Add bailingmoe-v2 support

convert_hf_to_gguf.py

@@ -864,6 +864,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "96a5f08be6259352137b512d4157e333e21df7edd3fcd152990608735a65b224":
             # ref: https://huggingface.co/inclusionAI/Ling-lite
             res = "bailingmoe"
+        if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
+            # ref: https://huggingface.co/inclusionAI/Ling-mini-2.0
+            res = "bailingmoe-v2"
         if chkhsh == "d353350c764d8c3b39c763113960e4fb4919bea5fbf208a0e3b22e8469dc7406":
             # ref: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
             res = "llama4"
@@ -7879,6 +7882,143 @@ def prepare_tensors(self):
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
 
+@ModelBase.register("BailingMoeV2ForCausalLM")
+class BailingMoeV2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.BAILINGMOE_V2
+
+    @staticmethod
+    def permute(
+        weights: Tensor, n_head: int, n_head_kv: int | None, rope_dim: int | None
+    ):
+        if n_head_kv is not None and n_head != n_head_kv:
+            n_head = n_head_kv
+        if rope_dim is None:
+            rope_dim = weights.shape[0] // n_head
+        weights_rope, weights_nope = weights.reshape(
+            n_head, weights.shape[0] // n_head, *weights.shape[1:]
+        ).split([rope_dim, weights.shape[0] // n_head - rope_dim], dim=1)
+        return torch.cat(
+            [
+                weights_rope.reshape(
+                    n_head, 2, rope_dim // 2, *weights_rope.shape[2:]
+                )
+                .swapaxes(1, 2)
+                .reshape(weights_rope.shape),
+                weights_nope,
+            ], dim=1
+        ).reshape(weights.shape)
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        rope_dim = int(hparams['partial_rotary_factor'] * hparams['head_dim'])
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+        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_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_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']}")
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # skip Multi-Token Prediction (MTP) layers
+        block_count = self.hparams["num_hidden_layers"]
+        match = re.match(r"model.layers.(\d+)", name)
+        if match and int(match.group(1)) >= block_count:
+            return []
+
+        rope_dim = int(self.hparams['partial_rotary_factor'] * self.hparams['head_dim'])
+        if name.endswith("query_key_value.weight"):
+            n_head = self.hparams["num_attention_heads"]
+            n_kv_head = self.hparams.get("num_key_value_heads")
+            n_embd = self.hparams["hidden_size"]
+            if (head_dim := self.hparams.get("head_dim")) is None:
+                head_dim = n_embd // n_head
+            q, k, v = data_torch.split([n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim], dim=-2)
+
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), BailingMoeV2Model.permute(q, n_head, n_head, rope_dim)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), BailingMoeV2Model.permute(k, n_head, n_kv_head, rope_dim)),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
+            ]
+        elif "attention.key_layernorm" in name or "attention.query_layernorm" in name:
+            mapping = {
+                "attention.key_layernorm": "self_attn.key_layernorm",
+                "attention.query_layernorm": "self_attn.query_layernorm",
+            }
+            for k, v in mapping.items():
+                name = name.replace(k, v)
+            return [(self.map_tensor_name(name), BailingMoeV2Model.permute(data_torch, 1, 1, rope_dim))]
+        elif name.find("mlp.experts") != -1:
+            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
+
+        pre_tensor_name_mapping = {
+            "attention.dense": "self_attn.dense",
+            "mlp.gate.expert_bias": "mlp.gate.e_score_correction.bias",
+        }
+        for k, v in pre_tensor_name_mapping.items():
+            name = name.replace(k, v)
+        new_name = self.map_tensor_name(name)
+
+        return [(new_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

convert_hf_to_gguf_update.py

@@ -131,6 +131,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"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", },
     {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
+    {"name": "bailingmoe-v2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-2.0", },
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },

gguf-py/gguf/constants.py

@@ -387,6 +387,7 @@ class MODEL_ARCH(IntEnum):
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
     BAILINGMOE       = auto()
+    BAILINGMOE_V2    = auto()
     DOTS1            = auto()
     ARCEE            = auto()
     ERNIE4_5         = auto()
@@ -723,6 +724,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
+    MODEL_ARCH.BAILINGMOE_V2:    "bailingmoe-v2",
     MODEL_ARCH.DOTS1:            "dots1",
     MODEL_ARCH.ARCEE:            "arcee",
     MODEL_ARCH.ERNIE4_5:         "ernie4_5",
@@ -2502,6 +2504,31 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.BAILINGMOE_V2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_NORM,
+        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.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE,
+    ],
     MODEL_ARCH.DOTS1: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2767,6 +2794,9 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.BAILINGMOE: [
         MODEL_TENSOR.ROPE_FREQS,
     ],
+    MODEL_ARCH.BAILINGMOE_V2: [
+        MODEL_TENSOR.ROPE_FREQS,
+    ],
 }
 
 #

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_BAILINGMOE_V2,    "bailingmoe-v2"    },
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
@@ -1922,6 +1923,35 @@ 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_BAILINGMOE_V2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FREQS,         "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { 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_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_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { 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_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+        },
+    },
     {
         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_BAILINGMOE_V2,
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
     LLM_ARCH_ERNIE4_5,

src/llama-chat.cpp

@@ -63,6 +63,7 @@ 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-v2",        LLM_CHAT_TEMPLATE_BAILING           },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
     { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
     { "hunyuan-moe",       LLM_CHAT_TEMPLATE_HUNYUAN_MOE       },
@@ -189,7 +190,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_MEGREZ;
     } else if (tmpl_contains(" Ассистент:")) {
         return LLM_CHAT_TEMPLATE_YANDEX;
-    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("HUMAN")) {
         return LLM_CHAT_TEMPLATE_BAILING;
     } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA4;