- 22 Apr, 2025 1 commit
-
-
jberchtold-nvidia authored
* [JAX-Q] Single GPU current scaling for JAX Signed-off-by:
Jeremy Berchtold <jberchtold@nvidia.com> * Fix scale check dtype for MXFP8 scales affecting tests using assert_bitwise_scaled_tensors Signed-off-by:
-
- 21 Apr, 2025 2 commits
-
-
jberchtold-nvidia authored
Check CuDNN version and apply unfused norm if below a version with the fix Signed-off-by: -
Sudhakar Singh authored
* rtx5090 arch fix support Signed-off-by:
Sudhakar Singh <sudhakars@nvidia.com> * apprend `nvte` to the function name so that its visible in framework specific dirs Signed-off-by:
pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
-
- 19 Apr, 2025 1 commit
-
-
Tim Moon authored
Revert "Allow NVTEShape to own data. (#1674)" This reverts commit e61ce77c . Signed-off-by:
Tim Moon <tmoon@nvidia.com>
-
- 18 Apr, 2025 4 commits
-
-
Kunlun Li authored
* Add fp8_primary_weights support for blockwise scaling Signed-off-by:
kunlunl <kunlunl@nvidia.com> custom fsdp Signed-off-by:
Tim Moon <4406448+timmoon10@users.noreply.github.com>
-
Kirthi Shankar Sivamani authored
* Move jaxx cuda kernels to core Signed-off-by:
Kirthi Shankar Sivamani <ksivamani@nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by:
-
Hongbin Liu authored
* split wgrad for GroupedLinear Signed-off-by:
Hongbin Liu <hongbinl@nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci Signed-off-by:
-
Phuong Nguyen authored
rm pax/praxis Signed-off-by:
Phuong Nguyen <phuonguyen@nvidia.com> Co-authored-by:
-
- 17 Apr, 2025 5 commits
-
-
wdykas authored
* re merge request Signed-off-by:
Peter Dykas <wdykas@nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * add docstring Signed-off-by:
-
Xin Yao authored
* move swizzle scaling factor to cpp Signed-off-by:
Xin Yao <xiny@nvidia.com> * resolve comments Signed-off-by:
-
kwyss-nvidia authored
* Allow NVTEShape to own data. Signed-off-by:
Keith Wyss <kwyss@nvidia.com> * Convert repeated copy paths to nvte_make_shape calls. Signed-off-by:
-
jberchtold-nvidia authored
* Add a flag to support computing zero-centered gamma in weight dtype or compute dtype for CuDNN Signed-off-by:
-
Paweł Gadziński authored
* drop Signed-off-by:
Pawel Gadzinski <pgadzinski@nvidia.com> * fix Signed-off-by:
-
- 16 Apr, 2025 2 commits
-
-
Paweł Gadziński authored
* add Signed-off-by:
Przemyslaw Tredak <ptrendx@gmail.com> Signed-off-by:
Paweł Gadziński <62263673+pggPL@users.noreply.github.com> * setup fix Signed-off-by:
-
Kshitij Lakhani authored
* Add test cases for full coverage in jax/test_layer.py - causal and window size None - causal and window size default (-1,1) - no_mask and window size default (-1,1) - no_mask and window size default (2,2) - padding and window size None - padding_causal and window_size (2,2) Signed-off-by:
Kshitij Janardan Lakhani <klakhani@nvidia.com> * Correct the condition where padding_causal_mask was being mapped to scaled upper triangle Signed-off-by:
-
- 15 Apr, 2025 2 commits
-
-
Li Tao authored
* support adam bf16 state Signed-off-by:
XiaobingSuper <xiaobingzhangupc@gmail.com> * use fp32 kernel but keep bf16 optimizer states to save memory Signed-off-by:
lit <lit@nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by:
-
Paweł Gadziński authored
* fix Signed-off-by:
-
- 14 Apr, 2025 7 commits
-
-
Jianbin Chang authored
* Add fp8 weight transpose cache check in backward, and regenerated it if it does not exist Signed-off-by:
jianbinc <shjwudp@gmail.com> * Properly handle fsdp shard model weight input. Signed-off-by:
-
Tim Moon authored
* Avoid unnecessary tensor usages when caching for linear op backward Signed-off-by:
-
Xin Yao authored
* Enable MXFP8 and Per-Tensor Current Scaling for Grouped Linear Signed-off-by:
-
Johannes Reifferscheid authored
* Add experimental Shardy support. Production use is not yet recommended. --------- Signed-off-by:Johannes Reifferscheid <jreiffers@nvidia.com>
-
Hua Huang authored
