sglang 0.5.0rc1__py3-none-any.whl → 0.5.1__py3-none-any.whl

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]

sglang/srt/models/olmoe.py

@@ -89,7 +89,6 @@ class OlmoeMoE(nn.Module):
             intermediate_size=intermediate_size,
             reduce_results=True,
             quant_config=quant_config,
-            tp_size=tp_size,
             layer_id=layer_id,
             prefix=add_prefix("experts", prefix),
         )

sglang/srt/models/phi4mm.py

@@ -54,25 +54,6 @@ VISION_ENCODER_TO_PROCESSING_CONFIG = {
 }
 
 
-def get_navit_vision_model():
-    vision_config = {
-        "hidden_size": 1152,
-        "image_size": 448,
-        "intermediate_size": 4304,
-        "model_type": "siglip_vision_model",
-        "num_attention_heads": 16,
-        "num_hidden_layers": 26,  # Model is originally 27-layer, we only need the first 26 layers for feature extraction.
-        "patch_size": 14,
-    }
-    model_config = SiglipVisionConfig(**vision_config)
-
-    vision_model = Idefics2VisionTransformer(
-        config=model_config, require_post_norm=False
-    )
-
-    return vision_model
-
-
 class Phi4MMImageEncoder(nn.Module):
     """Image embedding."""
 
@@ -88,8 +69,9 @@ class Phi4MMImageEncoder(nn.Module):
         # n_embed or hidden_size
         hidden_size = config.n_embd if hasattr(config, "n_embd") else config.hidden_size
         self.type_feature = "patch"
-
-        self.img_processor = get_navit_vision_model()
+        self.img_processor = Idefics2VisionTransformer(
+            config=config.vision_config, require_post_norm=False
+        )
 
         pe_weight = self.img_processor.embeddings.position_embedding.weight
         L, D = pe_weight.size()

sglang/srt/models/qwen2.py

@@ -27,6 +27,7 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
 )
 from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.dp_attention import is_dp_attention_enabled
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
     MergedColumnParallelLinear,
@@ -43,7 +44,6 @@ 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,
@@ -273,7 +273,7 @@ class Qwen2Model(nn.Module):
                 config.vocab_size,
                 config.hidden_size,
                 quant_config=quant_config,
-                enable_tp=not global_server_args_dict["enable_dp_attention"],
+                enable_tp=not is_dp_attention_enabled(),
                 prefix=add_prefix("embed_tokens", prefix),
             )
         else:

sglang/srt/models/qwen2_5_vl.py

@@ -117,6 +117,7 @@ class Qwen2_5_VisionBlock(nn.Module):
         attn_implementation: Optional[str] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        num_dummy_heads: int = 0,
     ) -> None:
         super().__init__()
         if norm_layer is None:
@@ -157,6 +158,7 @@ class Qwen2_5_VisionBlock(nn.Module):
             flatten_batch=flatten_batch,
             quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
+            num_dummy_heads=num_dummy_heads,
         )
         self.mlp = Qwen2_5_VLMLP(
             dim,

sglang/srt/models/qwen2_moe.py

@@ -17,8 +17,6 @@
 """Inference-only Qwen2MoE model compatible with HuggingFace weights."""
 
 import logging
-from dataclasses import dataclass
-from enum import Enum, auto
 from typing import Any, Dict, Iterable, Optional, Tuple, Union
 
 import torch
@@ -31,10 +29,7 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
-from sglang.srt.eplb.expert_distribution import (
-    ExpertDistributionRecorder,
-    get_global_expert_distribution_recorder,
-)
+from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
 from sglang.srt.eplb.expert_location import ModelConfigForExpertLocation
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.communicator import (
@@ -45,7 +40,7 @@ from sglang.srt.layers.communicator import (
 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
 from sglang.srt.layers.linear import (
@@ -54,8 +49,8 @@ from sglang.srt.layers.linear import (
     ReplicatedLinear,
     RowParallelLinear,
 )
-from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorOutput
-from sglang.srt.layers.moe.ep_moe.layer import EPMoE, get_moe_impl_class
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.moe.topk import TopK
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -107,10 +102,14 @@ class Qwen2MoeMLP(nn.Module):
             )
         self.act_fn = SiluAndMul()
 
-    def forward(self, x):
+    def forward(
+        self,
+        x,
+        use_reduce_scatter: bool = False,
+    ):
         gate_up, _ = self.gate_up_proj(x)
         x = self.act_fn(gate_up)
-        x, _ = self.down_proj(x)
+        x, _ = self.down_proj(x, skip_all_reduce=use_reduce_scatter)
         return x
 
 
@@ -144,14 +143,6 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
             intermediate_size=config.moe_intermediate_size,
             quant_config=quant_config,
             prefix=add_prefix("experts", prefix),
-            # Additional args for FusedMoE
-            **(
-                dict(
-                    enable_flashinfer_cutlass_moe=True,
-                )
-                if global_server_args_dict["enable_flashinfer_cutlass_moe"]
-                else {}
-            ),
         )
 
         self.gate = ReplicatedLinear(
@@ -175,7 +166,10 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
 
     def forward(
-        self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None
+        self,
+        hidden_states: torch.Tensor,
+        forward_batch: Optional[ForwardBatch] = None,
+        use_reduce_scatter: bool = False,
     ) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
@@ -193,6 +187,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         final_hidden_states = self.experts(hidden_states, topk_output)
         if shared_output is not None:
             final_hidden_states = final_hidden_states + shared_output
+        if self.tp_size > 1 and not use_reduce_scatter:
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states.view(num_tokens, hidden_dim)
@@ -331,7 +326,6 @@ class Qwen2MoeDecoderLayer(nn.Module):
 
         self.attn_tp_size = get_attention_tp_size()
         self.attn_tp_rank = get_attention_tp_rank()
-        self.local_dp_size = get_local_attention_dp_size()
 
         # Qwen2MoE all layers are sparse and have no nextn now
         self.is_layer_sparse = True
@@ -367,6 +361,7 @@ class Qwen2MoeDecoderLayer(nn.Module):
             layer_scatter_modes=self.layer_scatter_modes,
             input_layernorm=self.input_layernorm,
             post_attention_layernorm=self.post_attention_layernorm,
+            allow_reduce_scatter=True,
         )
 
     def forward(
@@ -392,7 +387,12 @@ class Qwen2MoeDecoderLayer(nn.Module):
             hidden_states, residual, forward_batch
         )
 
-        hidden_states = self.mlp(hidden_states, forward_batch)
+        # For DP with padding, reduce scatter can be used instead of all-reduce.
+        use_reduce_scatter = self.layer_communicator.should_use_reduce_scatter(
+            forward_batch
+        )
+
+        hidden_states = self.mlp(hidden_states, forward_batch, use_reduce_scatter)
 
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             hidden_states, residual, forward_batch
@@ -420,7 +420,7 @@ class Qwen2MoeModel(nn.Module):
             self.embed_tokens = VocabParallelEmbedding(
                 config.vocab_size,
                 config.hidden_size,
-                enable_tp=not global_server_args_dict["enable_dp_attention"],
+                enable_tp=not is_dp_attention_enabled(),
                 prefix=add_prefix("embed_tokens", prefix),
             )
         else:

sglang/srt/models/qwen3.py

@@ -327,8 +327,8 @@ class Qwen3ForCausalLM(nn.Module):
         # For EAGLE3 support
         self.capture_aux_hidden_states = False
 
-    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
-        return self.model.get_input_embeddings(input_ids)
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.model.get_input_embeddings()
 
     @torch.no_grad()
     def forward(