Sync from v0.13

vllm/model_executor/models/minicpm.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/llama/modeling_llama.py
 # Copyright 2023 The vLLM team.
@@ -21,35 +23,54 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only MiniCPM model compatible with HuggingFace weights."""
+
 import math
-from typing import Any, Dict, Iterable, List, Optional, Tuple
+from collections.abc import Iterable
+from itertools import islice
+from typing import Any
 
 import torch
 from torch import nn
+from transformers import PretrainedConfig
 
-from vllm.attention import Attention, AttentionMetadata
-from vllm.config import LoRAConfig
-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.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_rank,
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_reduce,
+)
+from vllm.model_executor.layers.activation import FatreluAndMul, SiluAndMul
+from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
 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 (
-    DEFAULT_VOCAB_PADDING_SIZE, 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.model_executor.utils import set_weight_attrs
-from vllm.sequence import SamplerOutput
+from vllm.platforms import current_platform
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
+from .utils import (
+    AutoWeightsLoader,
+    is_pp_missing_parameter,
+    make_empty_intermediate_tensors_factory,
+    make_layers,
+    maybe_prefix,
+)
 
 
 class MiniCPMMoE(nn.Module):
@@ -67,8 +88,8 @@ class MiniCPMMoE(nn.Module):
         top_k: int,
         hidden_size: int,
         intermediate_size: int,
-        params_dtype: Optional[torch.dtype] = None,
-        tp_size: Optional[int] = None,
+        params_dtype: torch.dtype | None = None,
+        tp_size: int | None = None,
     ):
         super().__init__()
         self.tp_size = tp_size or get_tensor_model_parallel_world_size()
@@ -81,34 +102,53 @@ class MiniCPMMoE(nn.Module):
             params_dtype = torch.get_default_dtype()
         self.params_dtype = params_dtype
 
-        self.gate = ReplicatedLinear(self.hidden_size,
-                                     self.num_total_experts,
-                                     bias=False,
-                                     params_dtype=self.params_dtype,
-                                     quant_config=None)
+        self.gate = ReplicatedLinear(
+            self.hidden_size,
+            self.num_total_experts,
+            bias=False,
+            params_dtype=self.params_dtype,
+            quant_config=None,
+        )
 
         self.ws = nn.Parameter(
-            torch.empty(self.num_total_experts,
-                        2 * self.intermediate_size,
-                        self.hidden_size,
-                        device="cuda",
-                        dtype=self.params_dtype))
+            torch.empty(
+                self.num_total_experts,
+                2 * self.intermediate_size,
+                self.hidden_size,
+                device=current_platform.device_type,
+                dtype=self.params_dtype,
+            )
+        )
         self.w2s = nn.Parameter(
-            torch.empty(self.num_total_experts,
-                        self.hidden_size,
-                        self.intermediate_size,
-                        device="cuda",
-                        dtype=self.params_dtype))
+            torch.empty(
+                self.num_total_experts,
+                self.hidden_size,
+                self.intermediate_size,
+                device=current_platform.device_type,
+                dtype=self.params_dtype,
+            )
+        )
 
-        set_weight_attrs(self.ws, {
-            "weight_loader": self.weight_loader,
-        })
-        set_weight_attrs(self.w2s, {
-            "weight_loader": self.weight_loader,
-        })
+        set_weight_attrs(
+            self.ws,
+            {
+                "weight_loader": self.weight_loader,
+            },
+        )
+        set_weight_attrs(
+            self.w2s,
+            {
+                "weight_loader": self.weight_loader,
+            },
+        )
 
-    def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
-                      weight_name: str, expert_id: int):
+    def weight_loader(
+        self,
+        param: nn.Parameter,
+        loaded_weight: torch.Tensor,
+        weight_name: str,
+        expert_id: int,
+    ):
         tp_rank = get_tensor_model_parallel_rank()
         param_data = param.data
         shard_size = self.intermediate_size
@@ -116,8 +156,9 @@ class MiniCPMMoE(nn.Module):
         if weight_name.endswith("w1.weight"):
             param_data[expert_id, 0:shard_size, :] = loaded_weight[shard, :]
         if weight_name.endswith("w3.weight"):
-            param_data[expert_id,
-                       shard_size:2 * shard_size, :] = loaded_weight[shard, :]
+            param_data[expert_id, shard_size : 2 * shard_size, :] = loaded_weight[
+                shard, :
+            ]
         if weight_name.endswith("w2.weight"):
             param_data[expert_id, :, :] = loaded_weight[:, shard]
 
