Update benchmark scripts (#8)

benchmark/react/README.md

@@ -1,5 +1,7 @@
 ## Run benchmark
 
+NOTE: This is an implementation for replaying a given trace for throughput/latency benchmark purposes. It is not an actual ReAct agent implementation.
+
 ### Benchmark sglang
 ```
 python -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000

benchmark/react/bench_other.py

@@ -124,6 +124,9 @@ def main(args):
             ))
             return out["result"]
 
+        # warmup
+        call_generate("Hello,", 1.0, 8, ".")
+
     else:
         raise ValueError(f"Invalid backend: {args.backend}")
 

benchmark/react/bench_sglang.py

@@ -82,9 +82,10 @@ Action 3: Finish[yes]
 """ + question)
     for i in range(1, len(triplets) + 2):
         s += "Thought " + str(i) + ":"
+        # NOTE: This is an implementation for replaying a given trace for benchmark purposes. It is not an actual ReAct agent implementation.
         ss = s.fork(1)
         ss[0] += sgl.gen(name="thought_action", max_tokens=200, stop="Observation")
-        # ss.join()
+        ss.join()
         # to verify the correctness of output, this should be collected
         # print(ss[0]["thought_action"])
         if i > len(triplets):