sglang 0.3.1.post1__py3-none-any.whl → 0.3.1.post3__py3-none-any.whl

sglang/srt/models/dbrx.py

@@ -27,12 +27,6 @@ from vllm.distributed import (
     tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.fused_moe import fused_moe
-from vllm.model_executor.layers.linear import (
-    QKVParallelLinear,
-    ReplicatedLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE,
@@ -40,12 +34,18 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding,
 )
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.model_executor.utils import set_weight_attrs
 from vllm.transformers_utils.configs.dbrx import DbrxConfig
 
+from sglang.srt.layers.linear import (
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
+from sglang.srt.utils import set_weight_attrs
 
 
 class DbrxRouter(nn.Module):

sglang/srt/models/deepseek.py

@@ -28,13 +28,6 @@ from vllm.distributed import (
     tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.fused_moe import fused_moe
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    ReplicatedLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -44,7 +37,14 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/deepseek_v2.py

@@ -27,13 +27,6 @@ from vllm.distributed import (
     tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.fused_moe import FusedMoE
-from vllm.model_executor.layers.linear import (
-    ColumnParallelLinear,
-    MergedColumnParallelLinear,
-    ReplicatedLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -43,7 +36,14 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    ColumnParallelLinear,
+    MergedColumnParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
@@ -507,7 +507,7 @@ class DeepseekV2DecoderLayer(nn.Module):
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
-        if global_server_args_dict["enable_mla"]:
+        if not global_server_args_dict["disable_mla"]:
             self.self_attn = DeepseekV2AttentionMLA(
                 config=config,
                 hidden_size=self.hidden_size,
@@ -732,7 +732,7 @@ class DeepseekV2ForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
-        if global_server_args_dict["enable_mla"]:
+        if not global_server_args_dict["disable_mla"]:
             for layer_id in range(self.config.num_hidden_layers):
                 self_attn = self.model.layers[layer_id].self_attn
                 w_kc, w_vc = self_attn.kv_b_proj.weight.unflatten(

sglang/srt/models/exaone.py

@@ -23,12 +23,6 @@ import torch
 from torch import nn
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -38,7 +32,13 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/gemma.py

@@ -23,19 +23,19 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import GeluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/gemma2.py

@@ -22,12 +22,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 
 # from vllm.model_executor.layers.rotary_embedding import GemmaRotaryEmbedding
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
@@ -35,7 +29,13 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import GeluAndMul
 from sglang.srt.layers.layernorm import GemmaRMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/gpt_bigcode.py

@@ -23,17 +23,17 @@ from torch import nn
 from transformers import GPTBigCodeConfig
 from vllm.config import CacheConfig, LoRAConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    ColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import get_act_fn
+from sglang.srt.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/grok.py

@@ -28,12 +28,6 @@ from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
 )
-from vllm.model_executor.layers.linear import (
-    QKVParallelLinear,
-    ReplicatedLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -44,7 +38,13 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.fused_moe import FusedMoE
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/internlm2.py

@@ -23,12 +23,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -38,7 +32,13 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/llama.py

@@ -24,12 +24,6 @@ from torch import nn
 from transformers import LlamaConfig
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -39,7 +33,13 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.managers.schedule_batch import global_server_args_dict
@@ -305,8 +305,6 @@ class LlamaForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
-        self.param_dict = dict(self.named_parameters())
-
     @torch.no_grad()
     def forward(
         self,
@@ -374,7 +372,7 @@ class LlamaForCausalLM(nn.Module):
             (".gate_up_proj", ".gate_proj", 0),
             (".gate_up_proj", ".up_proj", 1),
         ]
-        params_dict = self.param_dict
+        params_dict = dict(self.named_parameters())
 
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name or "projector" in name:

sglang/srt/models/llama_classification.py

@@ -19,10 +19,10 @@ import torch
 from torch import nn
 from transformers import LlamaConfig
 from vllm.config import CacheConfig
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 from sglang.srt.models.llama import LlamaForCausalLM, LlamaModel
 
@@ -36,6 +36,7 @@ class LlamaForClassification(nn.Module):
     ) -> None:
         super().__init__()
         self.config = config
+        self.torchao_config = None
         self.quant_config = quant_config
         self.model = LlamaModel(config, quant_config=quant_config)
 
@@ -44,8 +45,6 @@ class LlamaForClassification(nn.Module):
         )
         self.eos_token_id = config.eos_token_id
 
