Sync from v0.13

vllm/benchmarks/sweep/cli.py

@@ -0,0 +1,41 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+
+from vllm.entrypoints.utils import VLLM_SUBCMD_PARSER_EPILOG
+
+from .plot import SweepPlotArgs
+from .plot import main as plot_main
+from .plot_pareto import SweepPlotParetoArgs
+from .plot_pareto import main as plot_pareto_main
+from .serve import SweepServeArgs
+from .serve import main as serve_main
+from .serve_sla import SweepServeSLAArgs
+from .serve_sla import main as serve_sla_main
+
+SUBCOMMANDS = (
+    (SweepServeArgs, serve_main),
+    (SweepServeSLAArgs, serve_sla_main),
+    (SweepPlotArgs, plot_main),
+    (SweepPlotParetoArgs, plot_pareto_main),
+)
+
+
+def add_cli_args(parser: argparse.ArgumentParser):
+    subparsers = parser.add_subparsers(required=True, dest="sweep_type")
+
+    for cmd, entrypoint in SUBCOMMANDS:
+        cmd_subparser = subparsers.add_parser(
+            cmd.parser_name,
+            description=cmd.parser_help,
+            usage=f"vllm bench sweep {cmd.parser_name} [options]",
+        )
+        cmd_subparser.set_defaults(dispatch_function=entrypoint)
+        cmd.add_cli_args(cmd_subparser)
+        cmd_subparser.epilog = VLLM_SUBCMD_PARSER_EPILOG.format(
+            subcmd=f"sweep {cmd.parser_name}"
+        )
+
+
+def main(args: argparse.Namespace):
+    args.dispatch_function(args)

vllm/benchmarks/sweep/param_sweep.py

@@ -0,0 +1,158 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import json
+import os
+from typing import Any
+
+
+class ParameterSweep(list["ParameterSweepItem"]):
+    @classmethod
+    def read_json(cls, filepath: os.PathLike):
+        with open(filepath, "rb") as f:
+            data = json.load(f)
+
+        # Support both list and dict formats
+        if isinstance(data, dict):
+            return cls.read_from_dict(data)
+
+        return cls.from_records(data)
+
+    @classmethod
+    def read_from_dict(cls, data: dict[str, dict[str, object]]):
+        """
+        Read parameter sweep from a dict format where keys are names.
+
+        Example:
+            {
+                "experiment1": {"max_tokens": 100, "temperature": 0.7},
+                "experiment2": {"max_tokens": 200, "temperature": 0.9}
+            }
+        """
+        records = [{"_benchmark_name": name, **params} for name, params in data.items()]
+        return cls.from_records(records)
+
+    @classmethod
+    def from_records(cls, records: list[dict[str, object]]):
+        if not isinstance(records, list):
+            raise TypeError(
+                f"The parameter sweep should be a list of dictionaries, "
+                f"but found type: {type(records)}"
+            )
+
+        # Validate that all _benchmark_name values are unique if provided
+        names = [r["_benchmark_name"] for r in records if "_benchmark_name" in r]
+        if names and len(names) != len(set(names)):
+            duplicates = [name for name in names if names.count(name) > 1]
+            raise ValueError(
+                f"Duplicate _benchmark_name values found: {set(duplicates)}. "
+                f"All _benchmark_name values must be unique."
+            )
+
+        return cls(ParameterSweepItem.from_record(record) for record in records)
+
+
+class ParameterSweepItem(dict[str, object]):
+    @classmethod
+    def from_record(cls, record: dict[str, object]):
+        if not isinstance(record, dict):
+            raise TypeError(
+                f"Each item in the parameter sweep should be a dictionary, "
+                f"but found type: {type(record)}"
+            )
+
+        return cls(record)
+
+    def __or__(self, other: dict[str, Any]):
+        return type(self)(super().__or__(other))
+
+    @property
+    def name(self) -> str:
+        """
+        Get the name for this parameter sweep item.
+
+        Returns the '_benchmark_name' field if present, otherwise returns a text
+        representation of all parameters.
+        """
+        if "_benchmark_name" in self:
+            return self["_benchmark_name"]
+        return self.as_text(sep="-")
+
+    # In JSON, we prefer "_"
+    def _iter_param_key_candidates(self, param_key: str):
+        # Inner config arguments are not converted by the CLI
+        if "." in param_key:
+            prefix, rest = param_key.split(".", 1)
+            for prefix_candidate in self._iter_param_key_candidates(prefix):
+                yield prefix_candidate + "." + rest
+
+            return
+
+        yield param_key
+        yield param_key.replace("-", "_")
+        yield param_key.replace("_", "-")
+
+    # In CLI, we prefer "-"
+    def _iter_cmd_key_candidates(self, param_key: str):
+        for k in reversed(tuple(self._iter_param_key_candidates(param_key))):
+            yield "--" + k
+
+    def _normalize_cmd_key(self, param_key: str):
+        return next(self._iter_cmd_key_candidates(param_key))
+
+    def has_param(self, param_key: str) -> bool:
+        return any(k in self for k in self._iter_param_key_candidates(param_key))
+
+    def _normalize_cmd_kv_pair(self, k: str, v: object) -> list[str]:
+        """
+        Normalize a key-value pair into command-line arguments.
+
+        Returns a list containing either:
+        - A single element for boolean flags (e.g., ['--flag'] or ['--flag=true'])
+        - Two elements for key-value pairs (e.g., ['--key', 'value'])
+        """
+        if isinstance(v, bool):
+            # For nested params (containing "."), use =true/false syntax
+            if "." in k:
+                return [f"{self._normalize_cmd_key(k)}={'true' if v else 'false'}"]
+            else:
+                return [self._normalize_cmd_key(k if v else "no-" + k)]
+        else:
+            return [self._normalize_cmd_key(k), str(v)]
+
+    def apply_to_cmd(self, cmd: list[str]) -> list[str]:
+        cmd = list(cmd)
+
+        for k, v in self.items():
+            # Skip the '_benchmark_name' field, not a parameter
+            if k == "_benchmark_name":
+                continue
+
+            # Serialize dict values as JSON
+            if isinstance(v, dict):
+                v = json.dumps(v)
+
+            for k_candidate in self._iter_cmd_key_candidates(k):
+                try:
+                    k_idx = cmd.index(k_candidate)
+
+                    # Replace existing parameter
+                    normalized = self._normalize_cmd_kv_pair(k, v)
+                    if len(normalized) == 1:
+                        # Boolean flag
+                        cmd[k_idx] = normalized[0]
+                    else:
+                        # Key-value pair
+                        cmd[k_idx] = normalized[0]
+                        cmd[k_idx + 1] = normalized[1]
+
+                    break
+                except ValueError:
+                    continue
+            else:
+                # Add new parameter
+                cmd.extend(self._normalize_cmd_kv_pair(k, v))
+
+        return cmd
+
+    def as_text(self, sep: str = ", ") -> str:
+        return sep.join(f"{k}={v}" for k, v in self.items() if k != "_benchmark_name")

