GME模型分别用vLLM部署和swift部署（使用vllm backend ）的服务，返回的Emb结果不一致

GPU：H20
vLLM版本：0.10.1

swift部署的Emb

• swift 启动部署的命令

swift deploy --model Alibaba-NLP/gme-Qwen2-VL-2B-Instruct --served_model_name gme_emb --vllm_max_model_len 2048 --port 8092 --host 0.0.0.0 --infer_backend vllm --torch_dtype float16 --task_type embedding --vllm_max_num_seqs 256 --max_length 2048 --vllm_gpu_memory_utilization 0.9

• 服务测试Curl

curl http://localhost:8092/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
      "model": "gme_emb",
      "input": "苹果手机"
  }'

vLLM部署的Emb

• vLLM启动部署的命令

vllm serve Alibaba-NLP/gme-Qwen2-VL-2B-Instruct --served-model-name gme_emb --port 9092 --max-model-len 2048 --host 0.0.0.0 --dtype float16 --trust-remote-code --task embedding --enable-prefix-caching --gpu-memory-utilization 0.9 --seed 42

• 服务测试Curl

curl http://localhost:9092/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
      "model": "gme_emb",
      "input": "苹果手机"
  }'

结果差异

• swift部署服务返回Emb：[3.931640625,-1.197265625,1.5869140625,-3.466796875,7.328125,-4.58984375,-1.0654296875, …]
• vllm部署服务返回Emb：[-0.0229949951171875,-0.0164642333984375,0.042449951171875,0.0245513916015625,-0.017364501953125, …]
• 两个返回向量cos相似度仅为0.8左右，明显不是归一化的原因【看前两个元素正负符号对不上】
• 且有个细节，vLLM部署的服务返回usage信息中，"prompt_tokens":2； swift部署服务返回usage信息中，"prompt_tokens":22；

Reproduce the two deployments using the documented swift deploy and vLLM serve commands, then send the same /v1/embeddings request to both endpoints. Compare the returned vectors, cosine similarity, and prompt_tokens values, focusing on the differing token counts and embedding scale. Done means the cause of the discrepancy is identified and the services produce consistent embedding results.