-        self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
-        self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
-        self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
+        if (moe_intermediate_size := hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        else:
+            self.gguf_writer.add_expert_feed_forward_length(hparams.get("intermediate_size", 0))
+
+        if (n_routed_experts := hparams.get("n_routed_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_routed_experts)
+
+        if (n_shared_experts := hparams.get("n_shared_experts")) is not None:
+            self.gguf_writer.add_expert_shared_count(n_shared_experts)
+        else:
+            self.gguf_writer.add_expert_shared_count(0)
+
+        if (routed_scaling_factor := hparams.get("routed_scaling_factor")) is not None:
+            self.gguf_writer.add_expert_weights_scale(routed_scaling_factor)
+        else:
+            self.gguf_writer.add_expert_weights_scale(1.0)
+
+        if (norm_topk_prob := hparams.get("norm_topk_prob")) is not None and norm_topk_prob:
+            self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
 
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
@@ -7231,10 +7250,17 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         # skip vision tensors and remove "language_model." for Kimi-VL
         if "vision_tower" in name or "multi_modal_projector" in name:
             return []
-
+        if name.startswith("siglip2.") or name.startswith("merger."):
+            return []
         if name.startswith("language_model."):
             name = name.replace("language_model.", "")
 
+        # skip lm_head.weight if tie_word_embeddings is True
+        if self.hparams.get("tie_word_embeddings", False):
+            if name == "lm_head.weight" or name == "model.lm_head.weight":
+                logger.info("Skipping tied output layer 'lm_head.weight' (will use token_embd.weight)")
+                return []
+
         # rename e_score_correction_bias tensors
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
@@ -7246,7 +7272,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
             return []
 
         # process the experts separately
-        if name.find("mlp.experts") != -1:
+        if name.find("mlp.experts") != -1 and self.hparams.get("n_routed_experts") is not None:
             n_experts = self.hparams["n_routed_experts"]
             assert bid is not None
 
@@ -7309,7 +7335,6 @@ def prepare_tensors(self):
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
 
-
 @ModelBase.register("MiniMaxM2ForCausalLM")
 class MiniMaxM2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.MINIMAXM2
@@ -10466,7 +10491,53 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
 
         return []
 
-
+@ModelBase.register("UtuVLForConditionalGeneration", "UTUVLForCausalLM")
+class UtuVLVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.hparams_vision["image_size"] = self.hparams_vision.get("image_size", 560)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.UTUVL)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
+
+        # Handle activation function
+        hidden_act = str(self.hparams.get("hidden_act", "gelu_pytorch_tanh")).lower()
+        if hidden_act in ("gelu", "gelu_pytorch_tanh", "gelu_fast", "gelu_new", "gelu_accurate"):
+            self.gguf_writer.add_vision_use_gelu(True)
+        elif hidden_act == "silu":
+            self.gguf_writer.add_vision_use_silu(True)
+        else:
+            raise ValueError(f"Unsupported activation function for UTUVL: {hidden_act}")
+
+        self.gguf_writer.add_vision_spatial_merge_size(self.hparams.get("spatial_merge_size", 2))
+
+        window_size = self.hparams.get("window_size")
+        if window_size is not None:
+            self.gguf_writer.add_vision_window_size(window_size)
+        fullatt_block_indexes = self.hparams.get("fullatt_block_indexes")
+        assert fullatt_block_indexes is not None, "fullatt_block_indexes is required for utuvl"
+        self.gguf_writer.add_vision_wa_layers(layers=fullatt_block_indexes)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # Skip language model tensors
+        skip_prefixes = ('lm_head.', 'model.layers.', 'model.embed_tokens.', 'model.norm.')
+        if name.startswith(skip_prefixes):
+            return []
+
+        # Try to map the tensor using TensorNameMap (handles vision encoder and projector)
+        try:
+            new_name = self.map_tensor_name(name)
+            return [(new_name, data_torch)]
+        except ValueError:
+            # If mapping fails, log warning and skip
+            logger.warning(f"Cannot map tensor: {name}")
+            return []  
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -145,6 +145,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
     {"name": "minimax-m2",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/MiniMaxAI/MiniMax-M2", },
     {"name": "kormo",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/KORMo-Team/KORMo-tokenizer", },
+    {"name": "utu-vl",           "tokt": TOKENIZER_TYPE.BPE, "repo": "./Youtu-VL", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

gguf-py/gguf/constants.py

@@ -294,7 +294,9 @@ class ClipVision:
         USE_GELU            = "clip.use_gelu"
         USE_SILU            = "clip.use_silu"
         N_WA_PATTERN        = "clip.vision.n_wa_pattern" # used by qwen2.5vl
+        WA_LAYERS           = "clip.vision.wa_layers" # used by utuvl
         IS_DEEPSTACK_LAYERS = "clip.vision.is_deepstack_layers"
+        WINDOW_SIZE         = "clip.vision.window_size"
 
