sglang 0.3.4.post1__py3-none-any.whl → 0.3.5__py3-none-any.whl

sglang/srt/models/gpt2.py

@@ -0,0 +1,287 @@
+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gpt2/modeling_gpt2.py
+# Copyright 2023 The vLLM team.
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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 GPT-2 model compatible with HuggingFace weights."""
+from typing import Iterable, List, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import GPT2Config
+from vllm.config import CacheConfig
+from vllm.distributed.parallel_state import get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import get_act_fn
+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.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+
+
+class GPT2Attention(nn.Module):
+
+    def __init__(
+        self,
+        layer_id: int,
+        config: GPT2Config,
+        cache_config = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        total_num_heads = config.num_attention_heads
+        tensor_model_parallel_world_size = (
+            get_tensor_model_parallel_world_size())
+        assert total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = total_num_heads // tensor_model_parallel_world_size
+        self.head_dim = self.hidden_size // total_num_heads
+        self.scale = self.head_dim**-0.5
+
+        self.c_attn = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            total_num_heads,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_attn",
+        )
+        self.c_proj = RowParallelLinear(
+            self.hidden_size,
+            self.hidden_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
+        )
+        self.attn = RadixAttention(self.num_heads,
+                              self.head_dim,
+                              scaling=self.scale,
+                              num_kv_heads=total_num_heads,
+                              layer_id=layer_id)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        qkv, _ = self.c_attn(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        attn_output = self.attn(q, k, v, forward_batch)
+        attn_output, _ = self.c_proj(attn_output)
+        return attn_output
+
+
+class GPT2MLP(nn.Module):
+
+    def __init__(
+        self,
+        intermediate_size: int,
+        config: GPT2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.c_fc = ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_fc",
+        )
+        self.c_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
+        )
+        self.act = get_act_fn(config.activation_function, quant_config,
+                              intermediate_size)
+
+    def forward(self, hidden_states: torch.Tensor,) -> torch.Tensor:
+        hidden_states, _ = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.c_proj(hidden_states)
+        return hidden_states
+
+
+class GPT2Block(nn.Module):
+
+    def __init__(
+        self,
+        layer_id: int,
+        config: GPT2Config,
+        cache_config = None,
+        
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        inner_dim = (config.n_inner if config.n_inner is not None else 4 *
+                     hidden_size)
+
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = GPT2Attention(layer_id,
+                                  config,
+                                  cache_config,
+                                  quant_config,
+                                  prefix=f"{prefix}.attn")
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.mlp = GPT2MLP(inner_dim,
+                           config,
+                           quant_config,
+                           prefix=f"{prefix}.mlp")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_output = self.attn(
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+        # residual connection
+        hidden_states = attn_output + residual
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+        return hidden_states
+
+
+
+class GPT2Model(nn.Module):
+
+    def __init__(
+        self,
+        config: GPT2Config,
+        cache_config = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        assert not config.add_cross_attention
+        assert not config.scale_attn_by_inverse_layer_idx
+        assert not config.reorder_and_upcast_attn
+        self.embed_dim = config.hidden_size
+        self.wte = VocabParallelEmbedding(config.vocab_size, self.embed_dim)
+        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
+        self.h = nn.ModuleList(
+            [
+                GPT2Block(i, config, cache_config, quant_config)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+
+        for i in range(len(self.h)):
+            layer = self.h[i]
+            hidden_states = layer(hidden_states, forward_batch)
+
+        hidden_states = self.ln_f(hidden_states)
+        return hidden_states
+
+
+class GPT2LMHeadModel(nn.Module):
+
+    def __init__(
+        self,
+        config: GPT2Config,
+        cache_config = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.transformer = GPT2Model(config,
+                                     cache_config,
+                                     quant_config,
+                                     prefix="transformer")
+        self.lm_head = self.transformer.wte
+
+        self.logits_processor = LogitsProcessor(config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        hidden_states = self.transformer(input_ids, positions, forward_batch)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, forward_batch
+        )
+
+
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        for name, loaded_weight in weights:
+            if "lm_head.weight" in name:
+                # GPT-2 ties the weights of the embedding layer and the final
+                # linear layer.
+                continue
+            if ".attn.bias" in name or ".attn.masked_bias" in name:
+                # Skip attention mask.
+                # NOTE: "c_attn.bias" should not be skipped.
+                continue
+            if not name.startswith("transformer."):
+                name = "transformer." + name
+
+            param = params_dict[name]
+            # The HF's GPT-2 implementation uses Conv1D instead of Linear.
+            # Because of this, we need to transpose the weights.
+            # Note(zhuohan): the logic below might break quantized models.
+            for conv1d_weight_name in ["c_attn", "c_proj", "c_fc"]:
+                if conv1d_weight_name not in name:
+                    continue
+                if not name.endswith(".weight"):
+                    continue
+                loaded_weight = loaded_weight.t()
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)
+
+EntryClass = GPT2LMHeadModel

