diff --git a/README.md b/README.md
index 327e965fe..d82567202 100644
--- a/README.md
+++ b/README.md
@@ -261,6 +261,7 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 - Exaone 3
 - BaiChuan2
 - MiniCPM / MiniCPM 3
+- XVERSE / XVERSE MoE
 
 
 **Embedding Models**
diff --git a/python/sglang/srt/models/xverse.py b/python/sglang/srt/models/xverse.py
new file mode 100644
index 000000000..460ae820f
--- /dev/null
+++ b/python/sglang/srt/models/xverse.py
@@ -0,0 +1,380 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+# Adapted from
+# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/xverse.py#L1
+"""Inference-only XVERSE model compatible with HuggingFace weights."""
+
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+from vllm.config import CacheConfig
+from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.sampler import Sampler
+from sglang.srt.model_executor.model_runner import InputMetadata
+
+
+class XverseMLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj",
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.down_proj",
+        )
+        if hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Only silu is supported for now."
+            )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class XverseAttention(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        rope_is_neox_style: bool = True,
+        max_position_embeddings: int = 8192,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+        self.head_dim = getattr(
+            config, "head_dim", self.hidden_size // self.total_num_heads
+        )
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.o_proj",
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            is_neox_style=rope_is_neox_style,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class XverseDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+            config, "original_max_position_embeddings", None
+        ):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings
+            )
+        rope_is_neox_style = getattr(config, "rope_is_neox_style", True)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        num_kv_heads = getattr(
+            config, "num_key_value_heads", config.num_attention_heads
+        )
+        self.self_attn = XverseAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=num_kv_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            rope_is_neox_style=rope_is_neox_style,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.mlp = XverseMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=f"{prefix}.mlp",
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class XverseModel(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                XverseDecoderLayer(
+                    config, i, quant_config=quant_config, prefix=f"model.layers.{i}"
+                )
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        if input_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_embeds
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                input_metadata,
+                residual,
+            )
+            # print(f"layer[{i}].hidden_states: {hidden_states}")
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class XverseForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: Optional[CacheConfig] = None,
+        efficient_weight_load=False,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = XverseModel(config, quant_config=quant_config)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.logits_processor = LogitsProcessor(config)
+        self.sampler = Sampler()
+
+        self.param_dict = dict(self.named_parameters())
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
+        # print(f"{hidden_states=}")
+        logits_output = self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
+        return sample_output, logits_output
+
+    def load_weights(
+        self, weights: Iterable[Tuple[str, torch.Tensor]], name=None, loaded_weight=None
+    ):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = self.param_dict
+
+        def load_weights_per_param(name, loaded_weight):
+            if "rotary_emb.inv_freq" in name or "projector" in name:
+                return
+            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                return
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if name.startswith("model.vision_tower") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    return
+                if name.startswith("model.vision_tower") and name not in params_dict:
+                    return
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+        if name is None or loaded_weight is None:
+            for name, loaded_weight in weights:
+                load_weights_per_param(name, loaded_weight)
+        else:
+            load_weights_per_param(name, loaded_weight)
+
+
+EntryClass = XverseForCausalLM
diff --git a/python/sglang/srt/models/xverse_moe.py b/python/sglang/srt/models/xverse_moe.py
new file mode 100644
index 000000000..3aeb9ea98
--- /dev/null
+++ b/python/sglang/srt/models/xverse_moe.py
@@ -0,0 +1,449 @@
+"""
+Copyright 2023-2024 SGLang Team
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+"""Inference-only XVERSE MoE model."""
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+from vllm.config import CacheConfig
+from vllm.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_reduce,
+)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import fused_moe
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.sampler import Sampler
+from sglang.srt.model_executor.forward_batch_info import InputMetadata
+
+
+class XverseMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        reduce_results: bool = True,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2, bias=False, quant_config=quant_config
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            reduce_results=reduce_results,
+        )
+        if hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Only silu is supported for now."
