Remove EnableOption::IdModel

[naoyam]

The option no longer makes any difference since #5828

[naoyam]

!test

[github-actions]

Review updated until commit 4641943

Description

• Remove EnableOption::IdModel from enum and available options

• Enable TensorIndexer by default in IdModelOptions constructor

• Remove explicit IdModel enable calls from test files

• Update Python bindings to remove id_model from default options

Changes walkthrough

	Relevant files
Enhancement	5 files options.h Remove IdModel from EnableOption enum                                        +0/-1      options.cpp Remove id_model from available options map                              +0/-1      id_model_options.h Enable TensorIndexer by default                                                    +2/-5      lower2device.cpp Remove explicit TensorIndexer enable call                                +0/-1      __init__.py Remove id_model from Python default options                            +2/-5
Tests	3 files utils.cpp Remove IdModel enable from test setup                                        +0/-1      conftest.py Remove id_model from pytest environment setup                        +2/-1      test_with_id_model_indexer.py Remove explicit id_model environment setting                          +2/-3
Additional files	57 files test_abstract_tensor.cpp +0/-1      test_allocation_domain.cpp +1/-7      test_allocation_order_inference.cpp +1/-7      test_argsort.cpp +1/-7      test_bfs.cpp +0/-1      test_circular_buffering.cpp +0/-14    test_circular_buffering_ping_pong.cpp +0/-1      test_cluster.cpp +0/-1      test_combined_inner_outer_reduction.cpp +0/-1      test_compute_with.cpp +1/-7      test_contiguity_id_model.cpp +0/-1      test_evaluator.cpp +1/-1      test_external_src.cpp +1/-1      test_gather.cpp +0/-1      test_gpu1.cpp +1/-10    test_gpu2.cpp +1/-7      test_gpu3.cpp +1/-9      test_greedy.cpp +1/-7      test_index_put.cpp +0/-1      test_index_select.cpp +1/-7      test_indexing.cpp +4/-57    test_indexing_advanced.cpp +21/-39  test_inlining.cpp +1/-7      test_matmul.cpp +0/-3      test_matmul_scheduler.cpp +0/-3      test_memory.cpp +1/-5      test_mma.cpp +0/-4      test_moe.cpp +0/-1      test_move_pad.cpp +0/-1      test_move_repeat_forward.cpp +1/-7      test_move_split_cat.cpp +1/-7      test_outer_reduction.cpp +1/-7      test_persistent_buffer.cpp +1/-8      test_pointwise.cpp +0/-1      test_polymorphic_value.cpp +1/-1      test_predicate_elimination.cpp +1/-7      test_reduction.cpp +1/-7      test_reduction_pointwise.cpp +1/-7      test_remove_bcast_squeeze.cpp +1/-7      test_replay.cpp +1/-7      test_reshape.cpp +1/-8      test_resize.cpp +0/-40    test_rope.cpp +0/-1      test_scalar_hoisting.cpp +1/-1      test_scatter.cpp +0/-2      test_select.cpp +1/-7      test_serial_gridreduce.cpp +1/-7      test_smem_reuse.cpp +1/-1      test_stream.cpp +1/-3      test_swizzle.cpp +1/-1      test_tensor_factories.cpp +1/-1      test_topk.cpp +1/-7      test_transpose.cpp +0/-2      test_unary.cpp +0/-1      test_vectorization.cpp +1/-8      test_welford.cpp +1/-7      utils.h +0/-2

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

Default Behavior Change The default value of tensor_indexer_enabled_ has been changed from false to true. This is a significant behavioral change that should be verified to not cause regressions in existing functionality. bool tensor_indexer_enabled_ = true; API Change Impact The removal of default "id_model" enable option may affect existing Python code that relies on this being explicitly set. Need to verify backward compatibility. # Note: id_model is now enabled by default in C++, so no need to add it here return self.fec.execute( inputs, device=self._get_device_index(device), _enable_options=_enable_options, _disable_options=_disable_options, ) Test Environment Configuration The change from "id_model,id_model_extra_validation" to just "id_model_extra_validation" in default options should be verified to maintain proper test coverage. # Note: id_model is now enabled by default, so we only need extra validation existing = os.environ.get("NVFUSER_ENABLE", "") new_options = "id_model_extra_validation" if existing: os.environ["NVFUSER_ENABLE"] = f"{existing},{new_options}"

