Skip to content

Add fused multi-tensor kernel for 1D blockwise FP8 quantization #3168

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
shangxiaokang wants to merge 3 commits into
base: release_v2.8
Choose a base branch
from
Draft

Add fused multi-tensor kernel for 1D blockwise FP8 quantization #3168

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
953f13d
blockwise fp8 fused multi-tensor kernel
shangxiaokang Jun 15, 2026
fcd4b38
debug print
shangxiaokang Jun 15, 2026
c437581
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jul 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 14 additions & 0 deletions transformer_engine/common/include/transformer_engine/transpose.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,6 +93,20 @@ void nvte_fp8_transpose_dbias(const NVTETensor input, NVTETensor transposed_outp
void nvte_multi_cast_transpose(size_t num_tensors, const NVTETensor* input_list,
NVTETensor* output_list, cudaStream_t stream);

/*! \brief Cast multiple input tensors to 1D block-scaled FP8 tensors.
*
* \param[in] num_tensors Number of tensors.
* \param[in] input_list List of input tensors.
* \param[in,out] output_list List of 1D block-scaled FP8 output tensors.
* \param[in] quant_config Quantization configuration.
* \param[in] stream CUDA stream used for the operation.
*/
void nvte_multi_quantize_transpose_vector_blockwise(size_t num_tensors,
const NVTETensor* input_list,
NVTETensor* output_list,
const NVTEQuantizationConfig quant_config,
cudaStream_t stream);

/*! \brief Compute backward of GeLU operation on the input, then cast and transpose.
* Additionally, reduce the result of the GeLU backward along the first dimension.
*
Expand Down
6 changes: 6 additions & 0 deletions transformer_engine/common/transpose/cast_transpose.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,6 +63,12 @@ void quantize_transpose_vector_blockwise(const SimpleTensor &input, SimpleTensor
const bool pow_2_scale, const SimpleTensor &noop_tensor,
cudaStream_t stream);

void multi_quantize_transpose_vector_blockwise(
const std::vector<Tensor *> &input_list, std::vector<Tensor *> &output_list,
const float epsilon, FP8BlockwiseRowwiseOption rowwise_option,
FP8BlockwiseColumnwiseOption columnwise_option, const bool pow_2_scale,
const SimpleTensor &noop_tensor, cudaStream_t stream);

void quantize_transpose_vector_blockwise_fp4(
const SimpleTensor &input, const SimpleTensor &global_amax, SimpleTensor &scale_inv,
SimpleTensor &scale_inv_t, SimpleTensor &output, SimpleTensor &output_t, const float epsilon,
Expand Down
Loading