Files
pytorch/tools/testing/test_selections.py
PyTorch MergeBot 9eb8fae02d Revert "Fix round robin sharding (#121022)"
This reverts commit effdea5fc62c6bf13cb8035f7bfcc205f05a8b6a.

Reverted https://github.com/pytorch/pytorch/pull/121022 on behalf of https://github.com/clee2000 due to made sharding really uneven ([comment](https://github.com/pytorch/pytorch/pull/121022#issuecomment-1986552662))
2024-03-08 23:16:24 +00:00

263 lines
9.5 KiB
Python

import math
import os
import subprocess
from pathlib import Path
from typing import Callable, Dict, FrozenSet, List, Optional, Sequence, Tuple
from tools.stats.import_test_stats import get_disabled_tests, get_slow_tests
from tools.testing.test_run import ShardedTest, TestRun
REPO_ROOT = Path(__file__).resolve().parent.parent.parent
IS_MEM_LEAK_CHECK = os.getenv("PYTORCH_TEST_CUDA_MEM_LEAK_CHECK", "0") == "1"
# NUM_PROCS_FOR_SHARDING_CALC must remain consistent across all shards of a job
# to ensure that sharding is consistent, NUM_PROCS is the actual number of procs
# used to run tests. If they are not equal, the only consequence should be
# unequal shards.
IS_ROCM = os.path.exists("/opt/rocm")
NUM_PROCS = 1 if IS_MEM_LEAK_CHECK else 2
NUM_PROCS_FOR_SHARDING_CALC = NUM_PROCS if not IS_ROCM or IS_MEM_LEAK_CHECK else 2
THRESHOLD = 60 * 10 # 10 minutes
# See Note [ROCm parallel CI testing]
# Special logic for ROCm GHA runners to query number of GPUs available.
# torch.version.hip was not available to check if this was a ROCm self-hosted runner.
# Must check for ROCm runner in another way. We look for /opt/rocm directory.
if IS_ROCM and not IS_MEM_LEAK_CHECK:
try:
# This is the same logic used in GHA health check, see .github/templates/common.yml.j2
lines = (
subprocess.check_output(["rocminfo"], encoding="ascii").strip().split("\n")
)
count = 0
for line in lines:
if " gfx" in line:
count += 1
assert count > 0 # there must be at least 1 GPU
# Limiting to 8 GPUs(PROCS)
NUM_PROCS = min(count, 8)
except subprocess.CalledProcessError as e:
# The safe default for ROCm GHA runners is to run tests serially.
NUM_PROCS = 1
class ShardJob:
def __init__(self) -> None:
self.serial: List[ShardedTest] = []
self.parallel: List[ShardedTest] = []
def get_total_time(self) -> float:
procs = [0.0 for _ in range(NUM_PROCS_FOR_SHARDING_CALC)]
for test in self.parallel:
min_index = procs.index(min(procs))
procs[min_index] += test.get_time()
time = max(procs) + sum(test.get_time() for test in self.serial)
return time
def convert_to_tuple(self) -> Tuple[float, List[ShardedTest]]:
return (self.get_total_time(), self.serial + self.parallel)
def get_with_pytest_shard(
tests: Sequence[TestRun],
test_file_times: Dict[str, float],
test_class_times: Optional[Dict[str, Dict[str, float]]],
) -> List[ShardedTest]:
sharded_tests: List[ShardedTest] = []
for test in tests:
duration = get_duration(test, test_file_times, test_class_times or {})
if duration and duration > THRESHOLD:
num_shards = math.ceil(duration / THRESHOLD)
for i in range(num_shards):
sharded_tests.append(
ShardedTest(test, i + 1, num_shards, duration / num_shards)
)
else:
sharded_tests.append(ShardedTest(test, 1, 1, duration))
return sharded_tests
def get_duration(
test: TestRun,
test_file_times: Dict[str, float],
test_class_times: Dict[str, Dict[str, float]],
) -> Optional[float]:
file_duration = test_file_times.get(test.test_file, None)
if test.is_full_file():
return file_duration
def get_duration_for_classes(
test_file: str, test_classes: FrozenSet[str]
) -> Optional[float]:
duration: float = 0
for test_class in test_classes:
class_duration = test_class_times.get(test_file, {}).get(test_class, None)
if class_duration is None:
return None
duration += class_duration
return duration
included = test.included()
excluded = test.excluded()
included_classes_duration = get_duration_for_classes(test.test_file, included)
excluded_classes_duration = get_duration_for_classes(test.test_file, excluded)
if included_classes_duration is None or excluded_classes_duration is None:
# Didn't get the time for all classes, so time is unknown
return None
if included:
return included_classes_duration
assert (
excluded
), f"TestRun {test} is not full file but doesn't have included or excluded classes"
if file_duration is None:
return None
return file_duration - excluded_classes_duration
def shard(
sharded_jobs: List[ShardJob],
tests: Sequence[TestRun],
test_file_times: Dict[str, float],
