File size: 10,218 Bytes
2facf1f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 | import dataclasses
import re
import time
import os
import sys
import math
from pathlib import Path
from typing import Any
from collections import OrderedDict
import torch.cuda
from utils import set_seed
try:
from task import TestSpec
except ImportError:
TestSpec = dict
from submission import custom_kernel
from reference import check_implementation, generate_input
WARMUP_RUNS = 10
TIMED_RUNS = 100
class PopcornOutput:
def __init__(self, fd: int):
self.file = os.fdopen(fd, 'w')
def __enter__(self):
return self
def __exit__(self, exc_type, exc_val, exc_tb):
self.file.close()
def print(self, *args, **kwargs):
print(*args, **kwargs, file=self.file, flush=True)
def log(self, key, value):
self.print(f"{key}: {value}")
@dataclasses.dataclass
class TestCase:
args: dict
spec: str
def copy_kv_cache(module, kv_cache_shape):
"""
Creates a copy of the KVCache module manually.
"""
copied_module = type(module)(kv_cache_shape)
# Copy parameters
params = OrderedDict()
for name, param in module.named_parameters():
params[name] = param.clone().requires_grad_(param.requires_grad).cuda()
# Copy buffers
buffers = OrderedDict()
for name, buff in module.named_buffers():
print(f"Buff name: {name}, shape: {buff.shape}")
buffers[name] = buff.clone().cuda()
# Assign params and buffers to copied module
copied_module.load_state_dict(params, strict=False)
copied_module.load_state_dict(buffers, strict=False)
copied_module.seq_len = module.seq_len
return copied_module.cuda()
def get_test_cases(file_name: str) -> list[TestCase]:
try:
content = Path(file_name).read_text()
except Exception as E:
print(f"Could not open test file`{file_name}`: {E}", file=sys.stderr)
exit(113)
tests = []
lines = content.splitlines()
match = r"\s*([a-zA-Z]+):\s*([a-zA-Z]+|[+-]?[0-9]+)\s*"
for line in lines:
parts = line.split(";")
case = {}
for part in parts:
matched = re.match(match, part)
if not re.fullmatch(match, part):
print(f"invalid test case: '{line}': '{part}'", file=sys.stderr)
exit(113)
key = matched[1]
val = matched[2]
try:
val = int(val)
except ValueError:
pass
case[key] = val
tests.append(TestCase(spec=line, args=case))
return tests
def warm_up(test: TestCase):
config, data, kv_cache = generate_input(**test.args)
config_copy = copy_config_weights(config)
start = time.perf_counter()
while time.perf_counter() - start < 0.2:
custom_kernel((config_copy, data, kv_cache))
torch.cuda.synchronize()
@dataclasses.dataclass
class Stats:
runs: int
mean: float
std: float
err: float
best: float
worst: float
def calculate_stats(durations: list[int]):
"""
Calculate statistical data from a list of durations.
@param durations: A list of durations in nanoseconds.
@return: A Stats object containing the number of runs, mean, standard deviation, error, best, and worst durations.
"""
runs = len(durations)
total = sum(durations)
best = min(durations)
worst = max(durations)
avg = total / runs
variance = sum(map(lambda x: (x - avg)**2, durations))
std = math.sqrt(variance / (runs - 1))
err = std / math.sqrt(runs)
return Stats(runs=runs, mean=avg, std=std, err=err, best=float(best),
worst=float(worst))
def copy_config_weights(config):
"""
Creates a copy of the Config object with cloned weight tensors.
"""
return dataclasses.replace(
config,
Q_proj_down_weight=config.Q_proj_down_weight.clone().cuda(),
Q_proj_up_weight=config.Q_proj_up_weight.clone().cuda(),
KV_proj_down_weight=config.KV_proj_down_weight.clone().cuda(),
KV_proj_up_weight=config.KV_proj_up_weight.clone().cuda()
)
def run_testing(logger: PopcornOutput, tests: list[TestCase]):
"""
Executes the actual test case code and checks for correctness.
@param logger: A PopcornOutput object used for logging test results.
@param tests: A list of TestCase objects representing the test cases to be executed.
@return: An integer representing the exit status: 0 if all tests pass, otherwise 112.
