Supporting Velvet model

convert_hf_to_gguf.py

@@ -777,6 +777,9 @@ 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 chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
             # ref: https://huggingface.co/Xenova/gpt-4o
             res = "gpt-4o"

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"},
     {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
     {"name": "superbpe",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", },
     {"name": "trillion",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", },

include/llama.h

@@ -114,6 +114,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
         LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
         LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
+        LLAMA_VOCAB_PRE_TYPE_VELVET         = 36,
     };
 
     enum llama_rope_type {

src/llama-chat.cpp

@@ -60,6 +60,7 @@ 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            },
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
@@ -177,6 +178,8 @@ 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;
     } else if (tmpl_contains(" Ассистент:")) {
         return LLM_CHAT_TEMPLATE_YANDEX;
     } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
@@ -583,6 +586,31 @@ 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 = false;
+        bool is_inside_turn = false;
+        std::string system_message = "";
+        std::string last_message(chat.back()->content);
+        ss << "<s>";
+        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" && content != " ") {
+                system_message = content + "\n\n";
+            } else if (role == "user") {
+                ss << (content==last_message ? system_message : "") << content << leading_space << "</instruction>";
+            } else {
+                ss << trailing_space << (trim_assistant_message ? trim(content) : content) << "</s>";
+                is_inside_turn = false;
+            }
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_YANDEX) {
         // Yandex template ("\n\n" is defined as EOT token)
 

src/llama-chat.h

@@ -39,6 +39,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_MEGREZ,
+    LLM_CHAT_TEMPLATE_VELVET,
     LLM_CHAT_TEMPLATE_YANDEX,
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,

src/llama-vocab.cpp

@@ -396,6 +396,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_VELVET:
+                // original regex from tokenizer.json
+                //"[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*|\\p{N}|[\\p{P}\\p{S}]{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
+                regex_exprs = {
+                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*|\\p{N}|[\\p{P}\\p{S}]{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_GPT4O:
                 regex_exprs = {
                     // original regex from tokenizer.json
@@ -1629,6 +1636,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                 tokenizer_pre == "megrez") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
+            } else if (
+                tokenizer_pre == "velvet") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_VELVET;
             } else if (
                     tokenizer_pre == "gpt-4o" ||
                     tokenizer_pre == "llama4") {

