resolving merge conflicts

This commit is contained in:
khuagarw
2025-11-26 01:22:54 +00:00
parent f290428926
commit 3447196c8a
4 changed files with 5 additions and 5 deletions

View File

@@ -505,7 +505,8 @@ int run_gemm_example_with_layouts(const ck_tile::ArgParser& arg_parser,
{
ck_tile::FillUniformDistribution<ADataType>{-2.0f, 3.0f, fill_seed(gen)}(a_m_k);
}
ck_tile::FillUniformDistribution<AQDataType>{0, 0.5f, fill_seed(gen)}(*aq_tensor_ptr);
ck_tile::FillUniformDistribution<AQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
*aq_tensor_ptr);
ck_tile::FillUniformDistribution<BDataType>{-5.0f, 5.0f, fill_seed(gen)}(b_k_n);
}
else
@@ -528,7 +529,7 @@ int run_gemm_example_with_layouts(const ck_tile::ArgParser& arg_parser,
if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
{
ck_tile::FillConstant<ADataType>{static_cast<ADataType>(0x38)}(a_m_k);
ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x38)}(b_k_n);
ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x22)}(b_k_n);
ck_tile::FillConstant<BQDataType>{static_cast<BQDataType>(0.5f)}(*bq_tensor_ptr);
}
else

View File

@@ -190,6 +190,7 @@ struct BlockGemmWeightPreshuffleBQuantARegBRegCReg
scale_reg_dword = static_cast<uint32_t>(scale_reg);
}
// cross lane ops to get the value of scale_reg.
int gathered_scale_reg = __builtin_amdgcn_ds_bpermute(
pull_from_lane << 2, __builtin_bit_cast(int, scale_reg_dword));

View File

@@ -390,7 +390,6 @@ struct QuantGroupedGemmKernel
const auto& c_block_tile = GemmPipeline{}.template operator()(a_block_window,
b_block_window,
bq_block_window,
kargs.N,
num_loop,
tail_num,
smem_ptr_0,
@@ -458,7 +457,6 @@ struct QuantGroupedGemmKernel
const auto& c_block_tile = GemmPipeline{}.template operator()(a_block_window,
b_block_window,
bq_block_window,
kargs.N,
num_loop,
has_hot_loop,
tail_num,

View File

@@ -70,7 +70,7 @@ struct WPQuantBPipelineAgBgCrV2 : public WeightPreshufflePipelineAGmemBGmemCRegV
static constexpr bool PreshuffleQuant = Problem::Traits::PreshuffleQuant;
static constexpr index_t KPerBlockBQ =
integer_divide_ceil(BlockGemmShape::kK, QuantGroupSize::kK); // 256/128
integer_divide_ceil(BlockGemmShape::kK, QuantGroupSize::kK);
static constexpr index_t QScalesPerBlockRow =
integer_divide_ceil(kKPerBlock, QuantGroupSize::kK);