-
Notifications
You must be signed in to change notification settings - Fork 26.3k
Closed
Closed
Copy link
Labels
module: aotdispatchumbrella label for AOTAutograd issuesumbrella label for AOTAutograd issuesmodule: compile-timeCompilation mechanism or time spent in (re)compilation, tracing, startupCompilation mechanism or time spent in (re)compilation, tracing, startupmodule: pt2-dispatcherPT2 dispatcher-related issues (e.g., aotdispatch, functionalization, faketensor, custom-op,PT2 dispatcher-related issues (e.g., aotdispatch, functionalization, faketensor, custom-op,oncall: pt2triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module
Description
🐛 Describe the bug
Repro:
TORCH_COMPILE_CPROFILE=1 PYTHONPATH=benchmarks/dynamo/pr_time_benchmarks/ python benchmarks/dynamo/pr_time_benchmarks/benchmarks/update_hint_benchmark.py /tmp/b.csv
with N in the benchmark set to 1000
Quadratic case is here:
outs_with_identical_metadata_that_require_grad = (
[]
if not isinstance(o, Tensor)
else [
curr
for curr in out_storage_to_tensors[curr_storage]
if has_same_metadata(o, curr)
and curr.requires_grad
and o is not curr
]
)
Here's the profile:
Versions
main
cc @chauhang @penguinwu @oulgen @jamesjwu @aorenste @anijain2305 @laithsakka @zou3519 @bdhirsh @yf225
Metadata
Metadata
Assignees
Labels
module: aotdispatchumbrella label for AOTAutograd issuesumbrella label for AOTAutograd issuesmodule: compile-timeCompilation mechanism or time spent in (re)compilation, tracing, startupCompilation mechanism or time spent in (re)compilation, tracing, startupmodule: pt2-dispatcherPT2 dispatcher-related issues (e.g., aotdispatch, functionalization, faketensor, custom-op,PT2 dispatcher-related issues (e.g., aotdispatch, functionalization, faketensor, custom-op,oncall: pt2triagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate moduleThis issue has been looked at a team member, and triaged and prioritized into an appropriate module