sglang/srt/models/gpt_bigcode.py

@@ -23,7 +23,6 @@ from torch import nn
 from transformers import GPTBigCodeConfig
 from vllm.config import LoRAConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
-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
@@ -35,6 +34,7 @@ from sglang.srt.layers.linear import (
 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.vocab_parallel_embedding import VocabParallelEmbedding
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/grok.py

@@ -28,10 +28,6 @@ from vllm.distributed import (
     get_tensor_model_parallel_world_size,
 )
 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.loader import DefaultModelLoader
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
@@ -45,6 +41,10 @@ from sglang.srt.layers.linear import (
 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.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/internlm2.py

@@ -23,10 +23,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 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 sglang.srt.layers.activation import SiluAndMul
@@ -39,6 +35,10 @@ from sglang.srt.layers.linear import (
 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.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/llama.py

@@ -24,10 +24,6 @@ from torch import nn
 from transformers import LlamaConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 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 sglang.srt.layers.activation import SiluAndMul
@@ -38,9 +34,14 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
+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.layers.torchao_utils import apply_torchao_config_
+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
 
@@ -303,6 +304,7 @@ class LlamaForCausalLM(nn.Module):
         self.model = LlamaModel(config, quant_config=quant_config)
         self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
         self.logits_processor = LogitsProcessor(config)
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
     @torch.no_grad()
     def forward(
@@ -311,11 +313,15 @@ class LlamaForCausalLM(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        get_embedding: bool = False,
     ) -> LogitsProcessorOutput:
         hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
-        return self.logits_processor(
-            input_ids, hidden_states, self.lm_head.weight, forward_batch
-        )
+        if not get_embedding:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head.weight, forward_batch
+            )
+        else:
+            return self.pooler(hidden_states, forward_batch)
 
     def get_hidden_dim(self, module_name):
         # return input_dim, output_dim
@@ -409,11 +415,13 @@ class LlamaForCausalLM(nn.Module):
         if (
             hasattr(self.config, "tie_word_embeddings")
             and self.config.tie_word_embeddings
+            and "lm_head.weight" in params_dict
         ):
             # Tie output embedding layer to input embedding layer, to solve issues where lm_head.weight is missing
             param = self.lm_head.weight
             weight_loader = getattr(param, "weight_loader", default_weight_loader)
             weight_loader(param, self.model.embed_tokens.weight)
+
         apply_torchao_config_(self, params_dict, set(["proj.weight"]))
 
 

sglang/srt/models/llama_embedding.py

@@ -36,9 +36,7 @@ class LlamaEmbeddingModel(nn.Module):
         hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
         return self.pooler(hidden_states, forward_batch)
 
-    def load_weights(
-        self, weights: Iterable[Tuple[str, torch.Tensor]], name=None, loaded_weight=None
-    ):
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
             ("qkv_proj", "q_proj", "q"),
@@ -49,7 +47,7 @@ class LlamaEmbeddingModel(nn.Module):
         ]
         params_dict = dict(self.model.named_parameters())
 
-        def load_weights_per_param(name, loaded_weight):
+        for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name or "projector" in name:
                 return
             if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
@@ -78,12 +76,6 @@ class LlamaEmbeddingModel(nn.Module):
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 weight_loader(param, loaded_weight)
 