vllm/benchmarks/sweep/plot.py

@@ -0,0 +1,675 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import json
+from abc import ABC, abstractmethod
+from concurrent.futures import ProcessPoolExecutor
+from dataclasses import dataclass
+from functools import partial
+from pathlib import Path
+from types import TracebackType
+from typing import ClassVar
+
+from typing_extensions import Self, override
+
+from vllm.utils.collection_utils import full_groupby
+from vllm.utils.import_utils import PlaceholderModule
+
+from .utils import sanitize_filename
+
+try:
+    import matplotlib.pyplot as plt
+    import pandas as pd
+    import seaborn as sns
+except ImportError:
+    plt = PlaceholderModule("matplotlib").placeholder_attr("pyplot")
+    pd = PlaceholderModule("pandas")
+    seaborn = PlaceholderModule("seaborn")
+
+
+@dataclass
+class PlotFilterBase(ABC):
+    var: str
+    target: str
+
+    @classmethod
+    def parse_str(cls, s: str):
+        for op_key in PLOT_FILTERS:
+            if op_key in s:
+                key, value = s.split(op_key)
+                return PLOT_FILTERS[op_key](
+                    key,
+                    value.removeprefix(op_key).strip("'").strip('"'),
+                )
+        else:
+            raise ValueError(
+                f"Invalid operator for plot filter '{s}'. "
+                f"Valid operators are: {sorted(PLOT_FILTERS)}",
+            )
+
+    @abstractmethod
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        """Applies this filter to a DataFrame."""
+        raise NotImplementedError
+
+
+@dataclass
+class PlotEqualTo(PlotFilterBase):
+    @override
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        try:
+            target = float(self.target)
+        except ValueError:
+            target = self.target
+
+        return df[df[self.var] == target]
+
+
+@dataclass
+class PlotNotEqualTo(PlotFilterBase):
+    @override
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        try:
+            target = float(self.target)
+        except ValueError:
+            target = self.target
+
+        return df[df[self.var] != target]
+
+
+@dataclass
+class PlotLessThan(PlotFilterBase):
+    @override
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        return df[df[self.var] < float(self.target)]
+
+
+@dataclass
+class PlotLessThanOrEqualTo(PlotFilterBase):
+    @override
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        return df[df[self.var] <= float(self.target)]
+
+
+@dataclass
+class PlotGreaterThan(PlotFilterBase):
+    @override
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        return df[df[self.var] > float(self.target)]
+
+
+@dataclass
+class PlotGreaterThanOrEqualTo(PlotFilterBase):
+    @override
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        return df[df[self.var] >= float(self.target)]
+
+
+# NOTE: The ordering is important! Match longer op_keys first
+PLOT_FILTERS: dict[str, type[PlotFilterBase]] = {
+    "==": PlotEqualTo,
+    "!=": PlotNotEqualTo,
+    "<=": PlotLessThanOrEqualTo,
+    ">=": PlotGreaterThanOrEqualTo,
+    "<": PlotLessThan,
+    ">": PlotGreaterThan,
+}
+
+
+class PlotFilters(list[PlotFilterBase]):
+    @classmethod
+    def parse_str(cls, s: str):
+        if not s:
+            return cls()
+
+        return cls(PlotFilterBase.parse_str(e) for e in s.split(","))
+
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        for item in self:
+            df = item.apply(df)
+
+        return df
+
+
+@dataclass
+class PlotBinner:
+    var: str
+    bin_size: float
+
+    @classmethod
+    def parse_str(cls, s: str):
+        for op_key in PLOT_BINNERS:
+            if op_key in s:
+                key, value = s.split(op_key)
+                return PLOT_BINNERS[op_key](key, float(value.removeprefix(op_key)))
+        else:
+            raise ValueError(
+                f"Invalid operator for plot binner '{s}'. "
+                f"Valid operators are: {sorted(PLOT_BINNERS)}",
+            )
+
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        """Applies this binner to a DataFrame."""
+        df = df.copy()
+        df[self.var] = df[self.var] // self.bin_size * self.bin_size
+        return df
+
+
+PLOT_BINNERS: dict[str, type[PlotBinner]] = {
+    "%": PlotBinner,
+}
+
+
+class PlotBinners(list[PlotBinner]):
+    @classmethod
+    def parse_str(cls, s: str):
+        if not s:
+            return cls()
+
+        return cls(PlotBinner.parse_str(e) for e in s.split(","))
+
+    def apply(self, df: "pd.DataFrame") -> "pd.DataFrame":
+        for item in self:
+            df = item.apply(df)
+
+        return df
+
+
+def _json_load_bytes(path: Path) -> list[dict[str, object]]:
+    with path.open("rb") as f:
+        return json.load(f)
+
+
+def _convert_inf_nan_strings(data: list[dict[str, object]]) -> list[dict[str, object]]:
+    """
+    Convert string values "inf", "-inf", and "nan" to their float equivalents.
+
+    This handles the case where JSON serialization represents inf/nan as strings.
+    """
+    converted_data = []
+    for record in data:
+        converted_record = {}
+        for key, value in record.items():
+            if isinstance(value, str):
+                if value in ["inf", "-inf", "nan"]:
+                    converted_record[key] = float(value)
+                else:
+                    converted_record[key] = value
+            else:
+                converted_record[key] = value
+        converted_data.append(converted_record)
+    return converted_data
+
+
+def _get_metric(run_data: dict[str, object], metric_key: str):
+    try:
+        return run_data[metric_key]
+    except KeyError as exc:
+        raise ValueError(f"Cannot find metric {metric_key!r} in {run_data=}") from exc
+
+
+def _get_group(run_data: dict[str, object], group_keys: list[str]):
+    return tuple((k, str(_get_metric(run_data, k))) for k in group_keys)
+
+
+def _get_fig_path(fig_dir: Path, group: tuple[tuple[str, str], ...], fig_name: str):
+    parts = list[str]()
+
+    # Start with figure name (always provided, defaults to "FIGURE")
+    parts.append(fig_name)
+
+    # Always append group data if present
+    if group:
+        parts.extend(f"{k}={v}" for k, v in group)
+
+    return fig_dir / sanitize_filename("-".join(parts) + ".png")
+
+
+class DummyExecutor:
+    map = map
+
+    def __enter__(self) -> Self:
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_value: BaseException | None,
+        exc_traceback: TracebackType | None,
+    ) -> None:
+        return None
+
+
+def _plot_fig(
+    fig_dir: Path,
+    fig_group_data: tuple[tuple[tuple[str, str], ...], list[dict[str, object]]],
+    row_by: list[str],
+    col_by: list[str],
+    curve_by: list[str],
+    *,
+    var_x: str,
+    var_y: str,
+    filter_by: PlotFilters,
+    bin_by: PlotBinners,
+    scale_x: str | None,
+    scale_y: str | None,
+    dry_run: bool,
+    fig_name: str,
+    error_bars: bool,
+    fig_height: float,
+    fig_dpi: int,
+):
+    fig_group, fig_data = fig_group_data
+
+    row_groups = full_groupby(
+        fig_data,
+        key=lambda item: _get_group(item, row_by),
+    )
+    num_rows = len(row_groups)
+    num_cols = max(
+        len(full_groupby(row_data, key=lambda item: _get_group(item, col_by)))
+        for _, row_data in row_groups
+    )
+
+    fig_path = _get_fig_path(fig_dir, fig_group, fig_name)
+
+    print("[BEGIN FIGURE]")
+    print(f"Group: {dict(fig_group)}")
+    print(f"Grid: {num_rows} rows x {num_cols} cols")
+    print(f"Output file: {fig_path}")
+
+    if dry_run:
+        print("[END FIGURE]")
+        return
+
+    # Convert string "inf", "-inf", and "nan" to their float equivalents
+    fig_data = _convert_inf_nan_strings(fig_data)
+    df = pd.DataFrame.from_records(fig_data)
+
+    if var_x not in df.columns:
+        raise ValueError(
+            f"Cannot find {var_x=!r} in parameter sweep results. "
+            f"Available variables: {df.columns.tolist()}"
+        )
+    if var_y not in df.columns:
+        raise ValueError(
+            f"Cannot find {var_y=!r} in parameter sweep results. "
+            f"Available variables: {df.columns.tolist()}"
+        )
+    for k in row_by:
+        if k not in df.columns:
+            raise ValueError(
+                f"Cannot find row_by={k!r} in parameter sweep results. "
+                f"Available variables: {df.columns.tolist()}"
+            )
+    for k in col_by:
+        if k not in df.columns:
+            raise ValueError(
+                f"Cannot find col_by={k!r} in parameter sweep results. "
+                f"Available variables: {df.columns.tolist()}"
+            )
+    for k in curve_by:
+        if k not in df.columns:
+            raise ValueError(
+                f"Cannot find curve_by={k!r} in parameter sweep results. "
+                f"Available variables: {df.columns.tolist()}"
+            )
+
+    df = filter_by.apply(df)
+    df = bin_by.apply(df)
+
+    # Sort by curve_by columns alphabetically for consistent legend ordering
+    if curve_by:
+        df = df.sort_values(by=curve_by)
+
+    df["row_group"] = (
+        pd.concat(
+            [k + "=" + df[k].astype(str) for k in row_by],
+            axis=1,
+        ).agg("\n".join, axis=1)
+        if row_by
+        else "(All)"
+    )
+
+    df["col_group"] = (
+        pd.concat(
+            [k + "=" + df[k].astype(str) for k in col_by],
+            axis=1,
+        ).agg("\n".join, axis=1)
+        if col_by
+        else "(All)"
+    )
+
+    g = sns.FacetGrid(df, row="row_group", col="col_group", height=fig_height)
+
+    if row_by and col_by:
+        g.set_titles("{row_name}\n{col_name}")
+    elif row_by:
+        g.set_titles("{row_name}")
+    elif col_by:
+        g.set_titles("{col_name}")
+    else:
+        g.set_titles("")
+
+    if scale_x:
+        g.set(xscale=scale_x)
+    if scale_y:
+        g.set(yscale=scale_y)
+
+    if len(curve_by) <= 3:
+        hue, style, size, *_ = (*curve_by, None, None, None)
+
+        g.map_dataframe(