test_class_times: Dict[str, Dict[str, float]],
estimated_time_limit: Optional[float] = None,
sort_by_time: bool = True,
serial: bool = False,
) -> None:
if len(sharded_jobs) == 0:
assert len(tests) == 0, "No shards provided but there are tests to shard"
return
# Modifies sharded_jobs in place
known_tests = tests
unknown_tests = []
if sort_by_time:
known_tests = [
x
for x in tests
if get_duration(x, test_file_times, test_class_times) is not None
]
unknown_tests = [x for x in tests if x not in known_tests]
assert (
unknown_tests == [] or serial
), f"Attmempting to parallelize unknown tests {unknown_tests}"
del tests
known_tests = get_with_pytest_shard(known_tests, test_file_times, test_class_times)
if sort_by_time:
known_tests = sorted(known_tests, key=lambda j: j.get_time(), reverse=True)
def _shard_serial(tests: List[ShardedTest], sharded_jobs: List[ShardJob]) -> None:
assert estimated_time_limit is not None, "Estimated time limit must be provided"
new_sharded_jobs = sharded_jobs
for test in tests:
if (
len(sharded_jobs) > 1
and sharded_jobs[-1].get_total_time() > estimated_time_limit
):
new_sharded_jobs = sharded_jobs[:-1]
min_sharded_job = min(new_sharded_jobs, key=lambda j: j.get_total_time())
min_sharded_job.serial.append(test)
def _shard_parallel(tests: List[ShardedTest], sharded_jobs: List[ShardJob]) -> None:
for test in tests:
min_sharded_job = min(sharded_jobs, key=lambda j: j.get_total_time())
min_sharded_job.parallel.append(test)
if serial:
_shard_serial(known_tests, sharded_jobs)
else:
_shard_parallel(known_tests, sharded_jobs)
# Round robin the unknown jobs starting with the smallest shard
num_shards = len(sharded_jobs)
index = min(range(num_shards), key=lambda i: sharded_jobs[i].get_total_time())
for unknown_test in unknown_tests:
sharded_jobs[index].serial.append(ShardedTest(unknown_test, 1, 1, None))
index = (index + 1) % num_shards
return
def calculate_shards(
num_shards: int,
tests: Sequence[TestRun],
test_file_times: Dict[str, float],
test_class_times: Optional[Dict[str, Dict[str, float]]],
must_serial: Optional[Callable[[str], bool]] = None,
sort_by_time: bool = True,
) -> List[Tuple[float, List[ShardedTest]]]:
must_serial = must_serial or (lambda x: True)
test_class_times = test_class_times or {}
serial_tests = [
test
for test in tests
if get_duration(test, test_file_times, test_class_times) is None
or must_serial(test.test_file)
]
parallel_tests = [test for test in tests if test not in serial_tests]
serial_time = sum(
get_duration(test, test_file_times, test_class_times) or 0
for test in serial_tests
)
parallel_time = sum(
get_duration(test, test_file_times, test_class_times) or 0
for test in parallel_tests
)
total_time = serial_time + parallel_time / NUM_PROCS_FOR_SHARDING_CALC
estimated_time_per_shard = total_time / num_shards
# Separate serial tests from parallel tests as much as possible to maximize
# parallelism by putting all the serial tests on the first num_serial_shards
# shards. The estimated_time_limit is the estimated time it should take for
# the least filled serial shard. Ex if we have 8 min of serial tests, 20 min
# of parallel tests, 6 shards, and 2 procs per machine, we would expect each
# machine to take 3 min and should aim for 3 serial shards, with shards 1
# and 2 taking 3 min and shard 3 taking 2 min. The estimated time limit
# would be 2 min. This ensures that the first few shard contains as many
# serial tests as possible and as few parallel tests as possible. The least
# filled/last (in the example, the 3rd) shard may contain a lot of both
# serial and parallel tests.
estimated_time_limit = 0.0
if estimated_time_per_shard != 0:
estimated_time_limit = serial_time % estimated_time_per_shard
if estimated_time_limit <= 0.01:
estimated_time_limit = estimated_time_per_shard
if total_time == 0:
num_serial_shards = num_shards
else:
num_serial_shards = max(math.ceil(serial_time / total_time * num_shards), 1)
sharded_jobs = [ShardJob() for _ in range(num_shards)]
shard(
sharded_jobs[:num_serial_shards],
serial_tests,
test_file_times,
test_class_times,
estimated_time_limit=estimated_time_limit,
sort_by_time=sort_by_time,
serial=True,
)
shard(
sharded_jobs,
parallel_tests,
test_file_times,
test_class_times,
sort_by_time=sort_by_time,
serial=False,
)
return [job.convert_to_tuple() for job in sharded_jobs]
def get_test_case_configs(dirpath: str) -> None:
get_slow_tests(dirpath=dirpath)
get_disabled_tests(dirpath=dirpath)