GitLab

Same intention of commit 3e38a2ea

.
This commit is to improve acc.
Signed-off-by: wenjh <wenjh@sugon.com>

[PyTorch] FP8 Subchannel Recipe With FP8 Gather And Configurable Scaling Factor Tensor Swizzling (#1707)

* functional kernel for columnwise + no-transpose option, still hacky
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* pass all quantizer unit tests
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* refactor, add gemm ready api
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* make format options private members, simplify api
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* swizzle scales right before gemm
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* bug fix of single layer test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* attempt to fix lint issue
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* fp8 gather pass, need minor refine
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* fix return_layernorm_output_gathered case
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* remove special cases, add sanity check before gemm
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

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

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



* Implement dequantize for compact 1D blocks.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* add more unit test with dequantize compact supported
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

* make ag for subchannel respect async
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* zero tolerance in distributed test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix zero tolerance test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* resolve rebase issues
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

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

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

* relax rtol for fp32 distributed test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix some ci issue
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* fix ci test failure in debug mode
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* Force row-wise and column-wise data to have same data format

Prototype "all-gather usage" in quantizer.
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Remove dead logic for high-precision AGs
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 FP8 block-wise tensor tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

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

* Handle case where LayerNormLinear returns gathered norm output
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* fix debug mode
Signed-off-by: zhongboz <zhongboz@nvidia.com>

---------
Signed-off-by: zhongboz <zhongboz@nvidia.com>
Signed-off-by: Keith Wyss <kwyss@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>
Co-authored-by: Keith Wyss <kwyss@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Add GEMM logic for blockwise quantized tensors.

GEMM test cases included in pytorch integration.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update NVTE_BLOCK_SCALING for GEMM.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Gate feature on CUDA 12.9
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Gemm typo.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Remove unecessary type converter change.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Reflect epilogue availability and test supported epilogues.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* GEMM simplifications from recipe branch.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Format py code.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update GEMM DGelu tests to match support depending on output dtype.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Force pow2Scales in GEMM
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Add GEMM test to pytorch test suite.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Add copyright to GEMM test.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update import for GEMM test.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Add license.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update test gemm supported predicate.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Use sgemm like interfaces and naming.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Rewrite GEMM comment.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* MR Feedback.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Refactor GEMM param canonicalization

Configure A and B matrices separately. Have separate code path for each scaling mode.
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



* Prune number of tests.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

---------
Signed-off-by: Keith Wyss <kwyss@nvidia.com>
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>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* 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: Keith Wyss <kwyss@nvidia.com>

* Apply linting changes.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Alignment for 1D scaling for GEMM edge case.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* MR feedback.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Change API name.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Fix merge conflict with name change.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Use common tensor map API.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Change API to use two scaling mode enums.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Fix typo.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update some call sites.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Tests for torch tensor API surface.

Since the quantized tensor is a tensor
subclass, these tests exercise torch hooks.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Reuse scale calculation between quantizer refs.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Save memory by dropping reference to saved tensors.

Issues previously observed are solved.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Remove constexpr parameters from kernel.

Code size is reduced with fewer constexpr params.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Merge conflict from rebase.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Add shape implementations for block scaling.

nvte_shape was added upstream. Logic added
for block scaled fp8.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Move benchmark to te_playground
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Remove amax_epsilon and pow_2_scales from tensor.

Hardcodes the default values.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Lint changes.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Fixup MR changes that broke.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Safer ifdef in kernel.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Documentation prose.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Reuse compute_scale function from Current Scaling.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Bugfix on inf_value scale refactor.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Remove qopt calls from test.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update pytest list.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Add copyright to reference scale calc.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Use ptx.cuh functions instead of cde.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update shape logic with allocation and reuse shape.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Usage defaults MR feedback.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Copyright and header guard.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Updating torch dispatch code.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Fix exception type.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Use TypeInfo
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* MR feedback.
Signed-off-by: Keith Wyss <kwyss@nvidia.com>

