model : add BailingMoeV2 support

README.md

@@ -138,6 +138,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 - [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 - [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
+- [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
 
 #### Multimodal
 

convert_hf_to_gguf.py

@@ -888,6 +888,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
             # ref: https://huggingface.co/JetBrains/Mellum-4b-base
             res = "mellum"
+        if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
+            # ref: https://huggingface.co/inclusionAI/Ling-mini-base-2.0
+            res = "bailingmoe2"
 
         if res is None:
             logger.warning("\n")
@@ -7877,6 +7880,103 @@ def prepare_tensors(self):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("BailingMoeV2ForCausalLM")
+class BailingMoeV2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.BAILINGMOE2
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if nextn_layers := self.hparams.get("num_nextn_predict_layers", 0):
+            self.block_count = self.hparams["num_hidden_layers"] + nextn_layers
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        if (rope_dim := hparams.get("head_dim")) is None:
+            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
+
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+        else:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_shared_feed_forward_length(hparams["moe_shared_expert_intermediate_size"])
+        self.gguf_writer.add_expert_weights_scale(hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_count(hparams["num_experts"])
+        self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"])
+        self.gguf_writer.add_expert_group_count(hparams["n_group"])
+        self.gguf_writer.add_expert_group_used_count(hparams["topk_group"])
+        self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"])
+
+        if hparams["score_function"] == "sigmoid":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        elif hparams["score_function"] == "softmax":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        else:
+            raise ValueError(f"Unsupported score_function value: {hparams['score_function']}")
+
+        if (nextn_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(nextn_layers)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if "mlp.experts" in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            tensors: list[tuple[str, Tensor]] = []
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+
+            return tensors
+
+        if name.endswith(".expert_bias"):
+            name = name.replace(".expert_bias", ".expert_bias.bias")
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("ChameleonForConditionalGeneration")
 @ModelBase.register("ChameleonForCausalLM")  # obsolete
 class ChameleonModel(TextModel):

convert_hf_to_gguf_update.py

@@ -139,6 +139,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
+    {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

gguf-py/gguf/constants.py

@@ -101,6 +101,8 @@ class LLM:
         EXPERT_COUNT                      = "{arch}.expert_count"
         EXPERT_USED_COUNT                 = "{arch}.expert_used_count"
         EXPERT_SHARED_COUNT               = "{arch}.expert_shared_count"
+        EXPERT_GROUP_COUNT                = "{arch}.expert_group_count"
+        EXPERT_GROUP_USED_COUNT           = "{arch}.expert_group_used_count"
         EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
         EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
         EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
@@ -387,6 +389,7 @@ class MODEL_ARCH(IntEnum):
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
     BAILINGMOE       = auto()
+    BAILINGMOE2      = auto()
     DOTS1            = auto()
     ARCEE            = auto()
     ERNIE4_5         = auto()
@@ -722,6 +725,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
+    MODEL_ARCH.BAILINGMOE2:      "bailingmoe2",
     MODEL_ARCH.DOTS1:            "dots1",
     MODEL_ARCH.ARCEE:            "arcee",
     MODEL_ARCH.ERNIE4_5:         "ernie4_5",
@@ -2500,6 +2504,35 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.BAILINGMOE2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
     MODEL_ARCH.DOTS1: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/gguf_writer.py

@@ -748,6 +748,12 @@ def add_expert_used_count(self, count: int) -> None:
     def add_expert_shared_count(self, count: int) -> None:
         self.add_uint32(Keys.LLM.EXPERT_SHARED_COUNT.format(arch=self.arch), count)
 
+    def add_expert_group_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_GROUP_COUNT.format(arch=self.arch), count)
+
+    def add_expert_group_used_count(self, count: int) -> None:
+        self.add_uint32(Keys.LLM.EXPERT_GROUP_USED_COUNT.format(arch=self.arch), count)
+
     def add_expert_weights_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -168,6 +168,7 @@ class TensorNameMap:
             "h.{bid}.self_attention.query_key_value",                              # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
             "model.layers.{bid}.self_attn.query_key_value",                        # persimmon
+            "model.layers.{bid}.attention.query_key_value",                        # bailingmoe2
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
             "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
@@ -254,6 +255,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.out_proj",                            # gpt-j
             "language_model.encoder.layers.{bid}.self_attention.dense",     # persimmon
             "model.layers.{bid}.self_attn.dense",                           # persimmon
+            "model.layers.{bid}.attention.dense",                           # bailingmoe2
             "h.{bid}.attn.c_proj",                                          # gpt2
             "transformer.h.{bid}.mixer.out_proj",                           # phi2
             "model.layers.layers.{bid}.self_attn.o_proj",                   # plamo
@@ -365,6 +367,7 @@ class TensorNameMap:
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
             "model.layers.{bid}.mlp.gate.e_score_correction",               # deepseek-v3 dots1
             "model.layers.{bid}.mlp.moe_statics.e_score_correction",        # ernie4.5-moe
+            "model.layers.{bid}.mlp.gate.expert_bias",                      # bailingmoe2
         ),
 
