sglang 0.5.0rc2__py3-none-any.whl → 0.5.1__py3-none-any.whl

sglang/srt/models/interns1.py

@@ -4,8 +4,9 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from sglang.srt.distributed import parallel_state
+from sglang.srt.layers.attention import vision_utils
 from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
+from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.managers.mm_utils import (
     MultiModalityDataPaddingPatternTokenPairs,
@@ -20,6 +21,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.internvl import InternVisionModel
 from sglang.srt.models.qwen2 import Qwen2ForCausalLM
+from sglang.srt.models.qwen3 import Qwen3ForCausalLM
 from sglang.srt.models.qwen3_moe import Qwen3MoeForCausalLM
 from sglang.utils import logger
 
@@ -34,7 +36,7 @@ class InternS1ForConditionalGeneration(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self._update_hf_config()
+        vision_utils.update_vit_attn_dummy_heads_config(self.config)
         image_size = (
             getattr(config, "force_image_size", None) or config.vision_config.image_size
         )
@@ -69,6 +71,10 @@ class InternS1ForConditionalGeneration(nn.Module):
             self.language_model = Qwen3MoeForCausalLM(
                 config=config.text_config, quant_config=quant_config
             )
+        elif config.text_config.architectures[0] == "Qwen3ForCausalLM":
+            self.language_model = Qwen3ForCausalLM(
+                config=config.text_config, quant_config=quant_config
+            )
         else:
             raise NotImplementedError(
                 f"{config.text_config.architectures[0]} is not implemented."
@@ -86,21 +92,6 @@ class InternS1ForConditionalGeneration(nn.Module):
             nn.Linear(llm_hidden_size, llm_hidden_size),
         )
 
-    def _update_hf_config(self):
-        """update hf config to support tp"""
-        world_size = parallel_state.get_tensor_model_parallel_world_size()
-        num_heads = self.config.vision_config.num_attention_heads
-        head_dim = self.config.vision_config.hidden_size // num_heads
-        num_dummy_heads = 0
-
-        if num_heads % world_size != 0:
-            num_dummy_heads = (
-                (num_heads + world_size) // world_size
-            ) * world_size - num_heads
-
-        setattr(self.config.vision_config, "head_dim", head_dim)
-        setattr(self.config.vision_config, "num_dummy_heads", num_dummy_heads)
-
     def pixel_shuffle(self, x, scale_factor=0.5):
         n, w, h, c = x.size()
         # N, W, H, C --> N, W, H * scale, C // scale
@@ -183,34 +174,6 @@ class InternS1ForConditionalGeneration(nn.Module):
 
         return helper.pad_input_tokens(input_ids, mm_inputs)
 
-    def _pad_vit_attn_dummy_heads(self, name: str, loaded_weight: torch.Tensor):
-        """pad attn qkv weights for dummy heads"""
-        num_dummy_heads = self.config.vision_config.num_dummy_heads
-        if num_dummy_heads == 0:
-            return loaded_weight
-        head_dim = self.config.vision_config.head_dim
-
-        if any([_ in name for _ in ["attn.q_proj", "attn.k_proj", "attn.v_proj"]]):
-            if name.endswith(".weight"):
-                dummy_shape = [num_dummy_heads, head_dim, loaded_weight.shape[-1]]
-            elif name.endswith(".bias"):
-                dummy_shape = [num_dummy_heads, head_dim]
-            else:
-                raise RuntimeError(f"Unsupported weight with name={name}")
-            padded_weight = loaded_weight.new_zeros(dummy_shape)
-            loaded_weight = torch.cat(
-                [loaded_weight.unflatten(0, (-1, head_dim)), padded_weight], dim=0
-            ).flatten(0, 1)
-        if "attn.proj.weight" in name:
-            padded_weight = loaded_weight.new_zeros(
-                loaded_weight.shape[0], head_dim * num_dummy_heads
-            )
-            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=-1)
-        if "attn.q_norm.weight" in name or "attn.k_norm.weight" in name:
-            padded_weight = loaded_weight.new_zeros(head_dim * num_dummy_heads)
-            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=0)
-        return loaded_weight
-
     def _mapping_interns1_name(self, name):
         names_map = {
             "lm_head.weight": "language_model.lm_head.weight",
@@ -254,7 +217,7 @@ class InternS1ForConditionalGeneration(nn.Module):
         ]
         expert_params_mapping = []
         if "Qwen3MoeForCausalLM" in self.config.text_config.architectures:
-            expert_params_mapping = get_moe_impl_class().make_expert_params_mapping(
+            expert_params_mapping = FusedMoE.make_expert_params_mapping(
                 ckpt_gate_proj_name="gate_proj",
                 ckpt_down_proj_name="down_proj",
                 ckpt_up_proj_name="up_proj",
@@ -269,7 +232,9 @@ class InternS1ForConditionalGeneration(nn.Module):
                 continue
             name = self._mapping_interns1_name(name)
             if "vision_model" in name:
-                loaded_weight = self._pad_vit_attn_dummy_heads(name, loaded_weight)
+                loaded_weight = vision_utils.pad_vit_attn_dummy_heads(
+                    self.config, name, loaded_weight
+                )
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:

sglang/srt/models/internvl.py

@@ -10,9 +10,9 @@ from transformers import PretrainedConfig, PreTrainedModel
 from transformers.activations import ACT2FN
 from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
 