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
@@ -3432,6 +3434,7 @@ class VisionProjectorType:
     JANUS_PRO = "janus_pro"
     LFM2A = "lfm2a" # audio
     GLM4V = "glm4v"
+    UTUVL = "utuvl"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/gguf_writer.py

@@ -1130,10 +1130,16 @@ def add_vision_projector_scale_factor(self, value: int) -> None:
 
     def add_vision_n_wa_pattern(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
+
+    def add_vision_wa_layers(self, layers: Sequence[int]) -> None:
+        self.add_array(Keys.ClipVision.WA_LAYERS, layers)
 
     def add_vision_is_deepstack_layers(self, layers: Sequence[bool]) -> None:
         self.add_array(Keys.ClipVision.IS_DEEPSTACK_LAYERS, layers)
 
+    def add_vision_window_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.WINDOW_SIZE, value)
+
     # audio models
 
     def add_audio_projection_dim(self, value: int) -> None:

gguf-py/gguf/tensor_mapping.py

@@ -1218,6 +1218,7 @@ class TensorNameMap:
         MODEL_TENSOR.V_MMPROJ: (
             "multi_modal_projector.linear_{bid}",
             "visual.merger.mlp.{bid}", # qwen2vl
+            "merger.mlp.{bid}",
         ),
 
         MODEL_TENSOR.V_MMPROJ_FC: (
@@ -1255,6 +1256,7 @@ class TensorNameMap:
             "visual.patch_embed.proj", # qwen2vl
             "vision_tower.patch_embed.proj", # kimi-vl
             "model.vision.patch_embedding.proj", # cogvlm
+            "siglip2.vision_model.embeddings.patch_embedding",
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_NORM: (
@@ -1288,6 +1290,7 @@ class TensorNameMap:
             "vision_encoder.transformer.layers.{bid}.attention.wq", # pixtral
             "visual.blocks.{bid}.attn.q", # qwen2vl, generated
             "vision_tower.encoder.blocks.{bid}.wq", # kimi-vl, generated
+            "siglip2.vision_model.encoder.layers.{bid}.self_attn.q_proj", # utuvl
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
@@ -1305,6 +1308,7 @@ class TensorNameMap:
             "vision_encoder.transformer.layers.{bid}.attention.wk", # pixtral
             "visual.blocks.{bid}.attn.k", # qwen2vl, generated
             "vision_tower.encoder.blocks.{bid}.wk", # kimi-vl, generated
+            "siglip2.vision_model.encoder.layers.{bid}.self_attn.k_proj",
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
@@ -1322,6 +1326,7 @@ class TensorNameMap:
             "vision_encoder.transformer.layers.{bid}.attention.wv", # pixtral
             "visual.blocks.{bid}.attn.v", # qwen2vl, generated
             "vision_tower.encoder.blocks.{bid}.wv", # kimi-vl, generated
+            "siglip2.vision_model.encoder.layers.{bid}.self_attn.v_proj",
         ),
 
         MODEL_TENSOR.V_ENC_INPUT_NORM: (
@@ -1336,6 +1341,7 @@ class TensorNameMap:
             "visual.blocks.{bid}.norm1", # qwen2vl
             "vision_tower.encoder.blocks.{bid}.norm0", # kimi-vl (norm0/norm1)
             "model.vision.transformer.layers.{bid}.input_layernorm", # cogvlm
+            "siglip2.vision_model.encoder.layers.{bid}.layer_norm1",
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_O: (
@@ -1351,6 +1357,7 @@ class TensorNameMap:
             "visual.blocks.{bid}.attn.proj", # qwen2vl
             "vision_tower.encoder.blocks.{bid}.wo", # kimi-vl
             "model.vision.transformer.layers.{bid}.attention.dense", # cogvlm
+            "siglip2.vision_model.encoder.layers.{bid}.self_attn.out_proj", # utuvl
         ),
 
         MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
@@ -1365,6 +1372,7 @@ class TensorNameMap:
             "visual.blocks.{bid}.norm2", # qwen2vl
             "vision_tower.encoder.blocks.{bid}.norm1", # kimi-vl (norm0/norm1)
             "model.vision.transformer.layers.{bid}.post_attention_layernorm", # cogvlm
+            "siglip2.vision_model.encoder.layers.{bid}.layer_norm2",
         ),
 
         MODEL_TENSOR.V_ENC_FFN_UP: (
@@ -1380,6 +1388,7 @@ class TensorNameMap:
             "visual.blocks.{bid}.mlp.linear_fc1", # qwen3vl
             "vision_tower.encoder.blocks.{bid}.mlp.fc0", # kimi-vl (fc0/fc1)
             "model.vision.transformer.layers.{bid}.mlp.fc1", # cogvlm
+            "siglip2.vision_model.encoder.layers.{bid}.mlp.fc1",
         ),
 
