Runner changes for TorchTune Llama3.2 vision text decoder

[jackzhxng]

Summary

Changes to eager (Python) and native (ET) runners to run TorchTune's llama3_2_vision text decoder without KV cache (KV cache in progress). Should extend to the regular TorchTune llama3_2 model as well, will add support in following PRs.

Native runner relies on #6670 to get in.

PR chain:

• Add kwarg example inputs to eager model base
• Llama2 model cleanup
• Accept model type parameter in export_llama
• Export TorchTune llama3_2_vision in ET
• YOU ARE HERE ~> Runner changes for TorchTune Llama3.2 vision text decoder

Test plan

Download the model from torchtune: tune download meta-llama/Llama-3.2-11B-Vision-Instruct --output-dir /tmp/Llama-3.2-11B-Vision-Instruct.

Run eager:

python -m examples.models.llama.runner.eager --model llama3_2_vision --checkpoint /tmp/Llama-3.2-11B-Vision-Instruct/original/consolidated.pth  --params examples/models/llama3_2_vision/text_decoder/params/demo_config.json --metadata '{"append_eos_to_prompt": 0, "get_bos_id":128000, "get_eos_ids":[128009, 128001], "get_n_bos": 0, "get_n_eos": 0}' --output_name="llama3_2_vision.pte" --tokenizer /tmp/Llama-3.2-11B-Vision-Instruct/original/tokenizer.model -d fp32 --verbose --prompt "What is the capital of USA?" --max_seq_length 32

Run executorch on portable lib (doesn't work until #6670 gets in, can test for now by adding pytorch/pytorch#137662 following to PyTorch installation):

# Export model to executorch.
python -m examples.models.llama.export_llama --model llama3_2_vision --checkpoint /tmp/Llama-3.2-11B-Vision-Instruct/original/consolidated.pth  --params examples/models/llama3_2_vision/text_decoder/params/demo_config.json -d fp32 --metadata '{"append_eos_to_prompt": 0, "get_bos_id":128000, "get_eos_ids":[128009, 128001], "get_n_bos": 0, "get_n_eos": 0}' --output_name="llama3_2_vision.pte"

# Run using native runner.
python -m examples.models.llama.runner.native --model llama3_2_vision --pte llama3_2_vision.pte  --tokenizer /tmp/Llama-3.2-11B-Vision-Instruct/original/tokenizer.model --prompt "How many calories are in bread?" --params examples/models/llama3_2_vision/text_decoder/params/demo_config.json --max_len 64

[tarun292]

I think we'd want the default to still be True?

[jackzhxng]

Oh yeah this was weird since "store_true" works by having a default of False, but the default is here set to True so it's just always True regardless of what you put