[pyTorch] Replace the make_empty implementation to use C++ implementation

[ptrendx]

Description

This PR unifies the implementation of the QuantizedTensor creation by using the C++ implementation of the create_tensor.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Replaced the Python implementations of the make_empty with the calls to C++ create_tensor

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[ptrendx]

/te-ci L1 pytorch

[ptrendx]

/te-ci L1 pytorch

[negvet]

This clear NotImplementedError is beneficial for custom quantizers that do not override make_empty().
Now, if custom quantizer does not have make_empty(), then it will fail at the C++ convert_quantizer, because there is no registered C++ converter. C++ failure with NVTE_ERROR("Unexpected type for quantizer") is not as clear as NotImplementedError.

What about making C++ error more clear or even better add a check at base Quantizer.make_empty:

def make_empty(...):
    if getattr(self, "custom", False):
        raise NotImplementedError(
            f"{self.__class__.__name__} does not implement make_empty"
        )
    # ... existing C++ path ...

[ptrendx]

/te-ci pytorch L1

[greptile-apps]

Greptile Summary

Unified quantized tensor creation by migrating Python make_empty implementations to a single C++ create_tensor interface, removing ~287 lines of duplicated Python code across 4 quantizer types while adding device and pin_memory parameter support.

Key changes:

• Added device and pin_memory parameters to all create_tensor C++ method signatures
• Created new create_empty_quantized_tensor function exposed via pybind11
• Replaced abstract Quantizer.make_empty() with concrete implementation calling C++ backend
• Removed redundant Python implementations from Float8Quantizer, Float8CurrentScalingQuantizer, Float8BlockQuantizer, MXFP8Quantizer, and NVFP4Quantizer
• Added device string-to-torch.device conversion in Python wrapper for better API ergonomics

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• Well-structured refactoring that consolidates duplicate code into a single C++ implementation without changing functionality. The changes are straightforward, removing ~287 lines of duplicated Python code while properly handling device and pin_memory parameters. Device string handling was correctly added to prevent type errors.
• No files require special attention

Important Files Changed

Filename	Overview
transformer_engine/pytorch/csrc/extensions/cast.cpp	Implemented create_empty_quantized_tensor function that calls quantizer_cpp->create_tensor with device and pin_memory parameters
transformer_engine/pytorch/csrc/quantizer.cpp	Updated all quantizer create_tensor implementations to accept and use device and pin_memory parameters for tensor allocation
transformer_engine/pytorch/quantized_tensor.py	Replaced abstract make_empty with concrete implementation calling C++ create_empty_quantized_tensor, added device string handling
transformer_engine/pytorch/tensor/float8_tensor.py	Removed 83 lines of Python make_empty implementations for Float8Quantizer and Float8CurrentScalingQuantizer
transformer_engine/pytorch/tensor/nvfp4_tensor.py	Removed 84 lines of Python make_empty implementation for NVFP4Quantizer

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Python: quantizer.make_empty] --> B[tex.create_empty_quantized_tensor]
    B --> C[C++: create_empty_quantized_tensor]
    C --> D[convert_quantizer]
    C --> E[GetTransformerEngineDType]
    D --> F[quantizer_cpp->create_tensor]
    E --> F
    F --> G{Quantizer Type}
    G --> H[Float8Quantizer::create_tensor]
    G --> I[Float8BlockQuantizer::create_tensor]
    G --> J[MXFP8Quantizer::create_tensor]
    G --> K[NVFP4Quantizer::create_tensor]
    G --> L[NoneQuantizer::create_tensor]
    H --> M[Create Float8Tensor with device/pin_memory]
    I --> N[Create Float8BlockwiseQTensor with device/pin_memory]
    J --> O[Create MXFP8Tensor with device/pin_memory]
    K --> P[Create NVFP4Tensor with device/pin_memory]
    L --> Q[Create unquantized Tensor with device/pin_memory]
    M --> R[Return QuantizedTensor]
    N --> R
    O --> R
    P --> R
    Q --> R
    R --> S[Python: Apply requires_grad if needed]

Loading

_{Last reviewed commit: 9cad6d0}

[greptile-apps]

_{10 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}