sglang 0.4.5.post1__py3-none-any.whl → 0.4.5.post2__py3-none-any.whl

sglang/srt/models/llama.py

@@ -362,11 +362,11 @@ class LlamaForCausalLM(nn.Module):
     column_parallel_weights_modules = [".down_proj.", ".o_proj."]
     bitsandbytes_stacked_params_mapping = {
         # shard_name, weight_name, index
-        "q_proj": ("qkv_proj", 0),
-        "k_proj": ("qkv_proj", 1),
-        "v_proj": ("qkv_proj", 2),
-        "gate_proj": ("gate_up_proj", 0),
-        "up_proj": ("gate_up_proj", 1),
+        ".q_proj": (".qkv_proj", 0),
+        ".k_proj": (".qkv_proj", 1),
+        ".v_proj": (".qkv_proj", 2),
+        ".gate_proj": (".gate_up_proj", 0),
+        ".up_proj": (".gate_up_proj", 1),
     }
 
     def __init__(

sglang/srt/models/minicpm3.py

@@ -93,158 +93,6 @@ def input_to_float8(x, dtype=torch.float8_e4m3fn):
     return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
 
 
-class MiniCPM3Attention(nn.Module):
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        hidden_size: int,
-        num_heads: int,
-        qk_nope_head_dim: int,
-        qk_rope_head_dim: int,
-        v_head_dim: int,
-        q_lora_rank: int,
-        kv_lora_rank: int,
-        rope_theta: float = 10000,
-        rope_scaling: Optional[Dict[str, Any]] = None,
-        max_position_embeddings: int = 8192,
-        quant_config: Optional[QuantizationConfig] = None,
-        layer_id=None,
-        prefix: str = "",
-    ) -> None:
-        super().__init__()
-        self.layer_id = layer_id
-        self.hidden_size = hidden_size
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.qk_rope_head_dim = qk_rope_head_dim
-        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
-        self.v_head_dim = v_head_dim
-        self.q_lora_rank = q_lora_rank
-        self.kv_lora_rank = kv_lora_rank
-        self.num_heads = num_heads
-        tp_size = get_tensor_model_parallel_world_size()
-        assert num_heads % tp_size == 0
-        self.num_local_heads = num_heads // tp_size
-        self.scaling = self.qk_head_dim**-0.5
-        self.rope_theta = rope_theta
-        self.max_position_embeddings = max_position_embeddings
-
-        if self.q_lora_rank is not None:
-            self.q_a_proj = ReplicatedLinear(
-                self.hidden_size,
-                self.q_lora_rank,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_a_proj", prefix),
-            )
-            self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
-            self.q_b_proj = ColumnParallelLinear(
-                q_lora_rank,
-                self.num_heads * self.qk_head_dim,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_b_proj", prefix),
-            )
-        else:
-            self.q_proj = ColumnParallelLinear(
-                self.hidden_size,
-                self.num_heads * self.qk_head_dim,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_proj", prefix),
-            )
-
-        self.kv_a_proj_with_mqa = ReplicatedLinear(
-            self.hidden_size,
-            self.kv_lora_rank + self.qk_rope_head_dim,
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("kv_a_proj_with_mqa", prefix),
-        )
-        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
-        self.kv_b_proj = ColumnParallelLinear(
-            self.kv_lora_rank,
-            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("kv_b_proj", prefix),
-        )
-        # O projection.
-        self.o_proj = RowParallelLinear(
-            self.num_heads * self.v_head_dim,
-            self.hidden_size,
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("o_proj", prefix),
-        )
-        self.rotary_emb = get_rope(
-            qk_rope_head_dim,
-            rotary_dim=qk_rope_head_dim,
-            max_position=max_position_embeddings,
-            base=rope_theta,
-            rope_scaling=rope_scaling,
-        )
-
-        # TODO support head_size 96
-        self.attn = RadixAttention(
-            self.num_local_heads,
-            128,
-            self.scaling,
-            num_kv_heads=self.num_local_heads,
-            layer_id=layer_id,
-            quant_config=quant_config,
-            prefix=add_prefix("attn", prefix),
-        )
-
-    def forward(
-        self,
-        positions: torch.Tensor,
-        hidden_states: torch.Tensor,
-        forward_batch: ForwardBatch,
-    ) -> torch.Tensor:
-        if self.q_lora_rank is not None:
-            q = self.q_a_proj(hidden_states)[0]
-            q = self.q_a_layernorm(q)
-            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
-        else:
-            q = self.q_proj(hidden_states)[0].view(
-                -1, self.num_local_heads, self.qk_head_dim
-            )
-        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
-        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
-        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
-        latent_cache = latent_cache.unsqueeze(1)
-        kv_a = self.kv_a_layernorm(kv_a.contiguous())
-        kv = self.kv_b_proj(kv_a)[0]
-        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
-        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
-        k_pe = latent_cache[:, :, self.kv_lora_rank :]
-        original_shapes = [q_pe.shape, k_pe.shape]
-        q_pe, k_pe = self.rotary_emb(
-            positions, q_pe.reshape(q_pe.shape[0], -1), k_pe.reshape(k_pe.shape[0], -1)
-        )
-        q_pe, k_pe = q_pe.view(original_shapes[0]), k_pe.view(original_shapes[1])
-        q[..., self.qk_nope_head_dim :] = q_pe
-        k = torch.empty_like(q)
-        k[..., : self.qk_nope_head_dim] = k_nope
-        k[..., self.qk_nope_head_dim :] = k_pe
-        q = torch.nn.functional.pad(q, [0, 128 - self.qk_head_dim], value=0).view(
-            -1, self.num_local_heads * 128
-        )
-        k = torch.nn.functional.pad(k, [0, 128 - self.qk_head_dim], value=0).view(
-            -1, self.num_local_heads * 128
-        )
-        v = torch.nn.functional.pad(v, [0, 128 - self.v_head_dim], value=0).view(
-            -1, self.num_local_heads * 128
-        )
-        attn_output = self.attn(q, k, v, forward_batch)
-        attn_output = attn_output.view(-1, self.num_local_heads, 128)[
-            ..., : self.v_head_dim
-        ].reshape(-1, self.num_local_heads * self.v_head_dim)
-        output, _ = self.o_proj(attn_output)
-        return output
-
-
 class MiniCPM3AttentionMLA(nn.Module):
 
