Skip to content

cuda: software-FP16 fallback for fattn dequantize_V<half> on CC < 600 #3

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

Merged
cchuter merged 1 commit into from
May 14, 2026
+52 −0
Merged

cuda: software-FP16 fallback for fattn dequantize_V<half> on CC < 600 #3

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
52 changes: 52 additions & 0 deletions ggml/src/ggml-cuda/fattn-common.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -400,6 +400,18 @@ static __device__ __forceinline__ void dequantize_V_q4_0(const void * __restrict
((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]);
}
} else
#else // FP16_AVAILABLE
// Software-FP16 fallback: compute in float, pack via __floats2half2_rn (RTNE).
// Required because lightning_indexer_kernel_vec instantiates <half, 4> for the
// full CC matrix (50/61/70/...), but FP16_AVAILABLE is undefined for CC < 600.
if constexpr (std::is_same_v<T, half>) {
const float d = __half2float(x[ib].d);

#pragma unroll
for (int l0 = 0; l0 < ne; l0 += 2) {
((half2 *) dst)[l0/2] = __floats2half2_rn(d * q8[l0 + 0], d * q8[l0 + 1]);
}
} else
#endif // FP16_AVAILABLE
if constexpr (std::is_same_v<T, float>) {
const float d = x[ib].d;
Expand Down Expand Up @@ -440,6 +452,16 @@ static __device__ __forceinline__ void dequantize_V_q4_1(const void * __restrict
((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]) + m;
}
} else
#else // FP16_AVAILABLE
// Software-FP16 fallback: see dequantize_V_q4_0 for the rationale.
if constexpr (std::is_same_v<T, half>) {
const float2 dm = __half22float2(x[ib].dm);

#pragma unroll
for (int l0 = 0; l0 < ne; l0 += 2) {
((half2 *) dst)[l0/2] = __floats2half2_rn(dm.x * q8[l0 + 0] + dm.y, dm.x * q8[l0 + 1] + dm.y);
}
} else
#endif // FP16_AVAILABLE
if constexpr (std::is_same_v<T, float>) {
const float2 dm = __half22float2(x[ib].dm);
Expand Down Expand Up @@ -490,6 +512,16 @@ static __device__ __forceinline__ void dequantize_V_q5_0(const void * __restrict
((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]);
}
} else
#else // FP16_AVAILABLE
// Software-FP16 fallback: see dequantize_V_q4_0 for the rationale.
if constexpr (std::is_same_v<T, half>) {
const float d = __half2float(x[ib].d);

#pragma unroll
for (int l0 = 0; l0 < ne; l0 += 2) {
((half2 *) dst)[l0/2] = __floats2half2_rn(d * q8[l0 + 0], d * q8[l0 + 1]);
}
} else
#endif // FP16_AVAILABLE
if constexpr (std::is_same_v<T, float>) {
const float d = x[ib].d;
Expand Down Expand Up @@ -540,6 +572,16 @@ static __device__ __forceinline__ void dequantize_V_q5_1(const void * __restrict
((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]) + m;
}
} else
#else // FP16_AVAILABLE
// Software-FP16 fallback: see dequantize_V_q4_0 for the rationale.
if constexpr (std::is_same_v<T, half>) {
const float2 dm = __half22float2(x[ib].dm);

#pragma unroll
for (int l0 = 0; l0 < ne; l0 += 2) {
((half2 *) dst)[l0/2] = __floats2half2_rn(dm.x * q8[l0 + 0] + dm.y, dm.x * q8[l0 + 1] + dm.y);
}
} else
#endif // FP16_AVAILABLE
if constexpr (std::is_same_v<T, float>) {
const float2 dm = __half22float2(x[ib].dm);
Expand Down Expand Up @@ -573,6 +615,16 @@ static __device__ __forceinline__ void dequantize_V_q8_0(const void * __restrict
((half2 *) dst)[l0/2] = d * make_half2(qs[l0 + 0], qs[l0 + 1]);
}
} else
#else // FP16_AVAILABLE
// Software-FP16 fallback: see dequantize_V_q4_0 for the rationale.
if constexpr (std::is_same<T, half>::value) {
const float d = __half2float(x[ib].d);

#pragma unroll
for (int l0 = 0; l0 < ne; l0 += 2) {
((half2 *) dst)[l0/2] = __floats2half2_rn(d * qs[l0 + 0], d * qs[l0 + 1]);
}
} else
#endif // FP16_AVAILABLE
if constexpr (std::is_same<T, float>::value) {
const float d = x[ib].d;
Expand Down
Loading