[misc] Unify benchmark deprecation shims and one_batch_server CLI entrypoint (#29108)

python/sglang/bench_one_batch_server.py

@@ -1,13 +1,15 @@
-"""Back-compat shim. The implementation now lives in
+"""Deprecated import path for ``sglang.benchmark.one_batch_server``.
-``sglang.benchmark.one_batch_server``; this module preserves the
+
-``python -m sglang.bench_one_batch_server`` entry point and the
+``python -m sglang.bench_one_batch_server`` and
-``from sglang.bench_one_batch_server import ...`` imports.
+``from sglang.bench_one_batch_server import ...`` still work, but the
+implementation now lives in ``sglang.benchmark.one_batch_server``.
+Update references to the new path.
 """
 
 import warnings
 
 from sglang.benchmark.one_batch_server import *  # noqa: F401,F403
-from sglang.benchmark.one_batch_server import main
+from sglang.benchmark.one_batch_server import cli_main
 
 warnings.warn(
     "`sglang.bench_one_batch_server` is deprecated and will be removed in a "
@@ -18,4 +20,4 @@ warnings.warn(
 )
 
 if __name__ == "__main__":
-    main()
+    cli_main()

python/sglang/bench_serving.py

@@ -1,20 +1,22 @@
-# SPDX-License-Identifier: Apache-2.0
+"""Deprecated import path for ``sglang.benchmark.serving``.
-"""Compatibility shim for the relocated serving benchmark entrypoint."""
+
+``python -m sglang.bench_serving`` and ``from sglang.bench_serving import ...``
+still work, but the implementation now lives in ``sglang.benchmark.serving``.
+Update references to the new path.
+"""
 
 import warnings
 
-from sglang.benchmark.serving import *  # noqa: F403
+from sglang.benchmark.serving import *  # noqa: F401,F403
-from sglang.benchmark.serving import (  # noqa: F401
+from sglang.benchmark.serving import cli_main
-    _create_bench_client_session,
-    cli_main,
-)
 
 warnings.warn(
-    "sglang.bench_serving is deprecated; use sglang.benchmark.serving instead.",
+    "`sglang.bench_serving` is deprecated and will be removed in a future "
+    "release; use `sglang.benchmark.serving` instead "
+    "(e.g. `python -m sglang.benchmark.serving`).",
     FutureWarning,
     stacklevel=1,
 )
 
-
 if __name__ == "__main__":
     cli_main()

python/sglang/benchmark/datasets/autobench.py

@@ -173,7 +173,7 @@ def _estimate_prompt_lens(
         prompt_len = len(prompt)
         return prompt_len, prompt_len, 0
 
-    # Multi-turn prompt lists are handled specially by bench_serving and do not
+    # Multi-turn prompt lists are handled specially by the serving benchmark and do not
     # contribute reliable static prompt lengths.
     return 0, 0, 0
 

python/sglang/benchmark/datasets/generated_shared_prefix.py

@@ -60,7 +60,7 @@ class GeneratedSharedPrefixDataset(BaseDataset):
 
         # Defensive validation for in-process callers that construct a
         # Namespace by hand and bypass the argparse boundary in
-        # bench_serving.py. The CLI hook enforces the same rules first.
+        # serving.py. The CLI hook enforces the same rules first.
         if group_distribution not in ("uniform", "zipf"):
             raise ValueError(
                 f"--gsp-group-distribution must be 'uniform' or 'zipf', "

python/sglang/benchmark/datasets/openai_dataset.py

@@ -97,7 +97,7 @@ def sample_openai_requests(
             tools_tokens = len(tokenizer.encode(tools_str))
             prompt_len += tools_tokens
 
-        # Pass messages list directly - bench_serving handles List[Dict] prompts
+        # Pass messages list directly - the serving benchmark handles List[Dict] prompts
         filtered_dataset.append(
             DatasetRow(
                 prompt=messages,

python/sglang/benchmark/datasets/speed_bench.py

@@ -1,4 +1,4 @@
-"""SPEED-Bench (nvidia/SPEED-Bench) dataset for SGLang bench_serving.
+"""SPEED-Bench (nvidia/SPEED-Bench) dataset for the SGLang serving benchmark.
 
 Reads the pre-downloaded throughput_1k JSONL produced by prepare_speed_bench.sh
 (or equivalent), optionally filtering by category (low_entropy / mixed /

python/sglang/benchmark/offline_throughput.py

@@ -1,13 +1,13 @@
 """
 Benchmark the throughput in the offline mode.
-It accepts server arguments (the same as launch_server.py) and benchmark arguments (the same as bench_serving.py).
+It accepts server arguments (the same as launch_server.py) and benchmark arguments (the same as serving.py).
 
