GitLab

* Make sure to set usages for linear op quantizers before forward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Avoid unsupported case for fused dbias+quantize kernel

Hopper does not support dbias + FP8 cast without FP8 transpose.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Add NVFP4 recipe
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Frank Sun <frsun@nvidia.com>
Co-authored-by: Oleg Goncharov <ogoncharov@nvidia.com>
Co-authored-by: Zhongbo Zhu <zhongboz@nvidia.com>
Co-authored-by: Evgeny Tsykunov <etsykunov@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Teddy Do <tdophung@nvidia.com>

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

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



* Add MathDx dependency to GitHub builds
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Suggestions from GitHub Copilot
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move 2x shape logic from core to PyTorch
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix compilation errors with CUDA 12.1
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



* SM 70 is not supported in CUDA 13
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Typo
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Revert "Move 2x shape logic from core to PyTorch"

This reverts commit f8b2a2d0111d9af690b43bb98ae448d9a430a185.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Added dequantize kernel for FP4
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

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



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

* Add NVFP4 support with fusible ops

Use logical tensor dims for PyTorch NVFP4 tensors. Temporarily add unfused dequantize impl. Fix bug where NVFP4 recipe was not configurable.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix logic for 2x shapes and move to PyTorch
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix CG test model config
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Debug NVFP4 tensor size function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Proper handling of the RNG state
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

* Test SR properly
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

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



* Fix workspace size for GEMM heuristic.
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix compile error in C++ NVFP4 test

Some some numeric errors when blocks are all zero.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix distrbuted test problem shape
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* proper assert dim for low precision AG TP
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* clean up duplicated code in nvfp4_utils.cuh
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* lint
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* pylint: disable=unused-argument
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* `nvte_cublas_gemm_v2` to take alpha pointer (#12)

* make nvte_cublas_gemm_v2 to take alpha/beta pointers
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* users are expected to pass a valid C_tensor
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* typos
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* API to have const float* alpha
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Minor tweaks

Support arbitrary beta scales. Increase workspace to be aligned to 128 bytes.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug IMA with alpha pointer
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

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

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



* Support fused amax kernels with NVFP4 quantization
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable fused amax with cuDNN LayerNorm kernel
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add NVFP4 cases to distributed tests for TE ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Change assert to NVTE_CHECK in the hadamard cast fusion
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

* Use global thread IDs for Philox subsequences
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 shape checks for NVFP4 cast kernels
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



* Do not fuse amax if cuDNN normalization is forced by envvar
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Przemek Tredak <ptredak@nvidia.com>
Signed-off-by: zhongboz <zhongboz@nvidia.com>
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Frank Sun <frsun@nvidia.com>
Co-authored-by: Oleg Goncharov <ogoncharov@nvidia.com>
Co-authored-by: Zhongbo Zhu <zhongboz@nvidia.com>
Co-authored-by: Evgeny Tsykunov <etsykunov@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Teddy Do <tdophung@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Przemek Tredak <ptredak@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Add CPU offloading logic to ops. Fix test to compute dgrad.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Make sure grads are contiguous in op backwards
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add op-based MLP to CPU offloading tests
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



* Handle different weight cache behavior on Hopper/Blackwell

Add MXFP8 to CPU offload tests.
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



* Remove MXFP8 test
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add dropout kernel with 8-bit RNG
Co-authored-by: Vasudevan Rengasamy <vrengasamy@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
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 license
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Avoid ambiguous types
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Do not enforce dropout prob is representable in 8 bits
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Expand error message
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 small statistical bug from using less-equal instead of less-than

Refactor kernel implementations and add comments. Interpret masks as bytes rather than 16-bit uints.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warning
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



* Remove unnecessary helper function in PyTorch extensions
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



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

expose more activation functions
Signed-off-by: Xin Yao <xiny@nvidia.com>

* Return dummy wgrad tensors when requested by Mcore
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Apply suggestions from code review
Co-authored-by: Jan Bielak <janekb04@icloud.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

---------
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Jan Bielak <janekb04@icloud.com>

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

* Add BackwardAddRMSNorm fused operation
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Try to optimize register usage in kernels
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add separate BackwardAdd stage for the fused backward add
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

---------
Signed-off-by: Jan Bielak <jbielak@nvidia.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>

* Unfused impl for dbias-quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Unfused impl for dact-dbias-quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable fused bgrad-quantize for unsupported recipes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unfused dbias-quantize impls

Not supported in the core lib.
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



* Support unfused impls in tex functions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Tweaks
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



* Remove unused imports
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>

Avoid registering FP8 recipe update in ops without backward pass
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Register weight/bias params in linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Make sure linear op checkpoint is backward-compatible
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 linter warning
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Check for invalid case before setting bias
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 bug when deducing dtype in linear functional API
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update amax pointers when reallocating amax history in fusible ops
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



* Update weight tensor quantizer when recipe state is reset
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: 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>

Fix merge conflict bug with clearing op outputs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

Signed-off-by: Jan Bielak <jbielak@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.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>

* Add ops for dropout and constant scale
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



