Add exported program runner (#6969)

Summary:
Add an `ExportedProgram` runner for TorchTune Llama.


Test Plan:
```
# Download resources
tune download meta-llama/Llama-3.2-11B-Vision-Instruct --output-dir /tmp/Llama-3.2-11B-Vision-Instruct

# Export model
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  --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.pt2" -d fp32 --verbose --max_seq_length 64 --export_only -kv

# Run ExportedProgram
python -m examples.models.llama3_2_vision.runner.exported --model llama3_2_vision --pt2 llama3_2_vision.pt2  --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_seq_length 64 -kv
```

Output:
```
The number of calories in bread can vary greatly depending on the type of bread, its ingredients, and its size. Here are the approximate calorie counts for different types of bread:                                                                                                        
White bread: 80-100 calories per slice
```

Differential Revision: D66186052

Pulled By: dvorjackz

Author: jackzhxng

examples/models/llama3_2_vision/runner/exported.py

@@ -0,0 +1,94 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import argparse
+import json
+from typing import Optional
+
+import torch
+
+from executorch.examples.models.llama.export_llama_lib import (
+    build_args_parser as _build_args_parser,
+)
+from executorch.examples.models.llama3_2_vision.runner.generation import (
+    TorchTuneLlamaRunner,
+)
+
+
+class ExportedLlamaRunner(TorchTuneLlamaRunner):
+    """
+    Runs a torch-exported .pt2 Llama.
+    """
+
+    def __init__(self, args):
+        with open(args.params, "r") as f:
+            params = json.loads(f.read())
+        super().__init__(
+            tokenizer_path=args.tokenizer_path,
+            max_seq_len=args.max_seq_length,
+            max_batch_size=1,
+            use_kv_cache=args.use_kv_cache,
+            vocab_size=params["vocab_size"],
+            device="cuda" if torch.cuda.is_available() else "cpu",
+        )
+        print(f"Loading model from {args.pt2}")
+        self.model = torch.export.load(args.pt2).module()
+        print("Model loaded")
+
+    def forward(
+        self,
+        tokens: Optional[torch.LongTensor] = None,
+        input_pos: Optional[torch.LongTensor] = None,
+        mask: Optional[torch.LongTensor] = None,
+    ) -> torch.Tensor:
+        if self.use_kv_cache:
+            return self.model(tokens, input_pos=input_pos, mask=mask)
+        else:
+            return self.model(tokens)
+
+
+def build_args_parser() -> argparse.ArgumentParser:
+    parser = _build_args_parser()
+
+    parser.add_argument(
+        "--prompt",
+        type=str,
+        default="Hello",
+    )
+
+    parser.add_argument(
+        "--pt2",
+        type=str,
+        required=True,
+    )
+
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=0,
+    )
+
+    return parser
+
+
+def main() -> None:
+    parser = build_args_parser()
+    args = parser.parse_args()
+
+    runner = ExportedLlamaRunner(args)
+    result = runner.text_completion(
+        prompt=args.prompt,
+        temperature=args.temperature,
+    )
+    print(
+        "Response: \n{response}\n Tokens:\n {tokens}".format(
+            response=result["generation"], tokens=result["tokens"]
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()  # pragma: no cover