Tim Gu

Diable L2_cahce_size display on ROCm platforms

36

This is not supported by clr and hsa and will return wrong results. Fixes https://github.com/ROCm/ROCm/issues/4203

Removed unused code paths that referrs to unimplemented addInst method.

4

Related issue: https://github.com/ROCm/hipBLASLt/issues/1223 Removed sections of code that contains `addInst` & replaced some code paths containing addInst() `raise NotImplementedError`.

Patched incompatible primitives constructor but from upstream NCCL.

2

## Details ___Do not mention proprietary info or link to internal work items in this PR.___ **Work item:** _"Internal", or link to GitHub issue (if applicable)._ **What were the changes?**...

[Issue]: ROCm 6.4 -> ROCm 7.1 Performance regression on gfx110x

1

### Problem Description This is forked from https://github.com/ROCm/ROCm/issues/5674#issuecomment-3668793386. Opened on behalf of @briansp2020 rocm/pytorch:rocm7.1.1_ubuntu24.04_py3.12_pytorch_release_2.7.1 text processing is slow. Performance of ROCm 7.1 ```` root@rocm:~# PYTORCH_MIOPEN_SUGGEST_NHWC=0 python /pwd/Downloads/quickstart.py /opt/venv/lib/python3.12/site-packages/apex/transformer/functional/fused_rope.py:54: UserWarning: Using...