GitLab

Signed-off-by: wenjh <wenjh@sugon.com>

Signed-off-by: wenjh <wenjh@sugon.com>

Signed-off-by: wenjh <wenjh@sugon.com>

Signed-off-by: wenjh <wenjh@sugon.com>

* 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: Jan Bielak <jbielak@nvidia.com>

* Support use of `alpha` and `beta` in `general_gemm`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Support use of `alpha` and `beta` in `BasicLinear._functional_forward` and `BasicLinear._functional_backward`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add `ForwardLinearScaleAdd` fusion
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add `BackwardLinearScale` fusion
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Apply suggestions from code review
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Remove calls to `validate_gemm_scale` from `BasicLinear`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

---------
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

Add launch bounds to swizzle kernel, use empty scale inv
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* reduce driver calls
Signed-off-by: Xin Yao <xiny@nvidia.com>

* reduce driver calls
Signed-off-by: Xin Yao <xiny@nvidia.com>

* adjust tests to capture this
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* fix: update grad_output quant to avoid redundant work
Signed-off-by: kshitij12345 <kshitijkalambarkar@gmail.com>

* add test
Signed-off-by: kshitij12345 <kshitijkalambarkar@gmail.com>

* don't keep only columnwise quant if requires_dgrad=False
Signed-off-by: kshitij12345 <kshitijkalambarkar@gmail.com>

* fix stray merge
Signed-off-by: kshitij12345 <kshitijkalambarkar@gmail.com>

* fix for ctx.use_bias is True case
Signed-off-by: kshitij12345 <kshitijkalambarkar@gmail.com>

* Skip if FP8 not available
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: kshitij12345 <kshitijkalambarkar@gmail.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Signed-off-by: wenjh <wenjh@sugon.com>

Signed-off-by: wenjh <wenjh@sugon.com>

* enable PDL for blockwise qunatization kernels
Signed-off-by: Xin Yao <xiny@nvidia.com>

* add comment
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Xin Yao <xiny@nvidia.com>

* Compute amax in normalization forward in current scaling in untuned kernels
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Signed-off-by: wenjh <wenjh@sugon.com>

Signed-off-by: wenjh <wenjh@sugon.com>

* 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: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Working userbuffer overlapping API
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix overwriting bulk overlap UB object for layernormLinear
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Update external overlap to use tp size instead of nvsize to determine number of copies
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix linter error
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Explanatory comments of overlap logic
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix the UB fused ops tests
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* fix: Fix linter errors
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

---------
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* PDL for MXFP8 Quantize
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* for loop
Signed-off-by: Xin Yao <xiny@nvidia.com>

* bulk alloc
Signed-off-by: Xin Yao <xiny@nvidia.com>

* multi-tensor swizzle
Signed-off-by: Xin Yao <xiny@nvidia.com>

* pad zeros in swizzle kernels
Signed-off-by: Xin Yao <xiny@nvidia.com>

* unify single- and multi-tensor swizzle
Signed-off-by: Xin Yao <xiny@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix empty tensor list
Signed-off-by: Xin Yao <xiny@nvidia.com>

* fix bug for col swizzle
Signed-off-by: Xin Yao <xiny@nvidia.com>

* check context & fix signifiers
Signed-off-by: Xin Yao <xiny@nvidia.com>

---------
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add basic kernel for swapping first two tensor dims
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add NVRTC kernel for swapping first dims
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add PyTorch extension for swap first dims kernel
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Tweak variable names
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Tune kernel
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Make sure writes are contiguous
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Fix tensor numel() return dtype

The original dytpe int would be an issue if the tensor element numbers are larger than int32's range (which is not a super large number in real workloads)
Signed-off-by: Shang Zhang <shangz@nvidia.com>

* fix bug if all values<0
Signed-off-by: tongliu <tongliu@nvidia.com>

* minor fix
Signed-off-by: tongliu <tongliu@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: tongliu <tongliu@nvidia.com>
Co-authored-by: tongliu <tongliu@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xin Yao <xiny@nvidia.com>

* Compute amax in normalization kernels as long as the pointer is provided, even if using non quantized output
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fuse amax computation into normalization forward
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Use TE lahyernorm kernel instead of raising error about unsupported cuDNN feature
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Compute amax in activation kernels when the output pointer is provided, even for non-fp8 outputs
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 9f13fe2fefc58cae93bc467d87d01ecf792a0381)

* Initialize metatensor values
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fuse computation of amax into the activation kernel for fp8 current scaling
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 2b54327ac9c931a5340983a79e99de5caa0399dd)
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Zero out amax in `create_hp_tensor_with_amax` instead of relying on `Float8CurrentScalingQuantizer.__init__` to zero-initialize it
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* remove reciprocal op
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* Refactor Quantizer::create_tensor function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Fix bug when constructing FP8 tensor
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Add quantize function to C++ quantizers
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Prototype function to coerce Python quantized tensors to match quantizer
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use quantizer class in tex.quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 current scaling support for activation backward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable quantized GEMM output with FP8 current scaling
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add coerce_tensor functions for MXFP8 and DSv3
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Avoid quantizing empty tensors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use consistent shapes for FP8 transposes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* In attention impl, construct FP8 tensors with pre-initialized scale-invs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Initialize MXFP8 scales to zero
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Store copy of quantizer when creating quantized tensors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Make sure quantized tensors have private quantizer

Avoid problems with in-place ops after quantizer usages are changed externally.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename "coerce_tensor" to "convert_and_update_tensor"
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Make sure CUDA context is available when launching NVRTC kernel
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Expose CUDA context creation function externally
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: zhongboz <zhongboz@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: zhongboz <zhongboz@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Remove deprecated device arg
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Remove test
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fixed integer overflow when computing offsets
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Fix cudnn versioning in support in PyTorch DPA and Fused attn
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

Fix cuDNN lib runtime loading and simplify
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* Fixed conflicts
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Minor code refactoring to avoid unnecessary checks
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Fixed typo
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Fixed dBias accumulation error due to initialization. Minor code refactoring
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Test case to reproduce the init error
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Fixed rowwise dbias error
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Changed ptx API
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Added a struct for two packed FP8 values
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Rolled back to scalar code for columnwise scaling due to its better performance
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Minor corrections
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Rebased on main
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Fixes per code review
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Removed constexpr in C++ test suite to build faster
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Computed activations are now numerically truncated to InputType before scaling. Improved test suite.
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Minor refactoring
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Minor refactoring
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Modified mismatches checks of MXFP8 to address FP8 numerics
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Implemented Jeremy's fixes to JAX test suite with an intermediate downcast
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Reduced the dims of the test tensors to improve CI runtime
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* Fixed memory alignment issue. Compute dbias without downcast.
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Fixed misaligned memory issue also in gated kernels. Reduced size of MXFP8 gated tests
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



---------
Signed-off-by: Oleg Goncharov <ogoncharov@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>