DeepEP Benchmark
DeepEP Benchmark workload (test_template_name is DeepEP) allows users to execute DeepEP (Deep Expert Parallelism) MoE (Mixture of Experts) benchmarks within the CloudAI framework.
Overview
DeepEP is a benchmark for measuring the performance of MoE models with distributed expert parallelism. It supports:
-
Two operation modes: Standard and Low-Latency
-
Multiple data types: bfloat16 and FP8
-
Flexible network configurations: With or without NVLink
-
Configurable model parameters: Experts, tokens, hidden size, top-k
-
Performance profiling: Kineto profiler support
Usage Example
Test TOML example (Standard Mode):
Test TOML example (Low-Latency Mode):
Test Scenario example:
Test-in-Scenario example:
API Documentation
Command Arguments
class cloudai.workloads.deepep.deepep.DeepEPCmdArgs(
*,
docker_image_url: str,
subtest_name: Literal[‘test_internode’, ‘test_intranode’, ‘test_low_latency’, ‘test_ep’] = ‘test_internode’,
deep_ep_root: str = ‘/workspace/DeepEP’,
legacy_tests_root: str | None = None,
elastic_tests_root: str | None = None,
python_executable: str = ‘python’,
num_processes: int = 8,
num_tokens: int = 4096,
hidden: int = 7168,
num_topk: int = 8,
num_experts: int = 256,
num_topk_groups: int | None = None,
allow_mnnvl: bool = False,
test_ll_compatibility: bool = False,
pressure_test_mode: int = 0,
pressure_test: bool = False,
shrink_test: bool = False,
disable_nvlink: bool = False,
use_logfmt: bool = False,
num_sms: int = 0,
num_qps: int = 0,
num_allocated_qps: int = 0,
num_gpu_timeout_secs: int = 100,
num_cpu_timeout_secs: int = 100,
sl_idx: int = 0,
do_cpu_sync: int = 1,
allow_hybrid_mode: int = 1,
allow_multiple_reduction: int = 1,
prefer_overlap_with_compute: int = 0,
deterministic: bool = False,
seed: int = 0,
skip_check: bool = False,
skip_perf_test: bool = False,
do_pressure_test: bool = False,
reuse_elastic_buffer: bool = False,
test_first_only: bool = False,
unbalanced_ratio: float = 1.0,
precise_unbalanced_ratio: bool = False,
masked_ratio: float = 0.0,
dump_profile_traces: str = ”,
ignore_local_traffic: bool = False,
**extra_data: Any,
)
Bases: CmdArgs
Command arguments for the official DeepEP test scripts.
Test Definition
class cloudai.workloads.deepep.deepep.DeepEPTestDefinition(*, name: str, description: str, test_template_name: str, cmd_args: ~cloudai.workloads.deepep.deepep.DeepEPCmdArgs, dse_excluded_args: list[str] = <factory>, extra_env_vars: dict[str, str | ~typing.List[str]] = {}, extra_cmd_args: dict[str, str] = {}, extra_container_mounts: list[str] = [], git_repos: list[~cloudai._core.installables.git_repo.GitRepo] = [], nsys: ~cloudai.models.workload.NsysConfiguration | None = None, predictor: ~cloudai.models.workload.PredictorConfig | None = None, training_report: ~cloudai.models.workload.TrainingReportConfig | None = None, agent: str = ‘grid_search’, agent_steps: int = 1, agent_metrics: list[str] = [‘default’], agent_reward_function: str = ‘inverse’, agent_config: dict[str, ~typing.Any] | None = None, env_params: dict[str, ~cloudai.configurator.env_params.EnvParamSpec] = <factory>)
Bases: TestDefinition
Test object for official DeepEP v1/v2 test scripts.
property is_domain_randomization_enabled: bool**
at least one env_params annotation.
Type:
Whether the config declares domain randomization
is_dse_excluded_arg(path: str) → bool
Return whether a dot-separated cmd_args path should be ignored by DSE.
is_env_sampled(cmd_args_path: str) → bool
Whether a cmd_args field is env-sampled (env draws it per trial, not the agent).
validator validate_env_params » * all fields*
Validate env_params annotations against cmd_args.
env_params is an annotation: each key names a cmd_args field whose value is the candidate set (the single source of truth), and the entry carries only how to sample. So each key must name a real cmd_args field whose value is a candidate list; a scalar is already fixed, so annotating it is a meaningless label and is rejected here. When weights are declared, the list needs >= 2 values and the weights must align 1:1 with it. Sampling, persistence, the per-trial cmd_args overlay, and the cache key all live in CloudAIGymEnv; keeping this shape check in core lets the overlay stay agent- and workload-agnostic rather than re-implemented per workload.