Test failures (partial, pipeline still running)

• (Medium, 2) Thunder NVFuser nanoGPT autograd scalar mismatch in thunder/tests/test_networks on CUDA

  Test Name	A100	H100	Source
  thunder.tests.test_networks.test_nanogpt_complete_autograd_nvfuser_cuda_thunder.dtypes.float32	❌	❌

[naoyam]

!test

[greptile-apps]

Greptile Overview

Greptile Summary

This PR removes the EnableOption::IdModel feature flag, making the IdModel indexer (TensorIndexer) enabled by default instead of opt-in.

Key Changes:

• Removed EnableOption::IdModel enum value and string mapping from options system
• Changed IdModelOptions::tensor_indexer_enabled_ default from false to true
• Removed constructor that checked isOptionEnabled(EnableOption::IdModel)
• Removed redundant setTensorIndexer(true) call in lower2device.cpp (now default)
• Updated Python bindings to remove id_model from default enable options
• Cleaned up 60+ test files by removing explicit EnableOption::IdModel enablement
• Converted parameterized tests (IdModel vs Legacy) to regular tests (IdModel only)

Impact:
This is a significant architectural change that makes IdModel the default and only indexing approach. The TensorIndexer can still be disabled for unsupported fusions via TensorIndexer::isSupported() check. This simplifies the codebase by removing the dual-mode indexing system.

Confidence Score: 5/5

• This PR is safe to merge - it's a clean refactoring that makes IdModel the default
• All changes are consistent and follow a clear pattern: removing opt-in flag and making IdModel default. The fallback mechanism for unsupported fusions is preserved. Test changes are mechanical and thorough.
• No files require special attention

Important Files Changed

Filename	Overview
csrc/options.h	Removed IdModel enum value from EnableOption - clean removal
csrc/options.cpp	Removed id_model string mapping from enable options - clean removal
csrc/device_lower/id_model_options.h	Removed constructor, changed default to enable TensorIndexer by default
csrc/device_lower/lower2device.cpp	Removed redundant setTensorIndexer(true) call (now default)
python/nvfuser_direct/init.py	Removed id_model from default enable options (now enabled by default in C++)
tests/cpp/utils.h	Removed EnableOption::IdModel from test setup (BlackwellBase, TmaBase)
tests/cpp/test_indexing.cpp	Removed EnableOption::IdModel calls, converted parameterized tests to regular tests
tests/cpp/test_indexing_advanced.cpp	Removed parameterized test instantiation, converted to regular tests

Sequence Diagram

sequenceDiagram
    participant User
    participant Options
    participant IdModelOptions
    participant Lower2Device
    participant TensorIndexer

    Note over Options,TensorIndexer: Before: IdModel was an opt-in feature
    User->>Options: Set EnableOption::IdModel
    Options->>IdModelOptions: Constructor checks isOptionEnabled()
    IdModelOptions->>IdModelOptions: tensor_indexer_enabled = false by default
    Lower2Device->>IdModelOptions: setTensorIndexer(true)
    IdModelOptions->>TensorIndexer: Use TensorIndexer

    Note over Options,TensorIndexer: After: IdModel is always enabled
    User->>Lower2Device: Compile fusion
    Lower2Device->>IdModelOptions: getIdModelOptions()
    IdModelOptions->>IdModelOptions: tensor_indexer_enabled = true by default
    alt Fusion not supported
        Lower2Device->>IdModelOptions: setTensorIndexer(false)
    end
    IdModelOptions->>TensorIndexer: Use TensorIndexer (unless unsupported)

Loading

[greptile-apps]

_{8 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}