Sync from v0.13

vllm/model_executor/models/qwen2_moe.py

@@ -1,4 +1,6 @@
-# coding=utf-8
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 # Adapted from
 # https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
 # Copyright 2024 The Qwen team.
@@ -22,100 +24,124 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only Qwen2MoE model compatible with HuggingFace weights."""
-from typing import Any, Dict, Iterable, List, Optional, Tuple
+
+from collections.abc import Iterable
+from itertools import islice
+from typing import Any
 
 import torch
 import torch.nn.functional as F
 from torch import nn
-from transformers import PretrainedConfig
+from transformers import Qwen2MoeConfig
 
-from vllm.attention import Attention, AttentionMetadata
-from vllm.distributed import (get_tensor_model_parallel_rank,
-                              get_tensor_model_parallel_world_size,
-                              tensor_model_parallel_all_reduce)
+from vllm.attention.layer import Attention
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import SiluAndMul
-from vllm.model_executor.layers.fused_moe import fused_moe
+from vllm.model_executor.layers.fused_moe import SharedFusedMoE
 from vllm.model_executor.layers.layernorm import RMSNorm
-from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
-                                               QKVParallelLinear,
-                                               ReplicatedLinear,
-                                               RowParallelLinear)
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
-from vllm.model_executor.layers.quantization.base_config import (
-    QuantizationConfig)
+from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import Sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
-    ParallelLMHead, VocabParallelEmbedding)
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import SamplerOutput
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsLoRA, SupportsPP
+from .utils import (
+    AutoWeightsLoader,
+    extract_layer_index,
+    is_pp_missing_parameter,
+    make_empty_intermediate_tensors_factory,
+    make_layers,
+    maybe_prefix,
+)
+
+logger = init_logger(__name__)
 
 
 class Qwen2MoeMLP(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
         intermediate_size: int,
         hidden_act: str,
-        quant_config: Optional[QuantizationConfig] = None,
+        quant_config: QuantizationConfig | None = None,
         reduce_results: bool = True,
+        expert_gate: torch.nn.Linear | None = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
-            hidden_size, [intermediate_size] * 2,
+            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)
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj",
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            reduce_results=reduce_results,
+            prefix=f"{prefix}.down_proj",
+        )
         if hidden_act != "silu":
-            raise ValueError(f"Unsupported activation: {hidden_act}. "
-                             "Only silu is supported for now.")
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
+            )
         self.act_fn = SiluAndMul()
+        self.expert_gate = expert_gate
 
     def forward(self, x):
         gate_up, _ = self.gate_up_proj(x)
-        x = self.act_fn(gate_up)
-        x, _ = self.down_proj(x)
-        return x
+        out = self.act_fn(gate_up)
+        out, _ = self.down_proj(out)
+
+        if self.expert_gate is not None:
+            out = F.sigmoid(self.expert_gate(x)) * out
+
+        return out
 
 
 class Qwen2MoeSparseMoeBlock(nn.Module):
-
     def __init__(
         self,
-        config: PretrainedConfig,
-        quant_config: Optional[QuantizationConfig] = None,
+        config: Qwen2MoeConfig,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
     ):
         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.num_experts_per_tok
-        if self.tp_size > self.n_routed_experts:
+
+        if self.tp_size > config.num_experts:
             raise ValueError(
                 f"Tensor parallel size {self.tp_size} is greater than "
-                f"the number of experts {self.n_routed_experts}.")
+                f"the number of experts {config.num_experts}."
+            )
 
-        self.experts = nn.ModuleList([
-            Qwen2MoeMLP(hidden_size=config.hidden_size,
-                        intermediate_size=config.moe_intermediate_size,
-                        hidden_act=config.hidden_act,
-                        quant_config=quant_config,
-                        reduce_results=False)
-            for idx in range(self.n_routed_experts)
-        ])
-        self.pack_params()
+        self.gate = ReplicatedLinear(
+            config.hidden_size,
+            config.num_experts,
+            bias=False,
+            quant_config=None,
+            prefix=f"{prefix}.gate",
+        )
+
+        self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
 