-from sglang.srt.distributed import parallel_state
+from sglang.srt.layers.attention import vision_utils
 from sglang.srt.layers.attention.vision import SingletonCache, VisionAttention
-from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
+from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.managers.mm_utils import (
     MultiModalityDataPaddingPatternTokenPairs,
@@ -412,7 +412,7 @@ class InternVLChatModel(nn.Module):
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self._update_vision_config()
+        vision_utils.update_vit_attn_dummy_heads_config(self.config)
         image_size = config.force_image_size or config.vision_config.image_size
         patch_size = config.vision_config.patch_size
         self.patch_size = patch_size
@@ -462,21 +462,6 @@ class InternVLChatModel(nn.Module):
             nn.Linear(llm_hidden_size, llm_hidden_size),
         )
 
-    def _update_vision_config(self):
-        """update vision config to support tp"""
-        world_size = parallel_state.get_tensor_model_parallel_world_size()
-        num_heads = self.config.vision_config.num_attention_heads
-        head_dim = self.config.vision_config.hidden_size // num_heads
-        num_dummy_heads = 0
-
-        if num_heads % world_size != 0:
-            num_dummy_heads = (
-                (num_heads + world_size) // world_size
-            ) * world_size - num_heads
-
-        setattr(self.config.vision_config, "head_dim", head_dim)
-        setattr(self.config.vision_config, "num_dummy_heads", num_dummy_heads)
-
     def pixel_shuffle(self, x, scale_factor=0.5):
         n, w, h, c = x.size()
         # N, W, H, C --> N, W, H * scale, C // scale
@@ -559,36 +544,6 @@ class InternVLChatModel(nn.Module):
 
         return helper.pad_input_tokens(input_ids, mm_inputs)
 
-    def _pad_vit_attn_dummy_heads(self, name: str, loaded_weight: torch.Tensor):
-        """pad attn qkv weights for dummy heads"""
-        num_dummy_heads = self.config.vision_config.num_dummy_heads
-        if num_dummy_heads == 0:
-            return loaded_weight
-        head_dim = self.config.vision_config.head_dim
-
-        if "attn.qkv_proj" in name:
-            wq, wk, wv = loaded_weight.chunk(3, dim=0)
-            if name.endswith(".weight"):
-                dummy_shape = [num_dummy_heads, head_dim, wq.shape[-1]]
-            elif name.endswith(".bias"):
-                dummy_shape = [num_dummy_heads, head_dim]
-            else:
-                raise RuntimeError(f"Unsupported weight with name={name}")
-            pad_func = lambda x: torch.cat(
-                [x.unflatten(0, (-1, head_dim)), x.new_zeros(dummy_shape)], dim=0
-            ).flatten(0, 1)
-            wq, wk, wv = pad_func(wq), pad_func(wk), pad_func(wv)
-            loaded_weight = torch.cat([wq, wk, wv], dim=0)
-        if "attn.proj.weight" in name:
-            padded_weight = loaded_weight.new_zeros(
-                loaded_weight.shape[0], head_dim * num_dummy_heads
-            )
-            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=-1)
-        if "attn.q_norm.weight" in name or "attn.k_norm.weight" in name:
-            padded_weight = loaded_weight.new_zeros(head_dim * num_dummy_heads)
-            loaded_weight = torch.cat([loaded_weight, padded_weight], dim=0)
-        return loaded_weight
-
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         expert_params_mapping = []
         if "InternLM2ForCausalLM" in self.config.llm_config.architectures:
@@ -616,7 +571,7 @@ class InternVLChatModel(nn.Module):
                 ("gate_up_proj", "up_proj", 1),
             ]
 
