[Bugfix] Generate exactly input_len tokens in benchmark_throughput (#9592)

benchmarks/benchmark_throughput.py

@@ -233,7 +233,16 @@ def main(args: argparse.Namespace):
         args.tokenizer, trust_remote_code=args.trust_remote_code)
     if args.dataset is None:
         # Synthesize a prompt with the given input length.
-        prompt = "hi" * (args.input_len - 1)
+        # As tokenizer may add additional tokens like BOS, we need to try
+        # different lengths to get the desired input length.
+        for i in range(-10, 10):
+            prompt = "hi " * (args.input_len + i)
+            tokenized_prompt = tokenizer(prompt).input_ids
+            if len(tokenized_prompt) == args.input_len:
+                break
+        else:
+            raise ValueError(
+                f"Failed to synthesize a prompt with {args.input_len} tokens.")
         requests = [(prompt, args.input_len, args.output_len)
                     for _ in range(args.num_prompts)]
     else: