sync v0.15.1 (models)

vllm/model_executor/models/afmoe.py

@@ -425,7 +425,7 @@ class AfmoeModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -675,7 +675,7 @@ class AfmoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -694,4 +694,4 @@ class AfmoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
         return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
-        return self.model.get_expert_mapping()
+        return self.model.get_expert_mapping()
\ No newline at end of file

vllm/model_executor/models/apertus.py

@@ -542,7 +542,7 @@ class ApertusForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -564,4 +564,4 @@ class ApertusForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/arcee.py

@@ -394,7 +394,7 @@ class ArceeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -425,4 +425,4 @@ class ArceeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         )
         # AutoWeightLoader handles weight name remapping, including fusing
         # separate q_proj, k_proj, v_proj into qkv_proj
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/arctic.py

@@ -406,7 +406,7 @@ class ArcticModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -460,7 +460,7 @@ class ArcticForCausalLM(nn.Module, SupportsPP, SupportsQuant):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -591,4 +591,4 @@ class ArcticForCausalLM(nn.Module, SupportsPP, SupportsQuant):
                         )
                         weight_loader(param, loaded_weight)
             loaded_params.add(name)
-        return loaded_params
+        return loaded_params
\ No newline at end of file

vllm/model_executor/models/aria.py

@@ -629,7 +629,7 @@ class AriaForConditionalGeneration(nn.Module, SupportsMultiModal):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -656,4 +656,4 @@ class AriaForConditionalGeneration(nn.Module, SupportsMultiModal):
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
         loader = AutoWeightsLoader(self)
-        loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+        loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
\ No newline at end of file

vllm/model_executor/models/audioflamingo3.py

@@ -609,7 +609,7 @@ class AudioFlamingo3ForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -634,4 +634,4 @@ class AudioFlamingo3ForConditionalGeneration(
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/aya_vision.py

@@ -420,7 +420,7 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsP
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -441,4 +441,4 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsP
         self,
         hidden_states: torch.Tensor,
     ) -> torch.Tensor | None:
-        return self.language_model.compute_logits(hidden_states)
+        return self.language_model.compute_logits(hidden_states)
\ No newline at end of file

vllm/model_executor/models/bagel.py

@@ -507,7 +507,7 @@ class BagelForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -581,4 +581,4 @@ class BagelForConditionalGeneration(
 
         # Skip vit_pos_embed.pos_embed as it's handled by PositionEmbedding module
         loader = AutoWeightsLoader(self, skip_prefixes=["vit_pos_embed.pos_embed"])
-        return loader.load_weights(filtered_weights, mapper=self.hf_to_vllm_mapper)
+        return loader.load_weights(filtered_weights, mapper=self.hf_to_vllm_mapper)
\ No newline at end of file

vllm/model_executor/models/baichuan.py

@@ -334,7 +334,7 @@ class BaiChuanModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -534,7 +534,7 @@ class BaiChuanBaseForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/bailing_moe.py

@@ -440,7 +440,7 @@ class BailingMoeModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         position_ids: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -611,7 +611,7 @@ class BailingMoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -640,4 +640,4 @@ class BailingMoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
 
 
 class BailingMoeV2ForCausalLM(BailingMoeForCausalLM):
-    pass
+    pass
\ No newline at end of file

vllm/model_executor/models/bamba.py

@@ -311,7 +311,7 @@ class BambaModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -493,7 +493,7 @@ class BambaForCausalLM(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -514,4 +514,4 @@ class BambaForCausalLM(
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/bert_with_rope.py

@@ -475,7 +475,7 @@ class BertWithRope(nn.Module, SupportsQuant):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -726,4 +726,4 @@ class GteNewForSequenceClassification(nn.Module, SupportsCrossEncoding):
             positions=positions,
             inputs_embeds=inputs_embeds,
             intermediate_tensors=intermediate_tensors,
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/blip2.py

@@ -641,7 +641,7 @@ class Blip2ForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -727,4 +727,4 @@ class Blip2ForConditionalGeneration(
             "the number of tokens per image."
         )
         num_images = num_vision_tokens / self._vision_tokens_per_image
-        return num_images * self.config.num_query_tokens
+        return num_images * self.config.num_query_tokens
\ No newline at end of file

vllm/model_executor/models/bloom.py

@@ -294,7 +294,7 @@ class BloomModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         position_ids: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -412,7 +412,7 @@ class BloomForCausalLM(nn.Module, SupportsPP, SupportsQuant):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/chameleon.py