 # Usage
 ## Sharegpt dataset with default args
-python -m sglang.bench_offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --num-prompts 10
+python -m sglang.benchmark.offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --num-prompts 10
 
 ## Random dataset with default args
-python -m sglang.bench_offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --dataset-name random --random-input 1024 --random-output 1024
+python -m sglang.benchmark.offline_throughput --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --dataset-name random --random-input 1024 --random-output 1024
 """
 
 import argparse

python/sglang/benchmark/one_batch.py

@@ -6,18 +6,18 @@ It accepts server arguments (the same as launch_server.py) and benchmark argumen
 
 # Usage (latency test)
 ## with dummy weights:
-python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --load-format dummy
+python -m sglang.benchmark.one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --load-format dummy
 ## sweep through multiple data points and store (append) the results in a jsonl file:
-python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 12 14 --input-len 256 512 --output-len 32 256 --run-name test_run
+python -m sglang.benchmark.one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 12 14 --input-len 256 512 --output-len 32 256 --run-name test_run
 ## run with profiling:
-python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 12 14 --input-len 256 512 --profile
+python -m sglang.benchmark.one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 12 14 --input-len 256 512 --profile
 ## run with profiling to custom directory:
 export SGLANG_TORCH_PROFILER_DIR=/root/sglang/profile_log
-python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 --input-len 256 --profile
+python -m sglang.benchmark.one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 --input-len 256 --profile
 ## run with CUDA profiler (nsys):
-nsys profile --force-overwrite=true -o bench_one_batch python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 --input-len 256 --profile --profile-activities CUDA_PROFILER
+nsys profile --force-overwrite=true -o bench_one_batch python -m sglang.benchmark.one_batch --model-path meta-llama/Meta-Llama-3-8B-Instruct --batch-size 1 --input-len 256 --profile --profile-activities CUDA_PROFILER
 # Usage (correctness test):
-python -m sglang.bench_one_batch --model-path TinyLlama/TinyLlama-1.1B-Chat-v0.4 --correct
+python -m sglang.benchmark.one_batch --model-path TinyLlama/TinyLlama-1.1B-Chat-v0.4 --correct
 
 ## Reference output (of the correctness test above, can be gpu dependent):
 input_ids=[[1, 450, 7483, 310, 3444, 338], [1, 450, 7483, 310, 278, 3303, 13187, 290, 338], [1, 20628, 338, 263, 6575, 1460, 2462, 322, 306, 763]]

python/sglang/benchmark/one_batch_server.py

@@ -5,11 +5,11 @@ This script launches a server and uses the HTTP interface.
 It accepts server arguments (the same as launch_server.py) and benchmark arguments (e.g., batch size, input lengths).
 
 Usage:
-python3 -m sglang.bench_one_batch_server --model meta-llama/Meta-Llama-3.1-8B --batch-size 1 16 64 --input-len 1024 --output-len 8
+python3 -m sglang.benchmark.one_batch_server --model meta-llama/Meta-Llama-3.1-8B --batch-size 1 16 64 --input-len 1024 --output-len 8
 
-python3 -m sglang.bench_one_batch_server --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8
+python3 -m sglang.benchmark.one_batch_server --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8
-python3 -m sglang.bench_one_batch_server --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8 --show-report --profile --profile-by-stage
+python3 -m sglang.benchmark.one_batch_server --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8 --show-report --profile --profile-by-stage
-python3 -m sglang.bench_one_batch_server --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8 --result-filename results.jsonl --profile
+python3 -m sglang.benchmark.one_batch_server --model None --base-url http://localhost:30000 --batch-size 16 --input-len 1024 --output-len 8 --result-filename results.jsonl --profile
 """
 
 import argparse
@@ -309,7 +309,7 @@ class BenchArgs:
             default=BenchArgs.lora_request_distribution,
             choices=["uniform", "distinct", "skewed"],
             help="How to sample a LoRA adapter per prompt when more than one "
-            "is listed in --lora-name. Mirrors bench_serving.py. "
+            "is listed in --lora-name. Mirrors serving.py. "
             "'uniform' picks uniformly at random, 'distinct' round-robins so "
             "consecutive prompts get different adapters, 'skewed' samples "
             "from a Zipf distribution over --lora-name (alpha controls the "
@@ -507,7 +507,7 @@ def run_one_case(
     else:
         _flush_cache_with_retry(url, "/flush_cache")
 
-    # Load input token ids via bench_serving.get_dataset
+    # Load input token ids via benchmark.datasets.get_dataset
     supported_datasets = ("random", "random-ids", "mmmu", "generated-shared-prefix")
     if dataset_name not in supported_datasets:
         raise ValueError(
@@ -968,7 +968,7 @@ def run_benchmark_internal(
                     f"to actually exercise multi-batch."
                 )
 
-    # LoRA distribution args: mirror bench_serving.py semantics so multi-LoRA
+    # LoRA distribution args: mirror serving.py semantics so multi-LoRA
     # benchmarks behave consistently across harnesses.
     if bench_args.lora_request_distribution in ("distinct", "skewed"):
         assert bench_args.lora_name is not None and len(bench_args.lora_name) > 1, (
@@ -1152,7 +1152,7 @@ def run_benchmark(server_args: ServerArgs, bench_args: BenchArgs):
     return results, server_info
 
 
-def main():
+def cli_main():
     parser = argparse.ArgumentParser()
     ServerArgs.add_cli_args(parser)
     BenchArgs.add_cli_args(parser)
@@ -1165,4 +1165,4 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    cli_main()