[Feature] SigLib2 Model implementation

[khalil-Hennara]

For new models, have you tried:

from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    "microsoft/Phi-4-multimodal-instruct",
    trust_remote_code = True,
)
from transformers import AutoModelForSequenceClassification
model, tokenizer = FastModel.from_pretrained(
    auto_model = AutoModelForSequenceClassification,
)

Hi, I am working on VLM project, and I found that you've already support VLM like Gemma3, Llama4. I want to ask about vision Encoder implementation, is it worthy to work on such model, would the kernel implemented in triton like fast_layernorm, fast_linear_forward, provide a post in the model performance.

I've try to implement something like the model in the system gemma, mistral, etc. but I found it a slightly confusing, as for Language model you've build a FastLlamaModel that every other model inherent forms. I found that re-implement the SigLib model using your custom kernel but implemented as classes instead of your implementation style by just write the forward function which is quite amazing by the way easier for me.

My Questions are:

First do you think this model would benefit from your custom kernels if it's implemented within unsloth.

Second does this model might be needed within the framework, so I will keep working on it following your guide or it would better to just implemented for me.

of course the second question related to the first so if the first one is no, the second will be defiantly no.

Thanks in advance.