support SmolLM-3

Author: foldl

CMakeLists.txt

@@ -60,6 +60,7 @@ set(core_files src/backend.cpp
     models/llama.cpp
     models/qwen.cpp
     models/pangu.cpp
+    models/smol.cpp
     )
 
 add_library(libchatllm SHARED EXCLUDE_FROM_ALL src/main.cpp ${core_files})

README.md

@@ -13,6 +13,7 @@ pure C++ implementation based on [@ggerganov](https://github.com/ggerganov)'s [g
 
 **What's New:**
 
+* 2025-07-10: SmolLM-3
 * 2025-07-05: Pangu-Pro-MoE
 * 2025-07-04: ERNIE-MoE
 * 2025-06-30: Hunyuan-A13B, ERNIE-Dense
@@ -28,21 +29,12 @@ pure C++ implementation based on [@ggerganov](https://github.com/ggerganov)'s [g
 * 2025-05-17: [I can speak](./docs/multimodal.md): Orpheus-TTS
 * 2025-05-11: Seed-Coder
 * 2025-04-30: QWen3, MiMo
-* 2025-04-23: [ggml updated](https://github.com/ggml-org/llama.cpp/tree/658987cfc9d752dca7758987390d5fb1a7a0a54a) again
-* 2025-04-18: Apriel
-* 2025-04-15: GLM-4-0414
-* 2025-04-10: LlaMA4 (Language model)
-* 2025-03-27: Ling (Bailing)
-* 2025-03-26: AquilaChat2
-* 2025-03-25: DeepSeek v1 & GigaChat
 * 2025-03-24: [GGMM](./docs/ggmm.md) file format
 * 2025-02-21: [Distributed inference](./docs/rpc.md)
-* 2025-02-17: [ggml updated](https://github.com/ggml-org/llama.cpp/tree/0f2bbe656473177538956d22b6842bcaa0449fab) again
 * 2025-02-10: [GPU acceleration](./docs/gpu.md) 🔥
 * 2024-12-09: [Reversed role](./docs/fun.md#reversed-role)
 * 2024-11-21: [Continued generation](./docs/fun.md#continued-generation)
 * 2024-11-01: [generation steering](./docs/fun.md#generation-steering)
-* 2024-07-14: [ggml updated](https://github.com/ggerganov/ggml/tree/3e7e5e26f90fecf4f7c2808df7d94454630b219c)
 * 2024-06-15: [Tool calling](./docs/tool_calling.md)
 * 2024-05-29: [ggml](https://github.com/ggerganov/ggml) is forked instead of submodule
 * 2024-05-14: [OpenAI API](./docs/binding.md#openai-compatible-api), CodeGemma Base & Instruct supported

convert.py

@@ -201,6 +201,8 @@ class ModelType(Enum):
 
     PenguMoE        = 0x2600
 
+    SmolLM3         = 0x2700
+
     BCE_Embedding           = 0x10000100
     BCE_ReRanker            = 0x10000101
     BGE_M3                  = 0x10000102
@@ -1880,6 +1882,31 @@ def get_weight_names(config):
         r = Llama3Converter.get_weight_names(config)
         return r[:-1]
 
+class SmolLM3Converter(BaseConverter):
+    MODEL_TYPE = ModelType.SmolLM3
+    tie_word_embeddings = True
+
+    @staticmethod
+    def dump_config(f, config, ggml_type):
+        assert config.rope_scaling is None
+        assert (config.layer_types.count('full_attention') == config.num_hidden_layers) or \
+               (config.use_sliding_window is None) or (not config.use_sliding_window)
+        SmolLM3Converter.tie_word_embeddings = (config.tie_word_embeddings is None) or (config.tie_word_embeddings)
+
+        dump_llama_like_config(f, config, ggml_type)
+        config_values = [
+            config.num_key_value_heads,
+            config.no_rope_layer_interval,
+            1 if SmolLM3Converter.tie_word_embeddings else 0,
+        ]
+        f.write(struct.pack("i" * len(config_values), *config_values))
+        f.write(struct.pack("<f", config.rope_theta))
+
+    @staticmethod
+    def get_weight_names(config):
+        r = Llama3Converter.get_weight_names(config)
+        return r[:-1] if SmolLM3Converter.tie_word_embeddings else r
+
 class SmolVLMConverter(BaseConverter):
     MODEL_TYPE = ModelType.SmolVLM
 
@@ -7404,6 +7431,8 @@ def main():
         Llama3Converter.convert(config, model_files, vocab, ggml_type, args.save_path)
     elif arch == 'smollm':
         SmolLMConverter.convert(config, model_files, vocab, ggml_type, args.save_path)
+    elif arch == 'SmolLM3ForCausalLM':
+        SmolLM3Converter.convert(config, model_files, vocab, ggml_type, args.save_path)
     elif arch == 'SmolVLMForConditionalGeneration':
         SmolVLMConverter.convert(config, model_files, vocab, ggml_type, args.save_path)
     elif arch == 'XverseForCausalLM':

docs/models.md

@@ -237,6 +237,9 @@
     * [x] Confucius3-Math: [14B](https://huggingface.co/netease-youdao/Confucius3-Math/tree/62621490d5dccf5fea997be9df62dd8dc017f777) (`-a DeepSeek-R1-Distill-QWen`)
     * [x] Jan-Nano: [4B](https://huggingface.co/Menlo/Jan-nano/tree/5f4e450c127322db9477400890a0dd951c9f6ab7)
 
+* SmolLM-3 (`SmolLM3ForCausalLM`)
+    * [x] [3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B/tree/297fd6336cf21656d5f9d30a1db612ceeca67619)
+
 * Solor (`SolarForCausalLM`)
     * [x] [Pro](https://huggingface.co/upstage/solar-pro-preview-instruct/tree/dd4bcf7006df9b1ce3f87711e702e4063832aae3)
 

models/qwen.cpp

@@ -995,11 +995,7 @@ namespace chatllm::qwen::ds_r1_distill_v3
     Tokenizer::Tokenizer(BaseConfig config)
         : ds_r1_distill::Tokenizer(config)
     {
-        std::time_t now = std::time(nullptr);
-        std::tm* timeinfo = std::localtime(&now);
-        char buffer[1000];
-        std::strftime(buffer, sizeof(buffer), "%Y-%m-%d, %A", timeinfo);
-        std::string date_str(buffer);
+        std::string date_str = utils::now("%Y-%m-%d, %A");
         sys_prompt = "该助手为DeepSeek-R1，由深度求索公司创造。\n今天是" + date_str + "。";
     }
 }