add missing batch_size consideration

[xin3he]

Batch_size is not considered before, quantizing llama3.3 70b will only use 1 card and got OOM.

[wenhuach21]

the solution is incorrect

[xin3he]

Okay, I can share the reproduce step to you so that we can find out a correct solution. @wenhuach21
auto-round --model /models/Llama-3.3-70B-Instruct/ --scheme "MXFP4" --device_map 4,5,6

[xin3he]

BTW, only considering the weight size is not correct, for example, the large in_features of down_proj requires more memory to hold the activation.

(mlp): LlamaMLP(
          (gate_proj): Linear(in_features=8192, out_features=28672, bias=False)
          (up_proj): Linear(in_features=8192, out_features=28672, bias=False)
          (down_proj): Linear(in_features=28672, out_features=8192, bias=False)