Fix Cuda out of memory issue for eager runner (pytorch#6935)

Pull Request resolved: pytorch#6866

This PR updates the eager runner to disable grad and save memory usage.

It also update the prompt format to not include bos.
ghstack-source-id: 254139542

Differential Revision: [D65962743](https://our.internmc.facebook.com/intern/diff/D65962743/)

Co-authored-by: Lunwen He <[email protected]>

Author: pytorchbot

examples/models/llama/runner/eager.py

@@ -80,18 +80,20 @@ def build_args_parser() -> argparse.ArgumentParser:
 def execute_runner(runner_class: Type[LlamaRunner]) -> None:
     parser = build_args_parser()
     args = parser.parse_args()
-    runner = runner_class(args)  # pyre-ignore: Missing argument [20]
-    generated_tokens = (
-        runner.chat_completion(temperature=args.temperature)
-        if args.chat
-        else runner.text_completion(
-            prompt=args.prompt,
-            temperature=args.temperature,
-            echo=True,
+
+    with torch.no_grad():
+        runner = runner_class(args)  # pyre-ignore: Missing argument [20]
+        generated_tokens = (
+            runner.chat_completion(temperature=args.temperature)
+            if args.chat
+            else runner.text_completion(
+                prompt=args.prompt,
+                temperature=args.temperature,
+                echo=True,
+            )
         )
-    )
-    if args.show_tokens:
-        print(f"Generated {len(generated_tokens)} tokens: {generated_tokens}")
+        if args.show_tokens:
+            print(f"Generated {len(generated_tokens)} tokens: {generated_tokens}")
 
 
 def main() -> None:

examples/models/llama/runner/generation.py

@@ -199,15 +199,14 @@ def chat_completion(
                 temperature=temperature,
                 top_p=top_p,
                 echo=True,
-                pos_base=len(tokens),
+                pos_base=len(tokens) - 1 if len(tokens) > 0 else 0,
             )
             tokens.extend(new_tokens)
             prompt = input("Me: ")
         return tokens
 
     def _format_prompt(self, prompt: str) -> str:
-        return f"""
-<|begin_of_text|><|start_header_id|>system<|end_header_id|>
+        return f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>
 
 You are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>