Sync from v0.13

vllm/model_executor/models/qwen2.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/modeling_qwen2.py
 # Copyright 2024 The Qwen team.
@@ -22,53 +24,82 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
-from typing import 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 Qwen2Config
 
-from vllm.attention import Attention, AttentionMetadata
-from vllm.config import LoRAConfig
-from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
+from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
+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.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.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    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)
-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
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader,
+    maybe_remap_kv_scale_name,
+)
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.config import is_interleaved, set_default_rope_theta
+
+from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP
+from .utils import (
+    AutoWeightsLoader,
+    PPMissingLayer,
+    extract_layer_index,
+    is_pp_missing_parameter,
+    make_empty_intermediate_tensors_factory,
+    make_layers,
+    maybe_prefix,
+)
 
 
 class Qwen2MLP(nn.Module):
-
     def __init__(
         self,
         hidden_size: int,
         intermediate_size: int,
         hidden_act: str,
-        quant_config: Optional[QuantizationConfig] = None,
+        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)
+            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.")
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
+            )
         self.act_fn = SiluAndMul()
 
     def forward(self, x):
@@ -79,16 +110,21 @@ class Qwen2MLP(nn.Module):
 
 
 class Qwen2Attention(nn.Module):
-
-    def __init__(self,
-                 hidden_size: int,
-                 num_heads: int,
-                 num_kv_heads: int,
-                 max_position: int = 4096 * 32,
-                 rope_theta: float = 10000,
-                 use_sliding_window: bool = False,
-                 quant_config: Optional[QuantizationConfig] = None,
-                 sliding_window: Optional[int] = None) -> None:
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_parameters: dict[str, Any],
+        max_position: int = 4096 * 32,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
+        attn_type: str = AttentionType.DECODER,
+        dual_chunk_attention_config: dict[str, Any] | None = None,
+        qk_norm: bool = False,
+        rms_norm_eps: float = 1e-6,
+    ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
         tp_size = get_tensor_model_parallel_world_size()
@@ -109,8 +145,8 @@ class Qwen2Attention(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.sliding_window = sliding_window if use_sliding_window else None
+        self.dual_chunk_attention_config = dual_chunk_attention_config
+        self.qk_norm = qk_norm
 
         self.qkv_proj = QKVParallelLinear(
             hidden_size,
@@ -119,149 +155,371 @@ class Qwen2Attention(nn.Module):
             self.total_num_kv_heads,
             bias=True,
             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",
         )
 
+        # QK Normalization support (used in BAGEL and some other models)
+        if self.qk_norm:
+            self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+            self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+
         self.rotary_emb = get_rope(
             self.head_dim,
-            rotary_dim=self.head_dim,
             max_position=max_position,
-            base=self.rope_theta,
+            rope_parameters=rope_parameters,
+            dual_chunk_attention_config=dual_chunk_attention_config,
+        )
+        attn_cls = (
+            EncoderOnlyAttention
+            if attn_type == AttentionType.ENCODER_ONLY
+            else Attention
+        )
+        self.attn = attn_cls(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            attn_type=attn_type,
+            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,
-                              sliding_window=self.sliding_window)
 