         # Feed-forward up
@@ -528,6 +531,7 @@ class TensorNameMap:
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
             "model.layers.{bid}.self_attn.query_layernorm",                   # hunyuan
+            "model.layers.{bid}.attention.query_layernorm",                   # bailingmoe2
             "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo2
             "layers.{bid}.self_attn.q_norm",                                  # embeddinggemma
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
@@ -541,6 +545,7 @@ class TensorNameMap:
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
             "model.layers.{bid}.self_attn.key_layernorm",                     # hunyuan
+            "model.layers.{bid}.attention.key_layernorm",                     # bailingmoe2
             "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo2
             "layers.{bid}.self_attn.k_norm",                                  # embeddinggemma
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
@@ -561,6 +566,7 @@ class TensorNameMap:
             "transformer.decoder_layer.{bid}.rms_norm_3",   # Grok
             "encoder.layer.{bid}.mlp.layernorm",            # jina-bert-v2
             "encoder.layer.{bid}.layer_norm_2",             # jina-v2-code
+            "model.layers.{bid}.final_layernorm",           # bailingmoe2
         ),
 
         MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: (

src/llama-arch.cpp

@@ -84,6 +84,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
+    { LLM_ARCH_BAILINGMOE2,      "bailingmoe2"      },
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
@@ -129,6 +130,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_COUNT,                      "%s.expert_count"                      },
     { LLM_KV_EXPERT_USED_COUNT,                 "%s.expert_used_count"                 },
     { LLM_KV_EXPERT_SHARED_COUNT,               "%s.expert_shared_count"               },
+    { LLM_KV_EXPERT_GROUP_COUNT,                "%s.expert_group_count"                },
+    { LLM_KV_EXPERT_GROUP_USED_COUNT,           "%s.expert_group_used_count"           },
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
@@ -1921,6 +1924,38 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_BAILINGMOE2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_QKV,           "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_NEXTN_EH_PROJ,      "blk.%d.nextn.eh_proj" },
+            { LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
+            { LLM_TENSOR_NEXTN_ENORM,        "blk.%d.nextn.enorm" },
+            { LLM_TENSOR_NEXTN_HNORM,        "blk.%d.nextn.hnorm" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
+            { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
+            { LLM_TENSOR_LAYER_OUT_NORM,     "blk.%d.layer_output_norm" },
+        },
+    },
     {
         LLM_ARCH_DOTS1,
         {

src/llama-arch.h

@@ -88,6 +88,7 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_BAILINGMOE2,
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
     LLM_ARCH_ERNIE4_5,
@@ -133,6 +134,8 @@ enum llm_kv {
     LLM_KV_EXPERT_COUNT,
     LLM_KV_EXPERT_USED_COUNT,
     LLM_KV_EXPERT_SHARED_COUNT,
+    LLM_KV_EXPERT_GROUP_COUNT,
+    LLM_KV_EXPERT_GROUP_USED_COUNT,
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,

src/llama-chat.cpp

@@ -63,6 +63,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
+    { "bailing-think",     LLM_CHAT_TEMPLATE_BAILING_THINK     },
+    { "bailing2",          LLM_CHAT_TEMPLATE_BAILING2          },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
     { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
     { "hunyuan-moe",       LLM_CHAT_TEMPLATE_HUNYUAN_MOE       },
@@ -191,6 +193,10 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_YANDEX;
     } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
         return LLM_CHAT_TEMPLATE_BAILING;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("\"HUMAN\"") && tmpl_contains("<think>")) {
+        return LLM_CHAT_TEMPLATE_BAILING_THINK;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("<role>HUMAN</role>") && tmpl_contains("<|role_end|>")) {
+        return LLM_CHAT_TEMPLATE_BAILING2;
     } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
     } else if (tmpl_contains("<|endofuserprompt|>")) {
@@ -644,8 +650,8 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << " Ассистент:[SEP]";
         }
-    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) {
-        // Bailing (Ling) template
+    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING || tmpl == LLM_CHAT_TEMPLATE_BAILING_THINK) {
+        // Bailing (Ling/Ring) template
         for (auto message : chat) {
             std::string role(message->role);
 
@@ -658,6 +664,33 @@ int32_t llm_chat_apply_template(
             ss << "<role>" << role << "</role>" << message->content;
         }
 
+        if (add_ass) {
+            ss << "<role>ASSISTANT</role>";
+
+            if (tmpl == LLM_CHAT_TEMPLATE_BAILING_THINK) {
+                ss << "<think>";
+            }
+        }
+    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING2) {
+        // Bailing2 (Ling 2.0) template
+        bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
+
+        if (!has_system) {
+            ss << "<role>SYSTEM</role>detailed thinking off<|role_end|>";
+        }
+
+        for (auto message : chat) {
+            std::string role(message->role);
+
+            if (role == "user") {
+                role = "HUMAN";
+            } else {
+                std::transform(role.begin(), role.end(), role.begin(), ::toupper);
+            }
+
+            ss << "<role>" << role << "</role>" << message->content << "<|role_end|>";
+        }
+
         if (add_ass) {
             ss << "<role>ASSISTANT</role>";
         }