model : add EXAONE MoE (#18543)

* Add EXAONE MoE implementations Co-authored-by: Junwon Hwang <nuclear1221@gmail.com> * Address PR feedback * Address PR feedback * [WIP] Add MTP for EXAONE-MoE * Address PR feedback * Address PR feedback * Address PR feedback * Address PR feedback * Address PR feedback * Address PR feedback * Address PR feedback --------- Co-authored-by: LG-AI-EXAONE <exaonemodels@lgresearch.ai>
2026-01-14 07:28:38 +09:00
parent e4832e3ae4
commit 60591f01d4
17 changed files with 656 additions and 2 deletions
@@ -1252,6 +1252,9 @@ class TextModel(ModelBase):
        if chkhsh == "16389f0a1f51ee53e562ffd51c371dc508639ab0e4261502071836e50e223e91":
            # ref: https://huggingface.co/upstage/Solar-Open-100B
            res = "solar-open"
+        if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f":
+            # ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
+            res = "exaone-moe"

        if res is None:
            logger.warning("\n")
@@ -8748,6 +8751,106 @@ class Exaone4Model(TextModel):
                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))


+@ModelBase.register("ExaoneMoEForCausalLM")
+class ExaoneMoEModel(Exaone4Model):
+    model_arch = gguf.MODEL_ARCH.EXAONE_MOE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+        moe_intermediate_size = self.hparams["moe_intermediate_size"]
+        num_shared_experts = self.hparams["num_shared_experts"]
+        self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        self.gguf_writer.add_expert_shared_count(num_shared_experts)
+        self.gguf_writer.add_expert_shared_feed_forward_length(moe_intermediate_size * num_shared_experts)
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+        n_dense_layer = self.hparams.get("first_k_dense_replace", self.hparams.get("first_last_k_dense_replace", 0))
+        self.gguf_writer.add_leading_dense_block_count(n_dense_layer)
+        # For here, we hard-code the number of NextN/MTP layers to 1 for K-EXAONE,
+        # so that we can convert MTP weights to GGUF format for speculative decoding.
+        # This is because HF config of K-EXAONE does not have `num_nextn_predict_layers` at now.
+        # Will be updated when HF config is updated.
+        self.gguf_writer.add_nextn_predict_layers(self.hparams.get("num_nextn_predict_layers", 1))
+
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("mtp."):
+            if name.find("layers.") != -1:
+                # `mtp.layers.0.[module_name]` format
+                name = name.replace(f"mtp.layers.{bid}", f"model.layers.{bid + self.hparams['num_hidden_layers']}")
+            else:
+                # mtp fc/norm weights
+                remapper = {
+                    "mtp.fc": "model.layers.{bid}.eh_proj",
+                    "mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm",
+                    "mtp.pre_fc_norm_hidden": "model.layers.{bid}.hnorm",
+                    "mtp.norm": "model.layers.{bid}.shared_head.norm",
+                }
+                _n = Path(name)
+                new_name = remapper[_n.stem] + _n.suffix
+
+                # set shared weights for all NextN/MTP layers
+                tensors = []
+                for bid in range(self.hparams['num_hidden_layers'], self.block_count):
+                    new_name = new_name.format(bid=bid)
+                    tensors.append((self.map_tensor_name(new_name), data_torch))
+                return tensors
+
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            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:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # 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
+            else:
+                return []
+
+        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("GraniteForCausalLM")
 class GraniteModel(LlamaModel):
    """Conversion for IBM's GraniteForCausalLM"""