sglang 0.4.3.post2__py3-none-any.whl → 0.4.3.post4__py3-none-any.whl

sglang/srt/models/phi3_small.py

@@ -24,7 +24,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 )
 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 make_layers
+from sglang.srt.utils import add_prefix, make_layers
 
 
 @torch.jit.script
@@ -70,13 +70,14 @@ class Phi3SmallMLP(nn.Module):
             2 * [self.intermediate_size],
             bias=True,
             quant_config=quant_config,
-            prefix=f"{prefix}.up_proj",
+            prefix=add_prefix("up_proj", prefix),
         )
         self.down_proj = RowParallelLinear(
             self.intermediate_size,
             self.hidden_size,
             bias=True,
             quant_config=quant_config,
+            prefix=add_prefix("down_proj", prefix),
         )
 
     def forward(self, x):
@@ -140,7 +141,7 @@ class Phi3SmallSelfAttention(nn.Module):
             self.num_key_value_heads,
             bias=True,
             quant_config=quant_config,
-            prefix=f"{prefix}.qkv_proj",
+            prefix=add_prefix("qkv_proj", prefix),
         )
 
         self.dense = RowParallelLinear(
@@ -148,7 +149,7 @@ class Phi3SmallSelfAttention(nn.Module):
             self.hidden_size,
             bias=True,
             quant_config=quant_config,
-            prefix=f"{prefix}.o_proj",
+            prefix=add_prefix("o_proj", prefix),
         )
 
         if getattr(self.config, "rope_scaling", None) is not None:
@@ -201,6 +202,7 @@ class Phi3SmallSelfAttention(nn.Module):
             self.scale,
             num_kv_heads=self.num_kv_heads_per_partion,
             layer_id=layer_id,
+            prefix=add_prefix("attn", prefix),
         )
 
     def forward(
@@ -234,13 +236,21 @@ class Phi3SmallDecoderLayer(nn.Module):
         config: PretrainedConfig,
         layer_id: int,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.hidden_size = config.hidden_size
         self.self_attn = Phi3SmallSelfAttention(
-            config, layer_id, quant_config=quant_config
+            config,
+            layer_id,
+            quant_config=quant_config,
+            prefix=add_prefix("self_attn", prefix),
+        )
+        self.mlp = Phi3SmallMLP(
+            config,
+            quant_config,
+            prefix=add_prefix("mlp", prefix),
         )
-        self.mlp = Phi3SmallMLP(config, quant_config)
 
         self.input_layernorm = nn.LayerNorm(
             config.hidden_size, eps=config.layer_norm_epsilon
@@ -284,15 +294,20 @@ class Phi3SmallModel(nn.Module):
 
         self.config = config
         self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size, config.hidden_size
+            config.vocab_size,
+            config.hidden_size,
+            prefix=add_prefix("embed_tokens", prefix),
         )
         self.mup_embedding_multiplier = config.mup_embedding_multiplier
         self.start_layer, self.end_layer, self.layers = make_layers(
             config.num_hidden_layers,
             lambda prefix: Phi3SmallDecoderLayer(
-                config, int(prefix.split(".")[-1]), quant_config
+                config,
+                int(prefix.split(".")[-1]),
+                quant_config,
+                prefix=prefix,
             ),
-            prefix=f"{prefix}.layers",
+            prefix=add_prefix("layers", prefix),
         )
 
         self.final_layernorm = nn.LayerNorm(
@@ -335,6 +350,7 @@ class Phi3SmallForCausalLM(nn.Module):
         self,
         config: Phi3Config,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
 
         super().__init__()
@@ -344,7 +360,7 @@ class Phi3SmallForCausalLM(nn.Module):
         self.model = Phi3SmallModel(
             config=config,
             quant_config=quant_config,
-            prefix="model",
+            prefix=add_prefix("model", prefix),
         )
         self.vocab_size = config.vocab_size
         self.mup_width_multiplier = config.mup_width_multiplier
@@ -354,6 +370,7 @@ class Phi3SmallForCausalLM(nn.Module):
             org_num_embeddings=config.vocab_size,
             padding_size=DEFAULT_VOCAB_PADDING_SIZE,
             quant_config=quant_config,
+            prefix=add_prefix("lm_head", prefix),
         )
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
@@ -433,6 +450,8 @@ class Phi3SmallForCausalLM(nn.Module):
                 continue
             if name.endswith(".bias") and name not in params_dict:
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
 
