- 27 Aug, 2025 1 commit
-
-
yuguo authored
-
- 26 Aug, 2025 2 commits
- 25 Aug, 2025 1 commit
-
-
yuguo authored
-
- 23 Aug, 2025 2 commits
- 21 Aug, 2025 2 commits
- 19 Aug, 2025 1 commit
-
-
evt_fugx1 authored
-
- 18 Aug, 2025 3 commits
-
-
Phuong Nguyen authored
* fix fsdp Signed-off-by:
Phuong Nguyen <phuonguyen@nvidia.com> --------- Signed-off-by:
-
Xin Yao authored
* check if the given recipe is supported in fp8_autocast Signed-off-by:
Xin Yao <xiny@nvidia.com> * resolve comments Signed-off-by:
Tim Moon <4406448+timmoon10@users.noreply.github.com>
-
jberchtold-nvidia authored
Fix flax variables when creating quantizers directly from a recipe Signed-off-by:Jeremy Berchtold <jberchtold@nvidia.com>
-
- 16 Aug, 2025 1 commit
-
-
jomitchellnv authored
fix: fixes multi head attention for context parallel: rotary embedding to use padded cu_seq_lens (#2077) fix: fixes mha to use padded cu_seq_lens during cp Signed-off-by:Jonathan Mitchell <jomitchell@nvidia.com>
-
- 15 Aug, 2025 1 commit
-
-
Jan Bielak authored
* Add `nvte_cublas_gemm_scaled` Signed-off-by:
Jan Bielak <jbielak@nvidia.com> * Support use of `alpha` and `beta` in `tex.generic_gemm` Signed-off-by:
-
- 14 Aug, 2025 6 commits
-
-
Kirthi Shankar Sivamani authored
Add launch bounds to swizzle kernel, use empty scale inv Signed-off-by:Kirthi Shankar Sivamani <ksivamani@nvidia.com>
-
Tim Moon authored
* Unfused impl for dbias-quantize Signed-off-by:
Tim Moon <tmoon@nvidia.com> * Unfused impl for dact-dbias-quantize Signed-off-by:
pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
-
Tim Moon authored
Avoid registering FP8 recipe update in ops without backward pass Signed-off-by:
-
Tim Moon authored
* Register weight/bias params in linear op Signed-off-by:
-
Xin Yao authored
* reduce driver calls Signed-off-by:
-
Kshiteej K authored
* fix: update grad_output quant to avoid redundant work Signed-off-by:
kshitij12345 <kshitijkalambarkar@gmail.com> * add test Signed-off-by:
-
- 13 Aug, 2025 7 commits
-
-
Phuong Nguyen authored
* fix pspec check Signed-off-by:
-
Phuong Nguyen authored
* add manual axis filer to sharding_constraint impl Signed-off-by:
-
Kate Cheng authored
* Remove if-else and torch.tensor to meet cudagraph requirement Signed-off-by:
Kate Cheng <yunhsuanc@nvidia.com> * Add is_cg_capturable flag to guard the if-else statement Signed-off-by:
-
Paweł Gadziński authored
* code drop Signed-off-by:
Pawel Gadzinski <pgadzinski@nvidia.com> * fix Signed-off-by:
-
wenjh authored
Signed-off-by:wenjh <wenjh@sugon.com>
-
wenjh authored
Signed-off-by: -
Xin Yao authored
* enable PDL for blockwise qunatization kernels Signed-off-by:
-
- 12 Aug, 2025 5 commits
-
-
Jan Bielak authored
* Compute amax in normalization forward in current scaling in untuned kernels Signed-off-by:
-
jberchtold-nvidia authored
* Support setting collection name for quantizer set Flax variables in TransformerEngineBase flax module Signed-off-by:
-
wenjh authored
Signed-off-by: -
Tim Moon authored
Fix bug when deducing dtype in linear functional API Signed-off-by: -
Tim Moon authored
* Update amax pointers when reallocating amax history in fusible ops Signed-off-by:
-
- 11 Aug, 2025 2 commits
-
-
Kirthi Shankar Sivamani authored
* Fix high-precision dtype for MXFP8 AG Signed-off-by:
-
wenjh authored
Signed-off-by:
-
- 09 Aug, 2025 1 commit
-
-
Daniel Stokes authored
* fix: Add stream synchronization before destroying MPI communicator (#1979) Signed-off-by:
djns99 <40156487+djns99@users.noreply.github.com> * feat: Implement column-wise userbuffer overlap for comm+GEMM operations Add support for overlapping column-wise allgather communication with GEMM operations to improve training performance: * **Core infrastructure changes:** - Update bulk_overlap_columnwise_ag() to accept explicit stream parameter - Modify userbuffers send/recv loops to use rank-ordered iteration - Add userbuffers_send_all/recv_all function declarations * **Python integration:** - Add bulk_overlap_ag_with_external_gemm() C++ extension function - Expose new overlap function via pybind11 bindings - Update overlap method configurations to include more ring_exchange ops * **LayerNorm MLP optimization:** - Enable column-wise quantization for FC2 gradient output - Implement overlap of allgather communication with FC2 DGRAD GEMM - Use fill_userbuffers_buffer_for_all_gather for efficient buffering This optimization allows overlapping communication and computation phases more effectively, reducing training wall-clock time by hiding allgather latency behind GEMM execution. Signed-off-by:
-
- 08 Aug, 2025 5 commits
-
-
Phuong Nguyen authored
* enabled TE GEMM for all recipes Signed-off-by:
-
Paweł Gadziński authored
* turn on userbuffers for layers without debug Signed-off-by:
-
Phuong Nguyen authored
* rm cudaGraph compatible trait from GroupedGEMM and groupedQuantize Signed-off-by:
-
yuguo authored
-
yuguo authored
-
