[Fix] Reduce memory usage for loading llava model & Remove EntryClassRemapping (#1308)

python/sglang/srt/models/llama_embedding.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Tuple
+from typing import Iterable, Tuple
 
 import torch
 from torch import nn
@@ -7,7 +7,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from sglang.srt.layers.pooler import EmbeddingPoolerOutput, Pooler, PoolingType
 from sglang.srt.model_executor.model_runner import InputMetadata
-from sglang.srt.models.llama2 import LlamaForCausalLM, LlamaModel
+from sglang.srt.models.llama import LlamaModel
 
 
 class LlamaEmbeddingModel(nn.Module):
@@ -16,7 +16,6 @@ class LlamaEmbeddingModel(nn.Module):
         config: LlamaConfig,
         quant_config=None,
         cache_config=None,
-        efficient_weight_load=False,
     ) -> None:
         super().__init__()
         self.model = LlamaModel(config, quant_config=quant_config)
@@ -86,6 +85,8 @@ class LlamaEmbeddingModel(nn.Module):
             load_weights_per_param(name, loaded_weight)
 
 
-EntryClass = LlamaEmbeddingModel
-# compat: e5-mistral model.config class == MistralModel
-EntryClassRemapping = [("MistralModel", LlamaEmbeddingModel)]
+class MistralModel(LlamaEmbeddingModel):
+    pass
+
+
+EntryClass = [LlamaEmbeddingModel, MistralModel]