sglang 0.3.4.post2__py3-none-any.whl → 0.3.5__py3-none-any.whl

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

sglang/srt/models/qwen2_moe.py

@@ -29,10 +29,6 @@ from vllm.distributed import (
 )
 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 (
-    ParallelLMHead,
-    VocabParallelEmbedding,
-)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.activation import SiluAndMul
@@ -47,6 +43,10 @@ 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
 

sglang/srt/models/qwen2_vl.py

@@ -23,7 +23,7 @@
 # limitations under the License.
 """Inference-only Qwen2-VL model compatible with HuggingFace weights."""
 from functools import lru_cache, partial
-from typing import Iterable, List, Mapping, Optional, Tuple, Type, TypedDict, Union
+from typing import Iterable, List, Optional, Tuple, Type, TypedDict
 
 import numpy as np
 import torch
@@ -35,9 +35,7 @@ from vllm.distributed import parallel_state
 from vllm.distributed import utils as dist_utils
 from vllm.logger import init_logger
 from vllm.model_executor.layers.activation import QuickGELU
-from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.model_executor.models.interfaces import SupportsMultiModal
 
 from sglang.srt.configs import Qwen2VLConfig, Qwen2VLVisionConfig
 from sglang.srt.hf_transformers_utils import get_processor
@@ -47,6 +45,7 @@ from sglang.srt.layers.attention.triton_ops.prefill_attention import (
 from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
 from sglang.srt.managers.schedule_batch import ImageInputs
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.models.qwen2 import Qwen2Model
@@ -486,7 +485,7 @@ class Qwen2VisionTransformer(nn.Module):
 cached_get_processor = lru_cache(get_processor)
 
 
-class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal):
+class Qwen2VLForConditionalGeneration(nn.Module):
     def calculate_num_image_tokens(self, image_grid_thw: Tuple[int, int, int]):
         processor = cached_get_processor(self.config._name_or_path)
         grid_t, grid_h, grid_w = image_grid_thw
@@ -536,15 +535,12 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal):
     def __init__(
         self,
         config: Qwen2VLConfig,
-        multimodal_config: MultiModalConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
 
         self.config = config
-        self.multimodal_config = multimodal_config
-
         self.visual = Qwen2VisionTransformer(
             config.vision_config,
             norm_eps=getattr(config, "rms_norm_eps", 1e-6),
@@ -622,7 +618,7 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal):
             extend_start_loc_cpu = forward_batch.extend_start_loc.cpu().numpy()
             prefix_lens_cpu = forward_batch.extend_prefix_lens.cpu().numpy()
             for i, image in enumerate(forward_batch.image_inputs):
-                if image == None:
+                if image is None:
                     continue
                 start_idx = extend_start_loc_cpu[i]
                 prefix_len = prefix_lens_cpu[i]

sglang/srt/models/stablelm.py

@@ -24,10 +24,6 @@ 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/torch_native_llama.py

@@ -26,10 +26,6 @@ from torch.nn.parameter import Parameter
 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
@@ -38,6 +34,10 @@ from sglang.srt.layers.logits_processor import LogitsProcessor, LogitsProcessorO
 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
 

sglang/srt/models/xverse.py

@@ -31,15 +31,15 @@ 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.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.model_runner import ForwardBatch
 
 

sglang/srt/models/xverse_moe.py

@@ -34,15 +34,15 @@ 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.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