+            sns.lineplot,
+            x=var_x,
+            y=var_y,
+            hue=hue,
+            style=style,
+            size=size,
+            markers=True,
+            errorbar="sd" if error_bars else None,
+        )
+
+        g.add_legend(title=hue)
+    else:
+        df["curve_group"] = (
+            pd.concat(
+                [k + "=" + df[k].astype(str) for k in curve_by],
+                axis=1,
+            ).agg("\n".join, axis=1)
+            if curve_by
+            else "(All)"
+        )
+
+        g.map_dataframe(
+            sns.lineplot,
+            x=var_x,
+            y=var_y,
+            hue="curve_group",
+            markers=True,
+            errorbar="sd" if error_bars else None,
+        )
+
+        g.add_legend()
+
+    g.savefig(fig_path, dpi=fig_dpi)
+    plt.close(g.figure)
+
+    print("[END FIGURE]")
+
+
+def plot(
+    output_dir: Path,
+    fig_dir: Path,
+    fig_by: list[str],
+    row_by: list[str],
+    col_by: list[str],
+    curve_by: list[str],
+    *,
+    var_x: str,
+    var_y: str,
+    filter_by: PlotFilters,
+    bin_by: PlotBinners,
+    scale_x: str | None,
+    scale_y: str | None,
+    dry_run: bool,
+    fig_name: str = "FIGURE",
+    error_bars: bool = True,
+    fig_height: float = 6.4,
+    fig_dpi: int = 300,
+):
+    all_data = [
+        run_data
+        for path in output_dir.rglob("**/summary.json")
+        for run_data in _json_load_bytes(path)
+    ]
+
+    if not all_data:
+        raise ValueError(f"Did not find any parameter sweep results under {output_dir}")
+
+    fig_dir.mkdir(parents=True, exist_ok=True)
+
+    fig_groups = full_groupby(
+        all_data,
+        key=lambda item: _get_group(item, fig_by),
+    )
+
+    with DummyExecutor() if len(fig_groups) <= 1 else ProcessPoolExecutor() as executor:
+        # Resolve the iterable to ensure that the workers are run
+        all(
+            executor.map(
+                partial(
+                    _plot_fig,
+                    fig_dir,
+                    row_by=row_by,
+                    col_by=col_by,
+                    curve_by=curve_by,
+                    var_x=var_x,
+                    var_y=var_y,
+                    filter_by=filter_by,
+                    bin_by=bin_by,
+                    scale_x=scale_x,
+                    scale_y=scale_y,
+                    dry_run=dry_run,
+                    fig_name=fig_name,
+                    error_bars=error_bars,
+                    fig_height=fig_height,
+                    fig_dpi=fig_dpi,
+                ),
+                fig_groups,
+            )
+        )
+
+
+@dataclass
+class SweepPlotArgs:
+    output_dir: Path
+    fig_dir: Path
+    fig_by: list[str]
+    row_by: list[str]
+    col_by: list[str]
+    curve_by: list[str]
+    var_x: str
+    var_y: str
+    filter_by: PlotFilters
+    bin_by: PlotBinners
+    scale_x: str | None
+    scale_y: str | None
+    dry_run: bool
+    fig_name: str = "FIGURE"
+    error_bars: bool = True
+    fig_height: float = 6.4
+    fig_dpi: int = 300
+
+    parser_name: ClassVar[str] = "plot"
+    parser_help: ClassVar[str] = "Plot performance curves from parameter sweep results."
+
+    @classmethod
+    def from_cli_args(cls, args: argparse.Namespace):
+        output_dir = Path(args.OUTPUT_DIR)
+        if not output_dir.exists():
+            raise ValueError(f"No parameter sweep results under {output_dir}")
+
+        curve_by = [] if not args.curve_by else args.curve_by.split(",")
+        row_by = [] if not args.row_by else args.row_by.split(",")
+        col_by = [] if not args.col_by else args.col_by.split(",")
+        fig_by = [] if not args.fig_by else args.fig_by.split(",")
+
+        return cls(
+            output_dir=output_dir,
+            fig_dir=output_dir / args.fig_dir,
+            fig_by=fig_by,
+            row_by=row_by,
+            col_by=col_by,
+            curve_by=curve_by,
+            var_x=args.var_x,
+            var_y=args.var_y,
+            filter_by=PlotFilters.parse_str(args.filter_by),
+            bin_by=PlotBinners.parse_str(args.bin_by),
+            scale_x=args.scale_x,
+            scale_y=args.scale_y,
+            dry_run=args.dry_run,
+            fig_name=args.fig_name,
+            error_bars=not args.no_error_bars,
+            fig_height=args.fig_height,
+            fig_dpi=args.fig_dpi,
+        )
+
+    @classmethod
+    def add_cli_args(cls, parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
+        parser.add_argument(
+            "OUTPUT_DIR",
+            type=str,
+            default="results",
+            help="The directory containing the results to plot, "
+            "i.e., the `--output-dir` argument to the parameter sweep script.",
+        )
+        parser.add_argument(
+            "--fig-dir",
+            type=str,
+            default="",
+            help="The directory to save the figures, relative to `OUTPUT_DIR`. "
+            "By default, the same directory is used.",
+        )
+        parser.add_argument(
+            "--fig-by",
+            type=str,
+            default="",
+            help="A comma-separated list of variables, such that a separate figure "
+            "is created for each combination of these variables.",
+        )
+        parser.add_argument(
+            "--row-by",
+            type=str,
+            default="",
+            help="A comma-separated list of variables, such that a separate row "
+            "is created for each combination of these variables.",
+        )
+        parser.add_argument(
+            "--col-by",
+            type=str,
+            default="",
+            help="A comma-separated list of variables, such that a separate column "
+            "is created for each combination of these variables.",
+        )
+        parser.add_argument(
+            "--curve-by",
+            type=str,
+            default=None,
+            help="A comma-separated list of variables, such that a separate curve "
+            "is created for each combination of these variables.",
+        )
+        parser.add_argument(
+            "--var-x",
+            type=str,
+            default="request_throughput",
+            help="The variable for the x-axis.",
+        )
+        parser.add_argument(
+            "--var-y",
+            type=str,
+            default="p99_e2el_ms",
+            help="The variable for the y-axis",
+        )
+        parser.add_argument(
+            "--filter-by",
+            type=str,
+            default="",
+            help="A comma-separated list of statements indicating values to filter by. "
+            "This is useful to remove outliers. "
+            "Example: `max_concurrency<1000,max_num_batched_tokens<=4096` means "
+            "plot only the points where `max_concurrency` is less than 1000 and "
+            "`max_num_batched_tokens` is no greater than 4096.",
+        )
+        parser.add_argument(
+            "--bin-by",
+            type=str,
+            default="",
+            help="A comma-separated list of statements indicating values to bin by. "
+            "This is useful to avoid plotting points that are too close together. "
+            "Example: `request_throughput%%1` means "
+            "use a bin size of 1 for the `request_throughput` variable.",
+        )
+        parser.add_argument(
+            "--scale-x",
+            type=str,
+            default=None,
+            help="The scale to use for the x-axis. "
+            "Currently only accepts string values such as 'log' and 'sqrt'. "
+            "See also: https://seaborn.pydata.org/generated/seaborn.objects.Plot.scale.html",
+        )
+        parser.add_argument(
+            "--scale-y",
+            type=str,
+            default=None,
+            help="The scale to use for the y-axis. "
+            "Currently only accepts string values such as 'log' and 'sqrt'. "
+            "See also: https://seaborn.pydata.org/generated/seaborn.objects.Plot.scale.html",
+        )
+        parser.add_argument(
+            "--fig-name",
+            type=str,
+            default="FIGURE",
+            help="Name prefix for the output figure file. "
+            "Group data is always appended when present. "
+            "Default: 'FIGURE'. Example: --fig-name my_performance_plot",
+        )
+        parser.add_argument(
+            "--no-error-bars",
+            action="store_true",
+            help="If set, disables error bars on the plot. "
+            "By default, error bars are shown.",
+        )
+        parser.add_argument(
+            "--fig-height",
+            type=float,
+            default=6.4,
+            help="Height of each subplot in inches. Default: 6.4",
+        )
+        parser.add_argument(
+            "--fig-dpi",
+            type=int,
+            default=300,
+            help="Resolution of the output figure in dots per inch. Default: 300",
+        )
+        parser.add_argument(
+            "--dry-run",
+            action="store_true",
+            help="If set, prints the information about each figure to plot, "
+            "then exits without drawing them.",
+        )
+
+        return parser
+
+
+def run_main(args: SweepPlotArgs):
+    return plot(
+        output_dir=args.output_dir,
+        fig_dir=args.fig_dir,
+        fig_by=args.fig_by,
+        row_by=args.row_by,
+        col_by=args.col_by,
+        curve_by=args.curve_by,
+        var_x=args.var_x,
+        var_y=args.var_y,
+        filter_by=args.filter_by,
+        bin_by=args.bin_by,
+        scale_x=args.scale_x,
+        scale_y=args.scale_y,
+        dry_run=args.dry_run,
+        fig_name=args.fig_name,
+        error_bars=args.error_bars,
+        fig_height=args.fig_height,
+        fig_dpi=args.fig_dpi,
+    )
+
+
+def main(args: argparse.Namespace):
+    run_main(SweepPlotArgs.from_cli_args(args))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=SweepPlotArgs.parser_help)
+    SweepPlotArgs.add_cli_args(parser)
+
+    main(parser.parse_args())