tests/test-chat-template.cpp

@@ -274,6 +274,14 @@ int main(void) {
             /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
+        {
+            /* .name= */ "velvet",
+            /* .template_str= */ "{{ bos_token }}{%- if messages[0][\"role\"] == \"system\" %}\n    {%- set system_message = messages[0][\"content\"] %}\n    {%- set loop_messages = messages[1:] %}\n{%- else %}\n    {%- set loop_messages = messages %}\n{%- endif %}\n{%- if not tools is defined %}\n    {%- set tools = none %}\n{%- endif %}\n{%- set user_messages = loop_messages | selectattr(\"role\", \"equalto\", \"user\") | list %}\n\n{#- This block checks for alternating user/assistant messages, skipping tool calling messages #}\n{%- set ns = namespace() %}\n{%- set ns.index = 0 %}\n{%- for message in loop_messages %}\n    {%- if not (message.role == \"tool\" or message.role == \"tool_results\" or (message.tool_calls is defined and message.tool_calls is not none)) %}\n        {%- if (message[\"role\"] == \"user\") != (ns.index % 2 == 0) %}\n            {{- raise_exception(\"After the optional system message, conversation roles must alternate user/assistant/user/assistant/...\") }}\n        {%- endif %}\n        {%- set ns.index = ns.index + 1 %}\n    {%- endif %}\n{%- endfor %}\n\n{{- bos_token }}\n{%- for message in loop_messages %}\n    {%- if message[\"role\"] == \"user\" %}\n        {%- if message == user_messages[-1] and system_message is defined and tools is not none %}\n            {{- \"<instruction>\" + system_message + \"\\n\\n\" + \"<list_tools>[\" }}\n            {%- for tool in tools %}\n{{ tool }}\n {%- if not loop.last %}\n                    {{- \", \" }}\n                {%- else %}\n                    {{- \"]\" }}\n {%- endif %}\n    {%- endfor %}\n            {{- \"</list_tools>\" + \"\\n\\n\" + message[\"content\"] + \"</instruction>\" }}\n            {%- elif message == user_messages[-1] and system_message is defined  and tools is none %}\n            {{- \"<instruction>\" + system_message + \"\\n\\n\" + message[\"content\"] + \"</instruction>\" }}\n        {%- else %}\n            {{- \"<instruction>\" + message[\"content\"] + \"</instruction>\" }}\n        {%- endif %}\n    {%- elif (message.tool_calls is defined and message.tool_calls is not none) %}\n        {{- \"<tool_call>[\" }}\n        {%- for tool_call in message.tool_calls %}\n{{ tool_call }}\n {%- if not loop.last %}\n                    {{- \", \" }}\n                {%- else %}\n                    {{- \"]\" }}\n {%- endif %}\n {%- endfor %}      {{- \"</tool_call>\" + eos_token }}\n            {%- elif message[\"role\"] == \"assistant\" %}\n        {{- message[\"content\"] + eos_token}}\n    {%- elif message[\"role\"] == \"tool_results\" or message[\"role\"] == \"tool\" %}\n        {%- if message.content is defined and message.content.content is defined %}\n            {%- set content = message.content.content %}\n        {%- else %}\n            {%- set content = message.content %}\n        {%- endif %}\n        {{- '<tool_result>{\"content\": ' + content|string }}\n    {{- '}</tool_result>' }}\n    {%- else %}\n        {{- raise_exception(\"Only user and assistant roles are supported, with the exception of an initial optional system message!\") }}\n    {%- endif %}\n{%- endfor %}\n",
+            /* .expected_output= */ "<s><instruction>Hello</instruction>Hi there</s><instruction>Who are you</instruction>   I am an assistant   </s><instruction>You are a helpful assistant\n\nAnother question</instruction>",
+            /* .expected_output_jinja= */ "<s><instruction>Hello</instruction>Hi there</s><instruction>Who are you</instruction>   I am an assistant   </s><instruction>You are a helpful assistant\n\nAnother question</instruction>",
+            /* .bos_token= */ "<s>",
+            /* .eos_token= */ "</s>",
+        },
         {
             /* .name= */ "yandex/YandexGPT-5-Lite-8B-instruct",
             /* .template_str= */ "<s>{%- set names = {'assistant': ' Ассистент:', 'user': ' Пользователь:'} %}\n{%- set tools_prefix = 'Тебе доступны следующие функции:' %}\n{%- macro __render_tool(tool) %}\n    {%- set name = tool.function.name %}\n    {%- set description = tool.function.description|default('') %}\n    {%- set parameters = tool.function.parameters|tojson %}\n    {{- '\\n' }}function {{ '{' }}'name':'{{ name }}',\n    {%- if tool.function.description %}'description':'{{ description }}',{% endif %}\n'parameters':{{ parameters }}\n    {{- '}' }}\n{%- endmacro %}\n{%- macro __render_tools(tools) %}\n    {{- tools_prefix }}\n    {%- for tool in tools %}\n        {{- __render_tool(tool) }}\n    {%- endfor %}\n    {{- '\\n\\n' }}\n{%- endmacro %}\n{%- macro __render_tool_message(message) %}\n    {{- '\\n\\nРезультат вызова' }} {{ message.name }}: {{ message.content }} {{ '\\n\\n' }}\n{%- endmacro %}\n{%- if tools -%}\n    {{- __render_tools(tools) }}\n{%- endif -%}\n{%- macro __render_user_message(message) %}\n{{ names.user }} {{ message.content + '\\n\\n' }}\n{%- endmacro %}\n{%- macro __render_assistant_message(message) %}\n    {{- names.assistant }}\n    {%- set call = message['function_call'] %}\n    {%- if call %}\n        {{- '\\n[TOOL_CALL_START]' }}{{ call.name }}{{ '\\n' }}{{ call.arguments|tojson }}\n    {%- else %}\n        {{- ' ' + message.content + '\\n\\n' }}\n    {%- endif %}\n{%- endmacro %}\n{%- if not add_generation_prompt is defined %}\n{%- set add_generation_prompt = false %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- __render_user_message(message) }}\n    {%- endif %}\n    {%- if message.role == 'assistant' and not loop.last %}\n        {{- __render_assistant_message(message) }}\n    {%- endif %}\n    {%- if message.role == 'tool' %}\n        {{- __render_tool_message(message) }}\n    {%- endif %}\n    {%- if loop.last %}\n        {{- ' Ассистент:[SEP]' }}\n    {%- endif %}\n{%- endfor %}\n",