-            expert_params_mapping = get_moe_impl_class().make_expert_params_mapping(
+            expert_params_mapping = FusedMoE.make_expert_params_mapping(
                 ckpt_gate_proj_name="gate_proj",
                 ckpt_down_proj_name="down_proj",
                 ckpt_up_proj_name="up_proj",
@@ -699,8 +654,8 @@ class InternVLChatModel(nn.Module):
                             param, "weight_loader", default_weight_loader
                         )
                         if "vision_model" in name:
-                            loaded_weight = self._pad_vit_attn_dummy_heads(
-                                name, loaded_weight
+                            loaded_weight = vision_utils.pad_vit_attn_dummy_heads(
+                                self.config, name, loaded_weight
                             )
                         weight_loader(param, loaded_weight)
 

sglang/srt/models/llama4.py

@@ -31,7 +31,6 @@ from sglang.srt.layers.communicator import LayerCommunicator, LayerScatterModes
 from sglang.srt.layers.dp_attention import (
     get_attention_tp_rank,
     get_attention_tp_size,
-    get_local_attention_dp_size,
     is_dp_attention_enabled,
 )
 from sglang.srt.layers.layernorm import RMSNorm
@@ -364,7 +363,6 @@ class Llama4DecoderLayer(nn.Module):
         rope_theta = config.rope_theta
         rope_scaling = config.rope_scaling
         max_position_embeddings = config.max_position_embeddings
-        self.local_dp_size = get_local_attention_dp_size()
         self.attn_tp_size = get_attention_tp_size()
         self.attn_tp_rank = get_attention_tp_rank()
 

sglang/srt/models/minicpm3.py

@@ -37,7 +37,6 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, is_cuda

sglang/srt/models/mixtral.py

@@ -47,7 +47,6 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, make_layers
@@ -104,7 +103,6 @@ class MixtralMoE(nn.Module):
             intermediate_size=intermediate_size,
             params_dtype=params_dtype,
             quant_config=quant_config,
-            tp_size=tp_size,
             prefix=add_prefix("experts", prefix),
         )
 

