Fix TypeError when loading model in MoE router without bitsandbytes

[umran666]

I was trying to run mergekit-moe to merge some Qwen models on my CPU. Even though I wasn't using any 4-bit or 8-bit quantization flags, the merge kept crashing with:
TypeError: Qwen2ForCausalLM.__init__() got an unexpected keyword argument 'load_in_4bit'

It turns out that mergekit was passing load_in_4bit=False and load_in_8bit=False by default to the Hugging Face from_pretrained method. In newer versions of transformers, passing these options when you don't have a CUDA GPU or bitsandbytes configured causes a crash because the library doesn't know what to do with them.

Instead of always passing load_in_4bit and load_in_8bit as False, this change constructs the arguments dynamically. They are only sent to from_pretrained if they are set to True.
This keeps CPU merges (and any merge not using quantization) running smoothly without crashing!

---

Note

Low Risk
Narrow change to optional kwargs for model loading; behavior unchanged when 4/8-bit flags are explicitly enabled.

Overview
Fixes CPU / non-quantized MoE routing by only passing load_in_4bit and load_in_8bit to AutoModelForCausalLM.from_pretrained when those flags are True, instead of always forwarding them as False.

In get_gate_params (hidden / hidden_avg / hidden_last modes), quantization options are collected in a kwargs dict and spread into from_pretrained, so models like Qwen2 no longer receive unsupported keyword arguments on CPU or without bitsandbytes.

^{Reviewed by Cursor Bugbot for commit 07bae36. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[umran666]

I have read the CLA Document and I hereby sign the CLA

[cg123]

Thanks for the PR!