- 10 May, 2023 1 commit
-
-
Shriya Palsamudram authored
userbuffer pushsend/recv fix with atomicAdd_system Signed-off-by:
Sangkug Lym <slym@nvidia.com> Co-authored-by:
-
- 09 May, 2023 2 commits
-
-
Kirthi Shankar Sivamani authored
* Initial refactor Signed-off-by:
Kirthi Shankar Sivamani <ksivamani@nvidia.com> * refactor attention out of transformer.py Signed-off-by:
-
Neta Zmora authored
* ONNX export refactoring * Remove infer_ort (to enable more testing) * Add BF16 ORT tests for Q/DQ ops and GELU. * Use FP32 i/o instead of BF16 (because ORT doesn't support BF16 i/o) and add casts from FP32 to BF16 (this is only for subgraph inputs and outputs). * We'll need to add more BF16 testing. * GEMM: * Add cast after DQ to achieve better performance (matmul at sub-fp32 precisions). * Fold bias into Gemm operation (=> smaller graphs) * Wrap GEMM-GELU with FP32 (TE implements GELU in FP32) * Enable tests for cross attention (test_export_multihead_attention) * Reduce test thresholds for test_export_layernorm_mlp, test_export_layernorm_linear, test_export_layernorm Signed-off-by:
Neta Zmora <nzmora@nvidia.com> * Loosen MHA export validation thresholds for FP16 Signed-off-by:
-
- 03 May, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
DDP support for no-bias option Signed-off-by:
-
- 02 May, 2023 3 commits
-
-
cyanguwa authored
* move dbias from input list to output list for bwd Signed-off-by:
Charlene Yang <8636796+cyanguwa@users.noreply.github.com> * split asserts into three for bias checks Signed-off-by:
Tim Moon <4406448+timmoon10@users.noreply.github.com> Signed-off-by:
-
Sangkug Lym authored
* using different strems for pushsend and pushrecv Signed-off-by:
-
Kirthi Shankar Sivamani authored
Use only built-ins for setup Signed-off-by:
-
- 01 May, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Fixes in nested autocast Signed-off-by:
-
- 30 Apr, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Remove distutils Signed-off-by:
-
- 28 Apr, 2023 3 commits
-
-
Neta Zmora authored
* iFix LN ONNX export When exporting LayerNorm make sure that the weights and bias inputs have the same type as the LN input. Also: * Add a regression test. * Add environment variable to override directory of generated test artifacts Signed-off-by:
-
Kirthi Shankar Sivamani authored
Fix bug in Sequence Parallelism Signed-off-by: -
Kirthi Shankar Sivamani authored
Signed-off-by:
-
- 26 Apr, 2023 1 commit
-
-
Przemyslaw Tredak authored
* Faster split of QKV for FlashAttention Signed-off-by:
Przemek Tredak <ptredak@nvidia.com> * CI fixes Signed-off-by:
-
- 22 Apr, 2023 1 commit
-
-
asfiyab-nvidia authored
add cast for BF16 input/output types for Q/DQ ONNX ops Signed-off-by:
Asfiya Baig <asfiyab@nvidia.com> Co-authored-by:
-
- 21 Apr, 2023 4 commits
-
-
Kirthi Shankar Sivamani authored
Signed-off-by: -
Sangkug Lym authored
Signed-off-by:
-
cyanguwa authored
* Add FP8 fused attention to TE for PyTorch Signed-off-by:
Charlene Yang <charleney@nvidia.com> * add license for cudnn-frontend, modify installation requirements, and refactor some headers for aesthetics Signed-off-by:
-
Kirthi Shankar Sivamani authored
* Initial refactor; linker error Signed-off-by:
Tim Moon <tmoon@nvidia.com> * Lint Signed-off-by:
-
- 19 Apr, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* Port initial changes Co-authored-by:
Vasudevan Rengasamy <vrengasamy@nvidia.com> Signed-off-by:
-
- 18 Apr, 2023 1 commit
-
-
Sangkug Lym authored
* amax reduction internval Signed-off-by:
-
- 17 Apr, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* use upstream flash-attn Signed-off-by:
-
- 14 Apr, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Signed-off-by:
-
- 13 Apr, 2023 1 commit
-
-
Neta Zmora authored
* Fix model load exception when state resides on GPU - Whenever converting a torch.tensor to numpy, we need to first migrate the tensor storage to the host CPU. - Add a warning not to do contant-folding when exporting to ONNX. This is due to a torch.onnx export bug. - Refactor compare_outputs Signed-off-by:
-
- 07 Apr, 2023 2 commits
-
-
ngoyal2707 authored
* made bias configurable Signed-off-by:
Naman Goyal <naman@fb.com> * removed commented lines Signed-off-by:
ngoyal2707 <ngoyal2707@users.noreply.github.com> * Update transformer_engine/pytorch/jit.py Co-authored-by:
-
Kirthi Shankar Sivamani authored
* small cleanup before starting Signed-off-by:
-
- 05 Apr, 2023 1 commit
-
-
Sangkug Lym authored
* async amax reduction add env knob to enable async amax reduction Signed-off-by:
slym <slym@login-preos01.a51.clusters.nvidia.com> * Style fixes Signed-off-by:
-
- 04 Apr, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
* Add FP8 support for Ada Signed-off-by:
-
- 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:
-
- 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:
-
- 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:
-
Tim Moon authored
* Remove zombie process from querying TE install path Co-authored-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:
-
- 15 Mar, 2023 1 commit
-
-
Kirthi Shankar Sivamani authored
Use updated comm API PyTorch Signed-off-by:
-