-        self.gate = ReplicatedLinear(config.hidden_size,
-                                     self.n_routed_experts,
-                                     bias=False,
-                                     quant_config=None)
         if config.shared_expert_intermediate_size > 0:
             self.shared_expert = Qwen2MoeMLP(
                 hidden_size=config.hidden_size,
@@ -123,71 +149,57 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
                 hidden_act=config.hidden_act,
                 quant_config=quant_config,
                 reduce_results=False,
+                expert_gate=self.shared_expert_gate,
+                prefix=f"{prefix}.shared_expert",
             )
         else:
             self.shared_expert = None
-        self.shared_expert_gate = torch.nn.Linear(config.hidden_size,
-                                                  1,
-                                                  bias=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)
+        self.experts = SharedFusedMoE(
+            shared_experts=self.shared_expert,
+            num_experts=config.num_experts,
+            top_k=config.num_experts_per_tok,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.moe_intermediate_size,
+            reduce_results=False,
+            renormalize=config.norm_topk_prob,
+            quant_config=quant_config,
+            prefix=f"{prefix}.experts",
+        )
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        num_tokens, hidden_dim = hidden_states.shape
+        # NOTE: hidden_states can have either 1D or 2D shape.
+        orig_shape = hidden_states.shape
+        hidden_dim = hidden_states.shape[-1]
         hidden_states = hidden_states.view(-1, hidden_dim)
-        shared_output = None
-        if self.shared_expert is not None:
-            shared_output = self.shared_expert(hidden_states)
-            if self.shared_expert_gate is not None:
-                shared_output = F.sigmoid(
-                    self.shared_expert_gate(hidden_states)) * shared_output
 
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
-        final_hidden_states = fused_moe(hidden_states,
-                                        self.w1,
-                                        self.w2,
-                                        router_logits,
-                                        self.top_k,
-                                        renormalize=self.config.norm_topk_prob,
-                                        inplace=True)
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states, router_logits=router_logits
+        )
+        if self.shared_expert is not None:
+            final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
+        if self.tp_size > 1:
+            final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(  # noqa E501
+                final_hidden_states
+            )
 
-        if shared_output 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)
+        return final_hidden_states.view(orig_shape)
 
 
 class Qwen2MoeAttention(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
         num_heads: int,
         num_kv_heads: int,
-        rope_theta: float = 10000,
-        rope_scaling: Optional[Dict[str, Any]] = None,
+        rope_parameters: dict[str, Any] | None = None,
         max_position_embeddings: int = 8192,
-        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
+        dual_chunk_attention_config: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -209,8 +221,8 @@ class Qwen2MoeAttention(nn.Module):
         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.dual_chunk_attention_config = dual_chunk_attention_config
 
         self.qkv_proj = QKVParallelLinear(
             hidden_size,
@@ -219,6 +231,7 @@ class Qwen2MoeAttention(nn.Module):
             self.total_num_kv_heads,
             bias=True,
             quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
         )
 
         self.o_proj = RowParallelLinear(
@@ -226,185 +239,191 @@ class Qwen2MoeAttention(nn.Module):
             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,
+            rope_parameters=rope_parameters,
+            dual_chunk_attention_config=dual_chunk_attention_config,
+        )
+        self.attn = Attention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
+            **{
+                "layer_idx": extract_layer_index(prefix),
+                "dual_chunk_attention_config": dual_chunk_attention_config,
+            }
+            if dual_chunk_attention_config
+            else {},
         )
-        self.attn = Attention(self.num_heads,
-                              self.head_dim,
-                              self.scaling,
-                              num_kv_heads=self.num_kv_heads)
 
     def forward(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
-        kv_cache: torch.Tensor,
-        attn_metadata: AttentionMetadata,
     ) -> 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, kv_cache, attn_metadata)
+        attn_output = self.attn(q, k, v)
         output, _ = self.o_proj(attn_output)
         return output
 