@@ -126,43 +167,55 @@ class MiniCPMMoE(nn.Module):
         hidden_states = hidden_states.view(-1, self.hidden_size)
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
-        final_hidden_states = fused_moe(hidden_states,
-                                        self.ws,
-                                        self.w2s,
-                                        router_logits,
-                                        self.top_k,
-                                        renormalize=True,
-                                        inplace=True)
+
+        topk_weights, topk_ids, _ = fused_topk(
+            hidden_states, router_logits, self.top_k, renormalize=True
+        )
+
+        final_hidden_states = fused_experts(
+            hidden_states, self.ws, self.w2s, topk_weights, topk_ids, inplace=True
+        )
 
         if self.tp_size > 1:
-            final_hidden_states = tensor_model_parallel_all_reduce(
-                final_hidden_states)
+            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(num_tokens, hidden_size)
 
 
 class MiniCPMMLP(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
         intermediate_size: int,
         hidden_act: str,
-        quant_config: Optional[QuantizationConfig] = None,
+        hidden_act_param: float,
+        quant_config: QuantizationConfig | 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)
-        if hidden_act != "silu":
-            raise ValueError(f"Unsupported activation: {hidden_act}. "
-                             "Only silu is supported for now.")
-        self.act_fn = SiluAndMul()
+            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":
+            self.act_fn = SiluAndMul()
+        elif hidden_act == "fatrelu":
+            self.act_fn = FatreluAndMul(threshold=hidden_act_param)
+        else:
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. "
+                "Only silu and fatrelu are supported for now."
+            )
 
     def forward(self, x):
         gate_up, _ = self.gate_up_proj(x)
@@ -172,16 +225,16 @@ class MiniCPMMLP(nn.Module):
 
 
 class MiniCPMAttention(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 = "",
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -203,7 +256,6 @@ class MiniCPMAttention(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.qkv_proj = QKVParallelLinear(
@@ -213,35 +265,36 @@ class MiniCPMAttention(nn.Module):
             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,
+            rope_parameters=rope_parameters,
         )
-        # set rope as fp32 instead of bf16
-        self.rotary_emb.cos_sin_cache = self.rotary_emb._compute_cos_sin_cache(
+
+        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",
         )
-        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)
@@ -249,109 +302,138 @@ class MiniCPMAttention(nn.Module):
         q, k = q.float(), k.float()
         q, k = self.rotary_emb(positions, q, k)
         q, k = q.to(orig_dtype), k.to(orig_dtype)
-        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 MiniCPMDecoderLayer(nn.Module):
-
     def __init__(
         self,
-        config,
-        quant_config: Optional[QuantizationConfig] = None,
+        config: PretrainedConfig,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.config = config
+        self.cache_config = cache_config
+        self.quant_config = quant_config
         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)
+        self.max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        self.prefix = prefix
+        self._init_attn_block()
+        self._init_ffn_block()
+
+    def _init_attn_block(self):
+        self.input_layernorm = RMSNorm(
+            self.config.hidden_size, eps=self.config.rms_norm_eps
+        )
         self.self_attn = MiniCPMAttention(
             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,
-            max_position_embeddings=max_position_embeddings,
-            quant_config=quant_config,
+            num_heads=self.config.num_attention_heads,
+            num_kv_heads=self.config.num_key_value_heads,
+            rope_parameters=self.config.rope_parameters,
+            max_position_embeddings=self.max_position_embeddings,
+            cache_config=self.cache_config,
+            quant_config=self.quant_config,
+            prefix=f"{self.prefix}.self_attn",
+        )
+
+    def _init_ffn_block(self):
+        self.post_attention_layernorm = RMSNorm(
+            self.config.hidden_size, eps=self.config.rms_norm_eps
         )
         self.num_experts = getattr(self.config, "num_experts", 0)
         if self.num_experts == 0:
             self.mlp = MiniCPMMLP(
                 hidden_size=self.hidden_size,
-                intermediate_size=config.intermediate_size,
-                hidden_act=config.hidden_act,
-                quant_config=quant_config,
+                intermediate_size=self.config.intermediate_size,
+                hidden_act=self.config.hidden_act,
+                hidden_act_param=getattr(self.config, "hidden_act_param", 0.0),
+                quant_config=self.quant_config,
             )
         else:
-            self.mlp = MiniCPMMoE(num_experts=config.num_experts,
-                                  top_k=config.num_experts_per_tok,
-                                  hidden_size=config.hidden_size,
-                                  intermediate_size=config.intermediate_size)
-        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.mlp = MiniCPMMoE(
+                num_experts=self.config.num_experts,
+                top_k=self.config.num_experts_per_tok,
+                hidden_size=self.config.hidden_size,
+                intermediate_size=self.config.intermediate_size,
+            )
 
     def forward(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
-        kv_cache: torch.Tensor,
-        attn_metadata: AttentionMetadata,
-        residual: Optional[torch.Tensor],
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        residual: torch.Tensor | None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         # Self Attention
         residual = hidden_states
         hidden_states = self.input_layernorm(hidden_states)
         hidden_states = self.self_attn(
             positions=positions,
             hidden_states=hidden_states,
-            kv_cache=kv_cache,
-            attn_metadata=attn_metadata,
         )
-        hidden_states = residual + hidden_states * \
-            (self.config.scale_depth / math.sqrt(self.config.num_hidden_layers))
+        hidden_states = residual + hidden_states * (
+            self.config.scale_depth / math.sqrt(self.config.num_hidden_layers)
+        )
 