+            )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class XverseMoE(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.rank = get_tensor_model_parallel_rank()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.n_routed_experts = config.num_experts
+        self.top_k = config.moe_top_k
+        if self.tp_size > self.n_routed_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {self.n_routed_experts}."
+            )
+
+        self.experts = nn.ModuleList(
+            [
+                XverseMLP(
+                    hidden_size=config.hidden_size,
+                    intermediate_size=config.intermediate_size,
+                    hidden_act=config.hidden_act,
+                    quant_config=quant_config,
+                    reduce_results=False,
+                )
+                for _ in range(self.n_routed_experts)
+            ]
+        )
+        self.pack_params()
+
+        self.router = ReplicatedLinear(
+            config.hidden_size, self.n_routed_experts, bias=False, quant_config=None
+        )
+
+        if config.num_shared_experts is not None:
+            intermediate_size = config.intermediate_size * config.num_shared_experts
+            self.shared_experts = XverseMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+            )
+
+    def pack_params(self):
+        w1 = []
+        w2 = []
+        for expert in self.experts:
+            w1.append(expert.gate_up_proj.weight)
+            w2.append(expert.down_proj.weight)
+        self.w1 = torch._utils._flatten_dense_tensors(w1)
+        w1s = torch._utils._unflatten_dense_tensors(self.w1, w1)
+        for data, param in zip(w1s, w1):
+            param.data = data
+        self.w1 = self.w1.view(len(w1), *w1s[0].shape)
+
+        self.w2 = torch._utils._flatten_dense_tensors(w2)
+        w2s = torch._utils._unflatten_dense_tensors(self.w2, w2)
+        for data, param in zip(w2s, w2):
+            param.data = data
+
+        self.w2 = self.w2.view(len(w2), *w2s[0].shape)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        if self.config.num_shared_experts is not None:
+            shared_output = self.shared_experts(hidden_states)
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.router(hidden_states)
+        final_hidden_states = fused_moe(
+            hidden_states,
+            self.w1,
+            self.w2,
+            router_logits,
+            self.top_k,
+            renormalize=getattr(self.config, "norm_topk_prob", False),
+            inplace=True,
+        )
+
+        if self.config.num_shared_experts is not None:
+            final_hidden_states = final_hidden_states + shared_output
+        final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
+
+        return final_hidden_states.view(num_tokens, hidden_dim)
+
+
+class XverseAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = hidden_size // self.total_num_heads
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class XverseDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        num_key_value_heads = getattr(
+            config, "num_key_value_heads", config.num_attention_heads
+        )
+        self.self_attn = XverseAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+        )
+        if config.num_experts is not None:
+            self.mlp = XverseMoE(config=config, quant_config=quant_config)
+        else:
+            self.mlp = XverseMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+            )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class XverseModel(nn.Module):
+
+    fall_back_to_pt_during_load = False
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                XverseDecoderLayer(
+                    config, layer_id, cache_config, quant_config=quant_config
+                )
+                for layer_id in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions, hidden_states, input_metadata, residual
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class XverseMoeForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = XverseModel(config, cache_config, quant_config)
+        self.lm_head = ParallelLMHead(
+            config.vocab_size, config.hidden_size, quant_config=quant_config
+        )
+        self.logits_processor = LogitsProcessor(config)
+        self.sampler = Sampler()
+
+        self.param_dict = dict(self.named_parameters())
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata)
+        logits_output = self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+        sample_output = self.sampler(logits_output, input_metadata.sampling_info)
+        return sample_output, logits_output
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        params_dict = self.param_dict
+
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                # Skip experts that are not assigned to this worker.
+                if (
+                    "mlp.experts." in name or "mlp.shared_experts." in name
+                ) and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                # Skip experts that are not assigned to this worker.
+                if (
+                    "mlp.experts." in name or "mlp.shared_experts." in name
+                ) and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+
+EntryClass = XverseMoeForCausalLM