[Bug] group size=64的awq量化精度损失明显

### Checklist

- [ ] 1. I have searched related issues but cannot get the expected help.
- [ ] 2. The bug has not been fixed in the latest version.

### Describe the bug

通过修改源码gemm_s4_f16.cu中的硬编码
' /// TODO: add more group sizes
Generate<128, Ops>(kernels_);
group_sizes_.push_back(128);'
为
‘ /// TODO: add more group sizes
Generate<64, Ops>(kernels_);
group_sizes_.push_back(64);’
并使用group size=64进行量化。发现模型的精度内眼可见的差。
请问除了修改这些，我还需要做些什么来适配group size为64的awq量化呢

### Reproduction

‘ /// TODO: add more group sizes
Generate<64, Ops>(kernels_);
group_sizes_.push_back(64);’

### Environment

```Shell
lmdeploy v0.1.0
```

### Error traceback

_No response_