Skip to content

Enable TensorIndexer with all python_direct tests #5781

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

Merged
naoyam merged 14 commits into from
Jan 14, 2026
+202 −406
Merged

Enable TensorIndexer with all python_direct tests #5781

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
562d1f1
Enable TensorIndexer with all python_direct tests
naoyam Jan 9, 2026
f0c8bbe
fix test
naoyam Jan 9, 2026
af1b4b7
fix
naoyam Jan 9, 2026
e4b6f22
cleanup
naoyam Jan 9, 2026
6fb3cc6
test fix
naoyam Jan 9, 2026
0578f53
Simplify IdModel options
naoyam Jan 13, 2026
1d7b328
update
naoyam Jan 13, 2026
ae0dfac
fix
naoyam Jan 13, 2026
e13ad84
Merge branch 'main' into tensorindexer_pytest
naoyam Jan 13, 2026
32362d3
cleanup
naoyam Jan 13, 2026
2b91a06
Merge branch 'main' into tensorindexer_pytest
naoyam Jan 13, 2026
b1965c2
Merge branch 'id_model_option_cleanup' into tensorindexer_pytest
naoyam Jan 13, 2026
be3515d
Merge branch 'main' into id_model_option_cleanup
naoyam Jan 13, 2026
5c9050d
Merge branch 'id_model_option_cleanup' into tensorindexer_pytest
naoyam Jan 13, 2026
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
39 changes: 4 additions & 35 deletions csrc/id_model/utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,52 +30,21 @@ enum class IdModelEnableOption {
inline std::unordered_set<IdModelEnableOption> getIdModelEnabledOptions() {
std::unordered_set<IdModelEnableOption> opts;

if (hasEnableOptionArgument(EnableOption::IdModel, "consumer_index") ||
hasEnableOptionArgument(EnableOption::IdModel, "index") ||
hasEnableOptionArgument(EnableOption::IdModel, "all")) {
if (!hasEnableOptionArgument(EnableOption::IdModel, "predicate_only")) {
opts.insert(IdModelEnableOption::ConsumerIndex);
}

if (hasEnableOptionArgument(EnableOption::IdModel, "producer_index") ||
hasEnableOptionArgument(EnableOption::IdModel, "index") ||
hasEnableOptionArgument(EnableOption::IdModel, "all")) {
opts.insert(IdModelEnableOption::ProducerIndex);
}

if (hasEnableOptionArgument(EnableOption::IdModel, "inline_predicate") ||
hasEnableOptionArgument(EnableOption::IdModel, "predicate") ||
hasEnableOptionArgument(EnableOption::IdModel, "all")) {
if (!hasEnableOptionArgument(EnableOption::IdModel, "index_only")) {
opts.insert(IdModelEnableOption::InlinePredicate);
}

if (hasEnableOptionArgument(EnableOption::IdModel, "unswitch_predicate") ||
hasEnableOptionArgument(EnableOption::IdModel, "predicate") ||
hasEnableOptionArgument(EnableOption::IdModel, "all")) {
opts.insert(IdModelEnableOption::UnswitchPredicate);
}

if (hasEnableOptionArgument(EnableOption::IdModel, "loop") ||
hasEnableOptionArgument(EnableOption::IdModel, "all")) {
if (!hasEnableOptionArgument(EnableOption::IdModel, "predicate_only") &&
!hasEnableOptionArgument(EnableOption::IdModel, "index_only")) {
opts.insert(IdModelEnableOption::Loop);
}

// Loop requires ConsumerIndex, ProducerIndex, InlinePredicate and
// UnswitchPredicate
if (opts.find(IdModelEnableOption::Loop) != opts.end()) {
NVF_ERROR(
opts.find(IdModelEnableOption::ConsumerIndex) != opts.end(),
"ConsumerIndex required for Loop");
NVF_ERROR(
opts.find(IdModelEnableOption::ProducerIndex) != opts.end(),
"ProducerIndex required for Loop");
NVF_ERROR(
opts.find(IdModelEnableOption::InlinePredicate) != opts.end(),
"InlinePredicate required for Loop");
NVF_ERROR(
opts.find(IdModelEnableOption::UnswitchPredicate) != opts.end(),
"UnswitchPredicate required for Loop");
}

return opts;
}

Expand Down
7 changes: 6 additions & 1 deletion python/nvfuser_direct/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -366,10 +366,15 @@ def execute(
# A copy of fusion is created after construction FusionExecutorCache
# Delete the _fusion and reference the fusion inside FusionExecutorCache
del self._fusion

# Add "id_model" as a default enable option
default_enable_options = ["id_model"]
merged_enable_options = default_enable_options + _enable_options
Comment on lines +370 to +372
Copy link
Contributor

@greptile-apps greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Critical Logic Issue: Adding "id_model" to enable options without arguments has no effect.

The IdModel system requires explicit sub-options to be passed as arguments (e.g., "all", "index", "consumer_index", "producer_index", etc.). These are checked via hasEnableOptionArgument(EnableOption::IdModel, "arg") in csrc/id_model/utils.h:30-80.

When you add "id_model" here without arguments, the Python bindings call EnableOptionsGuard::getCurOptions().set(opt.value()) with an empty vector (see python/python_direct/runtime.cpp:284). This means no IdModel features are actually enabled.

Evidence from codebase:

  1. C++ tests use: EnableOption::IdModel, {"all"} (tests/cpp/utils.cpp)
  2. Python tests that need IdModel use: set_env(NVFUSER_ENABLE="id_model(all)") (tests/python/direct/test_with_id_model_indexer.py:183)

Proposed solution:
The Python _enable_options API doesn't support passing arguments. You need to either:

  1. Modify the Python bindings to support syntax like ["id_model:all"] or similar
  2. Use environment variable NVFUSER_ENABLE="id_model(all)" in test setup
  3. Add a new Python API parameter specifically for IdModel options

Without this fix, TensorIndexer will NOT be enabled with python_direct tests as the PR title claims.

Comment on lines +371 to +372
Copy link
Contributor

@greptile-apps greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P2] Potential duplicate options if user passes "id_model" in _enable_options. Consider checking if "id_model" is already in _enable_options before adding it, or use a set to avoid duplicates. While this doesn't cause functional issues (the second set() call just overwrites), it's cleaner to avoid duplicates.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

Comment on lines +370 to +372
Copy link
Contributor

@greptile-apps greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P2] No opt-out mechanism provided. Users cannot disable id_model when using execute() even if they want to. Consider either: (1) checking if "id_model" is in _disable_options and skipping the default, or (2) providing a parameter to control default options. This limits flexibility for users who may need to test without id_model or work around potential id_model bugs.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

Comment on lines +370 to +372
Copy link
Contributor

@greptile-apps greptile-apps bot Jan 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Adding "id_model" as a default enable option means users cannot opt-out of this behavior through the Python API. Consider whether this is intentional or if there should be a mechanism to allow users to disable id_model for testing/debugging purposes.

Potential implications:

  • Users performing A/B testing between old and new indexing cannot easily do so
  • Debugging issues specific to id_model becomes harder without a way to disable it
  • This is a behavioral change that affects all python_direct users

If this is intentional (to force id_model adoption), consider documenting this breaking change clearly. If not, consider checking if "id_model" is already in _enable_options or providing an opt-out mechanism.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!


return self.fec.execute(
inputs,
device=self._get_device_index(device),
_enable_options=_enable_options,
_enable_options=merged_enable_options,
_disable_options=_disable_options,
)

Expand Down