ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fp8 block scale quantization for fmha fwd (#3330)

Browse Source 
* add block scale parameters to kernel

* add block scale to kernel

* add smoke test

* format

* Revert "format"

This reverts commit 356c3c9706.

* only format my code

* format py

* fix auto not allowd in function prototype

* change instance tttt to ttff

* fix structured binding issue

* change s_acc elementwise op

* async pipeline add block scale

* add quantation P using shift exp2

* precompute (m - shift) once per row

* change blk scale seqstrt ptr name

* fix some name

* fix for  deduction guide

* fix some comments

* add P scale to qr_ksvs_pipeline

* add comment to idx_identity

* change the method of calculating descale block index

* unify naming style: use block_scale_ as name prefix

* unify naming style

* update the CHANGELOG.md

* Add FP8 block scale quantization support for FMHA forward kernel

---------

Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
This commit is contained in:
ltqin
2026-01-22 12:58:26 +08:00
committed by GitHub
parent 4c2c18ef48
commit dd0b4294af
14 changed files with 667 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
include/ck_tile/host/reference/reference_batched_gemm.hpp
View File

@@ -47,4 +47,44 @@ CK_TILE_HOST void reference_batched_gemm(const HostTensor<ADataType>& a_b_m_k,
make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
std::thread::hardware_concurrency());
}
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename CDataType,
typename AElementOp = ck_tile::idx_identity,
typename BElementOp = ck_tile::idx_identity,
typename ACCElementOp = ck_tile::idx_identity>
CK_TILE_HOST void reference_batched_quant_gemm(const HostTensor<ADataType>& a_b_m_k,
const HostTensor<BDataType>& b_b_n_k,
HostTensor<CDataType>& c_b_m_n,
const AElementOp& a_element_op = {},
const BElementOp& b_element_op = {},
const ACCElementOp& acc_element_op = {})
{
const int N = b_b_n_k.mDesc.get_lengths()[1];
const int K = b_b_n_k.mDesc.get_lengths()[2];
auto f = [&](auto batch, auto m) {
for(int n = 0; n < N; ++n)
{
AccDataType v_acc = 0;
for(int k = 0; k < K; ++k)
{
AccDataType v_a = ck_tile::type_convert<AccDataType>(
a_element_op(std::make_tuple(batch, m, k), a_b_m_k(batch, m, k)));
AccDataType v_b = ck_tile::type_convert<AccDataType>(
b_element_op(std::make_tuple(batch, n, k), b_b_n_k(batch, n, k)));
v_acc += v_a * v_b;
}
c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(
acc_element_op(std::make_tuple(batch, m, n), v_acc));
}
};
make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
std::thread::hardware_concurrency());
}
} // namespace ck_tile