- 03 Apr, 2023 1 commit
-
-
galagam authored
* Bugfix - compute scale_inv when loading checkpoint Signed-off-by:
Gal Hubara Agam <ghubaraagam@nvidia.com> * Save inverse scale in extra state tensor + minor CR fixes Signed-off-by:
Kirthi Shankar Sivamani <ksivamani@nvidia.com> --------- Signed-off-by:
-
- 30 Mar, 2023 2 commits
-
-
Kirthi Shankar Sivamani authored
* Fix segfault during GeLU export Signed-off-by:
-
Kirthi Shankar Sivamani authored
* Change FP8 recipe defaults Signed-off-by:
Neta Zmora <nzmora@nvidia.com> Signed-off-by:
-
- 29 Mar, 2023 1 commit
-
-
tcherckez-nvidia authored
Signed-off-by:
Tal Cherckez <tcherckez@nvidia.com> Signed-off-by:
-
- 28 Mar, 2023 3 commits
-
-
vasunvidia authored
* Add support for fp8 GEMM BIAS AUX GELU fusion Signed-off-by:
Vasudevan Rengasamy <vrengasamy@nvidia.com> * Fix Lint error Signed-off-by:
-
Tim Moon authored
* Remove zombie process from querying TE install path Co-authored-by:
Naman Goyal <naman@fb.com> Signed-off-by:
Tim Moon <tmoon@nvidia.com> * Fix FA version checking Signed-off-by:
-
Kirthi Shankar Sivamani authored
* fix usage of return_bias argument Signed-off-by:
-
- 23 Mar, 2023 1 commit
-
-
Neta Zmora authored
* Fix GELU ONNX export * Wrap GELU export with cast to/from FP32 to achieve same compute precision as TE. * Increase GELU export test thresholds. * Change export to ONNX opset 17 for smaller representation of LN (single node instead of subgraph). * Remove the need for LN work-around for ORT Signed-off-by:
-
- 22 Mar, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
FA doesn't support compute 8.6 with head_dim>64 Signed-off-by:
-
- 21 Mar, 2023 1 commit
-
-
vasunvidia authored
* Initial commit for fp8_transpose_dbias kernel Signed-off-by:
-
- 16 Mar, 2023 2 commits
-
-
Kirthi Shankar Sivamani authored
Signed-off-by: -
Neta Zmora authored
* Add a temporary workaround to layernorm export Seems like ORT is performing template-matching for LN and incorrectly concludes that it doesn't have a kernel for FP32 LN. The work-around adds the addition of fake_zero which is meant to prevent the template matching while keeping the graph virtually unchanged. This also requires `do_constant_folding=False` in `torch.onnx.export`. Signed-off-by:
Tim Moon <4406448+timmoon10@users.noreply.github.com> * Fix unused import Signed-off-by:
-
- 15 Mar, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Use updated comm API PyTorch Signed-off-by:
-
- 14 Mar, 2023 2 commits
-
-
Ming-Xu Huang authored
* Updated TE/JAX docs Signed-off-by:
Ming Huang <mingh@nvidia.com> * Adding TE/JAX docs' rst files Signed-off-by:
-
Kirthi Shankar Sivamani authored
* Catch FP8 modulo16 error before cublas and fp8 kernels Signed-off-by:
-
- 13 Mar, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* catch incorrect usage of fp8_autocast Signed-off-by:
-
- 11 Mar, 2023 2 commits
-
-
Przemyslaw Tredak authored
* Change from AutoDoc to AutoAPI Signed-off-by:
Przemek Tredak <ptredak@nvidia.com> * Fixes Signed-off-by:
-
Kirthi Shankar Sivamani authored
* deprecate qk layer scaling and fp32 softmax args Signed-off-by:
-
- 07 Mar, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* ignore self attention mask for causal type Signed-off-by:
-
- 02 Mar, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* fix qkv weight unfused path Signed-off-by:
-
- 25 Feb, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Signed-off-by:
-
- 24 Feb, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* Remove redundant amax AR for SP case Signed-off-by:
-
- 23 Feb, 2023 1 commit
-
-
Tim Moon authored
* Deprecate fp32_output option for PyT linear layers Automatically detect dtype for user-provided output tensors. Signed-off-by:
-
- 22 Feb, 2023 1 commit
-
-
cyanguwa authored
* add flash attention to TransformerLayer Signed-off-by:
Charlene Yang <charleney@nvidia.com> * Add docs for FP8 calibration (#61) Signed-off-by:
cyanguwa <cyang.uwa@gmail.com> Signed-off-by:
-
- 16 Feb, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* Fix no reduce_amax option for SP case Signed-off-by:
-
- 15 Feb, 2023 1 commit
-
-
Przemyslaw Tredak authored
* C++ implementation of LayerNorm1P Signed-off-by:
-
- 13 Feb, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
fix use of PyTorch training flag Signed-off-by:
-
- 10 Feb, 2023 2 commits
-
-
Kirthi Shankar Sivamani authored
retain grad related attrs while casting Signed-off-by: -
Kirthi Shankar Sivamani authored
* Bug fixes from PR 22 Signed-off-by:
-
- 07 Feb, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* Bug fixes from PR 22 Signed-off-by:
-
- 03 Feb, 2023 2 commits
-
-
vasunvidia authored
* Increase number of FP8 tensors per GEMM Signed-off-by:
-
Kirthi Shankar Sivamani authored
* remove d2d copies Signed-off-by:
-
- 31 Jan, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Signed-off-by:
-
- 27 Jan, 2023 2 commits
-
-
Przemyslaw Tredak authored
Signed-off-by: -
Kirthi Shankar Sivamani authored
Signed-off-by:
-
- 26 Jan, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Stats and dgrad bug fix Signed-off-by:
-
- 25 Jan, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* Fix conflict from #40 Signed-off-by:
-
- 24 Jan, 2023 2 commits
-
-
schetlur-nv authored
* Initial commit for fp8 calibration. Signed-off-by:
Sharan Chetlur <schetlur@dlcluster.nvidia.com> * Fixes to make unit tests pass Signed-off-by:
Sharan Chetlur <schetlur@nvidia.com> * Cleaning up handling of save_for_backward in Linear Signed-off-by:
schetlur-nv <116769508+schetlur-nv@users.noreply.github.com> Co-authored-by:
-
Kirthi Shankar Sivamani authored
* Add margin for LayerNorm kernel SM usage Signed-off-by:
-
- 20 Jan, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Signed-off-by:
-
