Skip to content

Support for a4w4 (fp4) in block scale gemm AB quant #3603

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
ErwinTerpstra wants to merge 12 commits into
base: develop
Choose a base branch
from
Draft

Support for a4w4 (fp4) in block scale gemm AB quant #3603

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
9891903
chore: split block scale example instances in more separate files to …
ErwinTerpstra Jan 14, 2026
e43fd33
wip: fp4 scaffolding for abquant
ErwinTerpstra Jan 14, 2026
6dea234
feat: add fp4 decoding-while-loading to abquant pipeline
ErwinTerpstra Jan 14, 2026
d0cd610
feat: add support for fp4 CPU verification in abquant
ErwinTerpstra Jan 15, 2026
58088a5
chore: add time tracking to reference calculation
ErwinTerpstra Jan 15, 2026
3d8bfdb
feat: add a4w4 test for blockscale gemm
ErwinTerpstra Jan 16, 2026
761ba1b
feat: optimize reference calculation by preconverting values to AccType
ErwinTerpstra Jan 19, 2026
a477fb8
feat: add fp4 to fp8 look-up table
ErwinTerpstra Jan 19, 2026
72a94bd
fix: reference to wrong ComputeDataType field in QuantProblem
ErwinTerpstra Jan 19, 2026
7563031
Merge branch 'develop' into eterpstr/206-block-scale-gemm-fp4-support
ErwinTerpstra Jan 19, 2026
e76d18e
feat: type utilities for determining MFMA compute types
ErwinTerpstra Jan 20, 2026
37af217
feat: packed fp4 for abquant weight preshuffle
ErwinTerpstra Jan 20, 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
3 changes: 3 additions & 0 deletions example/ck_tile/38_block_scale_gemm/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,9 @@ if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx12")
add_executable(${EXE_NAME}
gemm_quant.cpp
gemm_abquant_quantgrouped.cpp
gemm_abquant_quantgrouped_fp8.cpp
gemm_abquant_quantgrouped_bf8.cpp
gemm_abquant_quantgrouped_fp4.cpp
gemm_aquant_quantgrouped.cpp
gemm_aquant_quantgrouped_preshufflequant.cpp
gemm_bquant_quantgrouped_bf8i4.cpp
Expand Down
130 changes: 10 additions & 120 deletions example/ck_tile/38_block_scale_gemm/gemm_abquant_quantgrouped.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,127 +6,17 @@
template <typename T>
using GemmConfig = GemmConfigQuantPrefill<T>;

void abquant_quantgrouped_fp4_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut);
void abquant_quantgrouped_fp8_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut);
void abquant_quantgrouped_bf8_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut);

void abquant_quantgrouped_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut)
{
lut[hash_multiple_strings({"fp8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"bf8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"bf8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"bf8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"bf8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
abquant_quantgrouped_fp4_instance_factory(lut);
abquant_quantgrouped_fp8_instance_factory(lut);
abquant_quantgrouped_bf8_instance_factory(lut);
}
72 changes: 72 additions & 0 deletions example/ck_tile/38_block_scale_gemm/gemm_abquant_quantgrouped_bf8.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,72 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT

#include "run_gemm_quant_example.inc"

template <typename T>
using GemmConfig = GemmConfigQuantPrefill<T>;

void abquant_quantgrouped_bf8_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut)
{
lut[hash_multiple_strings({"bf8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"bf8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"bf8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"bf8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::bf8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
}
27 changes: 27 additions & 0 deletions example/ck_tile/38_block_scale_gemm/gemm_abquant_quantgrouped_fp4.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,27 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT

#include "run_gemm_quant_example.inc"

template <typename T>
using GemmConfig = GemmConfigQuantPrefill<T>;

void abquant_quantgrouped_fp4_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut)
{
lut[hash_multiple_strings(
{"fp4", "abquant", "non-preshuffleb", "non-preshufflequant", "1x128x128"})] =
[](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_fp4_t,
ck_tile::pk_fp4_t,
ck_tile::half_t,
float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_fp4_raw_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
}
72 changes: 72 additions & 0 deletions example/ck_tile/38_block_scale_gemm/gemm_abquant_quantgrouped_fp8.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,72 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT

#include "run_gemm_quant_example.inc"

template <typename T>
using GemmConfig = GemmConfigQuantPrefill<T>;

void abquant_quantgrouped_fp8_instance_factory(
std::unordered_map<size_t, std::function<int(const ck_tile::ArgParser&)>>& lut)
{
lut[hash_multiple_strings({"fp8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"abquant",
"non-preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x1x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
lut[hash_multiple_strings({"fp8",
"abquant",
"preshuffleb",
"non-preshufflequant",
"1x128x128"})] = [](const ck_tile::ArgParser& arg_parser) {
using AQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 1, 128>>;
using BQuantGroupSize = ck_tile::QuantGroupShape<ck_tile::sequence<1, 128, 128>>;
using TypeConfig =
decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
return run_gemm_example_prec_type<GemmConfigPreshuffleB_BQuant_Prefill<ck_tile::fp8_t>,
TypeConfig,
AQuantGroupSize,
BQuantGroupSize,
ck_tile::QuantType::ABQuantGrouped>(arg_parser);
};
}
2 changes: 1 addition & 1 deletion example/ck_tile/38_block_scale_gemm/gemm_quant.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,7 @@ auto create_args(int argc, char* argv[])
.insert("prec",
"fp8",
"Data type. For AQuant: fp8, bf8, i4fp8, or i4bf8; for Bquant: fp8, bf8, fp8i4, "
"or bf8i4; for ABQuant: fp8, bf8")
"or bf8i4; for ABQuant: fp8, bf8, fp4")
.insert("warmup", "50", "Number of iterations before benchmarking the kernel")
.insert("repeat", "1000", "Number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
Expand Down
Loading