 
 class Qwen2MoeDecoderLayer(nn.Module):
-
     def __init__(
         self,
-        config: PretrainedConfig,
-        layer_idx: int,
-        quant_config: Optional[QuantizationConfig] = None,
+        config: Qwen2MoeConfig,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = 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)
-        max_position_embeddings = getattr(config, "max_position_embeddings",
-                                          8192)
+        dual_chunk_attention_config = getattr(
+            config, "dual_chunk_attention_config", None
+        )
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = Qwen2MoeAttention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
             num_kv_heads=config.num_key_value_heads,
-            rope_theta=rope_theta,
-            rope_scaling=rope_scaling,
+            rope_parameters=config.rope_parameters,
             max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
             quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+            dual_chunk_attention_config=dual_chunk_attention_config,
         )
-        if (config.num_experts is not None
-                and (layer_idx + 1) % config.decoder_sparse_step == 0):
-            self.mlp = Qwen2MoeSparseMoeBlock(config=config,
-                                              quant_config=quant_config)
+
+        # Note: Qwen/Qwen2-57B-A14B-Instruct does not have
+        # `mlp_only_layers` in the config.
+        layer_idx = extract_layer_index(prefix)
+        mlp_only_layers = (
+            [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
+        )
+        if (layer_idx not in mlp_only_layers) and (
+            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
+        ):
+            self.mlp = Qwen2MoeSparseMoeBlock(
+                config=config, quant_config=quant_config, prefix=f"{prefix}.mlp"
+            )
         else:
             self.mlp = Qwen2MoeMLP(
                 hidden_size=config.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)
+        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,
-        kv_cache: torch.Tensor,
-        attn_metadata: AttentionMetadata,
-        residual: Optional[torch.Tensor],
+        residual: torch.Tensor | None,
     ) -> 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, residual = self.input_layernorm(hidden_states, residual)
         hidden_states = self.self_attn(
             positions=positions,
             hidden_states=hidden_states,
-            kv_cache=kv_cache,
-            attn_metadata=attn_metadata,
         )
 
         # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(
-            hidden_states, residual)
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
         hidden_states = self.mlp(hidden_states)
         return hidden_states, residual
 
 
+@support_torch_compile
 class Qwen2MoeModel(nn.Module):
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        quant_config: Optional[QuantizationConfig] = None,
-    ) -> None:
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
-        self.padding_idx = config.pad_token_id
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
         self.vocab_size = config.vocab_size
+        self.config = config
 
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
             config.hidden_size,
+            quant_config=quant_config,
+            prefix=f"{prefix}.embed_tokens",
+        )
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: Qwen2MoeDecoderLayer(
+                config=config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=f"{prefix}.layers",
         )
-        self.layers = nn.ModuleList([
-            Qwen2MoeDecoderLayer(config, layer_idx, quant_config=quant_config)
-            for layer_idx in range(config.num_hidden_layers)
-        ])
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
+            ["hidden_states", "residual"], config.hidden_size
+        )
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
 
     def forward(
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
-    ) -> 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,
-                                            kv_caches[i], attn_metadata,
-                                            residual)
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> torch.Tensor | IntermediateTensors:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.embed_input_ids(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
+            hidden_states, residual = layer(positions, hidden_states, residual)
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors(
+                {"hidden_states": hidden_states, "residual": residual}
+            )
         hidden_states, _ = self.norm(hidden_states, residual)
         return hidden_states
 
+    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
+        # Params for weights, fp8 weight scales, fp8 activation scales
+        # (param_name, weight_name, expert_id, shard_id)
+        return SharedFusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=self.config.num_experts,
+        )
 