vllm/benchmarks/sweep/plot_pareto.py

@@ -0,0 +1,393 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import math
+from concurrent.futures import ProcessPoolExecutor
+from dataclasses import dataclass
+from functools import partial
+from pathlib import Path
+from typing import ClassVar
+
+from vllm.utils.collection_utils import full_groupby
+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
+    import pandas as pd
+    import seaborn as sns
+except ImportError:
+    plt = PlaceholderModule("matplotlib").placeholder_attr("pyplot")
+    pd = PlaceholderModule("pandas")
+    sns = PlaceholderModule("seaborn")
+
+
+def _first_present(run_data: dict[str, object], keys: list[str]):
+    for key in keys:
+        for candidate in {key, key.replace("_", "-"), key.replace("-", "_")}:
+            if candidate in run_data:
+                return run_data[candidate]
+    return None
+
+
+def _get_numeric(
+    run_data: dict[str, object],
+    keys: list[str],
+    *,
+    allow_zero: bool = True,
+) -> float | None:
+    value = _first_present(run_data, keys)
+    if value is None:
+        return None
+
+    try:
+        numeric = float(value)
+    except (TypeError, ValueError) as exc:
+        raise ValueError(
+            f"Expected numeric value for one of {keys}, "
+            f"but found {value!r} in {run_data=}"
+        ) from exc
+
+    if not allow_zero and numeric == 0:
+        return None
+
+    return numeric
+
+
+def _infer_user_count(
+    run_data: dict[str, object],
+    user_count_var: str | None,
+) -> float | None:
+    candidates = [user_count_var] if user_count_var else []
+    candidates.extend(["request_rate"])
+    user_count = _get_numeric(run_data, candidates, allow_zero=False)
+    if user_count is not None:
+        return user_count
+
+    # Fallback to the observed peak if configured value is missing.
+    return _get_numeric(run_data, ["max_concurrent_requests"], allow_zero=False)
+
+
+def _infer_gpu_count(
+    run_data: dict[str, object],
+    gpu_count_var: str | None,
+) -> float:
+    direct_candidates = [gpu_count_var] if gpu_count_var else []
+    direct_gpu_count = _get_numeric(run_data, direct_candidates, allow_zero=False)
+    if direct_gpu_count:
+        return direct_gpu_count
+
+    tp_size = _get_numeric(run_data, ["tensor_parallel_size", "tp"])
+    pp_size = _get_numeric(run_data, ["pipeline_parallel_size", "pp"])
+    dp_size = _get_numeric(run_data, ["data_parallel_size", "dp"])
+    world_size = 1.0
+    if tp_size:
+        world_size *= tp_size
+    if pp_size:
+        world_size *= pp_size
+    if dp_size:
+        world_size *= dp_size
+
+    return world_size
+
+
+def _get_throughput(
+    run_data: dict[str, object],
+    throughput_var: str,
+) -> float:
+    throughput = _get_numeric(run_data, [throughput_var])
+    if throughput is None:
+        raise ValueError(
+            f"Cannot find throughput metric {throughput_var!r} in run data. "
+            f"Available keys: {sorted(run_data)}"
+        )
+
+    return throughput
+
+
+def _prepare_records(
+    all_data: list[dict[str, object]],
+    *,
+    user_count_var: str | None,
+    gpu_count_var: str | None,
+) -> tuple[list[dict[str, object]], int]:
+    prepared = []
+    skipped_missing_users = 0
+
+    for record in all_data:
+        throughput = _get_throughput(record, "output_throughput")
+        user_count = _infer_user_count(record, user_count_var)
+        if user_count is None:
+            skipped_missing_users += 1
+            continue
+
+        gpu_count = _infer_gpu_count(record, gpu_count_var)
+        tokens_per_user = throughput / user_count
+        tokens_per_gpu = throughput / gpu_count
+
+        prepared.append(
+            {
+                **record,
+                "tokens_per_user": tokens_per_user,
+                "tokens_per_gpu": tokens_per_gpu,
+                "user_count_estimate": user_count,
+                "gpu_count": gpu_count,
+            }
+        )
+
+    return prepared, skipped_missing_users
+
+
+def _pareto_frontier(
+    df: "pd.DataFrame",
+    x_col: str,
+    y_col: str,
+    *,
+    epsilon: float = 1e-9,
+) -> "pd.DataFrame":
+    sorted_df = df.sort_values([x_col, y_col], ascending=[False, False])
+    frontier_indices = []
+    best_y = -math.inf
+
+    for idx, row in sorted_df.iterrows():
+        y_val = row[y_col]
+        if y_val >= best_y - epsilon:
+            frontier_indices.append(idx)
+            best_y = max(best_y, y_val)
+
+    return df.loc[frontier_indices]
+
+
+def _get_fig_path(
+    fig_dir: Path,
+    fig_group: tuple[tuple[str, str], ...],
+) -> Path:
+    parts = ["PARETO"]
+    if fig_group:
+        parts.extend(f"{k}={v}" for k, v in fig_group)
+    filename = sanitize_filename("-".join(parts) + ".png")
+    return fig_dir / filename
+
+
+def _plot_fig(
+    fig_dir: Path,
+    fig_group_data: tuple[tuple[tuple[str, str], ...], list[dict[str, object]]],
+    label_by: list[str],
+    *,
+    dry_run: bool,
+):
+    fig_group, fig_data = fig_group_data
+    fig_path = _get_fig_path(fig_dir, fig_group)
+
+    print("[BEGIN FIGURE]")
+    print(f"Group: {dict(fig_group)}")
+    print(f"Output file: {fig_path}")
+
+    if dry_run:
+        print("[END FIGURE]")
+        return
+
+    df = pd.DataFrame.from_records(fig_data)
+    df = df.dropna(subset=["tokens_per_user", "tokens_per_gpu"])
+
+    if df.empty:
+        print("No data points available after filtering; skipping.")
+        print("[END FIGURE]")
+        return
+
+    frontier = _pareto_frontier(df, "tokens_per_user", "tokens_per_gpu")
+    frontier = frontier.sort_values("tokens_per_user")
+
+    fig, ax = plt.subplots()
+    sns.scatterplot(
+        data=df,
+        x="tokens_per_user",
+        y="tokens_per_gpu",
+        color="0.5",
+        alpha=0.6,
+        ax=ax,
+        label="All runs",
+    )
+    sns.lineplot(
+        data=frontier,
+        x="tokens_per_user",
+        y="tokens_per_gpu",
+        marker="o",
+        ax=ax,
+        label="Pareto frontier",
+    )
+
+    if label_by:
+        for _, row in frontier.iterrows():
+            label_parts = []
+            for key in label_by:
+                if key in row:
+                    label_parts.append(f"{key}={row[key]}")
+            if label_parts:
+                ax.text(
+                    row["tokens_per_user"],
+                    row["tokens_per_gpu"],
+                    "\n".join(label_parts),
+                    fontsize=8,
+                )
+
+    ax.set_xlabel("Tokens/s/user")
+    ax.set_ylabel("Tokens/s/GPU")
+    ax.grid(True, linestyle="--", linewidth=0.5, alpha=0.6)
+    fig.tight_layout()
+    fig.savefig(fig_path)
+    plt.close(fig)
+
+    print(
+        f"Plotted {len(df)} points; Pareto frontier size: {len(frontier)}.",
+    )
+    print("[END FIGURE]")
+
+
+def plot_pareto(
+    output_dir: Path,
+    user_count_var: str | None,
+    gpu_count_var: str | None,
+    label_by: list[str],
+    *,
+    dry_run: bool,
+):
+    fig_dir = output_dir / "pareto"
+    raw_data = [
+        run_data
+        for path in output_dir.rglob("**/summary.json")
+        for run_data in _json_load_bytes(path)
+    ]
+
+    if not raw_data:
+        raise ValueError(f"Did not find any parameter sweep results under {output_dir}")
+
+    fig_dir.mkdir(parents=True, exist_ok=True)
+
+    prepared_data, skipped_missing_users = _prepare_records(
+        raw_data,
+        user_count_var=user_count_var,
+        gpu_count_var=gpu_count_var,
+    )
+
+    if skipped_missing_users:
+        print(
+            f"Skipped {skipped_missing_users} runs without a user count "
+            "(`max_concurrency` or `max_concurrent_requests`).",
+        )
+
+    if not prepared_data:
+        raise ValueError(
+            "No data points with both throughput and user count available "
+            "to plot Pareto frontier.",
+        )
+
+    fig_groups = full_groupby(
+        prepared_data,
+        key=lambda item: tuple(),
+    )
+
+    with DummyExecutor() if len(fig_groups) <= 1 else ProcessPoolExecutor() as executor:
+        all(
+            executor.map(
+                partial(
+                    _plot_fig,
+                    fig_dir,
+                    label_by=label_by,
+                    dry_run=dry_run,
+                ),
+                fig_groups,
+            )
+        )
+
+
+@dataclass
+class SweepPlotParetoArgs:
+    output_dir: Path
+    user_count_var: str | None
+    gpu_count_var: str | None
+    label_by: list[str]
+    dry_run: bool
+
+    parser_name: ClassVar[str] = "plot_pareto"
+    parser_help: ClassVar[str] = (
+        "Plot Pareto frontier between tokens/s/user and tokens/s/GPU "
+        "from parameter sweep results."
+    )
+
+    @classmethod
+    def from_cli_args(cls, args: argparse.Namespace):
+        output_dir = Path(args.OUTPUT_DIR)
+        if not output_dir.exists():
+            raise ValueError(f"No parameter sweep results under {output_dir}")
+
+        label_by = [] if not args.label_by else args.label_by.split(",")
+
+        return cls(
+            output_dir=output_dir,
+            user_count_var=args.user_count_var,
+            gpu_count_var=args.gpu_count_var,
+            label_by=label_by,
+            dry_run=args.dry_run,
+        )
+
+    @classmethod
+    def add_cli_args(cls, parser: argparse.ArgumentParser):
+        parser.add_argument(
+            "OUTPUT_DIR",
+            type=str,
+            default="results",
+            help="The directory containing the sweep results to plot.",
+        )
+        parser.add_argument(
+            "--user-count-var",
+            type=str,
+            default="max_concurrency",
+            help="Result key that stores concurrent user count. "
+            "Falls back to max_concurrent_requests if missing.",
+        )
+        parser.add_argument(
+            "--gpu-count-var",
+            type=str,
+            default=None,
+            help="Result key that stores GPU count. "
+            "If not provided, falls back to num_gpus/gpu_count "
+            "or tensor_parallel_size * pipeline_parallel_size.",
+        )
+        parser.add_argument(
+            "--label-by",
+            type=str,
+            default="max_concurrency,gpu_count",
+            help="Comma-separated list of fields to annotate on Pareto frontier "
+            "points.",
+        )
+        parser.add_argument(
+            "--dry-run",
+            action="store_true",
+            help="If set, prints the figures to plot without drawing them.",
+        )
+
+        return parser
+
+
+def run_main(args: SweepPlotParetoArgs):
+    return plot_pareto(
+        output_dir=args.output_dir,
+        user_count_var=args.user_count_var,
+        gpu_count_var=args.gpu_count_var,
+        label_by=args.label_by,
+        dry_run=args.dry_run,
+    )
+
+
+def main(args: argparse.Namespace):
+    run_main(SweepPlotParetoArgs.from_cli_args(args))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=SweepPlotParetoArgs.parser_help)
+    SweepPlotParetoArgs.add_cli_args(parser)
+
+    main(parser.parse_args())