@@ -994,7 +994,7 @@ class ChameleonForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -1100,4 +1100,4 @@ class ChameleonForConditionalGeneration(
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
             loaded_params.add(name)
-        return loaded_params
+        return loaded_params
\ No newline at end of file

vllm/model_executor/models/chatglm.py

@@ -381,7 +381,7 @@ class ChatGLMModel(nn.Module, SupportsQuant):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -554,7 +554,7 @@ class ChatGLMForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP, SupportsQua
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/cohere2_vision.py

@@ -446,7 +446,7 @@ class Cohere2VisionForConditionalGeneration(nn.Module, SupportsMultiModal, Suppo
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -467,4 +467,4 @@ class Cohere2VisionForConditionalGeneration(nn.Module, SupportsMultiModal, Suppo
         self,
         hidden_states: torch.Tensor,
     ) -> torch.Tensor | None:
-        return self.language_model.compute_logits(hidden_states)
+        return self.language_model.compute_logits(hidden_states)
\ No newline at end of file

vllm/model_executor/models/commandr.py

@@ -312,7 +312,7 @@ class CohereModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -438,7 +438,7 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
     @torch.no_grad()
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -466,4 +466,4 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
         loader = AutoWeightsLoader(
             self, skip_prefixes=["lm_head", "rotary_emb.inv_freq"]
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/dbrx.py

@@ -361,7 +361,7 @@ class DbrxModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         position_ids: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -462,7 +462,7 @@ class DbrxForCausalLM(nn.Module, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -481,4 +481,4 @@ class DbrxForCausalLM(nn.Module, SupportsPP):
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/deepseek.py