-        self.param_dict = dict(self.named_parameters())
-
     @torch.no_grad()
     def forward(
         self,
@@ -77,7 +76,7 @@ class LlamaForClassification(nn.Module):
         return logits_output
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-        params_dict = self.param_dict
+        params_dict = dict(self.named_parameters())
 
         for name, loaded_weight in weights:
             if "classification_head" in name:

sglang/srt/models/llava.py

@@ -32,9 +32,9 @@ from transformers import (
 )
 from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
 from vllm.config import CacheConfig
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.mm_utils import (
     get_anyres_image_grid_shape,
     unpad_image,

sglang/srt/models/llavavid.py

@@ -23,9 +23,9 @@ from torch import nn
 from transformers import CLIPVisionModel, LlavaConfig
 from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
 from vllm.config import CacheConfig
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.model_executor.forward_batch_info import ForwardMode, InputMetadata
 from sglang.srt.models.llama import LlamaForCausalLM
 

sglang/srt/models/minicpm.py

@@ -22,12 +22,6 @@ import torch
 from torch import nn
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -37,7 +31,13 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/minicpm3.py

@@ -29,7 +29,6 @@ from vllm.model_executor.layers.linear import (
     ReplicatedLinear,
     RowParallelLinear,
 )
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -40,6 +39,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
@@ -419,7 +419,7 @@ class MiniCPM3DecoderLayer(nn.Module):
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
-        if global_server_args_dict["enable_mla"]:
+        if not global_server_args_dict["disable_mla"]:
             self.self_attn = MiniCPM3AttentionMLA(
                 config=config,
                 hidden_size=self.hidden_size,
@@ -653,7 +653,7 @@ class MiniCPM3ForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
-        if global_server_args_dict["enable_mla"]:
+        if not global_server_args_dict["disable_mla"]:
             for layer_id in range(self.config.num_hidden_layers):
                 self_attn = self.model.layers[layer_id].self_attn
                 w_kc, w_vc = self_attn.kv_b_proj.weight.unflatten(

sglang/srt/models/mixtral.py

@@ -24,12 +24,6 @@ from transformers import MixtralConfig
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.fused_moe import FusedMoE
-from vllm.model_executor.layers.linear import (
-    QKVParallelLinear,
-    ReplicatedLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE,
@@ -39,7 +33,13 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.managers.schedule_batch import global_server_args_dict

sglang/srt/models/mixtral_quant.py

@@ -29,12 +29,6 @@ from vllm.distributed import (
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
-from vllm.model_executor.layers.linear import (
-    QKVParallelLinear,
-    ReplicatedLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -43,7 +37,13 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/olmoe.py

@@ -35,7 +35,6 @@ from vllm.model_executor.layers.linear import (
     ReplicatedLinear,
     RowParallelLinear,
 )
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -47,6 +46,7 @@ from vllm.utils import print_warning_once
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/qwen.py

@@ -22,12 +22,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -37,7 +31,13 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/qwen2.py

@@ -22,12 +22,6 @@ import torch
 from torch import nn
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -37,8 +31,14 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.pooler import Pooler, PoolingType
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/qwen2_moe.py

@@ -29,13 +29,6 @@ from vllm.distributed import (
     tensor_model_parallel_all_reduce,
 )
 from vllm.model_executor.layers.fused_moe import FusedMoE
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    ReplicatedLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -45,7 +38,14 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.managers.schedule_batch import global_server_args_dict

sglang/srt/models/stablelm.py

@@ -24,12 +24,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.config import CacheConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.linear import (
-    MergedColumnParallelLinear,
-    QKVParallelLinear,
-    RowParallelLinear,
-)
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -38,7 +32,13 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 

sglang/srt/models/xverse.py

@@ -31,7 +31,6 @@ from vllm.model_executor.layers.linear import (
     QKVParallelLinear,
     RowParallelLinear,
 )
-from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead,
@@ -40,6 +39,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.model_runner import InputMetadata
 
@@ -307,8 +307,6 @@ class XverseForCausalLM(nn.Module):
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
 
-        self.param_dict = dict(self.named_parameters())
-
     @torch.no_grad()
     def forward(
         self,
@@ -333,7 +331,7 @@ class XverseForCausalLM(nn.Module):
             ("gate_up_proj", "gate_proj", 0),
             ("gate_up_proj", "up_proj", 1),
         ]
-        params_dict = self.param_dict
+        params_dict = dict(self.named_parameters())
 
         def load_weights_per_param(name, loaded_weight):
             if "rotary_emb.inv_freq" in name or "projector" in name: