[UX] Defer some imports on CLI paths to save ~2s (#40056)

Signed-off-by: mgoin <mgoin64@gmail.com>

vllm/benchmarks/sweep/plot.py

@@ -8,7 +8,7 @@ from dataclasses import dataclass
 from functools import partial
 from pathlib import Path
 from types import TracebackType
-from typing import ClassVar
+from typing import TYPE_CHECKING, ClassVar
 
 from typing_extensions import Self, override
 
@@ -17,20 +17,8 @@ from vllm.utils.import_utils import PlaceholderModule
 
 from .utils import sanitize_filename
 
-try:
-    import matplotlib.pyplot as plt
-except ImportError:
-    plt = PlaceholderModule("matplotlib").placeholder_attr("pyplot")
-
-try:
+if TYPE_CHECKING:
     import pandas as pd
-except ImportError:
-    pd = PlaceholderModule("pandas")
-
-try:
-    import seaborn as sns
-except ImportError:
-    seaborn = PlaceholderModule("seaborn")
 
 
 @dataclass
@@ -265,6 +253,20 @@ def _plot_fig(
     fig_height: float,
     fig_dpi: int,
 ):
+    # Lazy-import matplotlib/pandas/seaborn
+    try:
+        import matplotlib.pyplot as plt
+    except ImportError:
+        plt = PlaceholderModule("matplotlib").placeholder_attr("pyplot")
+    try:
+        import pandas as pd
+    except ImportError:
+        pd = PlaceholderModule("pandas")
+    try:
+        import seaborn as sns
+    except ImportError:
+        sns = PlaceholderModule("seaborn")
+
     fig_group, fig_data = fig_group_data
 
     row_groups = full_groupby(

vllm/benchmarks/sweep/plot_pareto.py

@@ -6,7 +6,7 @@ from concurrent.futures import ProcessPoolExecutor
 from dataclasses import dataclass
 from functools import partial
 from pathlib import Path
-from typing import ClassVar
+from typing import TYPE_CHECKING, ClassVar
 
 from vllm.utils.collection_utils import full_groupby
 from vllm.utils.import_utils import PlaceholderModule
@@ -14,20 +14,8 @@ from vllm.utils.import_utils import PlaceholderModule
 from .plot import DummyExecutor, _json_load_bytes
 from .utils import sanitize_filename
 
-try:
-    import matplotlib.pyplot as plt
-except ImportError:
-    plt = PlaceholderModule("matplotlib").placeholder_attr("pyplot")
-
-try:
+if TYPE_CHECKING:
     import pandas as pd
-except ImportError:
-    pd = PlaceholderModule("pandas")
-
-try:
-    import seaborn as sns
-except ImportError:
-    seaborn = PlaceholderModule("seaborn")
 
 
 def _first_present(run_data: dict[str, object], keys: list[str]):
@@ -195,6 +183,20 @@ def _plot_fig(
         print("[END FIGURE]")
         return
 
+    # Lazy-import matplotlib/pandas/seaborn
+    try:
+        import matplotlib.pyplot as plt
+    except ImportError:
+        plt = PlaceholderModule("matplotlib").placeholder_attr("pyplot")
+    try:
+        import pandas as pd
+    except ImportError:
+        pd = PlaceholderModule("pandas")
+    try:
+        import seaborn as sns
+    except ImportError:
+        sns = PlaceholderModule("seaborn")
+
     df = pd.DataFrame.from_records(fig_data)
     df = df.dropna(subset=["tokens_per_user", "tokens_per_gpu"])
 

vllm/entrypoints/cli/__init__.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from vllm.entrypoints.cli.benchmark.latency import BenchmarkLatencySubcommand
-from vllm.entrypoints.cli.benchmark.mm_processor import (
-    BenchmarkMMProcessorSubcommand,
-)
-from vllm.entrypoints.cli.benchmark.serve import BenchmarkServingSubcommand
-from vllm.entrypoints.cli.benchmark.startup import BenchmarkStartupSubcommand
-from vllm.entrypoints.cli.benchmark.sweep import BenchmarkSweepSubcommand
-from vllm.entrypoints.cli.benchmark.throughput import BenchmarkThroughputSubcommand
-
-__all__: list[str] = [
-    "BenchmarkLatencySubcommand",
-    "BenchmarkMMProcessorSubcommand",
-    "BenchmarkServingSubcommand",
-    "BenchmarkStartupSubcommand",
-    "BenchmarkSweepSubcommand",
-    "BenchmarkThroughputSubcommand",
-]

vllm/entrypoints/cli/benchmark/main.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
+import sys
 import typing
 
 from vllm.entrypoints.cli.benchmark.base import BenchmarkSubcommandBase
@@ -14,6 +15,17 @@ else:
     FlexibleArgumentParser = argparse.ArgumentParser
 
 
+def _import_bench_subcommand_modules() -> None:
+    # Imported lazily so `BenchmarkSubcommandBase` subclasses register only
+    # when `vllm bench` is actually invoked.
+    import vllm.entrypoints.cli.benchmark.latency  # noqa: F401
+    import vllm.entrypoints.cli.benchmark.mm_processor  # noqa: F401
+    import vllm.entrypoints.cli.benchmark.serve  # noqa: F401
+    import vllm.entrypoints.cli.benchmark.startup  # noqa: F401
+    import vllm.entrypoints.cli.benchmark.sweep  # noqa: F401
+    import vllm.entrypoints.cli.benchmark.throughput  # noqa: F401
+
+
 class BenchmarkSubcommand(CLISubcommand):
     """The `bench` subcommand for the vLLM CLI."""
 
@@ -38,6 +50,16 @@ class BenchmarkSubcommand(CLISubcommand):
         )
         bench_subparsers = bench_parser.add_subparsers(required=True, dest="bench_type")
 
+        # Only build the nested bench subparsers when the user is actually
+        # invoking `bench`; otherwise we'd drag in imports
+        # unnecessarily on every `vllm --help` and `vllm serve`.
+        # Scan for the first positional arg so global flags (e.g. `-v`)
+        # before the subcommand don't break detection.
+        first_positional = next(
+            (arg for arg in sys.argv[1:] if not arg.startswith("-")), None
+        )
+        if first_positional == self.name:
+            _import_bench_subcommand_modules()
             for cmd_cls in BenchmarkSubcommandBase.__subclasses__():
                 cmd_subparser = bench_subparsers.add_parser(
                     cmd_cls.name,

vllm/env_override.py

@@ -100,10 +100,9 @@ logger = init_logger(__name__)
 # it avoids unintentional cuda initialization from torch.cuda.is_available()
 os.environ["PYTORCH_NVML_BASED_CUDA_CHECK"] = "1"
 
-# see https://github.com/vllm-project/vllm/issues/10480
+# see https://github.com/vllm-project/vllm/issues/10480 and
+# https://github.com/vllm-project/vllm/issues/10619.
 os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1"
-# see https://github.com/vllm-project/vllm/issues/10619
-torch._inductor.config.compile_threads = 1
 
 # Enable Triton autotuning result caching to disk by default.
 # Without this, Triton re-runs autotuning on every process restart,