Skip to content

Refactor Dynamo benchmark parsing and phase resolution #1124

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
podkidyshev wants to merge 1 commit into from
+109 −97
Open

Refactor Dynamo benchmark parsing and phase resolution #1124

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
102 changes: 101 additions & 1 deletion src/cloudai/workloads/ai_dynamo/ai_dynamo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,9 +16,10 @@

import csv
import logging
import math
import shlex
from pathlib import Path
from typing import Literal, Optional, cast
from typing import ClassVar, Literal, Mapping, Optional, cast

from pydantic import (
AliasChoices,
Expand All @@ -29,6 +30,7 @@
model_validator,
)

import cloudai.util
from cloudai.core import (
DockerImage,
File,
Expand All @@ -41,6 +43,7 @@
)
from cloudai.models.workload import CmdArgs, TestDefinition
from cloudai.systems.slurm import SlurmSystem
from cloudai.workloads.common.llm_serving import LLMServingBenchReport

AIPERF_ARTIFACTS_DIR = "aiperf_artifacts"
AIPERF_ACCURACY_ARTIFACTS_DIR = "aiperf_accuracy_artifacts"
Expand Down Expand Up @@ -477,6 +480,19 @@ class AIDynamoCmdArgs(CmdArgs):
aiperf_accuracy: AIPerfAccuracy | None = None
workloads: str = "genai_perf.sh"

def resolve_aiperf_phase(self, phase: AIPerfPhase) -> AIPerf:
"""Apply phase overrides using the execution artifact paths."""
base_data = self.aiperf.model_dump(by_alias=True, exclude_none=True)
phase_data = phase.model_dump(by_alias=True, exclude_none=True, exclude_unset=True)
single_phase = self.aiperf_phases is None or len(self.aiperf_phases) == 1

if "artifact-dir-name" not in phase_data and not single_phase:
phase_data["artifact-dir-name"] = f"{self.aiperf.artifact_dir_name}/{phase.name}"
if "report-name" not in phase_data and not single_phase:
phase_data["report-name"] = f"aiperf_{phase.name}_report.csv"

return AIPerf.model_validate(cloudai.util.deep_merge(base_data, phase_data))

@field_validator("workloads", mode="before")
@classmethod
def validate_workloads(cls, v: str) -> str:
Expand Down Expand Up @@ -517,6 +533,16 @@ def installables(self) -> list[Installable]:
]


class AIDynamoBenchReport(LLMServingBenchReport):
"""Normalized AI Dynamo benchmark results used by comparison reports."""

output_throughput: float

@property
def throughput(self) -> float:
return self.output_throughput


class AIDynamoTestDefinition(TestDefinition):
"""Test definition for AI Dynamo."""

Expand All @@ -533,6 +559,80 @@ class AIDynamoTestDefinition(TestDefinition):
success_marker: str = "success-marker.txt"
failure_marker: str = "failure-marker.txt"

_TTFT_NAMES: ClassVar[tuple[str, ...]] = ("Time To First Token (ms)", "Time to First Token (ms)")
_TPOT_NAMES: ClassVar[tuple[str, ...]] = ("Inter Token Latency (ms)",)
_THROUGHPUT_NAMES: ClassVar[tuple[str, ...]] = ("Output Token Throughput (tokens/sec)",)
_REQUEST_COUNT_NAMES: ClassVar[tuple[str, ...]] = ("Request Count (count)", "Request Count", "Request count")

@staticmethod
def _parse_metric_rows(csv_file: Path) -> dict[str, dict[str, str]]:
"""Parse both single-table GenAI-Perf and multi-table AIPerf CSV output."""
metrics: dict[str, dict[str, str]] = {}
header: list[str] | None = None
with csv_file.open(newline="", encoding="utf-8-sig") as f:
for row in csv.reader(f):
if not row or not any(cell.strip() for cell in row):
continue
row = [cell.strip() for cell in row]
if row[0] == "Metric":
header = row
continue
if header is None:
continue
metrics[row[0]] = {column: value for column, value in zip(header[1:], row[1:], strict=False) if value}
return metrics

@staticmethod
def _find_value(
metrics: Mapping[str, Mapping[str, str]], metric_names: tuple[str, ...], value_names: tuple[str, ...]
) -> float | None:
for metric_name in metric_names:
row = metrics.get(metric_name)
if row is None:
continue
for value_name in value_names:
value = row.get(value_name)
if value is None:
continue
try:
return float(value.replace(",", ""))
except ValueError:
continue
return None