-# 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.
-# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
-#
-# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
-# and OPT implementations in this library. It has been modified from its
-# original forms to accommodate minor architectural differences compared
-# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Inference-only Deepseek model."""
-from collections.abc import Iterable
-from itertools import islice
-from typing import Any, Optional, Union
-
-import torch
-from torch import nn
-from transformers import PretrainedConfig
-
-from vllm.attention import Attention
-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 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.logits_processor import LogitsProcessor
-from vllm.model_executor.layers.quantization import QuantizationConfig
-from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.vocab_parallel_embedding import (
-    ParallelLMHead, VocabParallelEmbedding)
-from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from 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)
-import vllm.envs as envs
-
-
-class DeepseekMLP(nn.Module):
-
-    def __init__(
-        self,
-        hidden_size: int,
-        intermediate_size: int,
-        hidden_act: str,
-        quant_config: Optional[QuantizationConfig] = None,
-        reduce_results: bool = True,
-        prefix: str = "",
-    ) -> None:
-        super().__init__()
-        self.gate_up_proj = MergedColumnParallelLinear(
-            hidden_size, [intermediate_size] * 2,
-            bias=False,
-            quant_config=quant_config)
-        self.down_proj = RowParallelLinear(intermediate_size,
-                                           hidden_size,
-                                           bias=False,
-                                           quant_config=quant_config,
-                                           reduce_results=reduce_results)
-        if hidden_act != "silu":
-            raise ValueError(f"Unsupported activation: {hidden_act}. "
-                             "Only silu is supported for now.")
-        self.act_fn = SiluAndMul()
-
-    def forward(self, x):
-        gate_up, _ = self.gate_up_proj(x)
-        x = self.act_fn(gate_up)
-        x, _ = self.down_proj(x)
-        return x
-
-
-class DeepseekMoE(nn.Module):
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        quant_config: Optional[QuantizationConfig] = 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.n_routed_experts
-        self.top_k = config.num_experts_per_tok
-        if self.tp_size > self.n_routed_experts:
-            raise ValueError(
-                f"Tensor parallel size {self.tp_size} is greater than "
-                f"the number of experts {self.n_routed_experts}.")
-
-        self.experts = nn.ModuleList([
-            DeepseekMLP(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,
-                                     self.n_routed_experts,
-                                     bias=False,
-                                     quant_config=None)
-
-        if config.n_shared_experts is not None:
-            intermediate_size = (config.moe_intermediate_size *
-                                 config.n_shared_experts)
-            self.shared_experts = DeepseekMLP(
-                hidden_size=config.hidden_size,
-                intermediate_size=intermediate_size,
-                hidden_act=config.hidden_act,
-                quant_config=quant_config,
-                reduce_results=False,
-            )
-
-    def pack_params(self):
-        w1 = []
-        w2 = []
-        for expert in self.experts:
-            w1.append(expert.gate_up_proj.weight)
-            w2.append(expert.down_proj.weight)
-        self.w1 = torch._utils._flatten_dense_tensors(w1)
-        w1s = torch._utils._unflatten_dense_tensors(self.w1, w1)
-        for data, param in zip(w1s, w1):
-            param.data = data
-        self.w1 = self.w1.view(len(w1), *w1s[0].shape)
-
-        self.w2 = torch._utils._flatten_dense_tensors(w2)
-        w2s = torch._utils._unflatten_dense_tensors(self.w2, w2)
-        for data, param in zip(w2s, w2):
-            param.data = data
-
-        self.w2 = self.w2.view(len(w2), *w2s[0].shape)
-        
-        if envs.VLLM_USE_NN:
-            self.w1 = self.w1.permute(0,2,1).contiguous()   
-            for expert, w in zip(self.experts, self.w1):
-                expert.gate_up_proj.weight.data = w.permute(1,0)
-                
-            self.w2 = self.w2.permute(0, 2, 1).contiguous()
-            for expert, w in zip(self.experts, self.w2):
-                expert.down_proj.weight.data = w.permute(1, 0)
-
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        num_tokens, hidden_dim = hidden_states.shape
-        hidden_states = hidden_states.view(-1, hidden_dim)
-        if self.config.n_shared_experts is not None:
-            shared_output = self.shared_experts(hidden_states)
-        # router_logits: (num_tokens, n_experts)
-        router_logits, _ = self.gate(hidden_states)
-
-        topk_weights, topk_ids, _ = fused_topk(
-            hidden_states,
-            router_logits,
-            self.top_k,
-            renormalize=self.config.norm_topk_prob)
-
-        final_hidden_states = fused_experts(hidden_states,
-                                            self.w1,
-                                            self.w2,
-                                            topk_weights,
-                                            topk_ids,
-                                            inplace=True)
-
-        if self.config.n_shared_experts is not None:
-            final_hidden_states = final_hidden_states + shared_output
-        final_hidden_states = tensor_model_parallel_all_reduce(
-            final_hidden_states)
-
-        return final_hidden_states.view(num_tokens, hidden_dim)
-
-
-class DeepseekAttention(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,
-        max_position_embeddings: int = 8192,
-        cache_config: Optional[CacheConfig] = None,
-        quant_config: Optional[QuantizationConfig] = None,
-        prefix: str = "",
-    ) -> None:
-        super().__init__()
-        self.hidden_size = hidden_size
-        tp_size = get_tensor_model_parallel_world_size()
-        self.total_num_heads = num_heads
-        assert self.total_num_heads % tp_size == 0
-        self.num_heads = self.total_num_heads // tp_size
-        self.total_num_kv_heads = num_kv_heads
-        if self.total_num_kv_heads >= tp_size:
-            # Number of KV heads is greater than TP size, so we partition
-            # the KV heads across multiple tensor parallel GPUs.