-        if name is None or loaded_weight is None:
-            for name, loaded_weight in weights:
-                load_weights_per_param(name, loaded_weight)
-        else:
-            load_weights_per_param(name, loaded_weight)
-
 
 class MistralModel(LlamaEmbeddingModel):
     pass

sglang/srt/models/llama_reward.py

@@ -52,7 +52,12 @@ class LlamaForSequenceClassification(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        get_embedding: bool = True,
     ) -> EmbeddingPoolerOutput:
+        assert (
+            get_embedding
+        ), "LlamaForSequenceClassification is only used for embedding"
+
         hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
         scores = self.score(hidden_states)
 

sglang/srt/models/minicpm.py

@@ -22,10 +22,6 @@ import torch
 from torch import nn
 from vllm.distributed import get_tensor_model_parallel_world_size
 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 sglang.srt.layers.activation import SiluAndMul
@@ -38,6 +34,10 @@ from sglang.srt.layers.linear import (
 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.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/minicpm3.py

@@ -29,10 +29,6 @@ from vllm.model_executor.layers.linear import (
     RowParallelLinear,
 )
 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 sglang.srt.layers.activation import SiluAndMul
@@ -40,6 +36,10 @@ 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.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.utils import is_flashinfer_available

sglang/srt/models/mixtral.py

@@ -24,11 +24,6 @@ from transformers import MixtralConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.fused_moe import FusedMoE
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.vocab_parallel_embedding import (
-    DEFAULT_VOCAB_PADDING_SIZE,
-    ParallelLMHead,
-    VocabParallelEmbedding,
-)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.layernorm import RMSNorm
@@ -41,6 +36,10 @@ 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.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
 
@@ -369,6 +368,9 @@ class MixtralForCausalLM(nn.Module):
                     # Skip loading extra bias for GPTQ models.
                     if name.endswith(".bias") and name not in params_dict:
                         continue
+                    # Skip loading kv_scale from ckpts towards new design.
+                    if name.endswith(".kv_scale") and name not in params_dict:
+                        continue
                     if name is None:
                         continue
 

sglang/srt/models/mixtral_quant.py

@@ -29,10 +29,6 @@ from vllm.distributed import (
     tensor_model_parallel_all_reduce,
 )
 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 sglang.srt.layers.layernorm import RMSNorm
@@ -44,6 +40,10 @@ from sglang.srt.layers.linear import (
 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.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/mllama.py

@@ -15,11 +15,6 @@ from transformers.models.mllama.modeling_mllama import (
     _prepare_aspect_ratio_attention_mask,
 )
 from vllm.distributed import get_tensor_model_parallel_world_size
-from vllm.model_executor.layers.vocab_parallel_embedding import (
-    DEFAULT_VOCAB_PADDING_SIZE,
-    ParallelLMHead,
-    VocabParallelEmbedding,
-)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import get_act_fn
@@ -32,6 +27,11 @@ from sglang.srt.layers.linear import (
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.quantization import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.vocab_parallel_embedding import (
+    DEFAULT_VOCAB_PADDING_SIZE,
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.managers.schedule_batch import ImageInputs
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.models.llama import LlamaDecoderLayer, LlamaMLP

sglang/srt/models/olmo.py

@@ -23,10 +23,6 @@ from torch import nn
 from transformers import OlmoConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 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 sglang.srt.layers.activation import SiluAndMul
@@ -38,6 +34,10 @@ from sglang.srt.layers.linear import (
 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.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/olmoe.py

@@ -35,10 +35,6 @@ from vllm.model_executor.layers.linear import (
     RowParallelLinear,
 )
 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.utils import print_warning_once
 
@@ -47,6 +43,10 @@ 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.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/qwen.py

@@ -22,10 +22,6 @@ from torch import nn
 from transformers import PretrainedConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 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 sglang.srt.layers.activation import SiluAndMul
@@ -38,6 +34,10 @@ from sglang.srt.layers.linear import (
 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.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 

sglang/srt/models/qwen2.py

@@ -22,10 +22,6 @@ import torch
 from torch import nn
 from vllm.distributed import get_tensor_model_parallel_world_size
 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 sglang.srt.layers.activation import SiluAndMul
@@ -39,6 +35,10 @@ 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.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 Qwen2Config = None