         MODEL_TENSOR.V_ENC_FFN_GATE: (
@@ -1401,6 +1410,7 @@ class TensorNameMap:
             "visual.blocks.{bid}.mlp.linear_fc2", # qwen3vl
             "vision_tower.encoder.blocks.{bid}.mlp.fc1", # kimi-vl (fc0/fc1)
             "model.vision.transformer.layers.{bid}.mlp.fc2", # cogvlm
+            "siglip2.vision_model.encoder.layers.{bid}.mlp.fc2",
         ),
 
         MODEL_TENSOR.V_LAYER_SCALE_1: (
@@ -1427,6 +1437,7 @@ class TensorNameMap:
             "visual.merger.ln_q", # qwen2vl
             "vision_tower.encoder.final_layernorm", # kimi-vl
             "visual.post_layernorm", # glm4v
+            "siglip2.vision_model.post_layernorm",
         ),
 
         MODEL_TENSOR.V_MM_POST_NORM: (
@@ -1443,6 +1454,7 @@ class TensorNameMap:
             "multi_modal_projector.pre_norm",
             "pre_mm_projector_norm",
             "model.vision.linear_proj.norm1", # cogvlm
+            "merger.ln_q",
         ),
 
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (

src/llama-model.cpp

@@ -4699,7 +4699,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    // try to load output.weight, if not found, use token_embd (tied embeddings)
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (!output) {
+                        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];
@@ -4762,7 +4766,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    // try to load output.weight, if not found, use token_embd (tied embeddings)
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (!output) {
+                        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];

src/llama-vocab.cpp

@@ -314,6 +314,12 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_UTU_VL:
+                regex_exprs = {
+                    "[가-힣ㄱ-ㆎ]+|[！…“”‘’—：；，、-〿︰-﹏]+|[ㄅ-ㄯ]+|[一-龥぀-ゟ゠-ヿ]+",
+                    "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER:
                 regex_exprs = {
                     "[\r\n]",
@@ -1860,6 +1866,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "deepseek-v3") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM;
                 clean_spaces = false;
+            } else if (
+                    tokenizer_pre == "utu-vl") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_UTU_VL;
+                clean_spaces = false;
+                ignore_merges = true; 
             } else if (
                     tokenizer_pre == "falcon") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_FALCON;

src/llama-vocab.h

@@ -51,6 +51,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40,
     LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2      = 41,
     LLAMA_VOCAB_PRE_TYPE_AFMOE           = 42,
+    LLAMA_VOCAB_PRE_TYPE_UTU_VL          = 43,
 };
 
 struct LLM_KV;

src/unicode.cpp

@@ -964,6 +964,11 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
         { "\\p{P}", unicode_cpt_flags::PUNCTUATION },
         { "\\p{M}", unicode_cpt_flags::ACCENT_MARK },
         { "\\p{S}", unicode_cpt_flags::SYMBOL },
+        { "\\p{Lu}", unicode_cpt_flags::LETTER }, // Uppercase letter
+        { "\\p{Ll}", unicode_cpt_flags::LETTER }, // Lowercase letter
+        { "\\p{Lt}", unicode_cpt_flags::LETTER }, // Titlecase letter
+        { "\\p{Lm}", unicode_cpt_flags::LETTER }, // Modifier letter
+        { "\\p{Lo}", unicode_cpt_flags::LETTER }, // Other letter
     };
 
     static const std::map<int, int> k_ucat_cpt = {
@@ -1074,22 +1079,26 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                         continue;
                     }
 
-                    if (regex_expr[i + 0] == '\\' && i + 4 < regex_expr.size() &&
+                    // Match \p{...} Unicode properties of varying lengths
+                    if (regex_expr[i + 0] == '\\' && i + 3 < regex_expr.size() &&
                         regex_expr[i + 1] == 'p' &&
-                        regex_expr[i + 2] == '{' &&
-                        regex_expr[i + 4] == '}') {
-                        const std::string pat = regex_expr.substr(i, 5);
-                        if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
-                            if (!inside) {
-                                regex_expr_collapsed += '[';
+                        regex_expr[i + 2] == '{') {
+                        // Find the closing brace
+                        size_t closing_brace = regex_expr.find('}', i + 3);
+                        if (closing_brace != std::string::npos && closing_brace <= i + 10) { // reasonable limit
+                            const std::string pat = regex_expr.substr(i, closing_brace - i + 1);
+                            if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
+                                if (!inside) {
+                                    regex_expr_collapsed += '[';
+                                }
+                                regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
+                                regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
+                                if (!inside) {
+                                    regex_expr_collapsed += ']';
+                                }
+                                i = closing_brace;
+                                continue;
                             }
-                            regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
-                            regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
-                            if (!inside) {
-                                regex_expr_collapsed += ']';
-                            }
-                            i += 4;
-                            continue;
                         }
                     }
 

tools/mtmd/CMakeLists.txt

@@ -27,6 +27,7 @@ add_library(mtmd
             models/qwen3vl.cpp
             models/siglip.cpp
             models/whisper-enc.cpp
+            models/utuvl.cpp
             )
 
 set_target_properties(mtmd PROPERTIES