Greptile Overview

Greptile Summary

Enables the TensorIndexer (IdModel) by default for all Python direct tests by adding "id_model" as a default enable option in FusionDefinition.execute(). Simplifies the C++ option parsing logic in getIdModelEnabledOptions() by switching from an explicit opt-in model (checking for specific arguments like "consumer_index", "producer_index", etc.) to an opt-out model using "predicate_only" and "index_only" flags, removing redundant validation checks.

Confidence Score: 4/5

• Safe to merge - clean refactoring that enables IdModel by default for Python tests
• The changes are straightforward and follow the pattern established in the previous PR (#5724) that enabled TensorIndexer for C++ tests. The C++ logic simplification removes complex conditional checks and validation code, making it easier to maintain. The Python change simply adds "id_model" to the default enable options list, which is non-breaking since users can still override via _enable_options/_disable_options parameters. No logical errors or edge cases detected.
• No files require special attention

Important Files Changed

File Analysis

Sequence Diagram

sequenceDiagram
    participant Python as Python FusionDefinition.execute()
    participant FEC as FusionExecutorCache
    participant CPP as getIdModelEnabledOptions()
    
    Python->>Python: "Add 'id_model' to default_enable_options"
    Python->>FEC: "execute(merged_enable_options)"
    FEC->>CPP: "Check IdModel enable option"
    CPP->>CPP: "Check !predicate_only → Enable Index opts"
    CPP->>CPP: "Check !index_only → Enable Predicate opts"
    CPP->>CPP: "Check both unset → Enable Loop opt"
    CPP-->>FEC: "Return enabled options set"
    FEC-->>Python: "Execute with TensorIndexer enabled"

Greptile Overview

Greptile Summary

Enables TensorIndexer by default for all python_direct tests by making IdModel always build and removing the disable option. The PR removes the build_id_model_ flag from IdModelOptions and the DisableOption::IdModel enum, unconditionally building IdModel during lowering. Python's FusionDefinition.execute() now adds "id_model" as a default enable option.

Confidence Score: 2/5

• Breaking change to IdModel option semantics requires test updates
• The PR makes significant changes to how IdModel options are interpreted. The new getIdModelEnabledOptions() function changes from an opt-in model (explicit arguments like "all", "index") to a negation-based model ("predicate_only", "index_only"). This breaks existing usage in tests/python/direct/test_with_id_model_indexer.py which uses NVFUSER_ENABLE=id_model(all), but the "all" argument is no longer recognized. The old code would enable specific options based on explicit arguments, while the new code enables everything by default unless restricted. This semantic change warrants careful testing before merge.
• csrc/id_model/utils.h - breaks backward compatibility with option arguments; tests/python/direct/test_with_id_model_indexer.py - needs update for new semantics

Important Files Changed

File Analysis

Sequence Diagram

sequenceDiagram
    participant Python as Python FusionDefinition
    participant FEC as FusionExecutorCache
    participant Lower as GpuLower
    participant IdModel as IdModel
    participant Options as IdModelOptions
    
    Python->>Python: "Add 'id_model' to enable options"
    Python->>FEC: "execute(inputs, enable=['id_model'])"
    FEC->>Lower: "analysis(fusion)"
    Lower->>IdModel: "Unconditionally construct IdModel"
    IdModel->>Options: "Initialize IdModelOptions"
    Options->>Options: "Check EnableOption::IdModel"
    alt IdModel not enabled
        Options->>Options: "Return empty options"
    else IdModel enabled
        alt predicate_only arg
            Options->>Options: "Enable predicates only"
        else index_only arg
            Options->>Options: "Enable indexing only"
        else No restrictive args
            Options->>Options: "Enable all options"
        end
    end

Greptile Overview

Greptile Summary

This PR enables TensorIndexer with all Python direct tests by making IdModel always-on and simplifying its configuration.

Key Changes:

1. Removes DisableOption::IdModel - IdModel is now always built during GPU lowering, cannot be disabled
2. Simplifies IdModelOptions - Removes build_id_model_ flag and related methods, streamlines consistency checking
3. Updates enable option parsing - Replaces granular options (consumer_index, producer_index, etc.) with simpler predicate_only/index_only flags
4. Migrates from ComputeAtMap to IdModel APIs - Changes predicate_compute.cpp to use idModel().idGraph() instead of caMap() for getting concrete IDs and checking mappings
5. Enables id_model by default in Python API - Adds "id_model" to default enable options for FusionDefinition.execute()
6. Updates test expectations - Adjusts expected variable names and generated CUDA code to match new indexing behavior

The changes are part of a larger effort to transition from ComputeAtMap to IdModel/TensorIndexer as the primary indexing mechanism.

Confidence Score: 5/5

• Safe to merge - well-tested refactoring that enables TensorIndexer consistently across Python tests
• This PR removes the ability to disable IdModel and simplifies its configuration. All changes are internally consistent: DisableOption::IdModel is removed from enum, options map, and all conditional checks. The migration from caMap to idModel APIs in predicate_compute.cpp follows established patterns seen elsewhere in the codebase. Test updates reflect expected behavior changes from the new indexing approach. The Python API change safely adds id_model to default options (duplicates are handled gracefully by the options map).
• No files require special attention

Important Files Changed

File Analysis

Sequence Diagram

sequenceDiagram
    participant User
    participant PythonAPI as Python FusionDefinition
    participant Lower as GpuLower
    participant IdModelOpts as IdModelOptions
    participant IdModel
    participant PredicateCompute
    participant TensorIndexer

    User->>PythonAPI: execute(inputs)
    PythonAPI->>PythonAPI: Add "id_model" to enable_options
    PythonAPI->>Lower: FusionExecutorCache.execute()
    
    Lower->>IdModelOpts: Construct with options
    Note over IdModelOpts: No longer checks DisableOption::IdModel<br/>Always initializes for TensorIndexer
    
    Lower->>IdModel: Build IdModel (unconditional)
    Note over IdModel: Previously conditional on buildIdModel()
    IdModel-->>Lower: IdModel graphs built
    
    Lower->>PredicateCompute: Generate predicates
    PredicateCompute->>IdModel: getConcreteMappedId(id)
    Note over PredicateCompute: Uses idModel().idGraph().toGroup()->front()<br/>Instead of caMap().getConcreteMappedID()
    IdModel-->>PredicateCompute: concrete ID
    
    PredicateCompute->>IdModel: Check ID mapping
    Note over PredicateCompute: Uses idGraph().disjointValSets().strictAreMapped()<br/>Instead of caMap().areMapped()
    IdModel-->>PredicateCompute: mapping result
    
    PredicateCompute-->>Lower: Predicates with TensorIndexer
    Lower-->>User: Compiled kernel

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

_{No files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}