         # Fully Connected
         residual = hidden_states
         hidden_states = self.post_attention_layernorm(hidden_states)
         hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states * \
-            (self.config.scale_depth / math.sqrt(self.config.num_hidden_layers))
+        hidden_states = residual + hidden_states * (
+            self.config.scale_depth / math.sqrt(self.config.num_hidden_layers)
+        )
 
         return hidden_states, None
 
 
+@support_torch_compile
 class MiniCPMModel(nn.Module):
-
-    def __init__(
-        self,
-        config,
-        quant_config: Optional[QuantizationConfig] = None,
-        lora_config: Optional[LoRAConfig] = None,
-    ) -> None:
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
         self.config = config
-        self.padding_idx = config.pad_token_id
-        lora_vocab = (lora_config.lora_extra_vocab_size *
-                      (lora_config.max_loras or 1)) if lora_config else 0
-        self.vocab_size = config.vocab_size + lora_vocab
-        self.org_vocab_size = config.vocab_size
+        self.cache_config = cache_config
+        self.quant_config = quant_config
+
+        self.vocab_size = config.vocab_size
+
         self.embed_tokens = VocabParallelEmbedding(
             self.vocab_size,
             config.hidden_size,
-            org_num_embeddings=config.vocab_size,
         )
-        self.layers = nn.ModuleList([
-            MiniCPMDecoderLayer(config, quant_config)
-            for _ in range(config.num_hidden_layers)
-        ])
+        self.num_experts = getattr(self.config, "num_experts", 0)
+        self._init_layers(prefix, config, cache_config, quant_config)
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
-    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        self.aux_hidden_state_layers = tuple[int, ...]()
+
+        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
+            ["hidden_states", "residual"], self.config.hidden_size
+        )
+
+    def _init_layers(
+        self,
+        prefix: str,
+        config: PretrainedConfig,
+        cache_config: CacheConfig | None,
+        quant_config: QuantizationConfig | None,
+    ):
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: MiniCPMDecoderLayer(
+                config, cache_config, quant_config, prefix=prefix
+            ),
+            prefix=f"{prefix}.layers",
+        )
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
         embedding = self.embed_tokens(input_ids)
         return embedding * self.config.scale_emb
 
@@ -359,30 +441,115 @@ class MiniCPMModel(nn.Module):
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
-        inputs_embeds: Optional[torch.Tensor] = None,
-    ) -> torch.Tensor:
-        if inputs_embeds is not None:
-            hidden_states = inputs_embeds
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
+        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:
-            hidden_states = self.get_input_embeddings(input_ids)
-        residual = None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
 
-        for i in range(len(self.layers)):
-            layer = self.layers[i]
+        aux_hidden_states = []
+        for idx, layer in enumerate(
+            islice(self.layers, self.start_layer, self.end_layer)
+        ):
+            if idx in self.aux_hidden_state_layers:
+                aux_hidden_states.append(
+                    hidden_states + residual if residual is not None else hidden_states
+                )
             hidden_states, residual = layer(
                 positions,
                 hidden_states,
-                kv_caches[i],
-                attn_metadata,
                 residual,
             )
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors(
+                {"hidden_states": hidden_states, "residual": residual}
+            )
+
         hidden_states = self.norm(hidden_states)
+
+        if len(aux_hidden_states) > 0:
+            return hidden_states, aux_hidden_states
         return hidden_states
 
+    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"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        expert_params_mapping = [
+            # (param_name, weight_name, expert_id)
+            (
+                "ws" if weight_name in ["w1", "w3"] else "w2s",
+                f"experts.{expert_id}.{weight_name}.weight",
+                expert_id,
+            )
+            for expert_id in range(self.num_experts)
+            for weight_name in ["w1", "w2", "w3"]
+        ]
+        params_dict = dict(self.named_parameters())
 
-class MiniCPMForCausalLM(nn.Module):
+        loaded_params: set[str] = set()
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+            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.
+                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
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                for param_name, weight_name, expert_id in expert_params_mapping:
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+                    if is_pp_missing_parameter(name, self):
+                        continue
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    weight_loader(
+                        param, loaded_weight, weight_name, expert_id=expert_id
+                    )
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    if is_pp_missing_parameter(name, self):
+                        continue
+                    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 MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
@@ -396,136 +563,94 @@ class MiniCPMForCausalLM(nn.Module):
     }
 