@classmethod
def parse_benchmark_csv(cls, csv_file: Path, max_concurrency: int = 1) -> AIDynamoBenchReport | None:
"""Normalize a benchmark CSV into the shared LLM serving result shape."""
try:
metrics = cls._parse_metric_rows(csv_file)
except (OSError, csv.Error) as e:
logging.debug(f"Error parsing AI Dynamo benchmark output {csv_file}: {e}")
return None

ttft_mean = cls._find_value(metrics, cls._TTFT_NAMES, ("avg", "mean", "Value"))
tpot_mean = cls._find_value(metrics, cls._TPOT_NAMES, ("avg", "mean", "Value"))
throughput = cls._find_value(metrics, cls._THROUGHPUT_NAMES, ("avg", "Value"))
if ttft_mean is None or tpot_mean is None or throughput is None:
return None

request_count = cls._find_value(metrics, cls._REQUEST_COUNT_NAMES, ("avg", "Value"))
median_ttft = cls._find_value(metrics, cls._TTFT_NAMES, ("p50", "median"))
p99_ttft = cls._find_value(metrics, cls._TTFT_NAMES, ("p99",))
median_tpot = cls._find_value(metrics, cls._TPOT_NAMES, ("p50", "median"))
p99_tpot = cls._find_value(metrics, cls._TPOT_NAMES, ("p99",))
return AIDynamoBenchReport(
num_prompts=int(request_count or 0),
completed=int(request_count or 0),
mean_ttft_ms=ttft_mean,
median_ttft_ms=median_ttft if median_ttft is not None else math.nan,
p99_ttft_ms=p99_ttft if p99_ttft is not None else math.nan,
mean_tpot_ms=tpot_mean,
median_tpot_ms=median_tpot if median_tpot is not None else math.nan,
p99_tpot_ms=p99_tpot if p99_tpot is not None else math.nan,
output_throughput=throughput,
max_concurrency=max_concurrency,
)

@model_validator(mode="after")
def workload_scripts(self) -> "AIDynamoTestDefinition":
"""Populate prefill/decode args."""
Expand Down
92 changes: 7 additions & 85 deletions src/cloudai/workloads/ai_dynamo/report_generation_strategy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,16 +16,17 @@

from __future__ import annotations

import csv
import logging
import math
from pathlib import Path
from typing import ClassVar, Mapping
from typing import ClassVar

from cloudai.core import METRIC_ERROR, MetricValue, ReportGenerationStrategy
from cloudai.util.lazy_imports import lazy
from cloudai.workloads.ai_dynamo.ai_dynamo import AIDynamoTestDefinition, parse_aiperf_accuracy
from cloudai.workloads.common.llm_serving import LLMServingBenchReport
from cloudai.workloads.ai_dynamo.ai_dynamo import (
AIDynamoBenchReport,
AIDynamoTestDefinition,
parse_aiperf_accuracy,
)


class _AnyMetricSet(list):
Expand All @@ -35,63 +36,12 @@ def __contains__(self, item: object) -> bool:
return True


class AIDynamoBenchReport(LLMServingBenchReport):
"""Normalized AI Dynamo benchmark results used by comparison reports."""

output_throughput: float

@property
def throughput(self) -> float:
return self.output_throughput


class AIDynamoReportGenerationStrategy(ReportGenerationStrategy):
"""Strategy for generating reports from AI Dynamo run directories."""

# Accepts any metric string — get_metric parses "benchmark:metric_name:column" dynamically.
metrics: ClassVar[list[str]] = _AnyMetricSet()

_TTFT_NAMES = ("Time To First Token (ms)", "Time to First Token (ms)")
_TPOT_NAMES = ("Inter Token Latency (ms)",)
_THROUGHPUT_NAMES = ("Output Token Throughput (tokens/sec)",)
_REQUEST_COUNT_NAMES = ("Request Count (count)", "Request Count", "Request count")

@staticmethod
def _parse_metric_rows(csv_file: Path) -> dict[str, dict[str, str]]:
"""Parse both single-table GenAI-Perf and multi-table AIPerf CSV output."""
metrics: dict[str, dict[str, str]] = {}
header: list[str] | None = None
with csv_file.open(newline="", encoding="utf-8-sig") as f:
for row in csv.reader(f):
if not row or not any(cell.strip() for cell in row):
continue
row = [cell.strip() for cell in row]
if row[0] == "Metric":
header = row
continue
if header is None:
continue
metrics[row[0]] = {column: value for column, value in zip(header[1:], row[1:], strict=False) if value}
return metrics