vllm/benchmarks/sweep/serve.py

@@ -0,0 +1,450 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import contextlib
+import json
+import shlex
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import ClassVar
+
+from vllm.utils.import_utils import PlaceholderModule
+
+from .param_sweep import ParameterSweep, ParameterSweepItem
+from .server import ServerProcess
+from .utils import sanitize_filename
+
+try:
+    import pandas as pd
+except ImportError:
+    pd = PlaceholderModule("pandas")
+
+
+@contextlib.contextmanager
+def run_server(
+    serve_cmd: list[str],
+    after_bench_cmd: list[str],
+    *,
+    show_stdout: bool,
+    serve_overrides: ParameterSweepItem,
+    dry_run: bool,
+):
+    server_cmd = serve_overrides.apply_to_cmd(serve_cmd)
+
+    print("[BEGIN SERVER]")
+    print(f"Server overrides: {serve_overrides}")
+    print(f"Server command: {server_cmd}")
+
+    if dry_run:
+        yield None
+        print("[END SERVER]")
+        return
+
+    with ServerProcess(server_cmd, after_bench_cmd, show_stdout=show_stdout) as server:
+        yield server
+
+    print("[END SERVER]")
+
+
+def _update_run_data(
+    run_data: dict[str, object],
+    serve_overrides: ParameterSweepItem,
+    bench_overrides: ParameterSweepItem,
+    run_number: int,
+):
+    run_data["run_number"] = run_number
+    run_data.update(serve_overrides)
+    run_data.update(bench_overrides)
+
+    return run_data
+
+
+def run_benchmark(
+    server: ServerProcess | None,
+    bench_cmd: list[str],
+    *,
+    serve_overrides: ParameterSweepItem,
+    bench_overrides: ParameterSweepItem,
+    run_number: int,
+    output_path: Path,
+    dry_run: bool,
+):
+    benchmark_cmd = [
+        *bench_overrides.apply_to_cmd(bench_cmd),
+        "--percentile-metrics",
+        "ttft,tpot,itl,e2el",
+        "--save-result",
+        "--result-dir",
+        str(output_path.parent),
+        "--result-filename",
+        output_path.name,
+    ]
+
+    print("[BEGIN BENCHMARK]")
+    print(f"Benchmark overrides: {bench_overrides}")
+    print(f"Run Number: {run_number}")
+    print(f"Benchmark command: {benchmark_cmd}")
+    print(f"Output file: {output_path}")
+
+    run_data: dict[str, object]
+
+    if output_path.exists():
+        print("Found existing results. Skipping.")
+
+        with output_path.open("rb") as f:
+            run_data = json.load(f)
+            return _update_run_data(
+                run_data,
+                serve_overrides,
+                bench_overrides,
+                run_number,
+            )
+
+    if server is None:
+        if not dry_run:
+            raise ValueError(f"Cannot find results at {output_path}")
+
+        print("[END BENCHMARK]")
+        return None
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    server.run_subcommand(benchmark_cmd)
+    server.after_bench()
+
+    with output_path.open("rb") as f:
+        run_data = json.load(f)
+
+    run_data = _update_run_data(
+        run_data,
+        serve_overrides,
+        bench_overrides,
+        run_number,
+    )
+
+    with output_path.open("w") as f:
+        json.dump(run_data, f, indent=4)
+
+    print("[END BENCHMARK]")
+
+    return run_data
+
+
+def _get_comb_base_path(
+    output_dir: Path,
+    serve_comb: ParameterSweepItem,
+    bench_comb: ParameterSweepItem,
+):
+    parts = list[str]()
+    if serve_comb:
+        parts.extend(("SERVE-", serve_comb.name))
+    if bench_comb:
+        parts.extend(("BENCH-", bench_comb.name))
+
+    return output_dir / sanitize_filename("-".join(parts))
+
+
+def _get_comb_run_path(base_path: Path, run_number: int | None):
+    if run_number is None:
+        return base_path / "summary.json"
+
+    return base_path / f"run={run_number}.json"
+
+
+def _comb_needs_server(
+    serve_comb: ParameterSweepItem,
+    bench_combs: ParameterSweep,
+    output_dir: Path,
+):
+    for bench_comb in bench_combs:
+        base_path = _get_comb_base_path(output_dir, serve_comb, bench_comb)
+        if not _get_comb_run_path(base_path, run_number=None).exists():
+            return True
+
+    return False
+
+
+def run_comb(
+    server: ServerProcess | None,
+    bench_cmd: list[str],
+    *,
+    serve_comb: ParameterSweepItem,
+    bench_comb: ParameterSweepItem,
+    base_path: Path,
+    num_runs: int,
+    dry_run: bool,
+):
+    comb_data = list[dict[str, object]]()
+
+    for run_number in range(num_runs):
+        run_data = run_benchmark(
+            server,
+            bench_cmd,
+            serve_overrides=serve_comb,
+            bench_overrides=bench_comb,
+            run_number=run_number,
+            output_path=_get_comb_run_path(base_path, run_number),
+            dry_run=dry_run,
+        )
+
+        if run_data is not None:
+            comb_data.append(run_data)
+
+    if dry_run:
+        return None
+
+    with _get_comb_run_path(base_path, run_number=None).open("w") as f:
+        json.dump(comb_data, f, indent=4)
+
+    return comb_data
+
+
+def run_combs(
+    serve_cmd: list[str],
+    bench_cmd: list[str],
+    after_bench_cmd: list[str],
+    *,
+    show_stdout: bool,
+    serve_params: ParameterSweep,
+    bench_params: ParameterSweep,
+    output_dir: Path,
+    num_runs: int,
+    dry_run: bool,
+    links: list[tuple[str, str]],
+):
+    all_data = list[dict[str, object]]()
+    for serve_comb in serve_params:
+        with (
+            run_server(
+                serve_cmd,
+                after_bench_cmd,
+                show_stdout=show_stdout,
+                serve_overrides=serve_comb,
+                dry_run=dry_run,
+            )
+            if _comb_needs_server(serve_comb, bench_params, output_dir)
+            else contextlib.nullcontext()
+        ) as server:
+            for bench_comb in bench_params:
+                should_run = all(
+                    serve_key in serve_comb
+                    and bench_key in bench_comb
+                    and serve_comb[serve_key] == bench_comb[bench_key]
+                    for serve_key, bench_key in links
+                )
+                if not should_run:
+                    continue
+                base_path = _get_comb_base_path(output_dir, serve_comb, bench_comb)
+
+                comb_data = run_comb(
+                    server,
+                    bench_cmd,
+                    serve_comb=serve_comb,
+                    bench_comb=bench_comb,
+                    base_path=base_path,
+                    num_runs=num_runs,
+                    dry_run=dry_run,
+                )
+
+                if comb_data is not None:
+                    all_data.extend(comb_data)
+
+    if dry_run:
+        return None
+
+    combined_df = pd.DataFrame.from_records(all_data)
+    combined_df.to_csv(output_dir / "summary.csv")
+
+    return combined_df
+
+
+@dataclass
+class SweepServeArgs:
+    serve_cmd: list[str]
+    bench_cmd: list[str]
+    after_bench_cmd: list[str]
+    show_stdout: bool
+    serve_params: ParameterSweep
+    bench_params: ParameterSweep
+    output_dir: Path
+    num_runs: int
+    dry_run: bool
+    resume: str | None
+    link_vars: list[tuple[str, str]] | None
+
+    parser_name: ClassVar[str] = "serve"
+    parser_help: ClassVar[str] = "Run vLLM server benchmark under multiple settings."
+
+    @classmethod
+    def from_cli_args(cls, args: argparse.Namespace):
+        serve_cmd = shlex.split(args.serve_cmd)
+        bench_cmd = shlex.split(args.bench_cmd)
+        after_bench_cmd = (
+            [] if args.after_bench_cmd is None else shlex.split(args.after_bench_cmd)
+        )
+
+        if args.serve_params:
+            serve_params = ParameterSweep.read_json(args.serve_params)
+        else:
+            # i.e.: run serve_cmd without any modification
+            serve_params = ParameterSweep.from_records([{}])
+
+        if args.bench_params:
+            bench_params = ParameterSweep.read_json(args.bench_params)
+        else:
+            # i.e.: run bench_cmd without any modification
+            bench_params = ParameterSweep.from_records([{}])
+        link_vars = cls.parse_link_vars(args.link_vars)
+        num_runs = args.num_runs
+        if num_runs < 1:
+            raise ValueError("`num_runs` should be at least 1.")
+
+        return cls(
+            serve_cmd=serve_cmd,
+            bench_cmd=bench_cmd,
+            after_bench_cmd=after_bench_cmd,
+            show_stdout=args.show_stdout,
+            serve_params=serve_params,
+            bench_params=bench_params,
+            output_dir=Path(args.output_dir),
+            num_runs=num_runs,
+            dry_run=args.dry_run,
+            resume=args.resume,
+            link_vars=link_vars,
+        )
+
+    @classmethod
+    def add_cli_args(cls, parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
+        parser.add_argument(
+            "--serve-cmd",
+            type=str,
+            required=True,
+            help="The command used to run the server: `vllm serve ...`",
+        )
+        parser.add_argument(
+            "--bench-cmd",
+            type=str,
+            required=True,
+            help="The command used to run the benchmark: `vllm bench serve ...`",
+        )
+        parser.add_argument(
+            "--after-bench-cmd",
+            type=str,
+            default=None,
+            help="After a benchmark run is complete, invoke this command instead of "
+            "the default `ServerWrapper.clear_cache()`.",
+        )
+        parser.add_argument(
+            "--show-stdout",
+            action="store_true",
+            help="If set, logs the standard output of subcommands. "
+            "Useful for debugging but can be quite spammy.",
+        )
+        parser.add_argument(
+            "--serve-params",
+            type=str,
+            default=None,
+            help="Path to JSON file containing parameter combinations "
+            "for the `vllm serve` command. Can be either a list of dicts or a dict "
+            "where keys are benchmark names. "
+            "If both `serve_params` and `bench_params` are given, "
+            "this script will iterate over their Cartesian product.",
+        )
+        parser.add_argument(
+            "--bench-params",
+            type=str,
+            default=None,
+            help="Path to JSON file containing parameter combinations "
+            "for the `vllm bench serve` command. Can be either a list of dicts or "
+            "a dict where keys are benchmark names. "
+            "If both `serve_params` and `bench_params` are given, "
+            "this script will iterate over their Cartesian product.",
+        )
+        parser.add_argument(
+            "-o",
+            "--output-dir",
+            type=str,
+            default="results",
+            help="The directory to which results are written.",
+        )
+        parser.add_argument(
+            "--num-runs",
+            type=int,
+            default=3,
+            help="Number of runs per parameter combination.",
+        )
+        parser.add_argument(
+            "--dry-run",
+            action="store_true",
+            help="If set, prints the commands to run, "
+            "then exits without executing them.",
+        )
+        parser.add_argument(
+            "--resume",
+            type=str,
+            default=None,
+            help="Set this to the name of a directory under `output_dir` (which is a "
+            "timestamp) to resume a previous execution of this script, i.e., only run "
+            "parameter combinations for which there are still no output files.",
+        )
+
+        parser.add_argument(
+            "--link-vars",
+            type=str,
+            default="",
+            help=(
+                "Comma-separated list of linked variables between serve and bench, "
+                "e.g. max_num_seqs=max_concurrency,max_model_len=random_input_len"
+            ),
+        )
+
+        return parser
+
+    @staticmethod
+    def parse_link_vars(s: str) -> list[tuple[str, str]]:
+        if not s:
+            return []
+        pairs = []
+        for item in s.split(","):
+            a, b = item.split("=")
+            pairs.append((a.strip(), b.strip()))
+        return pairs
+
+
+def run_main(args: SweepServeArgs):
+    timestamp = args.resume or datetime.now().strftime("%Y%m%d_%H%M%S")
+    output_dir = args.output_dir / timestamp
+
+    if args.resume and not output_dir.exists():
+        raise ValueError(f"Cannot resume from non-existent directory ({output_dir})")
+
+    try:
+        return run_combs(
+            serve_cmd=args.serve_cmd,
+            bench_cmd=args.bench_cmd,
+            after_bench_cmd=args.after_bench_cmd,
+            show_stdout=args.show_stdout,
+            serve_params=args.serve_params,
+            bench_params=args.bench_params,
+            output_dir=output_dir,
+            num_runs=args.num_runs,
+            dry_run=args.dry_run,
+            links=args.link_vars,
+        )
+    except BaseException as exc:
+        raise RuntimeError(
+            f"The script was terminated early. Use `--resume {timestamp}` "
+            f"to continue the script from its last checkpoint."
+        ) from exc
+
+
+def main(args: argparse.Namespace):
+    run_main(SweepServeArgs.from_cli_args(args))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=SweepServeArgs.parser_help)
+    SweepServeArgs.add_cli_args(parser)
+
+    main(parser.parse_args())

