Wang Zhihao

Results 3 comments of


                                            Wang Zhihao

MaskedLM nan training loss

I've encountered this problem. Is there a solution available?

qwen2.5VL 7B模型推理，prompt和图片输入一致，hf推理结果与vllm推理结果很不一致。输出差异比较大。

我也看到了这个问题，但是我发现vllm效果更好一点，差了7、8个点了，你们有知道怎么解决吗

Are there any powerful llm knowledge distillation tools/frameworks currently available?

Do you find it?