     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)
+
+        # Apply QK normalization if enabled (before RoPE)
+        if self.qk_norm:
+            # Reshape to apply per-head normalization
+            # q shape: (total_tokens, q_size) -> (total_tokens, num_heads, head_dim)
+            total_tokens = q.shape[0]
+            q = q.view(total_tokens, self.num_heads, self.head_dim)
+            k = k.view(total_tokens, self.num_kv_heads, self.head_dim)
+
+            # Apply normalization
+            q = self.q_norm(q)
+            k = self.k_norm(k)
+
+            # Reshape back
+            q = q.view(total_tokens, self.q_size)
+            k = k.view(total_tokens, self.kv_size)
+
         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 Qwen2DecoderLayer(nn.Module):
-
     def __init__(
         self,
         config: Qwen2Config,
-        layer_idx: int,
-        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        # Requires transformers > 4.32.0
-        rope_theta = getattr(config, "rope_theta", 1000000)
-        use_sliding_window = (config.use_sliding_window
-                              and layer_idx < config.max_window_layers)
+        set_default_rope_theta(config, default_theta=1000000)
+        dual_chunk_attention_config = getattr(
+            config, "dual_chunk_attention_config", None
+        )
+
+        # By default, Qwen2 uses causal attention as it is a decoder-only model.
+        # You can override the HF config with `is_causal=False` to enable
+        # bidirectional attention, which is used in some embedding models
+        # (e.g. Alibaba-NLP/gte-Qwen2-7B-instruct)
+        if getattr(config, "is_causal", True):
+            attn_type = AttentionType.DECODER
+        else:
+            attn_type = AttentionType.ENCODER_ONLY
+
+        # Check if QK normalization is enabled (used in BAGEL and some other models)
+        qk_norm = getattr(config, "qk_norm", False)
+
         self.self_attn = Qwen2Attention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
             max_position=config.max_position_embeddings,
             num_kv_heads=config.num_key_value_heads,
-            rope_theta=rope_theta,
-            use_sliding_window=use_sliding_window,
+            cache_config=cache_config,
             quant_config=quant_config,
-            sliding_window=config.sliding_window)
+            rope_parameters=config.rope_parameters,
+            prefix=f"{prefix}.self_attn",
+            attn_type=attn_type,
+            dual_chunk_attention_config=dual_chunk_attention_config,
+            qk_norm=qk_norm,
+            rms_norm_eps=config.rms_norm_eps,
+        )
         self.mlp = Qwen2MLP(
             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
         )
-        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],
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        residual: torch.Tensor | None,
+    ) -> 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, 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
 
 
-class Qwen2Model(nn.Module):
+def qwen_2_model_invariants(
+    input_ids: torch.Tensor,
+    positions: torch.Tensor,
+    intermediate_tensors: IntermediateTensors | None = None,
+    inputs_embeds: torch.Tensor | None = None,
+):
+    """Shape invariants for Qwen2Model Model, those are translated to
+    runtime assertions for unbacked dynamic shapes and are compiled away for
+    backed"""
+    # All these should be equal.
+    # input_ids.size()[0]
+    # positions.size()[-1]
+    # intermediate_tensors["hidden_states"].size()[0]
+    # inputs_embeds.size()[0]
+    torch._check(input_ids.size()[0] == positions.size()[-1])
+    if intermediate_tensors is not None:
+        torch._check(
+            input_ids.size()[0] == intermediate_tensors["hidden_states"].size()[0]
+        )
 
+    if inputs_embeds is not None:
+        torch._check(input_ids.size()[0] == inputs_embeds.size()[0])
+
+    # Hidden dimensions should match (hidden_size)
+    # intermediate_tensors["hidden_states"].size()[1]
+    # inputs_embeds.size()[1]
+    if inputs_embeds is not None and intermediate_tensors is not None:
+        torch._check(
+            inputs_embeds.size()[1] == intermediate_tensors["hidden_states"].size()[1]
+        )
+
+
+@support_torch_compile(
+    dynamic_arg_dims={
+        "input_ids": 0,
+        # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl,
+        # otherwise (seq_len, ).
+        "positions": -1,
+        "intermediate_tensors": 0,
+        "inputs_embeds": 0,
+    },
+    shape_invariants=qwen_2_model_invariants,
+)
+class Qwen2Model(nn.Module):
     def __init__(
         self,
-        config: Qwen2Config,
-        quant_config: Optional[QuantizationConfig] = None,
-    ) -> None:
+        *,
+        vllm_config: VllmConfig,
+        prefix: str = "",
+        decoder_layer_type: type[nn.Module] = Qwen2DecoderLayer,
+    ):
         super().__init__()
+
+        config = vllm_config.model_config.hf_config.get_text_config()
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        # TODO (@robertgshaw2): see if this can be moved out
+        if is_interleaved(vllm_config.model_config.hf_text_config):
+            assert config.max_window_layers == config.num_hidden_layers, (
+                "Sliding window for some but all layers is not supported. "
+                "This model uses sliding window but `max_window_layers` = {} "
+                "is less than `num_hidden_layers` = {}. Please open an issue "
+                "to discuss this feature.".format(
+                    config.max_window_layers,
+                    config.num_hidden_layers,
+                )
+            )
+
         self.config = config
-        self.padding_idx = config.pad_token_id
+        self.quant_config = quant_config
         self.vocab_size = config.vocab_size
 