vllm/benchmarks/sweep/serve_sla.py

@@ -0,0 +1,492 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import contextlib
+import json
+import math
+from dataclasses import asdict, dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import ClassVar, Literal, get_args
+
+from typing_extensions import assert_never
+
+from vllm.utils.import_utils import PlaceholderModule
+
+from .param_sweep import ParameterSweep, ParameterSweepItem
+from .serve import SweepServeArgs, run_benchmark, run_server
+from .server import ServerProcess
+from .sla_sweep import SLASweep, SLASweepItem
+from .utils import sanitize_filename
+
+try:
+    import pandas as pd
+except ImportError:
+    pd = PlaceholderModule("pandas")
+
+
+def _get_sla_base_path(
+    output_dir: Path,
+    serve_comb: ParameterSweepItem,
+    bench_comb: ParameterSweepItem,
+):
+    parts = list[str]()
+    if serve_comb:
+        parts.extend(("SERVE-", serve_comb.as_text(sep="-")))
+    if bench_comb:
+        parts.extend(("BENCH-", bench_comb.as_text(sep="-")))
+
+    return output_dir / sanitize_filename("-".join(parts))
+
+
+def _get_sla_iter_path(
+    base_path: Path,
+    sla_comb: SLASweepItem,
+    sla_variable: str,
+    sla_value: int | None,
+):
+    if sla_value is None:
+        prefix = sla_comb.as_text(sep="-")
+        return base_path / f"SLA--{prefix}.json"
+
+    return base_path / f"{sla_variable}={sla_value}"
+
+
+def _get_sla_run_path(iter_path: Path, run_number: int | None):
+    if run_number is None:
+        return iter_path / "summary.json"
+
+    return iter_path / f"run={run_number}.json"
+
+
+def _sla_needs_server(
+    serve_comb: ParameterSweepItem,
+    bench_combs: ParameterSweep,
+    sla_combs: SLASweep,
+    sla_variable: str,
+    output_dir: Path,
+):
+    for bench_comb in bench_combs:
+        base_path = _get_sla_base_path(output_dir, serve_comb, bench_comb)
+        for sla_comb in sla_combs:
+            if not _get_sla_iter_path(
+                base_path,
+                sla_comb,
+                sla_variable,
+                sla_value=None,
+            ).exists():
+                return True
+
+    return False
+
+
+def run_sla(
+    server: ServerProcess | None,
+    bench_cmd: list[str],
+    *,
+    serve_comb: ParameterSweepItem,
+    bench_comb: ParameterSweepItem,
+    iter_path: Path,
+    num_runs: int,
+    dry_run: bool,
+):
+    iter_data = list[dict[str, object]]()
+
+    for run_number in range(num_runs):
+        run_data = run_benchmark(
+            server,
+            bench_cmd,
+            serve_overrides=serve_comb,
+            bench_overrides=bench_comb,
+            run_number=run_number,
+            output_path=_get_sla_run_path(iter_path, run_number),
+            dry_run=dry_run,
+        )
+
+        if run_data is not None:
+            iter_data.append(run_data)
+
+    if dry_run:
+        return None
+
+    with _get_sla_run_path(iter_path, run_number=None).open("w") as f:
+        json.dump(iter_data, f, indent=4)
+
+    return iter_data
+
+
+SLAVariable = Literal["request_rate", "max_concurrency"]
+
+
+def _estimate_sla_value(run_data: dict[str, object], sla_variable: SLAVariable):
+    request_throughput = float(run_data["request_throughput"])  # type: ignore
+    if sla_variable == "request_rate":
+        return request_throughput
+    if sla_variable == "max_concurrency":
+        mean_latency_ms = float(run_data["mean_e2el_ms"])  # type: ignore
+        return request_throughput * mean_latency_ms / 1000
+
+    assert_never(sla_variable)
+
+
+def _estimate_sla_bounds(
+    server: ServerProcess | None,
+    bench_cmd: list[str],
+    *,
+    serve_comb: ParameterSweepItem,
+    bench_comb: ParameterSweepItem,
+    sla_comb: SLASweepItem,
+    base_path: Path,
+    num_runs: int,
+    dry_run: bool,
+    sla_variable: SLAVariable,
+    init_value: int,
+    max_value: int,
+):
+    sla_data = list[dict[str, object]]()
+
+    max_passing: int = 0
+    min_failing: int = 0
+
+    val: int = init_value
+    assert val > 0
+
+    while True:
+        print(f"Testing {sla_variable}: {val} req/s")
+
+        iter_data = run_sla(
+            server,
+            bench_cmd,
+            serve_comb=serve_comb,
+            bench_comb=bench_comb | {sla_variable: val},
+            iter_path=_get_sla_iter_path(base_path, sla_comb, sla_variable, val),
+            num_runs=num_runs,
+            dry_run=dry_run,
+        )
+
+        assert iter_data is not None
+        sla_data.extend(iter_data)
+
+        iter_data_mean = {
+            k: sum(float(run_data[k]) for run_data in iter_data) / len(iter_data)  # type: ignore
+            for k in sla_comb
+        }
+
+        sla_results = [
+            criterion.print_and_validate(iter_data_mean, k)
+            for k, criterion in sla_comb.items()
+        ]
+
+        if all(sla_results):
+            print("SLA criteria are met.")
+            max_passing = val
+            val *= 2
+        else:
+            print("SLA criteria are not met.")
+            min_failing = val
+            break
+
+        if val >= max_value:
+            break
+
+    return sla_data, (max_passing, min_failing)
+
+
+def _find_sla_value(
+    server: ServerProcess | None,
+    bench_cmd: list[str],
+    *,
+    serve_comb: ParameterSweepItem,
+    bench_comb: ParameterSweepItem,
+    sla_comb: SLASweepItem,
+    base_path: Path,
+    num_runs: int,
+    dry_run: bool,
+    sla_variable: SLAVariable,
+    min_value: int,
+    max_value: int,
+):
+    sla_data = list[dict[str, object]]()
+
+    left: int = min_value
+    right: int = max_value
+
+    while True:
+        val = (left + right) // 2
+        print(f"Testing {sla_variable}: {val} req/s")
+
+        iter_data = run_sla(
+            server,
+            bench_cmd,
+            serve_comb=serve_comb,
+            bench_comb=bench_comb | {sla_variable: val},
+            iter_path=_get_sla_iter_path(base_path, sla_comb, sla_variable, val),
+            num_runs=num_runs,
+            dry_run=dry_run,
+        )
+
+        assert iter_data is not None
+        sla_data.extend(iter_data)
+
+        iter_data_mean = {
+            k: sum(float(run_data[k]) for run_data in iter_data) / len(iter_data)  # type: ignore
+            for k in sla_comb
+        }
+
+        sla_results = [
+            criterion.print_and_validate(iter_data_mean, k)
+            for k, criterion in sla_comb.items()
+        ]
+
+        if all(sla_results):
+            print("SLA criteria are met.")
+            left = val
+        else:
+            print("SLA criteria are not met.")
+            right = val
+
+        if right - left <= 1:
+            break
+
+    return sla_data, left
+
+
+def search_sla(
+    server: ServerProcess | None,
+    bench_cmd: list[str],
+    *,
+    serve_comb: ParameterSweepItem,
+    bench_comb: ParameterSweepItem,
+    sla_comb: SLASweepItem,
+    sla_variable: SLAVariable,
+    sla_inf_value: int = 65536,  # The value that represents infinite QPS
+    base_path: Path,
+    num_runs: int,
+    dry_run: bool,
+):
+    print("[SLA START]")
+    print(f"SLA criteria: {sla_comb.as_text()}")
+
+    sla_data_0 = run_sla(