-class Qwen2MoeForCausalLM(nn.Module):
-
-    fall_back_to_pt_during_load = False
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        quant_config: Optional[QuantizationConfig] = None,
-    ) -> None:
-        super().__init__()
-        self.config = config
-        self.quant_config = quant_config
-        self.model = Qwen2MoeModel(config, quant_config)
-        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
-        self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = Sampler()
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        positions: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
-    ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, kv_caches,
-                                   attn_metadata)
-        return hidden_states
-
-    def compute_logits(self, hidden_states: torch.Tensor,
-                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
-        logits = self.logits_processor(self.lm_head.weight, hidden_states,
-                                       sampling_metadata)
-        return logits
-
-    def sample(
-        self,
-        logits: Optional[torch.Tensor],
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
@@ -415,33 +434,164 @@ class Qwen2MoeForCausalLM(nn.Module):
         ]
 
         params_dict = dict(self.named_parameters())
+        loaded_params: set[str] = set()
+        expert_params_mapping = self.get_expert_mapping()
         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:
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name:
                     continue
+                # We have mlp.experts[0].gate_proj in the checkpoint.
+                # Since we handle the experts below in expert_params_mapping,
+                # we need to skip here BEFORE we update the name, otherwise
+                # name will be updated to mlp.experts[0].gate_up_proj, which
+                # will then be updated below in expert_params_mapping
+                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
+                if "mlp.experts" 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:
+                if (
+                    name.endswith(".bias") or 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_expert." in name)
-                        and name not in params_dict):
+                # Skip layers on other devices.
+                if is_pp_missing_parameter(name, self):
                     continue
+                if 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_expert." 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)
+                for mapping in expert_params_mapping:
+                    param_name, weight_name, expert_id, shard_id = mapping
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+
+                    # Skip layers on other devices.
+                    if is_pp_missing_parameter(name, self):
+                        continue
+                    # Skip loading extra bias for GPTQ models.
+                    if (
+                        name.endswith(".bias") or name.endswith("_bias")
+                    ) and name not in params_dict:
+                        continue
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    weight_loader(
+                        param,
+                        loaded_weight,
+                        name,
+                        shard_id=shard_id,
+                        expert_id=expert_id,
+                    )
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if (
+                        name.endswith(".bias") or name.endswith("_bias")
+                    ) and name not in params_dict:
+                        continue
+                    # Skip layers on other devices.
+                    if is_pp_missing_parameter(name, self):
+                        continue
+                    # Remapping the name of FP8 kv-scale.
+                    if name.endswith("kv_scale"):
+                        remapped_kv_scale_name = name.replace(
+                            ".kv_scale", ".attn.kv_scale"
+                        )
+                        if remapped_kv_scale_name not in params_dict:
+                            logger.warning_once(
+                                "Found kv_scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv_scale is not loaded.",  #  noqa: E501
+                                name,
+                                remapped_kv_scale_name,
+                            )
+                            continue
+                        else:
+                            name = remapped_kv_scale_name
+                    # GGUF: make sure that shared_expert_gate is a 2D tensor.
+                    if (
+                        "mlp.shared_expert_gate" in name
+                        and len(loaded_weight.shape) == 1
+                    ):
+                        loaded_weight = loaded_weight[None, :]
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
+
+class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
+    fall_back_to_pt_during_load = False
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ]
+    }
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        self.config = config
+        self.quant_config = quant_config
+        # Only perform the following mapping when Qwen2MoeMLP exists
+        if (
+            getattr(config, "mlp_only_layers", [])
+            or config.shared_expert_intermediate_size > 0
+        ):
+            self.packed_modules_mapping["gate_up_proj"] = ["gate_proj", "up_proj"]
+
+        self.model = Qwen2MoeModel(
+            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
+        )
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+            prefix=maybe_prefix(prefix, "lm_head"),
+        )
+        if self.config.tie_word_embeddings:
+            self.lm_head.weight = self.model.embed_tokens.weight
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors
+        )
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.embed_input_ids(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> torch.Tensor | IntermediateTensors:
+        hidden_states = self.model(
+            input_ids, positions, intermediate_tensors, inputs_embeds
+        )
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor | None:
+        logits = self.logits_processor(self.lm_head, hidden_states)
+        return logits
+
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(self)
+        return loader.load_weights(weights)
+
+    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
+        return self.model.get_expert_mapping()