GitLab

* 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>

* 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>

* 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>

* 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>

* Debug checkpointing with 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



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

Explicitly specify quantized tensor usages needed for linear op backward
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

Signed-off-by: Przemek Tredak <ptredak@nvidia.com>

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

* Add activation 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



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

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

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

* Respect PyTorch autograd dtype
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename CastFloat8 op to Quantize
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add support for fused dSwiGLU-cast-transpose
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>
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>

* Remove manual FP8 scale update for FP8 params
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



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

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

* Add layer norm op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 cast op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add tests for linear and layernorm with FP8 output
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* RMSNorm op
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>

* Replace LayerNorm module with LayerNorm op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Replace RMSNorm module with RMSNorm op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add AMP support
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 save autograd context if grad mode is disabled

Debugging ONNX export 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



* Forward args in pre_forward func to base op class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update to use QuantizedTensor class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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

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



* Review suggestions from @ptrendx

Rename "CastFloat8" op to "Quantize". Add more fine-grained control for SM margin. Add docs for legacy sequence_parallel kwarg.
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>

* Use weight dtype as default compute dtype
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



---------
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>
Co-authored-by: Przemyslaw Tredak <ptrendx@gmail.com>

* Handle Float8Tensor when casting module dtype

Keep data in Float8Tensor and only change nominal dtype. Monkey-patch PyTorch module casting functions to handle Float8Tensor. Add tests.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Respect autocast dtype in linear op
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



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

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

* Tweak comments

Review suggestion from @ptrendx
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 base class for tensor proxies
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move tensor detaching logic to tensor proxy base class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use Python wrappers to PyTorch extensions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Include transpose caching logic in proxy encode function
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug dimension mismatch with amax history
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move dequantize logic to proxy_decode func
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

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

* Include transpose cache in detach and clone funcs
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 FP8 workspaces with QuantizedTensor functions
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Move logic for FP8 transpose cache in FP8 workspaces to base class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove cast-transpose logic from linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unnecessary args for Float8Tensor when using FP8 attr dict
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove __torch_function__ to QuantizedTensor
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Update tests/pytorch/test_float8tensor.py
Signed-off-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Debug FP8 transpose test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Add op for in-place add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add op for in-place add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add op that adds extra output to fuser
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add fused op for GEMM+bias+add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add fused op for dgrad+add
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add documentation
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>

* Review suggestions from @ptrendx

Output tensor dtype and device take precedence over weight tensor in linear functional API. Move some index calculation to fuser constructor. Avoid some unnecessary dereferences.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Update transformer_engine/pytorch/ops/fuser.py
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 basic infrastructure for Sequential module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add FP8 support in linear op

Runs, but need to validate. Runtime errors with non-FP8 params and FP8 compute, or FP8 params and non-FP8 compute.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add reshape op and unit test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add unfused linear op

Test does not pass with FP8.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug unfused linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add test for linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add separate abstract classes for unfused and fused ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Consolidate unfused ops in submodule
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add linear-bias fused op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use fused cast-transpose in linear ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable GEMM+bias fusion with FP32 activations

Not supported by cuBLAS.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add parallel unit test for unfused linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Refactor parallel tests to reduce job launches
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add all-reduce, all-gather, and reduce-scatter ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unused file
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug multi-GPU FP8 test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add support for FP8 scale updates

Still need to implement amax reductions.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add license boilerplate
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fuse GEMM+bias in row TP

Add documentation for unfused ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename pipeline to fuser

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

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

* Preserve cached FP8 transpose between ops
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add option for fused wgrad accumulation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Directly output FP8 from linear if needed
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix cuDNN front-end commit
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use updated FP8 tensor API for transpose caching
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use updated API for FP8 scale updates
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add tests for non-default FP8 recipes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename UnfusedOperation to BasicOperation
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Add unit test to check amax reduction with fusable op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Operator autograd state no longer needs to be initialized
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Initial functional implementation of linear op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug fused linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove autograd context from functional linear impl
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Use functional linear impl in fused linear+bias op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Rename subdirectory from "fuser" to "ops"

Avoid confusion with kernel fusers and graph compilers.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update with Float8Tensor changes in #820
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove unnecessary CPU overheads
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Correctly pass FP8 metadata from next op
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Add convenience functions to manipulate Sequential class
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Clear saved tensor data in linear op after bprop
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix test name in QA script
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update name of PyTorch extensions module
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Run distributed tests even when only 1 GPU is available
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Only run distributed tests with 2 GPUs if there are >=2 GPUs
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



* Review suggestions from @sudhakarsingh27 and @ksivaman

Fix spelling of "fusible". Avoid "input" name in internal APIs.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update transformer_engine/pytorch/ops/__init__.py
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>