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

* Mark output tensors as not deletable in backward
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add `in_place` kwarg to `MakeExtraOutput`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Rename `AddInPlace` to `AddExtraInput` and add an `in_place` kwarg
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>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Increase intermediate precision and reuse tensors from fwd
Signed-off-by: Evgeny <etsykunov@nvidia.com>

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

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



* JIT warmup only when required
Signed-off-by: Evgeny <etsykunov@nvidia.com>

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

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



* Recompute only rsqrt_norm
Signed-off-by: Evgeny <etsykunov@nvidia.com>

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

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



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

* Fix bug where TE ops were not updating fp8_meta dicts
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename reset_recipe_state 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



* Update error message when initializing meta device quantized weight without recipe
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>

* Refactor _OperationFuserAutogradFunction.forward to use less parameters
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit f8f59b1bb184e89468058521df4cfff029ad909c)

* Rename `BackwardBiasActivation` to `BackwardActivationBias`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 397c58fc296f801fe4ad600aadc2daff3b78be45)

* Use forward operation order in backward fused operations
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 2d37a9385069b066e6cdeff3eb9173c2079cb791)

* Rename `prev_op_grad_input_quantizer` to `prev_op_grad_output_quantizer`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit d7ab5dfb23e216866f7f4fc4d7a99f625d329f1e)

* Make OperationFuser persistent
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 77984d9715d31e87519dc6ea1e02c483a81355a7)

* Distribute extra inputs to and collect extra outputs from multiple module groups in Sequential
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 0716aaad542e59f2c1ac4620167965a0334bbf71)

* Take requires_grad into account when fusing operations
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Change get_quantizer to return None if no quantization recipe is used
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Refactor pre_first_forward
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fix for failing `test_make_graphed_callables[fp8_recipe0-*-True-*-linear_op]`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fix linting errors
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

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

* Fix fp8 meta tensors in CUDA Graph capture
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



* Fix failing distributed userbuffers tests
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>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* some initial code
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* onnx support
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* mxfp8 support
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fixed returning layernorm etc
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* formatting
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* lint fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* license fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* tests passing
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* refactor
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* lint
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fixes
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* added pip install to test.sh
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Update transformer_engine/pytorch/export.py
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>

* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* float8currentscaling quantizer exception
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* added to wheels
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* onnx versions
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* installations in tests
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* fix
Signed-off-by: root <root@prenyx0221.a51.clusters.nvidia.com>

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

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



* lint fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* fix
Signed-off-by: root <pgadzinski@nvidia.com>

* fixes
Signed-off-by: root <pgadzinski@nvidia.com>

* fixes
Signed-off-by: root <pgadzinski@nvidia.com>

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

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



* fixes
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* fixes
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Update setup.py
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>

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

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



* onnxscript version chnage
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



* fixes
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

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

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



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

* fix
Signed-off-by: Pawel Gadzinski <pgadzinski@gmail.com>

* Update build.yml
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Update pytorch.py
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
Signed-off-by: Paweł Gadziński <62263673+pggPL@users.noreply.github.com>
Signed-off-by: root <root@prenyx0221.a51.clusters.nvidia.com>
Signed-off-by: root <pgadzinski@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: Pawel Gadzinski <pgadzinski@gmail.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: root <root@prenyx0221.a51.clusters.nvidia.com>
Co-authored-by: Pawel Gadzinski <pgadzinski@gmail.com>

* Fix clearing tensor data in backward removing is_first_op
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Misc fixes
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Use Linear weight dtype and device for compute consistently
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add backward dbias + quantize fusion
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Pass recipe to OperationFuser to allow recipe-dependent fusions
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Remove redundant view from activations
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Add bias activation backward fusion
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Apply suggestions from code review
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
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: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Change pre_forward to pre_first_forward
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fix passing invalid recipe with fp8 disabled
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 test for `LayerNormMLP` implementation using `te.ops.Sequential` to `test_fusible_ops.py` (#1924)

* Add e2e test for LayerNormMLP implemented with te.Sequential
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fix bugs uncovered by test
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



* Fix reshaping columnwise_data of MXFP8Tensor
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Fix taking dtype from weight or grad_output in BasicLinear._functional_backward
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

---------
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Added MCore fsdp support for TE
Signed-off-by: Selvaraj Anandaraj <selvaraja@cw-dfw-cs-001-login-01.cm.cluster>

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

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



* Refactored based on new MCore FSDP
Signed-off-by: Selvaraj Anandaraj <selvaraja@cw-dfw-cs-001-login-01.cm.cluster>

* Code cleanup and extended across modules
Signed-off-by: Selvaraj Anandaraj <selvaraja@cw-dfw-cs-001-login-01.cm.cluster>

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

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



* Added FSDP suport
Signed-off-by: Selvaraj Anandaraj <selvaraja@cw-dfw-cs-001-login-01.cm.cluster>

---------
Signed-off-by: Selvaraj Anandaraj <selvaraja@cw-dfw-cs-001-login-01.cm.cluster>
Co-authored-by: Selvaraj Anandaraj <selvaraja@cw-dfw-cs-001-login-01.cm.cluster>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Replace `is_float8_tensor` with `is_quantized_tensor`