-        self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size,
-            config.hidden_size,
+        if get_pp_group().is_first_rank or (
+            config.tie_word_embeddings and get_pp_group().is_last_rank
+        ):
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=f"{prefix}.embed_tokens",
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        # Use the provided decoder layer type or default to Qwen2DecoderLayer
+        decoder_layer_type = decoder_layer_type or Qwen2DecoderLayer
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: decoder_layer_type(
+                config=config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=f"{prefix}.layers",
         )
-        self.layers = nn.ModuleList([
-            Qwen2DecoderLayer(config, layer_idx, 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
+        )
+        if get_pp_group().is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer()
+
+        self.aux_hidden_state_layers = tuple[int, ...]()
+
+    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"]
+
+        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)
+            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)
+
+        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),
+        ]
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: set[str] = set()
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+            if self.quant_config is not None and (
+                scale_name := self.quant_config.get_cache_scale(name)
+            ):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                loaded_weight = (
+                    loaded_weight if loaded_weight.dim() == 0 else loaded_weight[0]
+                )
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_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
+                if is_pp_missing_parameter(name, self):
+                    continue
+                if name.endswith("scale"):
+                    # Remapping the name of FP8 kv-scale.
+                    name = maybe_remap_kv_scale_name(name, params_dict)
+                    if name is None:
+                        continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                if weight_loader == default_weight_loader:
+                    weight_loader(param, loaded_weight)
+                else:
+                    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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+                if is_pp_missing_parameter(name, self):
+                    continue
+                if name not in params_dict:
+                    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 Qwen2ForCausalLM(nn.Module):
+
+class Qwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
@@ -274,94 +532,69 @@ class Qwen2ForCausalLM(nn.Module):
         ],
     }
 
-    # LoRA specific attributes
-    supported_lora_modules = [
-        "qkv_proj",
-        "o_proj",
-        "gate_up_proj",
-        "down_proj",
-    ]
-    embedding_modules = {}
-    embedding_padding_modules = []
-
-    def __init__(
-        self,
-        config: Qwen2Config,
-        quant_config: Optional[QuantizationConfig] = None,
-        lora_config: Optional[LoRAConfig] = None,
-    ) -> None:
-        del lora_config
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
-        self.config = config
-        self.quant_config = quant_config
-        self.model = Qwen2Model(config, quant_config)
+        config = vllm_config.model_config.hf_config.get_text_config()
+        quant_config = vllm_config.quant_config
 
-        if config.tie_word_embeddings:
-            self.lm_head_weight = self.model.embed_tokens.weight
+        self.config = config
+
+        self.quant_config = quant_config
+        self.model = Qwen2Model(
+            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
+        )
+
+        if get_pp_group().is_last_rank:
+            if config.tie_word_embeddings:
+                self.lm_head = self.model.embed_tokens
+            else:
+                self.lm_head = ParallelLMHead(
+                    config.vocab_size,
+                    config.hidden_size,
+                    quant_config=quant_config,
+                    prefix=maybe_prefix(prefix, "lm_head"),
+                )
         else:
-            self.lm_head = ParallelLMHead(config.vocab_size,
-                                          config.hidden_size)
-            self.lm_head_weight = self.lm_head.weight
+            self.lm_head = PPMissingLayer()
 
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = Sampler()
+
+        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 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)
+        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,
-                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
-        logits = self.logits_processor(self.lm_head_weight, hidden_states,
-                                       sampling_metadata)
+    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),
-        ]
-        params_dict = dict(self.named_parameters(remove_duplicate=False))
-        for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
-            if self.config.tie_word_embeddings and "lm_head.weight" 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
-                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
-                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)