"""
passed = True
logger.log("test-count", len(tests))
for idx, test in enumerate(tests):
logger.log(f"test.{idx}.spec", test.spec)
config, data, kv_cache = generate_input(**test.args)
kv_cache_copy = copy_kv_cache(kv_cache, config.kv_cache_shape)
torch.cuda.synchronize()
submission_output = custom_kernel((config, data, kv_cache))
torch.cuda.synchronize()
error = check_implementation((config, data, kv_cache_copy), submission_output)
if error:
logger.log(f"test.{idx}.status", "fail")
logger.log(f"test.{idx}.error", error)
passed = False
else:
logger.log(f"test.{idx}.status", "pass")
if passed:
logger.log("check", "pass")
return 0
else:
logger.log("check", "fail")
return 112
def benchmark(test: TestCase, recheck: bool, max_repeats: int, max_time_ns: float) -> Stats | Any:
"""
For a particular test case, check correctness (if applicable) and grab runtime results.
@param test: TestCase object.
@param recheck: Flag for whether to explicitly check functional correctness.
@param max_repeats: Number of trials to repeat.
@param max_time_ns: Timeout time in nanoseconds.
@return: A Stats object for this particular benchmark case or an error if the test fails.
"""
durations = []
# generate input data once
config, data, kv_cache = generate_input(**test.args)
# first, one obligatory correctness check; also triggers triton compile for the given shape
kv_cache_copy = copy_kv_cache(kv_cache, config.kv_cache_shape)
config_copy = copy_config_weights(config)
with torch.no_grad():
output = custom_kernel((config, data, kv_cache))
error = check_implementation((config_copy, data, kv_cache_copy), output)
if error:
return error
# now, do multiple timing runs without further correctness testing
# there is an upper bound of 100 runs, and a lower bound of 3 runs;
# otherwise, we repeat until we either measure at least 10 full seconds,
# or the relative error of the mean is below 1%.
with torch.no_grad():
for i in range(max_repeats):
if recheck:
config, data, kv_cache = generate_input(**test.args)
kv_cache_copy = copy_kv_cache(kv_cache, config.kv_cache_shape)
config_copy = copy_config_weights(config)
torch.cuda.synchronize()
start = time.perf_counter_ns()
output = custom_kernel((config, data, kv_cache))
torch.cuda.synchronize()
end = time.perf_counter_ns()
if recheck:
error = check_implementation((config_copy, data, kv_cache_copy), output)
if error:
return error
del output
durations.append(end-start)
if i > 1:
stats = calculate_stats(durations)
if stats.err / stats.mean < 0.01 or stats.mean * stats.runs > max_time_ns:
break
return calculate_stats(durations)
def run_benchmarking(logger: PopcornOutput, tests: list[TestCase]):
"""
Executes benchmarking code for a CUDA Kernel and logs runtimes.
@param logger: A PopcornOutput object used for logging benchmark results.
@param tests: A list of TestCase objects representing the test cases to be benchmarked.
@return: An integer representing the exit status: 0 if all benchmarks pass, otherwise 112.
"""
warm_up(tests[0])
passed = True
logger.log("benchmark-count", len(tests))
for idx, test in enumerate(tests):
logger.log(f"benchmark.{idx}.spec", test.spec)
result = benchmark(test, False, 100, 10e9)
if isinstance(result, Stats):
for field in dataclasses.fields(Stats):
logger.log(f"benchmark.{idx}.{field.name}", getattr(result, field.name))
else:
passed = False
logger.log(f"benchmark.{idx}.status", "fail")
logger.log(f"benchmark.{idx}.error", result)
if passed:
logger.log("check", "pass")
return 0
else:
logger.log("check", "fail")
return 112
def main():
fd = os.getenv("POPCORN_FD")
if not fd:
return 111
if len(sys.argv) < 3:
return 2
mode = sys.argv[1]
tests = get_test_cases(sys.argv[2])
with PopcornOutput(int(fd)) as logger:
seed = os.getenv("POPCORN_SEED")
seed = int(seed) if seed else 42
set_seed(seed)
if mode == "test":
return run_testing(logger, tests)
if mode == "benchmark":
return run_benchmarking(logger, tests)
if mode == "leaderboard":
warm_up(tests[0])
result = benchmark(tests[-1], True, 100, 30e9)
if isinstance(result, Stats):
logger.log("benchmark-count", 1)
logger.log(f"benchmark.0.spec", tests[-1].spec)
logger.log(f"benchmark.0.runs", result.runs)
logger.log(f"benchmark.0.mean", result.mean)
logger.log(f"benchmark.0.std", result.std)
logger.log(f"benchmark.0.err", result.err)
logger.log("check", "pass")
else:
logger.log("test-count", 1)
logger.log("test.0.status", "fail")
logger.log("test.0.error", str(result)) #TODO: Make sure result implements __str__?
else:
# TODO: Implement script and profile mode
return 2
if __name__ == "__main__":
sys.exit(main()) |