* New GroupedGemmPrimitive using variadic args * Remove squeeze() to reduce D2D memcpy * Revert to the list append fashion to simplify code --------- Signed-off-by:
Hua Huang <huah@nvidia.com> Co-authored-by:
-
Autumn1998 authored
* add support for new recipe on permute_fusion, rm fp unpermute Signed-off-by:
tongliu <tongliu@nvidia.com> * fix lint Signed-off-by:
-
Selvaraj Anandaraj authored
* Added attention activation offloading support for TE v2.0 Signed-off-by:
Selvaraj Anandaraj <selvaraja@login-ptyche02.ptyche.clusters.nvidia.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Signed-off-by:
-
- 11 Apr, 2025 2 commits
-
-
Tim Moon authored
* Add option to cache activation input in FP8 Signed-off-by:
-
kwyss-nvidia authored
Repeated calls to nvte_shape should not invalidate previous data pointers. It would be possible to avoid unnecessary comparisons by duplicating some of the logic from shape() so that the cache is only relevant when columnwise shapes are involved. Whether this code duplication is preferable to the comparisons that arise from by value semantics of reusing shape is a judgment call. Signed-off-by:
-
- 10 Apr, 2025 1 commit
-
-
kwyss-nvidia authored
* Add GEMM logic for blockwise quantized tensors. GEMM test cases included in pytorch integration. Signed-off-by:
zhongboz <zhongboz@nvidia.com> * Edit and cleanup BF16 gather code. Signed-off-by:
Xin Yao <yaox12@outlook.com> * Update test_float8blockwisetensor.py Signed-off-by:
-
- 09 Apr, 2025 3 commits
-
-
Tim Moon authored
* Debug checkpointing with te.Sequential Signed-off-by:
-
Tim Moon authored
Explicitly specify quantized tensor usages needed for linear op backward Signed-off-by: -
Phuong Nguyen authored
* scaling enum abstract * rm NVTE_ from ScalingMode names * rework scaling mode enum in grouped gemm * fix norm sharding --------- Signed-off-by:
-
- 08 Apr, 2025 2 commits
-
-
Tim Moon authored
* Minor stylistic tweaks and typo fixes Review suggestions from @ptrendx Signed-off-by:
-
vasunvidia authored
* Use dummy wgrads for lower memory consumption Signed-off-by:
Vasudevan Rengasamy <vrengasamy@nvidia.com> * Bug fix to avoid sharing gradients. Signed-off-by:
-
- 07 Apr, 2025 5 commits
-
-
kwyss-nvidia authored
* Add GEMM logic for blockwise quantized tensors. GEMM test cases included in pytorch integration. Signed-off-by:
-
Phuong Nguyen authored
* rm no scaling enum Signed-off-by:
-
Jianbin Chang authored
Support fp8 primary weight in fsdp training Signed-off-by:
-
Xin Yao authored
* fix cpp warning Signed-off-by:
-
Xin Yao authored
* refactor to add cp support for sbhd/bshd Signed-off-by:
-
- 04 Apr, 2025 3 commits
-
-
Phuong Nguyen authored
* rename QuantizeAxis to QuantizeLayout, get_layout to get_data_layout, q_axis to q_layout * add fatten_axis option * added gated act to test encoder * sharding constraint fixes * fix padding when flattening first dim needs to be padded * update test sizes so that padding is tested * rm output sharding as it can be done in the flax module * sharding scale_inv for mxfp8 --------- Signed-off-by: -
jberchtold-nvidia authored
MXFP8 flax layer tests Signed-off-by: -
kwyss-nvidia authored
* Blockwise float8 quantizer and quantized tensor class. The classes are configurable for 128x128 blocksize and 1x128 blocksize via setting block_scaling_dim == 2,1 respectively. Scale tensors are stored in a format emenable for matrix multiplication, however the integration of matmul is deferred as a separate story. Fusions of quantization and DBIAS or activation functions are not yet implemented, and the dequantization is currently implemented in torch. Tests for quantization are included in C++ and pytorch layers, with exact comparison to reference quantizer behavior as well as an attempt to hit interesting branches through the API such as tensor creation in pytorch and CPP and dequantization of row and columnwise usage. Two CUDA kernels for quantization are included, and are direct ports of equivalents in the kitchen repository, where a subchannel recipe has been used for end to end training. Signed-off-by:
-