Replace free function `is_float8_tensor` with `is_quantized_tensor` in `_common.py` and use it throughout the `ops` codebase to check if a tensor is a (possibly internal) quantized tensor
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Pass next and previous op quantizers directly to op_forward and fuser_forward

Change interface of `fuser_forward` and `op_forward` to no longer take preceding and following ops and instead take the following op's input quantizer and preceding op's input gradient's quantizer directly
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Remove use redundant `detach` in `BasicLinear`

Remove use of `detach` in `BasicLinear` for improved performance (enabled by not passing prev_op to backward)
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Handle saving internal tensors

Handle saving internal tensors in `_OperationFuserAutogradFunction` using `prepare_for_saving` and `restore_from_saved`
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Use internal tensors

Enable use of internal tensors in `BasicLinear` quantizers and fix issues resulting from internal tensors not having methods that regular tensors have
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Apply suggestions from code review
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
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>

* Optimize _common.reshape by removing redundant operations
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Use view instead of reshape when possible
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Simplify convert_tensor (requires testing)
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Remove reshape
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Refactor existing code to use maybe_quantize where possible
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Check if tensor is any kind of quantized tensor
Signed-off-by: Jan Bielak <jbielak@nvidia.com>

* Revert "Check if tensor is any kind of quantized tensor"

This reverts commit cf09d61ffe41f38720d820ddc4f011f9dc1fb56e.
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 FP8 current scaling to te.Sequential tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Helper function for test/ref tensors does not produce quantized tensor by default
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 current scaling to distributed te.Sequential tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 current scaling to Userbuffers te.Sequential tests
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



* Debug MXFP8 tests
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>

* Flatten basic op params during fuser init
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit 949abe97070721b1da5117903067608250f5fb61)

* Add caching for is_non_tn_fp8_gemm_supported
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit fd830ae24ffbd2d0727010b1a8a119ca72f61ce5)

* Pass fuser to _OperationFuserAutogradFunction.forward and moving computation to __init__
Signed-off-by: Jan Bielak <jbielak@nvidia.com>
(cherry picked from commit fd808991993958b670726896254b82fcb967fa07)

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

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



* Pass basic_op_kwargs and is_grad_enabled as parameters rather than in fuser
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>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Add support for overlapping wgrad NCCL AG with dgrad GEMM
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* Remove unused wait on memcpy API from UB
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

* Add better commenting to MXFP8 overlap
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>

---------
Signed-off-by: djns99 <40156487+djns99@users.noreply.github.com>
Co-authored-by: dastokes <dastokes@dastokes-dvt-01.nvidia.com>

* Support L2Norm basic op
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* Add L2Norm module wrapper
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* Expose qk_norm to MHA nd transformer laayer
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* Move tests into separate file
Signed-off-by: Evgeny <etsykunov@nvidia.com>

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

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



* fix pass
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* Add license
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* Remove  module
Signed-off-by: Evgeny <etsykunov@nvidia.com>

* Resollve comments
Signed-off-by: Evgeny <etsykunov@nvidia.com>

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

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



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

* Linear op avoids saving input tensor if weight grad is not needed
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Linear op forward avoids producing quantized tensors with unnecessary usages
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Avoid unnecessary usages in fused linear ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

Activation ops support fusing backward pass with quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

fix saved_tensors
Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>

* Initial work toward restoring UB support in te.Sequential
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Forward UB linear runs, but has numerical error
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug UB forward tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Minor tweaks
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove Python checks for MXFP8 UB linear forward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add dim check for MXFP8 full tiles
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move QuantizedTensor logic out of UB comm and into Python helper function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Support MXFP8 AGs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Coalesce NCCL all-gathers for MXFP8 all-gather
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Initial impl of backward UB linear in te.Sequential
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug UB linear backward with no quantization

dgrad GEMM + dx RS is still broken.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix chunk dims for dgrad GEMM + dx RS
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debugging MXFP8 UB cases

Still failing with dy AG + wgrad GEMM
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use NCCL to overlap dy AG with dgrad GEMM
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug UB GEMM tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Initial refactoring of linear module forward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Refactor linear module backward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug linear module UB tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Tweak test tensor dims
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Do not store autograd context within wgrad GEMM closure
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings
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



* Update LayerNormLinear
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



* Update LayerNormMLP
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug UB tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix linter warnings
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



* Debug test failures
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Minor style tweaks
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 incorrect usage for GEMM input with block-scaled FP8
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix RS out dims
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable dgrad GEMM + UB AG + NCCL AG overlapping
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



* Disable dgrad GEMM + UB AG + NCCL AG overlap in te.Sequential
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



* Restore support for internal quantized tensors
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add tests for MXFP8 GEMM with UB
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 linter warnings
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug test failures
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug test failures
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>

* Avoid unnecessary tensor usages when caching for linear op backward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug test failure
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Add option to cache activation input in FP8
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Avoid casting to FP8 transpose
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Skip input caching if device is not supported
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add documentation that FP8 input caching is experimental
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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