+        server,
+        bench_cmd,
+        serve_comb=serve_comb,
+        bench_comb=bench_comb | {sla_variable: sla_inf_value},
+        iter_path=_get_sla_iter_path(base_path, sla_comb, sla_variable, sla_inf_value),
+        num_runs=num_runs,
+        dry_run=dry_run,
+    )
+    if sla_data_0 is None:
+        assert dry_run
+        print("Omitting SLA search.")
+        print("[SLA END]")
+        return None
+
+    sla_init_value = math.ceil(
+        sum(_estimate_sla_value(item, sla_variable) for item in sla_data_0)
+        / len(sla_data_0)
+    )
+    print(f"Initial {sla_variable} to search: {sla_init_value} req/s.")
+
+    sla_data_1, (sla_min, sla_max) = _estimate_sla_bounds(
+        server,
+        bench_cmd,
+        serve_comb=serve_comb,
+        bench_comb=bench_comb,
+        sla_comb=sla_comb,
+        base_path=base_path,
+        num_runs=num_runs,
+        dry_run=dry_run,
+        sla_variable=sla_variable,
+        init_value=sla_init_value,
+        max_value=sla_inf_value,
+    )
+    print(f"Range of {sla_variable} to search: [{sla_min}, {sla_max}] req/s.")
+
+    sla_data_2, sla_value = _find_sla_value(
+        server,
+        bench_cmd,
+        serve_comb=serve_comb,
+        bench_comb=bench_comb,
+        sla_comb=sla_comb,
+        base_path=base_path,
+        num_runs=num_runs,
+        dry_run=dry_run,
+        sla_variable=sla_variable,
+        min_value=sla_min,
+        max_value=sla_max,
+    )
+
+    sla_data = sla_data_0 + sla_data_1 + sla_data_2
+    print(f"Maximum {sla_variable} for SLA: {sla_value} req/s.")
+
+    with _get_sla_iter_path(
+        base_path,
+        sla_comb,
+        sla_variable,
+        sla_value=None,
+    ).open("w") as f:
+        json.dump(sla_data, f, indent=4)
+
+    print("[SLA END]")
+
+    return sla_data
+
+
+def run_slas(
+    serve_cmd: list[str],
+    bench_cmd: list[str],
+    after_bench_cmd: list[str],
+    *,
+    show_stdout: bool,
+    serve_params: ParameterSweep,
+    bench_params: ParameterSweep,
+    sla_params: SLASweep,
+    sla_variable: SLAVariable,
+    output_dir: Path,
+    num_runs: int,
+    dry_run: bool,
+):
+    if any(bench_comb.has_param(sla_variable) for bench_comb in bench_params):
+        raise ValueError(
+            f"You should not override `{sla_variable}` in `bench_params` in SLA mode, "
+            "since it is supposed to be determined automatically."
+        )
+
+    all_data = list[dict[str, object]]()
+    for serve_comb in serve_params:
+        with (
+            run_server(
+                serve_cmd,
+                after_bench_cmd,
+                show_stdout=show_stdout,
+                serve_overrides=serve_comb,
+                dry_run=dry_run,
+            )
+            if _sla_needs_server(
+                serve_comb,
+                bench_params,
+                sla_params,
+                sla_variable,
+                output_dir,
+            )
+            else contextlib.nullcontext()
+        ) as server:
+            for bench_comb in bench_params:
+                for sla_comb in sla_params:
+                    base_path = _get_sla_base_path(output_dir, serve_comb, bench_comb)
+
+                    comb_data = search_sla(
+                        server,
+                        bench_cmd,
+                        serve_comb=serve_comb,
+                        bench_comb=bench_comb,
+                        sla_comb=sla_comb,
+                        sla_variable=sla_variable,
+                        base_path=base_path,
+                        num_runs=num_runs,
+                        dry_run=dry_run,
+                    )
+
+                    if comb_data is not None:
+                        all_data.extend(comb_data)
+
+    if dry_run:
+        return None
+
+    combined_df = pd.DataFrame.from_records(all_data)
+    combined_df.to_csv(output_dir / "summary.csv")
+
+    return combined_df
+
+
+@dataclass
+class SweepServeSLAArgs(SweepServeArgs):
+    sla_params: SLASweep
+    sla_variable: SLAVariable
+
+    parser_name: ClassVar[str] = "serve_sla"
+    parser_help: ClassVar[str] = "Tune a variable to meet SLAs under multiple settings."
+
+    @classmethod
+    def from_cli_args(cls, args: argparse.Namespace):
+        # NOTE: Don't use super() as `from_cli_args` calls `cls()`
+        base_args = SweepServeArgs.from_cli_args(args)
+
+        if args.sla_params:
+            sla_params = SLASweep.read_json(args.sla_params)
+        else:
+            sla_params = SLASweep.from_records([])
+
+        return cls(
+            **asdict(base_args),
+            sla_params=sla_params,
+            sla_variable=args.sla_variable,
+        )
+
+    @classmethod
+    def add_cli_args(cls, parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
+        parser = super().add_cli_args(parser)
+
+        sla_group = parser.add_argument_group("sla options")
+        sla_group.add_argument(
+            "--sla-params",
+            type=str,
+            required=True,
+            help="Path to JSON file containing a list of SLA constraints to satisfy. "
+            'Each constraint is expressed in `{"<KEY>": "<OP><VALUE>"}` format, '
+            'e.g.: `{"p99_e2el_ms": "<=500"}` means that '
+            "the E2E latency should be less than 500ms 99%% of the time. "
+            "Setting this option runs this script in SLA mode, which searches for "
+            "the maximum `sla_variable` that satisfies the constraints for "
+            "each combination of `serve_params`, `bench_params`, and `sla_params`.",
+        )
+        sla_group.add_argument(
+            "--sla-variable",
+            type=str,
+            choices=get_args(SLAVariable),
+            default="request_rate",
+            help="Whether to tune request rate or maximum concurrency to satisfy "
+            "the SLA constraints.",
+        )
+
+        return parser
+
+
+def run_main(args: SweepServeSLAArgs):
+    timestamp = args.resume or datetime.now().strftime("%Y%m%d_%H%M%S")
+    output_dir = args.output_dir / timestamp
+
+    if args.resume and not output_dir.exists():
+        raise ValueError(f"Cannot resume from non-existent directory ({output_dir})")
+
+    try:
+        return run_slas(
+            serve_cmd=args.serve_cmd,
+            bench_cmd=args.bench_cmd,
+            after_bench_cmd=args.after_bench_cmd,
+            show_stdout=args.show_stdout,
+            serve_params=args.serve_params,
+            bench_params=args.bench_params,
+            sla_params=args.sla_params,
+            sla_variable=args.sla_variable,
+            output_dir=output_dir,
+            num_runs=args.num_runs,
+            dry_run=args.dry_run,
+        )
+    except BaseException as exc:
+        raise RuntimeError(
+            f"The script was terminated early. Use `--resume {timestamp}` "
+            f"to continue the script from its last checkpoint."
+        ) from exc
+
+
+def main(args: argparse.Namespace):
+    run_main(SweepServeSLAArgs.from_cli_args(args))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=SweepServeSLAArgs.parser_help)
+    SweepServeSLAArgs.add_cli_args(parser)
+
+    main(parser.parse_args())

