llama3.2 + cross attn test

[maleksan85]

Commands:
server

 VLLM_NO_TUNED_GEMM=1 vllm serve /data/models/Llama-3.2-90B-Vision-Instruct --tensor_parallel_size 2 --enforce-eager --limit-mm-per-prompt "image=2" --max-num-seqs 32 --max_model_len 8192

client (from https://huggingface.co/nltpt/VLLM-llama3.2):

root@banff-cyxtera-s82-5:~/workspace/VLLM-llama3.2# python openai_vision_api_client.py
Chat completion output: The image depicts a serene lake scene with a wooden dock extending into the water, surrounded by lush greenery and majestic mountains in the background. The overall atmosphere of the image exudes tranquility and natural beauty, inviting the viewer to step into its peaceful world.
remove me: testing done, exitting...

[shajrawi]

Great work!!
Few nits + one question.
Also, can you please make the linter happy? :)

[shajrawi]

Looks good - assuming performance does not regress due to reshape