Skip to content

Feat/swiglustep moe no quant #3731

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
LJ-underdog wants to merge 1 commit into
base: develop
Choose a base branch
from
+76 −0
Draft

Feat/swiglustep moe no quant #3731

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
76 changes: 76 additions & 0 deletions include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1450,6 +1450,29 @@ struct GridwiseMoeGemm : public GridwiseGemm_xdl_cshuffle_base<
tensor_operation::element_wise::Gelu{}(gate, gate);
c_thread_buf_fp32(cidx) = gate * up;
}
else if constexpr(ActivationOperation ==
Activation::swiglustep_and_mul)
{
const float scale_up =
p_scale_b[(n0 * NWave * NPerXdl + problem.N) *
PerTokenQuant];
float gate = scale_a * scale_b * c_thread_buf[cidx];
float up = scale_a * scale_up * c_thread_buf_up[cidx];
if constexpr(MulRoutedWeight)
{
gate = gate * topk_weights.template AsType<float>()[m4];
up = up * topk_weights.template AsType<float>()[m4];
}
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
{
gate *= 16;
up *= 16;
}
tensor_operation::element_wise::Silu{}(gate, gate);
gate = gate < 7.0f ? gate : 7.0f;
up = up < 7.0f ? (up > -7.0f ? up : -7.0f) : 7.0f;
c_thread_buf_fp32(cidx) = gate * up;
}
}
else
{
Expand Down Expand Up @@ -1511,6 +1534,21 @@ struct GridwiseMoeGemm : public GridwiseGemm_xdl_cshuffle_base<
tensor_operation::element_wise::Gelu{}(gate, gate);
c_thread_buf_fp32(cidx) = gate * up;
}
else if constexpr(ActivationOperation ==
Activation::swiglustep_and_mul)
{
float gate = c_thread_buf[cidx];
float up = c_thread_buf_up[cidx];
if constexpr(MulRoutedWeight)
{
gate = gate * topk_weights.template AsType<float>()[m4];
up = up * topk_weights.template AsType<float>()[m4];
}
tensor_operation::element_wise::Silu{}(gate, gate);
gate = gate < 7.0f ? gate : 7.0f;
up = up < 7.0f ? (up > -7.0f ? up : -7.0f) : 7.0f;
c_thread_buf_fp32(cidx) = gate * up;
}
}
else
{
Expand Down Expand Up @@ -1923,6 +1961,29 @@ struct GridwiseMoeGemm : public GridwiseGemm_xdl_cshuffle_base<
tensor_operation::element_wise::Gelu{}(gate, gate);
c_thread_buf_fp32(cidx) = gate * up;
}
else if constexpr(ActivationOperation ==
Activation::swiglustep_and_mul)
{
const float scale_up =
p_scale_b[(n0 * NWave * NPerXdl + problem.N) *
PerTokenQuant];
float gate = scale_a * scale_b * c_thread_buf[cidx];
float up = scale_a * scale_up * c_thread_buf_up[cidx];
if constexpr(MulRoutedWeight)
{
gate = gate * topk_weights.template AsType<float>()[m4];
up = up * topk_weights.template AsType<float>()[m4];
}
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
{
gate *= 16;
up *= 16;
}
tensor_operation::element_wise::Silu{}(gate, gate);
gate = gate < 7.0f ? gate : 7.0f;
up = up < 7.0f ? (up > -7.0f ? up : -7.0f) : 7.0f;
c_thread_buf_fp32(cidx) = gate * up;
}
}
else
{
Expand Down Expand Up @@ -1984,6 +2045,21 @@ struct GridwiseMoeGemm : public GridwiseGemm_xdl_cshuffle_base<
tensor_operation::element_wise::Gelu{}(gate, gate);
c_thread_buf_fp32(cidx) = gate * up;
}
else if constexpr(ActivationOperation ==
Activation::swiglustep_and_mul)
{
float gate = c_thread_buf[cidx];
float up = c_thread_buf_up[cidx];
if constexpr(MulRoutedWeight)
{
gate = gate * topk_weights.template AsType<float>()[m4];
up = up * topk_weights.template AsType<float>()[m4];
}
tensor_operation::element_wise::Silu{}(gate, gate);
gate = gate < 7.0f ? gate : 7.0f;
up = up < 7.0f ? (up > -7.0f ? up : -7.0f) : 7.0f;
c_thread_buf_fp32(cidx) = gate * up;
}
}
else
{
Expand Down