Add Idefics3/SmolVLM quant support via traceable class

[leon-seidel]

SUMMARY:
Adding a traceable Idefics3 class following the new guide to allow W4A16 quants of Idefics3 and SmolVLM (which share the same architecture). Idefics3 seems to require a max_sequence_length of 4096 and I copied the example from the Phi 3 Vision example as the dataset loading approach from the Llava example led to OOM on 64 GB RAM.

TEST PLAN:
Tested on A100 with Idefics3 @512 samples and on a 4060 Ti with SmolVLM @128 samples.

[kylesayrs]

Thank you for your contribution @leon-seidel! This looks great to me, I look forward to getting this landed!

[kylesayrs]

Hi @leon-seidel

I've run your example end to end and it looks good! All that's left is to fix the quality tests and this is good to land!

pip install -e ./[dev]
make style
make quality

[kylesayrs]

Base

hf-multimodal (pretrained=HuggingFaceM4/Idefics3-8B-Llama3,dtype=bfloat16,add_bos_token=True,convert_img_format=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 1
|     Tasks      |Version|Filter|n-shot|Metric|   |Value |   |Stderr|
|----------------|------:|------|-----:|------|---|-----:|---|-----:|
|Computer Science|      0|none  |     0|acc   |↑  |0.3333|±  |0.0875|

W4A16

hf-multimodal (pretrained=Idefics3-8B-Llama3-W4A16-G128,dtype=bfloat16,add_bos_token=True,convert_img_format=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 1
|     Tasks      |Version|Filter|n-shot|Metric|   |Value|   |Stderr|
|----------------|------:|------|-----:|------|---|----:|---|-----:|
|Computer Science|      0|none  |     0|acc   |↑  |  0.3|±  |0.0851|

[leon-seidel]

Quality should be fixed now, thanks for your help! Also great tutorial on making the models traceable in the first place. I tried it on my own finetuned Idefics3 model and can't see any deterioration in the outputs!