vllm/benchmarks/sweep/server.py

@@ -0,0 +1,114 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import contextlib
+import os
+import signal
+import subprocess
+from types import TracebackType
+
+import requests
+from typing_extensions import Self
+
+
+class ServerProcess:
+    def __init__(
+        self,
+        server_cmd: list[str],
+        after_bench_cmd: list[str],
+        *,
+        show_stdout: bool,
+    ) -> None:
+        super().__init__()
+
+        self.server_cmd = server_cmd
+        self.after_bench_cmd = after_bench_cmd
+        self.show_stdout = show_stdout
+
+    def __enter__(self) -> Self:
+        self.start()
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_value: BaseException | None,
+        exc_traceback: TracebackType | None,
+    ) -> None:
+        self.stop()
+
+    def start(self):
+        # Create new process for clean termination
+        self._server_process = subprocess.Popen(
+            self.server_cmd,
+            start_new_session=True,
+            stdout=None if self.show_stdout else subprocess.DEVNULL,
+            # Need `VLLM_SERVER_DEV_MODE=1` for `_reset_caches`
+            env=os.environ | {"VLLM_SERVER_DEV_MODE": "1"},
+        )
+
+    def stop(self):
+        server_process = self._server_process
+
+        if server_process.poll() is None:
+            # In case only some processes have been terminated
+            with contextlib.suppress(ProcessLookupError):
+                # We need to kill both API Server and Engine processes
+                os.killpg(os.getpgid(server_process.pid), signal.SIGKILL)
+
+    def run_subcommand(self, cmd: list[str]):
+        return subprocess.run(
+            cmd,
+            stdout=None if self.show_stdout else subprocess.DEVNULL,
+            check=True,
+        )
+
+    def after_bench(self) -> None:
+        if not self.after_bench_cmd:
+            self.reset_caches()
+            return
+
+        self.run_subcommand(self.after_bench_cmd)
+
+    def _get_vllm_server_address(self) -> str:
+        server_cmd = self.server_cmd
+
+        for host_key in ("--host",):
+            if host_key in server_cmd:
+                host = server_cmd[server_cmd.index(host_key) + 1]
+                break
+        else:
+            host = "localhost"
+
+        for port_key in ("-p", "--port"):
+            if port_key in server_cmd:
+                port = int(server_cmd[server_cmd.index(port_key) + 1])
+                break
+        else:
+            port = 8000  # The default value in vllm serve
+
+        return f"http://{host}:{port}"
+
+    def reset_caches(self) -> None:
+        server_cmd = self.server_cmd
+
+        # Use `.endswith()` to match `/bin/...`
+        if server_cmd[0].endswith("vllm"):
+            server_address = self._get_vllm_server_address()
+            print(f"Resetting caches at {server_address}")
+
+            res = requests.post(f"{server_address}/reset_prefix_cache")
+            res.raise_for_status()
+
+            res = requests.post(f"{server_address}/reset_mm_cache")
+            res.raise_for_status()
+        elif server_cmd[0].endswith("infinity_emb"):
+            if "--vector-disk-cache" in server_cmd:
+                raise NotImplementedError(
+                    "Infinity server uses caching but does not expose a method "
+                    "to reset the cache"
+                )
+        else:
+            raise NotImplementedError(
+                f"No implementation of `reset_caches` for `{server_cmd[0]}` server. "
+                "Please specify a custom command via `--after-bench-cmd`."
+            )

vllm/benchmarks/sweep/sla_sweep.py

@@ -0,0 +1,132 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import json
+import os
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+
+from typing_extensions import override
+
+
+@dataclass
+class SLACriterionBase(ABC):
+    target: float
+
+    @abstractmethod
+    def validate(self, actual: float) -> bool:
+        """Return `True` if this criterion is met; otherwise `False`."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def format_cond(self, lhs: str) -> str:
+        raise NotImplementedError
+
+    def print_and_validate(
+        self,
+        metrics: dict[str, float],
+        metrics_key: str,
+    ) -> bool:
+        metric = metrics[metrics_key]
+        result = self.validate(metric)
+
+        cond = self.format_cond(f"{metrics_key} = {metric:.2f}")
+        print(f"Validating SLA: {cond} | " + ("PASSED" if result else "FAILED"))
+
+        return result
+
+
+@dataclass
+class SLALessThan(SLACriterionBase):
+    @override
+    def validate(self, actual: float) -> bool:
+        return actual < self.target
+
+    @override
+    def format_cond(self, lhs: str) -> str:
+        return f"{lhs}<{self.target:.2f}"
+
+
+@dataclass
+class SLALessThanOrEqualTo(SLACriterionBase):
+    @override
+    def validate(self, actual: float) -> bool:
+        return actual <= self.target
+
+    @override
+    def format_cond(self, lhs: str) -> str:
+        return f"{lhs}<={self.target:.2f}"
+
+
+@dataclass
+class SLAGreaterThan(SLACriterionBase):
+    @override
+    def validate(self, actual: float) -> bool:
+        return actual > self.target
+
+    @override
+    def format_cond(self, lhs: str) -> str:
+        return f"{lhs}>{self.target:.2f}"
+
+
+@dataclass
+class SLAGreaterThanOrEqualTo(SLACriterionBase):
+    @override
+    def validate(self, actual: float) -> bool:
+        return actual >= self.target
+
+    @override
+    def format_cond(self, lhs: str) -> str:
+        return f"{lhs}>={self.target:.2f}"
+
+
+# NOTE: The ordering is important! Match longer op_keys first
+SLA_CRITERIA: dict[str, type[SLACriterionBase]] = {
+    "<=": SLALessThanOrEqualTo,
+    ">=": SLAGreaterThanOrEqualTo,
+    "<": SLALessThan,
+    ">": SLAGreaterThan,
+}
+
+
+class SLASweep(list["SLASweepItem"]):
+    @classmethod
+    def read_json(cls, filepath: os.PathLike):
+        with open(filepath, "rb") as f:
+            records = json.load(f)
+
+        return cls.from_records(records)
+
+    @classmethod
+    def from_records(cls, records: list[dict[str, str]]):
+        if not isinstance(records, list):
+            raise TypeError(
+                f"The SLA sweep should be a list of dictionaries, "
+                f"but found type: {type(records)}"
+            )
+
+        return cls(SLASweepItem.from_record(record) for record in records)
+
+
+class SLASweepItem(dict[str, SLACriterionBase]):
+    @classmethod
+    def from_record(cls, record: dict[str, str]):
+        sla_criteria: dict[str, SLACriterionBase] = {}
+
+        for metric_key, metric_value in record.items():
+            for op_key in SLA_CRITERIA:
+                if metric_value.startswith(op_key):
+                    sla_criteria[metric_key] = SLA_CRITERIA[op_key](
+                        float(metric_value.removeprefix(op_key))
+                    )
+                    break
+            else:
+                raise ValueError(
+                    f"Invalid operator for "
+                    f"SLA constraint '{metric_key}={metric_value}'. "
+                    f"Valid operators are: {sorted(SLA_CRITERIA)}",
+                )
+
+        return cls(sla_criteria)
+
+    def as_text(self, sep: str = ", ") -> str:
+        return sep.join(v.format_cond(k) for k, v in self.items())

vllm/benchmarks/sweep/utils.py

@@ -0,0 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+def sanitize_filename(filename: str) -> str:
+    return filename.replace("/", "_").replace("..", "__").strip("'").strip('"')