     def __init__(
@@ -434,44 +282,25 @@ 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 not global_server_args_dict["disable_mla"]:
-            self.self_attn = MiniCPM3AttentionMLA(
-                config=config,
-                hidden_size=self.hidden_size,
-                num_heads=config.num_attention_heads,
-                qk_nope_head_dim=config.qk_nope_head_dim,
-                qk_rope_head_dim=config.qk_rope_head_dim,
-                v_head_dim=self.hidden_size // config.num_attention_heads,
-                q_lora_rank=(
-                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
-                ),
-                kv_lora_rank=config.kv_lora_rank,
-                rope_theta=rope_theta,
-                rope_scaling=rope_scaling,
-                max_position_embeddings=max_position_embeddings,
-                quant_config=quant_config,
-                layer_id=layer_id,
-                prefix=add_prefix("self_attn", prefix),
-            )
-        else:
-            self.self_attn = MiniCPM3Attention(
-                config=config,
-                hidden_size=self.hidden_size,
-                num_heads=config.num_attention_heads,
-                qk_nope_head_dim=config.qk_nope_head_dim,
-                qk_rope_head_dim=config.qk_rope_head_dim,
-                v_head_dim=self.hidden_size // config.num_attention_heads,
-                q_lora_rank=(
-                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
-                ),
-                kv_lora_rank=config.kv_lora_rank,
-                rope_theta=rope_theta,
-                rope_scaling=rope_scaling,
-                max_position_embeddings=max_position_embeddings,
-                quant_config=quant_config,
-                layer_id=layer_id,
-                prefix=add_prefix("self_attn", prefix),
-            )
+        self.self_attn = MiniCPM3AttentionMLA(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            qk_nope_head_dim=config.qk_nope_head_dim,
+            qk_rope_head_dim=config.qk_rope_head_dim,
+            v_head_dim=self.hidden_size // config.num_attention_heads,
+            q_lora_rank=(
+                config.q_lora_rank if hasattr(config, "q_lora_rank") else None
+            ),
+            kv_lora_rank=config.kv_lora_rank,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            layer_id=layer_id,
+            prefix=add_prefix("self_attn", prefix),
+        )
+
         self.mlp = MiniCPM3MLP(
             hidden_size=self.hidden_size,
             intermediate_size=config.intermediate_size,
@@ -674,17 +503,16 @@ class MiniCPM3ForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
-        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(
-                    0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
-                ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
-                self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
-                self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
-                if hasattr(self_attn.kv_b_proj, "weight_scale"):
-                    self_attn.w_scale = self_attn.kv_b_proj.weight_scale
-                del self_attn.kv_b_proj
+        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(
+                0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
+            ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
+            self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
+            self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
+            if hasattr(self_attn.kv_b_proj, "weight_scale"):
+                self_attn.w_scale = self_attn.kv_b_proj.weight_scale
+            del self_attn.kv_b_proj
 
 
 EntryClass = MiniCPM3ForCausalLM

sglang/srt/models/qwen2.py

@@ -239,6 +239,7 @@ class Qwen2Model(nn.Module):
         config: Qwen2Config,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        decoder_layer_type: type[nn.Module] = Qwen2DecoderLayer,
     ) -> None:
         super().__init__()
         self.config = config
@@ -250,9 +251,11 @@ class Qwen2Model(nn.Module):
             quant_config=quant_config,
             prefix=add_prefix("embed_tokens", prefix),
         )
+        # Use the provided decoder layer type or default to Qwen2DecoderLayer
+        decoder_layer_type = decoder_layer_type or Qwen2DecoderLayer
         self.layers = make_layers(
             config.num_hidden_layers,
-            lambda idx, prefix: Qwen2DecoderLayer(
+            lambda idx, prefix: decoder_layer_type(
                 layer_id=idx,
                 config=config,
                 quant_config=quant_config,

sglang/srt/models/qwen2_moe.py

@@ -47,7 +47,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.managers.expert_distribution import ExpertDistributionRecorder
 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
+from sglang.srt.utils import add_prefix, make_layers
 
 expert_distribution_recorder = ExpertDistributionRecorder()
 