             param = params_dict[name]
             weight_loader = getattr(param, "weight_loader", default_weight_loader)

sglang/srt/models/qwen.py

@@ -39,6 +39,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 )
 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
 
 
 class QWenMLP(nn.Module):
@@ -48,6 +49,7 @@ class QWenMLP(nn.Module):
         intermediate_size: int,
         hidden_act: str = "silu",
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
@@ -56,6 +58,7 @@ class QWenMLP(nn.Module):
             bias=False,
             gather_output=False,
             quant_config=quant_config,
+            prefix=add_prefix("gate_up_proj", prefix),
         )
         self.c_proj = RowParallelLinear(
             intermediate_size,
@@ -63,6 +66,7 @@ class QWenMLP(nn.Module):
             bias=False,
             input_is_parallel=True,
             quant_config=quant_config,
+            prefix=add_prefix("c_proj", prefix),
         )
         if hidden_act != "silu":
             raise ValueError(
@@ -88,6 +92,7 @@ class QWenAttention(nn.Module):
         rope_theta: float = 10000,
         rope_scaling: Optional[Dict[str, Any]] = None,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.hidden_size = hidden_size
@@ -104,6 +109,7 @@ class QWenAttention(nn.Module):
             self.total_num_heads,
             bias=True,
             quant_config=quant_config,
+            prefix=add_prefix("c_attn", prefix),
         )
         self.c_proj = RowParallelLinear(
             self.total_num_heads * self.head_dim,
@@ -111,6 +117,7 @@ class QWenAttention(nn.Module):
             bias=False,
             input_is_parallel=True,
             quant_config=quant_config,
+            prefix=add_prefix("c_proj", prefix),
         )
         self.rotary_emb = get_rope(
             self.head_dim,
@@ -126,6 +133,7 @@ class QWenAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_heads,
             layer_id=layer_id,
+            prefix=add_prefix("attn", prefix),
         )
 
     def forward(
@@ -148,6 +156,7 @@ class QWenBlock(nn.Module):
         config: PretrainedConfig,
         layer_id,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
@@ -162,6 +171,7 @@ class QWenBlock(nn.Module):
             rope_scaling=rope_scaling,
             layer_id=layer_id,
             quant_config=quant_config,
+            prefix=add_prefix("attn", prefix),
         )
 
         self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
@@ -170,6 +180,7 @@ class QWenBlock(nn.Module):
             config.hidden_size,
             config.intermediate_size // 2,
             quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
         )
 
     def forward(
@@ -201,6 +212,7 @@ class QWenModel(nn.Module):
         self,
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.config = config
@@ -210,10 +222,16 @@ class QWenModel(nn.Module):
         self.wte = VocabParallelEmbedding(
             vocab_size,
             config.hidden_size,
+            prefix=add_prefix("wte", prefix),
         )
         self.h = nn.ModuleList(
             [
-                QWenBlock(config, i, quant_config=quant_config)
+                QWenBlock(
+                    config,
+                    i,
+                    quant_config=quant_config,
+                    prefix=add_prefix(f"h.{i}", prefix),
+                )
                 for i in range(config.num_hidden_layers)
             ]
         )
@@ -242,12 +260,17 @@ class QWenLMHeadModel(nn.Module):
         self,
         config: PretrainedConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.config = config
-        self.transformer = QWenModel(config, quant_config=quant_config)
+        self.transformer = QWenModel(
+            config, quant_config=quant_config, prefix=add_prefix("transformer", prefix)
+        )
         vocab_size = ((config.vocab_size + 63) // 64) * 64
-        self.lm_head = ParallelLMHead(vocab_size, config.hidden_size)
+        self.lm_head = ParallelLMHead(
+            vocab_size, config.hidden_size, prefix=add_prefix("lm_head", prefix)
+        )
         self.logits_processor = LogitsProcessor(config)
 
     @torch.no_grad()

sglang/srt/models/qwen2.py

@@ -15,7 +15,7 @@
 # Adapted from llama2.py
 # Modify details for the adaptation of Qwen2 model.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
-
+from readline import add_history
 from typing import Any, Dict, Iterable, Optional, Tuple
 
 import torch
@@ -46,7 +46,7 @@ from sglang.srt.model_loader.weight_utils import (
     default_weight_loader,
     kv_cache_scales_loader,
 )
-from sglang.srt.utils import make_layers
+from sglang.srt.utils import add_prefix, make_layers
 
