eqy

NVFuser JIT eager mode InstanceNorm3d

12

Prototype version with "cuDNN conv style" caching.

[cublas][cublasLt] Fall back to unfused `addmm` for 2-byte-aligned inputs

1

Fix for this issue surfaced from the discuss forum: https://discuss.pytorch.org/t/cuda-error-cublas-status-not-supported-when-calling-cublasltmatmul-from-torch-nn-functional-linear/170214 Note that PyTorch builds before #71200 should not be affected as there was no `cublasLt` dispatch path. Additionally, the provided...

[CUDA][CUDA 11] Remove more CUDA 11 version checks

5

Working on removing stragglers missed in previous CUDA version < 11.0 cleanup PRs.

[NCCL] Add experimental Nonblocking NCCL Fault Tolerance/Checking

5

Support for nonblocking NCCL communicators/fault tolerance/checking which was added in 2.14 as an experimental feature. Enabled via the environment variable: ``` NCCL_USE_COMM_NONBLOCKING=1 ``` CC @ptrblck

[cuDNN] Bump cuDNN frontend submodule to 1.1.1

1

Hopefully addresses the failure seen when trying to bump to 1.1.0 (#119642) CC @Skylion007 cc @csarofeen @ptrblck @xwang233

[CUDA][cuBLAS] Check if a context is present when grabbing a cuBLAS handle

1

cuBLAS has indicated that certain kernels will transition to using the driver API over the CUDA runtime API, which we've observed to break existing tests (e.g., DataParallel) that use multithreading...

The current cuDNN submodule is ancient, this shouldn't break the build right?

GNMT won't run without `dllogger`

[CUDA][Compex] `test_reference_numerics_large_jiterator_unary_cuda_complex64` broken after updating to `numpy >= 1.25.0`

2

### 🐛 Describe the bug Unsure why this hasn't surfaced in upstream CI as we've observed it on sm90, sm86, sm80, sm60, ... Doesn't surface on e.g., `numpy==1.24.4` but starts...

[cuDNN][SDPA] Remove `TORCH_CUDNN_SDPA_ENABLED=1`, enable cuDNN SDPA by default on H100 and 2nd on other archs >= sm80

55

Looks like one of the first failures seen is `test_causal_variants_compile_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` when `test_causal_variants_causal_variant_CausalVariant_LOWER_RIGHT_shape0_cuda` passes. What seems interesting here is that the `torch.compile` version fails while the eager version passes. Not sure...