@@ -262,8 +262,7 @@ class Qwen2MoeDecoderLayer(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)
-        # note: replace config.num_hidden_layers < 80 with True once its available in transformers 4.50.0
-        qkv_bias = getattr(config, "qkv_bias", config.num_hidden_layers < 80)
+        qkv_bias = getattr(config, "qkv_bias", True)
         self.self_attn = Qwen2MoeAttention(
             hidden_size=self.hidden_size,
             num_heads=config.num_attention_heads,
@@ -334,6 +333,7 @@ class Qwen2MoeModel(nn.Module):
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        decoder_layer_type: type[nn.Module] = Qwen2MoeDecoderLayer,
     ) -> None:
         super().__init__()
         self.padding_idx = config.pad_token_id
@@ -344,16 +344,17 @@ class Qwen2MoeModel(nn.Module):
             config.hidden_size,
             prefix=add_prefix("embed_tokens", prefix),
         )
-        self.layers = nn.ModuleList(
-            [
-                Qwen2MoeDecoderLayer(
-                    config,
-                    layer_id,
-                    quant_config=quant_config,
-                    prefix=add_prefix(f"layers.{layer_id}", prefix),
-                )
-                for layer_id in range(config.num_hidden_layers)
-            ]
+        # Use the provided decoder layer type or default to Qwen2MoeDecoderLayer
+        decoder_layer_type = decoder_layer_type or Qwen2MoeDecoderLayer
+        self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: decoder_layer_type(
+                layer_id=idx,
+                config=config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=add_prefix("layers", prefix),
         )
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 

sglang/srt/models/qwen3.py

@@ -0,0 +1,335 @@
+# Adapted from qwen2.py
+
+from functools import partial
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+
+from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    split_tensor_along_last_dim,
+    tensor_model_parallel_all_gather,
+)
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import 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.layers.rotary_embedding import get_rope
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+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.qwen2 import Qwen2MLP as Qwen3MLP
+from sglang.srt.models.qwen2 import Qwen2Model
+from sglang.srt.utils import add_prefix
+
+Qwen3Config = None
+
+
+class Qwen3Attention(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 1000000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        head_dim: Optional[int] = None,
+        max_position_embeddings: int = 32768,
+        quant_config: Optional[QuantizationConfig] = None,
+        rms_norm_eps: float = None,
+        attention_bias: bool = False,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % self.tp_size == 0
+        self.num_heads = self.total_num_heads // self.tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= self.tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % self.tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size)
+        self.head_dim = head_dim or hidden_size // self.total_num_heads
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+        self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+        self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("qkv_proj", prefix),
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            prefix=add_prefix("attn", prefix),
+        )
+
+    def _apply_qk_norm(
+        self, q: torch.Tensor, k: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        q_by_head = q.reshape(-1, self.head_dim)
+        q_by_head = self.q_norm(q_by_head)
+        q = q_by_head.view(q.shape)
+        k_by_head = k.reshape(-1, self.head_dim)
+        k_by_head = self.k_norm(k_by_head)
+        k = k_by_head.view(k.shape)
+        return q, k
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self._apply_qk_norm(q, k)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Qwen3DecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: Qwen3Config,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 1000000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
+        head_dim = getattr(config, "head_dim", None)
+        self.self_attn = Qwen3Attention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            head_dim=head_dim,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            rms_norm_eps=config.rms_norm_eps,
+            attention_bias=config.attention_bias,
+            prefix=add_prefix("self_attn", prefix),
+        )
+        self.mlp = Qwen3MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class Qwen3Model(Qwen2Model):
+    def __init__(
+        self,
+        config: Qwen3Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(
+            config=config,
+            quant_config=quant_config,
+            prefix=prefix,
+            decoder_layer_type=Qwen3DecoderLayer,
+        )
+
+
+class Qwen3ForCausalLM(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
+    def __init__(
+        self,
+        config: Qwen3Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Qwen3Model(
+            config, quant_config=quant_config, prefix=add_prefix("model", prefix)
+        )
+        if config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.lm_head = ParallelLMHead(
+                config.vocab_size,
+                config.hidden_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 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,
+        input_embeds: torch.Tensor = None,
+        get_embedding: bool = False,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
+        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)
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        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 or "projector" 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 name.startswith("model.vision_tower") and name not in params_dict:
+                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
+                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
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_embed_and_head(self, embed, head):
+        del self.model.embed_tokens.weight
+        del self.lm_head.weight
+        self.model.embed_tokens.weight = embed
+        self.lm_head.weight = head
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        self.model.load_kv_cache_scales(quantization_param_path)
+
+
+EntryClass = Qwen3ForCausalLM