 Qwen2Config = None
 
@@ -58,6 +58,7 @@ class Qwen2MLP(nn.Module):
         intermediate_size: int,
         hidden_act: str,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
@@ -65,12 +66,14 @@ class Qwen2MLP(nn.Module):
             [intermediate_size] * 2,
             bias=False,
             quant_config=quant_config,
+            prefix=add_prefix("gate_up_proj", prefix),
         )
         self.down_proj = RowParallelLinear(
             intermediate_size,
             hidden_size,
             bias=False,
             quant_config=quant_config,
+            prefix=add_prefix("down_proj", prefix),
         )
         if hidden_act != "silu":
             raise ValueError(
@@ -97,6 +100,7 @@ class Qwen2Attention(nn.Module):
         rope_scaling: Optional[Dict[str, Any]] = None,
         max_position_embeddings: int = 32768,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -128,12 +132,14 @@ class Qwen2Attention(nn.Module):
             self.total_num_kv_heads,
             bias=True,
             quant_config=quant_config,
+            prefix=add_prefix("qkv_proj", prefix),
         )
         self.o_proj = RowParallelLinear(
             self.total_num_heads * self.head_dim,
             hidden_size,
             bias=False,
             quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
         )
 
         self.rotary_emb = get_rope(
@@ -149,6 +155,7 @@ class Qwen2Attention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            prefix=add_prefix("attn", prefix),
         )
 
     def forward(
@@ -171,6 +178,7 @@ class Qwen2DecoderLayer(nn.Module):
         config: Qwen2Config,
         layer_id: int = 0,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -186,12 +194,14 @@ class Qwen2DecoderLayer(nn.Module):
             rope_scaling=rope_scaling,
             max_position_embeddings=max_position_embeddings,
             quant_config=quant_config,
+            prefix=add_prefix("self_attn", prefix),
         )
         self.mlp = Qwen2MLP(
             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(
@@ -228,6 +238,7 @@ class Qwen2Model(nn.Module):
         self,
         config: Qwen2Config,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.config = config
@@ -237,6 +248,7 @@ class Qwen2Model(nn.Module):
             config.vocab_size,
             config.hidden_size,
             quant_config=quant_config,
+            prefix=add_prefix("embed_tokens", prefix),
         )
         self.layers = make_layers(
             config.num_hidden_layers,
@@ -244,7 +256,9 @@ class Qwen2Model(nn.Module):
                 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)
 
@@ -325,16 +339,22 @@ class Qwen2ForCausalLM(nn.Module):
         self,
         config: Qwen2Config,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.model = Qwen2Model(config, quant_config=quant_config)
+        self.model = Qwen2Model(
+            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
+                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)
@@ -377,6 +397,8 @@ class Qwen2ForCausalLM(nn.Module):
                 # 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
 

sglang/srt/models/qwen2_5_vl.py

@@ -52,6 +52,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2 import Qwen2Model
 from sglang.srt.models.qwen2_vl import Qwen2VLImageInputs, Qwen2VLVideoInputs
+from sglang.srt.utils import add_prefix
 
 logger = logging.getLogger(__name__)
 
@@ -65,16 +66,29 @@ class Qwen2_5_VLMLP(nn.Module):
         bias: bool = True,
         hidden_act="silu",
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.gate_proj = ColumnParallelLinear(
-            in_features, hidden_features, bias=bias, quant_config=quant_config
+            in_features,
+            hidden_features,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=add_prefix("gate_proj", prefix),
         )
         self.up_proj = ColumnParallelLinear(
-            in_features, hidden_features, bias=bias, quant_config=quant_config
+            in_features,
+            hidden_features,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=add_prefix("up_proj", prefix),
         )
         self.down_proj = RowParallelLinear(
-            hidden_features, in_features, bias=bias, quant_config=quant_config
+            hidden_features,
+            in_features,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=add_prefix("down_proj", prefix),
         )
         self.act = ACT2FN[hidden_act]
 
