[PyTorch] Reduce CPU overheads

[ksivaman]

Description

Based on single GPU profiling of the GroupedLinear module, implement some optimizations in order to reduce CPU overhead due to PyTorch.

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
• Optimization/performance

Changes

• Consolidate creation and caching of workspace in the GEMM logic. Fix workspace device for cases where incorrectly cached tensor is used.
• Reduce number of arguments to PyTorch autograd function in order to not reduce overheads due to functions such as unwrap_dead_wrappers .
• Use nvtx context manager only when enabled via envvar.
• Remove torch.cuda.device context manager to C++.
• Minor refactors such that delayed scaling recipe checks are grouped together to reduce unnecessary checks for other recipes.
• Reduce number of calls to torch.is_grad_enabled and misc other torch calls.
• Make quantizer specific copy implementations in order to avoid copy.copy().

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

[greptile-apps]

Greptile Summary

• Consolidates workspace creation/caching and moves device context management to C++ to reduce CPU overhead in PyTorch operations
• Reduces PyTorch autograd function arguments by grouping non-tensor args into tuples, minimizing overhead from unwrap_dead_wrappers and arg validation
• Optimizes recipe checks by caching torch.is_grad_enabled() results and grouping delayed scaling checks, plus implements custom quantizer copy() methods to avoid copy.copy() overhead

Confidence Score: 3/5

• This PR has critical bugs that must be fixed before merging
• The workspace caching and CPU overhead optimizations are solid improvements, but the get_tensor_device() function has a critical bug where .device.index can return None for CUDA devices without explicit index (e.g., torch.device('cuda')), which will cause runtime errors when creating tensors
• Pay close attention to transformer_engine/pytorch/cpp_extensions/gemm.py - the device index handling must be fixed

Important Files Changed

Filename	Overview
transformer_engine/pytorch/cpp_extensions/gemm.py	Added workspace caching and device detection, but .device.index can return None causing issues with device specification

Sequence Diagram

sequenceDiagram
    participant User
    participant Linear/GroupedLinear
    participant AutogradFunction
    participant general_gemm
    participant CUDAGuard
    participant cuBLAS/cuDNN

    User->>Linear/GroupedLinear: forward(input)
    Linear/GroupedLinear->>Linear/GroupedLinear: Cache torch.is_grad_enabled()
    Linear/GroupedLinear->>Linear/GroupedLinear: Consolidate args into non_tensor_args tuple
    Linear/GroupedLinear->>AutogradFunction: forward(tensors, non_tensor_args)
    AutogradFunction->>general_gemm: gemm(A, B, quantization_params)
    general_gemm->>general_gemm: get_cublas_workspace(device, ub, grouped)
    general_gemm->>CUDAGuard: Set correct device context
    CUDAGuard->>cuBLAS/cuDNN: Execute GEMM with correct device
    cuBLAS/cuDNN-->>general_gemm: output
    general_gemm-->>AutogradFunction: result
    AutogradFunction-->>Linear/GroupedLinear: output
    Linear/GroupedLinear-->>User: result

Loading

[greptile-apps]

_{27 files reviewed, 5 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{27 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[vthumbe1503]

Reviewed with Kirthi Offline. Looks good to me. LGTM.

[vthumbe1503]

Just a minor further optimization, can be done in later PR as well. Instead of calling empty and then calling a repeat again which means there are 2 torch operations.Directly you can call

torch.empty(
            get_cublas_workspace_size_bytes() * _NUM_MAX_UB_STREAMS, dtype=torch.uint8,device=device)

[ksivaman]

/te-ci

[greptile-apps]

_{27 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile
React with 👍 or 👎 to share your feedback on this new summary format}