* Update CS scale update test to use updated ref impl
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Update JAX scaling mode enum
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Skip tests on Lovelace
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: Keith Wyss <kwyss@nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* Enable fp8_primary_weights for current scaling
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Use different cast_master_weights_to_fp8 functions depending on the type of quantizer
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* All amaxes of model_weights should participate in reduce-max
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Clear _high_precision_init_val automatically in cast_master_weights_to_fp8 function
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Merge all all-reduce on amaxes into one NCCL kernel
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Add unit tests for multi_tensor_compute_scale_and_scale_inv and preserve_high_precision_init_val
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Fix conflicts
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Add unit test for cast_master_weights_to_fp8
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* use mock group to initialize fp8_autocast to avoid reduction of amax_history by fp8_autocast_exit
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Remove with_computing_amax and with_computing_scale
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Move replace_raw_data from QuantizedTensor to utils.py
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Remove allow_empty_output argument from nvte_compute_amax and set it always be true
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Rename import guard of recipe_common.cuh to be align with other import guards
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Add unit test for replace_raw_data
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Add test_replace_raw_data into qa/L0_pytorch_unittest/test.sh
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Minor changes in comments
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Add randomness to the unit test of replace_raw_data
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* (Maybe need revert) Add tex.quantize_to_fragment
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* (Maybe needsto rrevert) Use nvte_quantize_noop in quantize_to_fragment
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Fix lint error
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Move high_precision_init_val test and replace_raw_data test to test_sanity.py
Signed-off-by: kunlunl <kunlunl@nvidia.com>

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

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



* Remove test_fp8_model_init.py and test_replace_raw_data.py
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Remove cast_master_weights_to_fp8 and replace_raw_data from __all__ of tensor.__init__.py
Signed-off-by: kunlunl <kunlunl@nvidia.com>

* Move FP8 casting logic back from C++ tex funcs to Python
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 unimplemented function from header
Signed-off-by: Tim Moon <tmoon@nvidia.com>

---------
Signed-off-by: kunlunl <kunlunl@nvidia.com>
Signed-off-by: Kunlun Li <94586211+kunlunl@users.noreply.github.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>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>

* check in per-tensor current scaling full recipe
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

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

setup basics of current scaling quantizer in python level
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

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

add test case for current scaling dequantize
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

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

finish linear layer fwd bwd test, determined error with bf16
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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

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

achieved zero tolerance for Linear by specify gemm use_split_accumulator config
Signed-off-by: zhongboz <zhongboz@nvidia.com>

enable layernormlinear with current scaling, pass bitwise test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

refactor test case code
Signed-off-by: zhongboz <zhongboz@nvidia.com>

make current scaling quantizers distrbuted, pass distributed linear&layernormlinear tests
Signed-off-by: zhongboz <zhongboz@nvidia.com>

bug fix: use cached fp8 recipe in backward
Signed-off-by: zhongboz <zhongboz@nvidia.com>

fix layernorm_mlp with current scaling, fix activation_helper with current scaling
Signed-off-by: zhongboz <zhongboz@nvidia.com>

support detailed numerical settings from recipe to quantization kernel
Signed-off-by: zhongboz <zhongboz@nvidia.com>

resolving MR comments
Signed-off-by: zhongboz <zhongboz@nvidia.com>

recipe naming
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* resolve mr comments, remove IS_CURRENT_SCALING template from kernels
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* resolve mr comments, make current scaling c++ test cases
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* add current scaling to test_numerics.py, skip act recomp and grouped linear
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* add benchmark for quantizer
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* add benchmarks for linear layer
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* bug fix, typo
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* resolve more mr comments
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* avoid potential race condition by not using from_blob to construct amax tensor in C++
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* resolve more comments
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* Debug linter warnings and license check
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 import error in FP8 tensor test
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Debug compilation error with CUDA 12.1 for Turing
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



* resolve mr comments, fix activation cast fusion
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* resolve comments, add NVTEQuantizationParams for compute scale
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* remove is_current_scaling check totally from common folder
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* remove benchmarks, will contribute in another repo
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* adjust cs default recipe config
Signed-off-by: zhongboz <zhongboz@nvidia.com>

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

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



* adjust comments in test
Signed-off-by: zhongboz <zhongboz@nvidia.com>

* Remove current scaling mode from core lib
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Refactor current-scaling-specific logic in core C++ lib

Move amax and scale update functions out of casting functions, and put into dedicated current-scaling source file. Add general API for accessing quantization config object.
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 missing header in C++ tests
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Disable test config with FP8 transpose on Blackwell
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* Fix compilation error in C++ test
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: 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: Tim Moon <tmoon@nvidia.com>