[CK TILE] Support fp8/fp16 with pk_int4_t as data types for tensors A and B (#2805)

- Add support for tensor A/B in both fp16+pk_int4_t and fp8+pk_int4_t formats
- Implement A(bf8) B(i4) support in universal GEMM
- Use new implementation for i4 to fp8 conversion in Block Scale

[ROCm/composable_kernel commit: 82890192dd]
This commit is contained in:
Cong Ma
2025-09-09 17:40:52 -06:00
committed by GitHub
parent 22490acf0b
commit f7ffd111ee
15 changed files with 320 additions and 135 deletions

View File

@@ -5,11 +5,8 @@
#include <cstring>
#include <iostream>
#include <sstream>
#include <string>
#include <tuple>
#include "ck_tile/host.hpp"
#include "gemm_utils.hpp"
#include "run_gemm_example.inc"
#include "run_gemm_example_common.hpp"
@@ -58,7 +55,7 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
ck_tile::int8_t,
ck_tile::int32_t>(a_layout, b_layout, arg_parser);
}
else if(data_type == "pk_int4_t")
else if(data_type == "fp16i4")
{
// TODO: Add support for bhalf_t ADataType
if constexpr(GemmConfig<ck_tile::half_t>::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
@@ -74,6 +71,36 @@ int run_gemm_example(ck_tile::ArgParser& arg_parser)
throw std::runtime_error("Unsupported pipeline for this operation !!!");
}
}
else if(data_type == "fp8i4")
{
if constexpr(GemmConfig<ck_tile::fp8_t>::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
{
return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
Invoker,
ck_tile::fp8_t,
ck_tile::pk_int4_t,
ck_tile::half_t>(a_layout, b_layout, arg_parser);
}
else
{
throw std::runtime_error("Unsupported pipeline for this operation !!!");
}
}
else if(data_type == "bf8i4")
{
if constexpr(GemmConfig<ck_tile::bf8_t>::Pipeline == CK_TILE_PIPELINE_COMPUTE_V3)
{
return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
Invoker,
ck_tile::bf8_t,
ck_tile::pk_int4_t,
ck_tile::half_t>(a_layout, b_layout, arg_parser);
}
else
{
throw std::runtime_error("Unsupported pipeline for this operation !!!");
}
}
else
{
throw std::runtime_error("Unsupported data type for this operation !!!");