@@ -98,6 +112,7 @@ class Qwen2_5_VisionBlock(nn.Module):
         norm_layer: Type[nn.Module] = None,
         attn_implementation: Optional[str] = "sdpa",
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         if norm_layer is None:
@@ -123,9 +138,14 @@ class Qwen2_5_VisionBlock(nn.Module):
             use_full_precision_softmax=use_full_precision_softmax,
             flatten_batch=True,
             quant_config=quant_config,
+            prefix=add_prefix("attn", prefix),
         )
         self.mlp = Qwen2_5_VLMLP(
-            dim, intermediate_dim, hidden_act=hidden_act, quant_config=quant_config
+            dim,
+            intermediate_dim,
+            hidden_act=hidden_act,
+            quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
         )
 
     def forward(
@@ -178,6 +198,7 @@ class Qwen2_5_VisionPatchMerger(nn.Module):
         context_dim: int,
         spatial_merge_size: int = 2,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.hidden_size = context_dim * (spatial_merge_size**2)
@@ -189,10 +210,15 @@ class Qwen2_5_VisionPatchMerger(nn.Module):
                     self.hidden_size,
                     bias=True,
                     quant_config=quant_config,
+                    prefix=add_prefix("mlp.0", prefix),
                 ),
                 nn.GELU(),
                 RowParallelLinear(
-                    self.hidden_size, dim, bias=True, quant_config=quant_config
+                    self.hidden_size,
+                    dim,
+                    bias=True,
+                    quant_config=quant_config,
+                    prefix=add_prefix("mlp.2", prefix),
                 ),
             ]
         )
@@ -250,6 +276,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
         vision_config: Qwen2_5_VLVisionConfig,
         norm_eps: float = 1e-6,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
 
@@ -286,8 +313,9 @@ class Qwen2_5_VisionTransformer(nn.Module):
                     norm_layer=norm_layer,
                     attn_implementation="sdpa",
                     quant_config=quant_config,
+                    prefix=add_prefix(f"blocks.{i}", prefix),
                 )
-                for _ in range(depth)
+                for i in range(depth)
             ]
         )
         self.merger = Qwen2_5_VisionPatchMerger(
@@ -295,6 +323,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
             context_dim=hidden_size,
             spatial_merge_size=spatial_merge_size,
             quant_config=quant_config,
+            prefix=add_prefix("merger", prefix),
         )
 
     def get_window_index(self, grid_thw):
@@ -447,6 +476,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
         self,
         config: Qwen2VLConfig,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
 
@@ -457,15 +487,23 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
             # NOTE: Qwen2-VL vision encoder does not support any
             # quantization method now.
             quant_config=None,
+            prefix=add_prefix("visual", prefix),
         )
 
-        self.model = Qwen2Model(config, quant_config)
+        self.model = Qwen2Model(
+            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
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("lm_head", prefix),
             )
 
         self.logits_processor = LogitsProcessor(config)

sglang/srt/models/qwen2_eagle.py

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 
+from sglang.srt.utils import add_prefix
+
 # Adapted from
 # https://github.com/SafeAILab/EAGLE/blob/main/eagle/model/cnets.py
 """Inference-only LLaMA-EAGLE model compatible with HuggingFace weights."""
@@ -42,7 +44,7 @@ class Qwen2DecoderLayer(Qwen2DecoderLayer):
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
-        super().__init__(config, layer_id, quant_config)
+        super().__init__(config, layer_id, quant_config, prefix=prefix)
 
         # Skip the input_layernorm
         # https://github.com/SafeAILab/EAGLE/blob/35c78f6cdc19a73e05cf5c330b4c358dad970c6a/eagle/model/cnets.py#L427
@@ -56,6 +58,7 @@ class Qwen2Model(nn.Module):
         self,
         config: Qwen2Config,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.config = config
@@ -63,11 +66,15 @@ class Qwen2Model(nn.Module):
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
             config.hidden_size,
+            prefix=add_prefix("embed_tokens", prefix),
         )
         self.layers = nn.ModuleList(
             [
                 Qwen2DecoderLayer(
-                    config, i, quant_config=quant_config, prefix=f"model.layers.{i}"
+                    config,
+                    i,
+                    quant_config=quant_config,
+                    prefix=add_prefix(f"layers.{i}", prefix),
                 )
                 for i in range(config.num_hidden_layers)
             ]
@@ -107,17 +114,22 @@ class Qwen2ForCausalLMEagle(Qwen2ForCausalLM):
         self,
         config: Qwen2Config,
         quant_config: Optional[QuantizationConfig] = None,
-        cache_config=None,
+        prefix: str = "",
     ) -> None:
         nn.Module.__init__(self)
         self.config = config
         self.quant_config = quant_config
-        self.model = Qwen2Model(config, quant_config=quant_config)
+        self.model = Qwen2Model(
+            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.lm_head = ParallelLMHead(
-                config.vocab_size, config.hidden_size, quant_config=quant_config
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("lm_head", prefix),
             )
         self.logits_processor = LogitsProcessor(config)