from sentence_transformers import CrossEncoder

# 原始HF格式的Qwen3-Reranker路径
src_model = "/PATH/models/Qwen/Qwen3-Reranker-4B-HF"

# 目标保存路径（sentence-transformers结构）
dst_model = "/PATH/models/Qwen/Qwen3-Reranker-4B"

# 加载HuggingFace格式模型
print(f"Loading HF model from {src_model} ...")
model = CrossEncoder(src_model)

# 保存为sentence-transformers结构
print(f"Saving as sentence-transformers CrossEncoder to {dst_model} ...")
model.save(dst_model)

print("✅ Done! You can now mount this folder to TEI and call /rerank")

from transformers import AutoTokenizer
import sys
import os

# 默认模型路径
default_model_path = "/PATH/models/Qwen/Qwen3-Reranker-4B"

# 获取模型路径，支持命令行传参，没传就用默认路径
if len(sys.argv) == 2:
    model_path = sys.argv[1]
else:
    model_path = default_model_path
    print(f"未传入模型目录，使用默认路径: {model_path}")

if not os.path.isdir(model_path):
    print(f"错误: 模型目录不存在: {model_path}")
    sys.exit(1)

print(f"正在修复模型目录: {model_path}")
tokenizer = AutoTokenizer.from_pretrained(model_path)

# 如果没有 pad_token，则用 eos_token 作为 pad
if tokenizer.pad_token is None:
    print(f"未检测到 pad_token，使用 eos_token 作为 pad_token: {tokenizer.eos_token}")
    tokenizer.pad_token = tokenizer.eos_token
else:
    print(f"已存在 pad_token: {tokenizer.pad_token}")

# 如果 pad_token_id 为空，则设置为 eos_token_id
if tokenizer.pad_token_id is None:
    tokenizer.pad_token_id = tokenizer.eos_token_id
    print(f"设置 pad_token_id = eos_token_id: {tokenizer.pad_token_id}")
else:
    print(f"已存在 pad_token_id: {tokenizer.pad_token_id}")

# 保存修改
tokenizer.save_pretrained(model_path)
print("✅ 修复完成，已保存到模型目录。")

{
  "pooling_mode_cls_token": false,
  "pooling_mode_mean_tokens": false,
  "pooling_mode_max_tokens": false,
  "pooling_mode_mean_sqrt_len_tokens": false,
  "pooling_mode_weightedmean_tokens": false,
  "pooling_mode_lasttoken": true,
  "word_embedding_dimension": 2560,
  "max_seq_length": 5120
}

{
  "architectures": [
    "Qwen3ForSequenceClassification"
  ],
  "attention_bias": false,
  "attention_dropout": 0.0,
  "bos_token_id": 151643,
  "eos_token_id": 151645,
  "pad_token_id": 151643,
  "head_dim": 128,
  "hidden_act": "silu",
  "hidden_size": 2560,
  "id2label": {
    "0": "LABEL_0"
  },
  "initializer_range": 0.02,
  "intermediate_size": 9728,
  "label2id": {
    "LABEL_0": 0
  },
  "layer_types": [
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention"
  ],
  "max_position_embeddings": 40960,
  "max_window_layers": 36,
  "model_type": "qwen3",
  "num_attention_heads": 32,
  "num_hidden_layers": 36,
  "num_key_value_heads": 8,
  "rms_norm_eps": 1e-06,
  "rope_scaling": null,
  "rope_theta": 1000000,
  "sentence_transformers": {
    "activation_fn": "torch.nn.modules.activation.Sigmoid",
    "version": "5.1.0"
  },
  "sliding_window": null,
  "tie_word_embeddings": true,
  "torch_dtype": "float32",
  "transformers_version": "4.53.3",
  "use_cache": true,
  "use_sliding_window": false,
  "vocab_size": 151669
}