sglang/srt/models/nemotron_nas.py

@@ -0,0 +1,435 @@
+# Copyright 2023-2025 SGLang Team
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/nemotron_nas.py
+
+"""Inference-only deci model compatible with HuggingFace weights."""
+from typing import Iterable, Optional, Tuple, Type, Union
+
+import torch
+from torch import nn
+from transformers import LlamaConfig
+
+from sglang.srt.distributed import get_pp_group
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.pooler import Pooler, PoolingType
+from sglang.srt.layers.quantization import QuantizationConfig
+from sglang.srt.layers.utils import PPMissingLayer
+from sglang.srt.layers.vocab_parallel_embedding import (
+    DEFAULT_VOCAB_PADDING_SIZE,
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
+from sglang.srt.model_loader.weight_utils import (
+    default_weight_loader,
+    maybe_remap_kv_scale_name,
+)
+from sglang.srt.models.llama import LlamaAttention, LlamaMLP
+from sglang.srt.utils import add_prefix, make_layers
+from sglang.utils import logger
+
+
+def _ffn_mult_to_intermediate_size(ffn_mult: float, n_embd: int) -> int:
+    # DeciLM-specific code
+    intermediate_size = int(2 * ffn_mult * n_embd / 3)
+    return _find_multiple(intermediate_size, 256)
+
+
+def _find_multiple(n: int, k: int) -> int:
+    # DeciLM-specific code
+    if n % k == 0:
+        return n
+    return n + k - (n % k)
+
+
+class DeciLMDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+        layer_idx: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        block_config = config.block_configs[layer_idx]
+        self._is_no_op_attention = block_config.attention.no_op
+        self._is_no_op_ffn = block_config.ffn.no_op
+
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+            config, "original_max_position_embeddings", None
+        ):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings
+            )
+        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
+        # Support abacusai/Smaug-72B-v0.1 with attention_bias
+        # Support internlm/internlm-7b with bias
+        rope_is_neox_style = getattr(config, "rope_is_neox_style", True)
+        attention_bias = getattr(config, "attention_bias", False) or getattr(
+            config, "bias", False
+        )
+        # support internlm/internlm3-8b with qkv_bias
+        if hasattr(config, "qkv_bias"):
+            attention_bias = config.qkv_bias
+
+        if not self._is_no_op_attention:
+            num_kv_heads = (
+                config.num_attention_heads // block_config.attention.n_heads_in_group
+            )
+            self.self_attn = LlamaAttention(
+                config=config,
+                hidden_size=self.hidden_size,
+                num_heads=config.num_attention_heads,
+                num_kv_heads=num_kv_heads,
+                layer_id=layer_idx,
+                rope_theta=rope_theta,
+                rope_scaling=rope_scaling,
+                rope_is_neox_style=rope_is_neox_style,
+                max_position_embeddings=max_position_embeddings,
+                quant_config=quant_config,
+                prefix=add_prefix("self_attn", prefix),
+                bias=attention_bias,
+            )
+            self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        if not self._is_no_op_ffn:
+            ffn_mult = block_config.ffn.ffn_mult
+            intermediate_size = _ffn_mult_to_intermediate_size(
+                ffn_mult, config.hidden_size
+            )
+            self.mlp = LlamaMLP(
+                hidden_size=self.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                prefix=add_prefix("mlp", prefix),
+            )
+            self.post_attention_layernorm = RMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+
+        if self._is_no_op_attention:
+            pass
+        else:
+            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,
+                forward_batch=forward_batch,
+            )
+
+        # Fully Connected
+        if not self._is_no_op_ffn:
+            hidden_states, residual = self.post_attention_layernorm(
+                hidden_states, residual
+            )
+            hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class DeciModel(nn.Module):
+    def __init__(
+        self,
+        *,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        layer_type: Type[DeciLMDecoderLayer] = DeciLMDecoderLayer,
+    ):
+        super().__init__()
+
+        lora_config = None
+        self.config = config
+        self.quant_config = quant_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
+        )
+        vocab_size = config.vocab_size + lora_vocab
+        if get_pp_group().is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                quant_config=quant_config,
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        def get_layer(idx: int, prefix: str):
+            return layer_type(
+                config,
+                layer_idx=idx,
+                quant_config=quant_config,
+                prefix=prefix,
+            )
+
+        self.layers, self.start_layer, self.end_layer = make_layers(
+            config.num_hidden_layers,
+            get_layer,
+            pp_rank=get_pp_group().rank_in_group,
+            pp_size=get_pp_group().world_size,
+            prefix=add_prefix("layers", prefix),
+        )
+        if get_pp_group().is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[torch.Tensor, PPProxyTensors]:
+        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:
+            assert pp_proxy_tensors is not None
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+
+        kv_cache_index = 0
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            if not layer._is_no_op_attention:
+                hidden_states, residual = layer(
+                    positions, hidden_states, forward_batch, residual
+                )
+                kv_cache_index += 1
+            else:
+                hidden_states, residual = layer(
+                    positions, hidden_states, forward_batch, residual
+                )
+
+        if not get_pp_group().is_last_rank:
+            return PPProxyTensors(
+                {"hidden_states": hidden_states, "residual": residual}
+            )
+
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class DeciLMForCausalLM(nn.Module):
+    packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+
+    # 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"]
+
+    # Mistral/Llama models can also be loaded with --load-format mistral
+    # from consolidated.safetensors checkpoints
+    mistral_mapping = {
+        "layers": "model.layers",
+        "attention": "self_attn",
+        "wq": "q_proj",
+        "wk": "k_proj",
+        "wv": "v_proj",
+        "wo": "o_proj",
+        "attention_norm": "input_layernorm",
+        "feed_forward": "mlp",
+        "w1": "gate_proj",
+        "w2": "down_proj",
+        "w3": "up_proj",
+        "ffn_norm": "post_attention_layernorm",
+        "tok_embeddings": "model.embed_tokens",
+        "output": "lm_head",
+        "norm": "model.norm",
+    }
+
+    def __init__(
+        self,
+        *,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        lora_config = None
+        self.config = config
+        self.lora_config = lora_config
+
+        self.model = self._init_model(
+            config=config, quant_config=quant_config, prefix=add_prefix("model", prefix)
+        )
+        if self.config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.unpadded_vocab_size = config.vocab_size
+            if lora_config:
+                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+            self.lm_head = ParallelLMHead(
+                self.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
+                ),
+                quant_config=quant_config,
+                prefix=add_prefix("lm_head", prefix),
+            )
+        self.logits_processor = LogitsProcessor(config)
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
+
+    def _init_model(
+        self,
+        config: LlamaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        return DeciModel(config=config, quant_config=quant_config, prefix=prefix)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        get_embedding: bool = False,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> LogitsProcessorOutput:
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            inputs_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
+        )
+        if get_pp_group().is_last_rank:
+            if not get_embedding:
+                return self.logits_processor(
+                    input_ids, hidden_states, self.lm_head, forward_batch
+                )
+            else:
+                return self.pooler(hidden_states, forward_batch)
+        else:
+            return hidden_states
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> None:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            (".qkv_proj", ".q_proj", "q"),
+            (".qkv_proj", ".k_proj", "k"),
+            (".qkv_proj", ".v_proj", "v"),
+            (".gate_up_proj", ".gate_proj", 0),
+            (".gate_up_proj", ".up_proj", 1),
+        ]
+
+        params_dict = 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
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+            if self.model.quant_config is not None and (
+                scale_name := self.model.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)
+                continue
+            if "scale" in name:
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    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 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
+                if name in params_dict.keys():
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+                else:
+                    logger.warning(f"Parameter {name} not found in params_dict")
+
+
+EntryClass = [DeciLMForCausalLM]