fix: handle INFONCE_HARD_NEGATIVES as integer if provided

[dlutwy]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

---

PR information

在训练 GME Embedding 时，有可能需要在环境变量中提供 INFONCE_HARD_NEGATIVES。原本的代码遗漏了将其转换为 int 类型。

In the training of GME Embedding, it may be necessary to provide INFONCE_HARD_NEGATIVES via environment variables. The original code missed converting this value to an integer.

---

Experiment results

如果没有这段修改，在 validation 完成后计算 metrics 时将会报错：

If this modification is not included, an error will occur when computing metrics after validation:

[rank0]:     output = eval_loop(
[rank0]:   File "/home/tuozhi.wy/.conda/envs/gme/lib/python3.10/site-packages/transformers/trainer.py", line 4463, in evaluation_loop
[rank0]:     metrics = self.compute_metrics(
[rank0]:   File "/home/tuozhi.wy/.conda/envs/gme/lib/python3.10/site-packages/swift/trainers/trainers.py", line 75, in calculate_metric
[rank0]:     return calculate_infonce_metrics(eval_prediction.predictions, eval_prediction.label_ids)
[rank0]:   File "/home/tuozhi.wy/.conda/envs/gme/lib/python3.10/site-packages/swift/plugin/loss.py", line 173, in calculate_infonce_metrics
[rank0]:     split_tensors = _parse_multi_negative_sentences(torch.tensor(embeddings), torch.tensor(labels), hard_negatives)
[rank0]:   File "/home/tuozhi.wy/.conda/envs/gme/lib/python3.10/site-packages/swift/plugin/loss.py", line 255, in _parse_multi_negative_sentences
[rank0]:     if negatives > hard_negatives:
[rank0]: TypeError: '>' not supported between instances of 'int' and 'str'

[dlutwy]

从 #3019 来看，我想此处可以 @tastelikefeet

[tastelikefeet]

感谢PR😁