Make EvalMode.eager the default

[rostan-t]

Category: New feature (non-breaking change which adds functionality)

Description:

Make EvalMode.eager the default eval mode instead of EvalMode.deferred. This PR also run some tests with all eval modes.

Additional information:

Affected modules and functionalities:

Key points relevant for the review:

Tests:

• Existing tests apply
• New tests added
  • Python tests
  • GTests
  • Benchmark
  • Other
• N/A

Modified tests to make them run with all eval modes:

• test_arithm_ops.py
• test_batch.py
• test_math_funcs.py
• test_tensor.py

Checklist

Documentation

• Existing documentation applies
• Documentation updated
  • Docstring
  • Doxygen
  • RST
  • Jupyter
  • Other
• N/A

DALI team only

Requirements

• Implements new requirements
• Affects existing requirements
• N/A

REQ IDs: N/A

JIRA TASK: DALI-4558

[greptile-apps]

Greptile Overview

Greptile Summary

This PR changes the default evaluation mode from deferred to eager and enhances test coverage by running tests across all evaluation modes.

Key Changes:

• Default EvalMode changed from deferred to eager for more immediate error detection
• Fixed slice handling bug where negative stop indices (e.g., [:-1]) were incorrectly filtered out
• Added @eval_modes() decorator utility to automatically test functions across all eval modes (deferred, eager, sync_cpu, sync_full)
• Updated 4 test files to use the new decorator, ensuring comprehensive eval mode coverage

Note: Copyright year updated to 2026 in all modified files.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• All changes are well-structured and focused: the default eval mode change is intentional and documented in JIRA-4558, the slice bug fix corrects existing behavior, and test enhancements improve coverage. The implementation of the @eval_modes() decorator is clean and correctly handles parameter composition with existing @params decorators.
• No files require special attention

Important Files Changed

File Analysis

Filename	Score	Overview
dali/python/nvidia/dali/experimental/dynamic/_eval_mode.py	5/5	Changed default eval mode from deferred to eager (line 43), updated copyright year to 2026
dali/python/nvidia/dali/experimental/dynamic/_tensor.py	5/5	Fixed slice handling by removing incorrect filter on negative stop indices (line 863), updated copyright year to 2026
dali/test/python/experimental_mode/eval_mode_utils.py	5/5	New test utility providing @eval_modes() decorator to parametrize tests across all eval modes

Sequence Diagram

sequenceDiagram
    participant User
    participant Test
    participant EvalMode
    participant ThreadLocalState
    participant Operation
    
    Note over User,Operation: Before PR: Default was deferred
    User->>Test: Run test
    Test->>ThreadLocalState: Initialize eval_mode_stack
    ThreadLocalState->>EvalMode: Set default = deferred
    Test->>Operation: Execute operation
    Operation->>Operation: Defer execution
    Note over Operation: Only executes when needed
    
    Note over User,Operation: After PR: Default is eager
    User->>Test: Run test with @eval_modes()
    loop For each EvalMode
        Test->>EvalMode: Enter context (mode)
        EvalMode->>ThreadLocalState: Push mode to stack
        Test->>Operation: Execute operation
        alt mode is eager
            Operation->>Operation: Execute immediately
            Note over Operation: Input validation happens now
        else mode is deferred
            Operation->>Operation: Defer execution
        else mode is sync_cpu/sync_full
            Operation->>Operation: Execute synchronously
        end
        EvalMode->>ThreadLocalState: Pop mode from stack
    end

Loading

[greptile-apps]

Greptile Overview

Greptile Summary

This PR changes the default evaluation mode from EvalMode.deferred to EvalMode.eager for DALI's experimental dynamic API. The eager mode provides immediate evaluation start and input validation, offering better error reporting at the cost of potentially less optimization compared to deferred mode.

Key Changes:

• Changed EvalMode.default = deferred to EvalMode.default = eager in _eval_mode.py
• Fixed a bug in tensor slicing where negative stop indices (e.g., t[:-1]) weren't being passed through to the slice operator
• Added comprehensive test coverage with new @eval_modes() decorator that runs tests across all 4 eval modes (deferred, eager, sync_cpu, sync_full)
• Updated 50+ test functions across 4 test files to verify behavior consistency across all eval modes

Test Coverage Enhancement:
The new eval_mode_utils.py provides a parameterization decorator that integrates with the existing @params decorator to multiply test coverage across all evaluation modes. This ensures the dynamic API works correctly regardless of eval mode.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The changes are well-structured and thoroughly tested. The default eval mode change is a behavioral change but non-breaking since all modes remain available. The tensor slicing bug fix is necessary and correct. The decorator pattern for test parameterization is elegant and properly handles both decorated and non-decorated test functions. All test files were updated consistently.
• No files require special attention

Important Files Changed

File Analysis

Filename	Score	Overview
dali/python/nvidia/dali/experimental/dynamic/_eval_mode.py	5/5	Changed default eval mode from deferred to eager on line 43, updated copyright year
dali/python/nvidia/dali/experimental/dynamic/_tensor.py	5/5	Fixed bug where negative slice indices weren't passed to _tensor_subscript, updated copyright year
dali/test/python/experimental_mode/eval_mode_utils.py	5/5	New utility file with @eval_modes() decorator for parameterizing tests across all eval modes

Sequence Diagram

sequenceDiagram
    participant User
    participant Test as Test Function
    participant Decorator as @eval_modes()
    participant EvalMode
    participant DynamicAPI as Dynamic API (ndd)
    
    User->>Test: Run test
    Test->>Decorator: Apply decorator
    Decorator->>Decorator: Read existing paramList
    Decorator->>Decorator: Generate mode combinations
    Note over Decorator: Creates params for all 4 modes:<br/>deferred, eager, sync_cpu, sync_full
    
    loop For each eval mode
        Decorator->>EvalMode: __enter__() with mode
        EvalMode->>EvalMode: Push mode to stack
        Note over EvalMode: If mode==default, use eager<br/>(changed from deferred)
        Decorator->>DynamicAPI: Execute test operations
        DynamicAPI->>DynamicAPI: Use current eval mode
        Note over DynamicAPI: Operations evaluated per mode:<br/>- eager: immediate start<br/>- deferred: lazy eval<br/>- sync_*: synchronous
        DynamicAPI-->>Decorator: Return results
        Decorator->>EvalMode: __exit__()
        EvalMode->>EvalMode: Pop mode from stack
    end
    
    Decorator-->>User: All modes tested

Loading