     # LoRA specific attributes
-    supported_lora_modules = [
-        "qkv_proj",
-        "o_proj",
-        "gate_up_proj",
-        "down_proj",
-        "embed_tokens",
-        "lm_head",
-    ]
     embedding_modules = {
         "embed_tokens": "input_embeddings",
         "lm_head": "output_embeddings",
     }
-    embedding_padding_modules = ["lm_head"]
 
-    def __init__(
-        self,
-        config,
-        quant_config: Optional[QuantizationConfig] = None,
-        lora_config: Optional[LoRAConfig] = None,
-    ) -> None:
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        parallel_config = vllm_config.parallel_config
+
+        self.prefix = prefix
+        self.vllm_config = vllm_config
         self.config = config
-        self.num_experts = getattr(self.config, "num_experts", 0)
+
+        self.cache_config = cache_config
         self.quant_config = quant_config
-        self.model = MiniCPMModel(config,
-                                  quant_config,
-                                  lora_config=lora_config)
-        unpadded_vocab_size = config.vocab_size
-        if lora_config:
-            unpadded_vocab_size += lora_config.lora_extra_vocab_size
-        if not self.config.tie_word_embeddings:
-            self.lm_head = ParallelLMHead(
-                unpadded_vocab_size,
-                config.hidden_size,
-                org_num_embeddings=config.vocab_size,
-                padding_size=DEFAULT_VOCAB_PADDING_SIZE
-                # We need bigger padding if using lora for kernel
-                # compatibility
-                if not lora_config else lora_config.lora_vocab_padding_size,
-            )
+
+        self.model = self._init_model(
+            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 config.tie_word_embeddings:
+            self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
         self.scale_width = self.config.hidden_size / self.config.dim_model_base
 
-        self.logits_processor = LogitsProcessor(unpadded_vocab_size,
-                                                config.vocab_size)
-        self.sampler = Sampler()
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors
+        )
+        if parallel_config.enable_eplb and getattr(config, "num_experts", 0) > 0:
+            raise NotImplementedError("EPLB is not supported for MiniCPM yet.")
+
+    def _init_model(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        return MiniCPMModel(vllm_config=vllm_config, prefix=prefix)
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.embed_input_ids(input_ids)
+
+    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
+        self.model.aux_hidden_state_layers = layers
+
+    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
+        num_layers = len(self.model.layers)
+        return (2, num_layers // 2, num_layers - 3)
 
     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
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]:
+        model_output = self.model(
+            input_ids, positions, intermediate_tensors, inputs_embeds
+        )
 
-    def compute_logits(self, hidden_states: torch.Tensor,
-                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
-        hidden_states = hidden_states / self.scale_width
-        if self.config.tie_word_embeddings:
-            lm_head_weight = self.model.embed_tokens.weight
+        if isinstance(model_output, tuple) and len(model_output) == 2:
+            # Aux hidden states are present.
+            hidden_states, aux_hidden_states = model_output
+            hidden_states = hidden_states / self.scale_width
+            return hidden_states, aux_hidden_states
         else:
-            lm_head_weight = self.lm_head.weight
-        logits = self.logits_processor(lm_head_weight, hidden_states,
-                                       sampling_metadata)
+            # Only hidden states or IntermediateTensors
+            if isinstance(model_output, IntermediateTensors):
+                return model_output
+            else:
+                hidden_states = model_output / self.scale_width
+                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 sample(
-        self,
-        logits: 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]]):
-        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),
-        ]
-        expert_params_mapping = [
-            # (param_name, weight_name, expert_id)
-            ("ws" if weight_name in ["w1", "w3"] else "w2s",
-             f"experts.{expert_id}.{weight_name}.weight", expert_id)
-            for expert_id in range(self.num_experts)
-            for weight_name in ["w1", "w2", "w3"]
-        ]
-        params_dict = dict(self.named_parameters())
-        for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
-            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.
-                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
-                param = params_dict[name]
-                weight_loader = param.weight_loader
-                weight_loader(param, loaded_weight, shard_id)
-                break
-            else:
-                for param_name, weight_name, expert_id in expert_params_mapping:
-                    if weight_name not in name:
-                        continue
-                    name = name.replace(weight_name, param_name)
-                    param = params_dict[name]
-                    weight_loader = param.weight_loader
-                    weight_loader(param,
-                                  loaded_weight,
-                                  weight_name,
-                                  expert_id=expert_id)
-                    break
-                else:
-                    # Skip loading extra bias for GPTQ models.
-                    if name.endswith(".bias") 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)
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(weights)