-            assert self.total_num_kv_heads % tp_size == 0
-        else:
-            # Number of KV heads is less than TP size, so we replicate
-            # the KV heads across multiple tensor parallel GPUs.
-            assert tp_size % self.total_num_kv_heads == 0
-        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
-        self.head_dim = hidden_size // self.total_num_heads
-        self.q_size = self.num_heads * self.head_dim
-        self.kv_size = self.num_kv_heads * self.head_dim
-        self.scaling = self.head_dim**-0.5
-        self.rope_theta = rope_theta
-        self.max_position_embeddings = max_position_embeddings
-
-        self.qkv_proj = QKVParallelLinear(
-            hidden_size,
-            self.head_dim,
-            self.total_num_heads,
-            self.total_num_kv_heads,
-            bias=False,
-            quant_config=quant_config,
-        )
-
-        self.o_proj = RowParallelLinear(
-            self.total_num_heads * self.head_dim,
-            hidden_size,
-            bias=False,
-            quant_config=quant_config,
-        )
-
-        self.rotary_emb = get_rope(
-            self.head_dim,
-            rotary_dim=self.head_dim,
-            max_position=max_position_embeddings,
-            base=rope_theta,
-            rope_scaling=rope_scaling,
-        )
-        self.attn = 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")
-
-    def forward(
-        self,
-        positions: torch.Tensor,
-        hidden_states: torch.Tensor,
-    ) -> 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)
-        output, _ = self.o_proj(attn_output)
-        return output
-
-
-class DeepseekDecoderLayer(nn.Module):
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        cache_config: Optional[CacheConfig] = None,
-        quant_config: Optional[QuantizationConfig] = None,
-        prefix: str = "",
-    ) -> None:
-        super().__init__()
-        layer_idx = extract_layer_index(prefix)
-        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.self_attn = DeepseekAttention(
-            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,
-            cache_config=cache_config,
-            quant_config=quant_config,
-            prefix=f"{prefix}.self_attn",
-        )
-        if (config.n_routed_experts is not None
-                and layer_idx >= config.first_k_dense_replace
-                and layer_idx % config.moe_layer_freq == 0):
-            self.mlp = DeepseekMoE(config=config,
-                                   quant_config=quant_config,
-                                   prefix=f"{prefix}.mlp")
-        else:
-            self.mlp = DeepseekMLP(
-                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)
-
-    def forward(
-        self,
-        positions: torch.Tensor,
-        hidden_states: torch.Tensor,
-        residual: Optional[torch.Tensor],
-    ) -> torch.Tensor:
-        # Self Attention
-        if residual is None:
-            residual = hidden_states
-            hidden_states = self.input_layernorm(hidden_states)
-        else:
-            hidden_states, residual = self.input_layernorm(
-                hidden_states, residual)
-        hidden_states = self.self_attn(
-            positions=positions,
-            hidden_states=hidden_states,
-        )
-
-        # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(
-            hidden_states, residual)
-        hidden_states = self.mlp(hidden_states)
-        return hidden_states, residual
-
-
-class DeepseekModel(nn.Module):
-
-    fall_back_to_pt_during_load = False
-
-    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.vocab_size = config.vocab_size
-
-        self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size,
-            config.hidden_size,
-        )
-        self.start_layer, self.end_layer, self.layers = make_layers(
-            config.num_hidden_layers,
-            lambda prefix: DeepseekDecoderLayer(
-                config, cache_config, quant_config=quant_config, prefix=prefix
-            ),
-            prefix=f"{prefix}.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 get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
-        return self.embed_tokens(input_ids)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        positions: torch.Tensor,
-        intermediate_tensors: Optional[IntermediateTensors],
-        inputs_embeds: Optional[torch.Tensor] = None,
-    ) -> Union[torch.Tensor, IntermediateTensors]:
-        if get_pp_group().is_first_rank:
-            if inputs_embeds is not None:
-                hidden_states = inputs_embeds
-            else:
-                hidden_states = self.get_input_embeddings(input_ids)
-            residual = None
-        else:
-            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 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())
-        loaded_params: set[str] = set()
-        for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
-            for (param_name, weight_name, shard_id) in stacked_params_mapping:
-                if weight_name not in name:
-                    continue
-                name = name.replace(weight_name, param_name)
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                # Skip experts that are not assigned to this worker.
-                if (("mlp.experts." in name or "mlp.shared_experts." in name)
-                        and name not in params_dict):
-                    continue
-                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:
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                # Skip experts that are not assigned to this worker.
-                if (("mlp.experts." in name or "mlp.shared_experts." in name)
-                        and name not in params_dict):
-                    continue
-                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 DeepseekForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
-    packed_modules_mapping = {
-        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
-        "gate_up_proj": ["gate_proj", "up_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
-        self.model = DeepseekModel(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 get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
-        return self.model.get_input_embeddings(input_ids)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        positions: torch.Tensor,
-        intermediate_tensors: Optional[IntermediateTensors] = None,
-        inputs_embeds: Optional[torch.Tensor] = None,
-    ) -> Union[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,
-    ) -> Optional[torch.Tensor]:
-        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)