Supporting Velvet model

convert_hf_to_gguf.py

@@ -699,6 +699,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "b3f499bb4255f8ca19fccd664443283318f2fd2414d5e0b040fbdd0cc195d6c5":
             # ref: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
             res = "deepseek-r1-qwen"
+        if chkhsh == "a3df2b8943e01cfd7d68c9f8446b294f3d8706d1d6853df65df7fda5d4fcb19f":
+            # ref: https://huggingface.co/Almawave/Velvet-14B
+            res = "velvet"    
 
         if res is None:
             logger.warning("\n")

convert_hf_to_gguf_update.py

@@ -109,6 +109,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "megrez",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"},
     {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
     {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
+    {"name": "velvet",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Almawave/Velvet-14B"}
 ]
 
 

include/llama.h

@@ -105,6 +105,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
         LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
         LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
+        LLAMA_VOCAB_PRE_TYPE_VELVET         = 29
     };
 
     enum llama_rope_type {

src/llama-chat.cpp

@@ -58,6 +58,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
+    { "velvet",            LLM_CHAT_TEMPLATE_VELVET            },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -167,6 +168,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_GIGACHAT;
     } else if (tmpl_contains("<|role_start|>")) {
         return LLM_CHAT_TEMPLATE_MEGREZ;
+    } else if (tmpl_contains("<instruction>")) {
+        return LLM_CHAT_TEMPLATE_VELVET;    
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -566,10 +569,32 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|role_start|>assistant<|role_end|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_VELVET) {
+        // Velvet template
+        std::string leading_space = "";
+        std::string trailing_space = "";
+        bool trim_assistant_message = true;
+        bool is_inside_turn = false;
+        for (auto message : chat) {
+            if (!is_inside_turn) {
+                ss << leading_space << "<instruction>" << trailing_space;
+                is_inside_turn = true;
+            }
+            std::string role(message->role);
+            std::string content(message->content);
+            if (role == "system") {
+                ss << content << "\n\n";
+            } else if (role == "user") {
+                ss << content << leading_space << "</instruction>";
+            } else {
+                ss << trailing_space << (trim_assistant_message ? trim(content) : content) << "</s>";
+                is_inside_turn = false;
+            }
+        }
     } else {
         // template not supported
         return -1;
-    }
+    }  
     dest = ss.str();
     return dest.size();
 }

src/llama-chat.h

@@ -39,6 +39,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_MEGREZ,
     LLM_CHAT_TEMPLATE_UNKNOWN,
+    LLM_CHAT_TEMPLATE_VELVET
 };
 
 struct llama_chat_message;