amazon-science/ammo / ai_cli_session/.codex/skills/ammo
ammo
UnverifiedProfile and optimize GPU kernels for vLLM inference on NVIDIA GPUs. Use when targeting a specific model, hardware, dtype, parallelism, and workload to improve latency.
tests/test_new_target_dp_ep.py
View on GitHub#!/usr/bin/env python3
# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: CC-BY-NC-4.0
"""TDD tests for DP / EP pass-through in new_target.py.
Covers:
- --data-parallel-size flag on new_target.py
- --enable-expert-parallel flag on new_target.py
- Injection of --distributed-executor-backend external_launcher when DP > 1
- Backward compatibility: defaults produce empty bench.extra_args
"""
from __future__ import annotations
import argparse
import json
import subprocess
import sys
from pathlib import Path
import pytest
# Ensure the scripts directory is importable (matches test_workload_matrix.py).
_SCRIPTS_DIR = Path(__file__).resolve().parent.parent / "scripts"
if str(_SCRIPTS_DIR) not in sys.path:
sys.path.insert(0, str(_SCRIPTS_DIR))
def _make_fields(**overrides):
"""Build a TargetFields via _default_target_fields with test-friendly defaults.
Mirrors the argparse.Namespace shape new_target.main() produces so we can
exercise the JSON-composition logic directly without invoking argparse.
"""
from new_target import _default_target_fields
ns = argparse.Namespace(
model_id="Qwen/Qwen3-Coder-480B",
hardware="H100",
dtype="fp8",
tp=1,
ep=1,
max_model_len=4096,
input_len=64,
output_len=512,
batch_sizes=[1, 8, 32],
num_iters=5,
noise_tolerance_pct=0.5,
catastrophic_regression_pct=5.0,
data_parallel_size=1,
enable_expert_parallel=False,
)
for k, v in overrides.items():
setattr(ns, k, v)
return _default_target_fields(ns)
def _target_extra_args(**overrides):
"""Build TargetFields + run _target_json, return bench.extra_args list."""
from new_target import _target_json
fields = _make_fields(**overrides)
out = _target_json(fields, Path("/tmp/fake_artifact_dir"))
return out["bench"]["extra_args"]
class TestDefaults:
"""Backward compatibility: old callers with no DP/EP flags."""
def test_defaults_no_dp_ep(self):
"""No DP/EP overrides → extra_args stays empty (backward compat)."""
assert _target_extra_args() == []
def test_dp_1_extra_args_empty(self):
"""Explicit --data-parallel-size 1 → no injection (DP=1 is single-process)."""
assert _target_extra_args(data_parallel_size=1) == []
class TestDpInjection:
"""DP > 1 triggers auto-injection of external_launcher backend."""
def test_dp_2_injects_backend(self):
"""--data-parallel-size 2 → DP flag + external_launcher backend, in order."""
assert _target_extra_args(data_parallel_size=2) == [
"--data-parallel-size", "2",
"--distributed-executor-backend", "external_launcher",
]
def test_dp_4_injects_backend(self):
"""Higher DP values stringify correctly."""
assert _target_extra_args(data_parallel_size=4) == [
"--data-parallel-size", "4",
"--distributed-executor-backend", "external_launcher",
]
class TestEpFlag:
"""--enable-expert-parallel is independent of DP."""
def test_enable_expert_parallel_only(self):
"""EP=True, DP=1 → only the EP flag in extra_args."""
assert _target_extra_args(enable_expert_parallel=True) == [
"--enable-expert-parallel",
]
class TestCombined:
"""DP and EP compose together. Order is deterministic: DP group first."""
def test_dp_and_ep_compose(self):
assert _target_extra_args(data_parallel_size=4, enable_expert_parallel=True) == [
"--data-parallel-size", "4",
"--distributed-executor-backend", "external_launcher",
"--enable-expert-parallel",
]
class TestExistingFieldsUnaffected:
"""Regression: existing workload / target fields are not disturbed by DP/EP."""
def test_dp_preserves_existing_tp_and_batch_sizes(self):
from new_target import _target_json
fields = _make_fields(tp=4, batch_sizes=[1, 8, 32], data_parallel_size=2)
out = _target_json(fields, Path("/tmp/fake_artifact_dir"))
assert out["target"]["tp"] == 4
assert out["workload"]["batch_sizes"] == [1, 8, 32]
assert out["bench"]["extra_args"] == [
"--data-parallel-size", "2",
"--distributed-executor-backend", "external_launcher",
]
class TestArgparseIntegration:
"""End-to-end: run new_target.py as a subprocess to catch argparse wiring bugs.
Direct imports of _target_json bypass argparse, so dest=/action= typos
in add_argument() would silently slip through. One subprocess invocation
guards against that.
"""
def test_cli_flags_produce_expected_extra_args(self, tmp_path):
script = _SCRIPTS_DIR / "new_target.py"
artifact_dir = tmp_path / "artifact"
cmd = [
sys.executable,
str(script),
"--artifact-dir", str(artifact_dir),
"--model-id", "test-model",
"--hardware", "H100",
"--dtype", "fp8",
"--tp", "2",
"--data-parallel-size", "2",
"--enable-expert-parallel",
]
result = subprocess.run(cmd, capture_output=True, text=True)
assert result.returncode == 0, f"stderr: {result.stderr}"
target = json.loads((artifact_dir / "target.json").read_text())
assert target["bench"]["extra_args"] == [
"--data-parallel-size", "2",
"--distributed-executor-backend", "external_launcher",
"--enable-expert-parallel",
]
class TestStateJsonDpPersistence:
"""state.json must record dp parallelism alongside tp / ep."""
def test_state_json_includes_dp_default_1(self):
from new_target import _state_json
fields = _make_fields()
state = _state_json(fields, Path("/tmp/fake_artifact_dir"))
assert state["target"]["dp"] == 1
assert state["campaign"]["schema_version"] == "4.2"
def test_state_json_includes_dp_when_set(self):
from new_target import _state_json
fields = _make_fields(tp=2, ep=1, data_parallel_size=4)
state = _state_json(fields, Path("/tmp/fake_artifact_dir"))
assert state["target"]["dp"] == 4
assert state["target"]["tp"] == 2
assert state["target"]["ep"] == 1
def test_state_json_dp_via_argparse_subprocess(self, tmp_path):
script = _SCRIPTS_DIR / "new_target.py"
artifact_dir = tmp_path / "artifact"
cmd = [
sys.executable,
str(script),
"--artifact-dir", str(artifact_dir),
"--model-id", "test-model",
"--hardware", "H100",
"--dtype", "fp8",
"--tp", "2",
"--data-parallel-size", "2",
]
result = subprocess.run(cmd, capture_output=True, text=True)
assert result.returncode == 0, f"stderr: {result.stderr}"
state = json.loads((artifact_dir / "state.json").read_text())
assert state["target"]["dp"] == 2
assert state["target"]["tp"] == 2
def test_state_schema_allows_legacy_missing_dp_but_defines_optional_dp(self):
schema_path = _SCRIPTS_DIR.parents[2] / "schemas" / "state.schema.json"
schema = json.loads(schema_path.read_text())
target_schema = schema["properties"]["target"]
assert "dp" not in target_schema["required"]
assert target_schema["properties"]["dp"]["minimum"] == 1
if __name__ == "__main__":
pytest.main([__file__, "-v"])
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |