adapt to sglang v0.5.2rc1 on dcu

benchmark/mtbench/bench_sglang.py

@@ -0,0 +1,99 @@
+import argparse
+import json
+import os
+import time
+import uuid
+
+import sglang as sgl
+from sglang.test.test_utils import (
+    add_common_sglang_args_and_parse,
+    select_sglang_backend,
+)
+
+
+def load_questions(filename):
+    questions = []
+    with open(filename, "r") as fin:
+        for line in fin:
+            obj = json.loads(line)
+            questions.append(obj)
+    return questions
+
+
+def write_answers(filename, model_id, questions, answers):
+    with open(os.path.expanduser(filename), "w") as fout:
+        for i in range(len(answers)):
+            ans_json = {
+                "question_id": questions[i]["question_id"],
+                "answer_id": uuid.uuid4().hex,
+                "model_id": model_id,
+                "choices": {
+                    "index": 0,
+                    "turns": [answers[i][0], answers[i][1]],
+                },
+                "tstamp": time.time(),
+            }
+            fout.write(json.dumps(ans_json) + "\n")
+
+
+@sgl.function
+def answer_mt_bench(s, question_1, question_2):
+    s += sgl.system()
+    s += sgl.user(question_1)
+    s += sgl.assistant(sgl.gen("answer_1"))
+    s += sgl.user(question_2)
+    s += sgl.assistant(sgl.gen("answer_2"))
+
+
+def main(args):
+    # Construct prompts
+    questions = load_questions(args.question_file)[: args.num_questions]
+    arguments = [
+        {"question_1": q["turns"][0], "question_2": q["turns"][1]} for q in questions
+    ]
+
+    # Select backend
+    backend = select_sglang_backend(args)
+    sgl.set_default_backend(backend)
+
+    # Run requests
+    tic = time.perf_counter()
+    rets = answer_mt_bench.run_batch(
+        arguments,
+        temperature=0,
+        max_new_tokens=256,
+        num_threads=args.parallel,
+        progress_bar=True,
+    )
+    answers = [[s["answer_1"], s["answer_2"]] for s in rets]
+    latency = time.perf_counter() - tic
+
+    print(f"#questions: {len(questions)}, Latency: {latency:.2f}")
+
+    # Write results
+    model_id = backend.model_info["model_path"]
+    answer_file = args.answer_file or f"tmp_output_{args.backend}.txt"
+    write_answers(answer_file, model_id, questions, answers)
+
+    with open(args.result_file, "a") as fout:
+        value = {
+            "task": "mtbench",
+            "backend": args.backend,
+            "num_gpus": 1,
+            "latency": round(latency, 3),
+            "num_requests": args.num_questions,
+            "other": {
+                "num_questions": args.num_questions,
+                "parallel": args.parallel,
+            },
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--question-file", type=str, default="question.jsonl")
+    parser.add_argument("--answer-file", type=str, default=None)
+    parser.add_argument("--num-questions", type=int, default=80)
+    args = add_common_sglang_args_and_parse(parser)
+    main(args)