@staticmethod
def _find_value(
metrics: Mapping[str, Mapping[str, str]], metric_names: tuple[str, ...], value_names: tuple[str, ...]
) -> float | None:
for metric_name in metric_names:
row = metrics.get(metric_name)
if row is None:
continue
for value_name in value_names:
value = row.get(value_name)
if value is None:
continue
try:
return float(value.replace(",", ""))
except ValueError:
continue
return None

def _benchmark_csv(self) -> Path | None:
tdef = self.test_run.test
if not isinstance(tdef, AIDynamoTestDefinition):
Expand Down Expand Up @@ -125,35 +75,7 @@ def parse_results(self) -> AIDynamoBenchReport | None:
csv_file = self._benchmark_csv()
if csv_file is None:
return None
try:
metrics = self._parse_metric_rows(csv_file)
except (OSError, csv.Error) as e:
logging.debug(f"Error parsing AI Dynamo benchmark output {csv_file}: {e}")
return None

ttft_mean = self._find_value(metrics, self._TTFT_NAMES, ("avg", "mean", "Value"))
tpot_mean = self._find_value(metrics, self._TPOT_NAMES, ("avg", "mean", "Value"))
throughput = self._find_value(metrics, self._THROUGHPUT_NAMES, ("avg", "Value"))
if ttft_mean is None or tpot_mean is None or throughput is None:
return None

request_count = self._find_value(metrics, self._REQUEST_COUNT_NAMES, ("avg", "Value"))
median_ttft = self._find_value(metrics, self._TTFT_NAMES, ("p50", "median"))
p99_ttft = self._find_value(metrics, self._TTFT_NAMES, ("p99",))
median_tpot = self._find_value(metrics, self._TPOT_NAMES, ("p50", "median"))
p99_tpot = self._find_value(metrics, self._TPOT_NAMES, ("p99",))
return AIDynamoBenchReport(
num_prompts=int(request_count or 0),
completed=int(request_count or 0),
mean_ttft_ms=ttft_mean,
median_ttft_ms=median_ttft if median_ttft is not None else math.nan,
p99_ttft_ms=p99_ttft if p99_ttft is not None else math.nan,
mean_tpot_ms=tpot_mean,
median_tpot_ms=median_tpot if median_tpot is not None else math.nan,
p99_tpot_ms=p99_tpot if p99_tpot is not None else math.nan,
output_throughput=throughput,
max_concurrency=self._max_concurrency(),
)
return AIDynamoTestDefinition.parse_benchmark_csv(csv_file, self._max_concurrency())

def used_gpus_count(self) -> int:
"""Match the GPU accounting used when AI Dynamo produces its result CSV."""
Expand Down
12 changes: 1 addition & 11 deletions src/cloudai/workloads/ai_dynamo/slurm_command_gen_strategy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,6 @@
import yaml
from pydantic import BaseModel, TypeAdapter, ValidationError

import cloudai.util
from cloudai.core import File, GitRepo
from cloudai.systems.slurm import SlurmCommandGenStrategy

Expand Down Expand Up @@ -323,16 +322,7 @@ def _render_aiperf_phase_args(self, resolved_phase: AIPerf, artifact_dir: str) -
return " ".join(part for part in parts if part)

def _resolve_aiperf_phase(self, phase: AIPerfPhase) -> AIPerf:
base_data = self.td.cmd_args.aiperf.model_dump(by_alias=True, exclude_none=True)
phase_data = phase.model_dump(by_alias=True, exclude_none=True, exclude_unset=True)
single_phase = self.td.cmd_args.aiperf_phases is None or len(self.td.cmd_args.aiperf_phases) == 1

if "artifact-dir-name" not in phase_data and not single_phase:
phase_data["artifact-dir-name"] = f"{self.td.cmd_args.aiperf.artifact_dir_name}/{phase.name}"
if "report-name" not in phase_data and not single_phase:
phase_data["report-name"] = f"aiperf_{phase.name}_report.csv"

return AIPerf.model_validate(cloudai.util.deep_merge(base_data, phase_data))
return self.td.cmd_args.resolve_aiperf_phase(phase)

def _render_aiperf_setup_blocks(self, log_message: str, setup_cmd: str | None) -> list[str]:
if not setup_cmd:
Expand Down
Loading