Skip to content

eval: hand kernel a fresh input each timed iteration (close result-replay in scored path) #501

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

Draft
nikhilbarhate99 wants to merge 1 commit into
base: main
Choose a base branch
from
+14 −0
Draft

eval: hand kernel a fresh input each timed iteration (close result-replay in scored path) #501

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
14 changes: 14 additions & 0 deletions examples/eval.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -256,6 +256,7 @@ def _run_single_benchmark(test: TestCase, recheck: bool, max_repeats: int, max_t
durations = []
# generate input data once
data = generate_input(**test.args)
bench_template = data # kept as a template for fresh per-iteration copies (anti-replay)
check_copy = _clone_data(data, 0)
# first, one obligatory correctness check
output = custom_kernel(data)
Expand All @@ -268,6 +269,14 @@ def _run_single_benchmark(test: TestCase, recheck: bool, max_repeats: int, max_t
# otherwise, we repeat until we either measure at least 10 full seconds,
# or the relative error of the mean is below 1%.

# ANTI-REPLAY: even in the non-recheck (ranked/benchmark) path, hand the kernel a FRESH
# tensor object every timed iteration. Without this, the same `data` object is reused, so a
# submission can cache on the first call (keyed on object identity / .data_ptr() / ._version)
# and return the cached output with zero compute for all timed iterations -- the dominant
# "result replay" hack family. The clone happens OUTSIDE the timed region (before
# start_event.record()), so it does not inflate the measured kernel time. We retain the two
# most recent inputs so the caching allocator cannot immediately recycle a freed data_ptr.
_recent_inputs = []
bm_start_time = time.perf_counter_ns()
for i in range(max_repeats):
if recheck:
Expand All @@ -277,6 +286,11 @@ def _run_single_benchmark(test: TestCase, recheck: bool, max_repeats: int, max_t

data = generate_input(**test.args)
check_copy = _clone_data(data, 0)
else:
data = _clone_data(bench_template, 0)
_recent_inputs.append(data)
if len(_recent_inputs) > 2:
_recent_inputs.pop(0)
torch.cuda.synchronize()
start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
Expand Down
Loading