Compiling error: ‘OVecT’ has not been declared

[cryoco]

Describe the bug

/TransformerEngine/transformer_engine/common/activation/./../cast/dispatch/../fp8/group_quantize_fp8.cuh:1075: error: template argument 2 is invalid
/TransformerEngine/transformer_engine/common/activation/./../cast/dispatch/../fp8/group_quantize_fp8.cuh:1075: error: ‘IVecT’ has not been declared
/TransformerEngine/transformer_engine/common/activation/./../cast/dispatch/../fp8/group_quantize_fp8.cuh:1075: error: ‘OVecT’ has not been declared

Steps/Code to reproduce bug

NCCL_HOME=/usr/local/miniconda3/lib/python3.12/site-packages/nvidia/nccl NVTE_BUILD_WITH_NCCL_EP=1 USE_NCCL=1 NVTE_CUDA_ARCHS="100a" NVTE_BUILD_THREADS_PER_JOB=8 NVTE_FRAMEWORK=pytorch VERBOSE=1 pip install -vvv --no-build-isolation --no-cache-dir -e .

Environment details

• B200
• PyTorch 2.11
• Python 3.12
• Transformer Engine version 4cd244e
• CUDA 13.0
• CUDNN 9.19

[sbhavani]

@cryoco I was able to reproduce this on a single Blackwell GPU using PyTorch 2.11.0, Python 3.12, CUDA 13.0 and TE commit 4cd244e.

The same source built successfully for me with CUDA 13.1, so this might be specific to CUDA 13.0.