Support Alibaba-NLP/gte-Qwen2-7B-instruct embedding Model (#1186)

Co-authored-by: Ying Sheng <sqy1415@gmail.com>

python/sglang/srt/models/qwen2.py

@@ -38,6 +38,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from sglang.srt.layers.activation import SiluAndMul
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.pooler import EmbeddingPoolerOutput, Pooler, PoolingType
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.model_executor.forward_batch_info import InputMetadata
 
@@ -275,6 +276,7 @@ class Qwen2ForCausalLM(nn.Module):
         self.model = Qwen2Model(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(
@@ -283,11 +285,15 @@ class Qwen2ForCausalLM(nn.Module):
         positions: torch.Tensor,
         input_metadata: InputMetadata,
         input_embeds: torch.Tensor = None,
+        get_embedding: bool = False,
     ) -> torch.Tensor:
         hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
-        return self.logits_processor(
-            input_ids, hidden_states, self.lm_head.weight, input_metadata
-        )
+        if not get_embedding:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head.weight, input_metadata
+            )
+        else:
+            return self.pooler(hidden_states, input_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [