upstream-archive/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-09-29 01:17:49 -05:00
Code Issues Packages Projects Releases Wiki Activity

metal: add the missing f32 x bf16 mul_mv variants (llama/28741)

Browse Source 
ggml_conv_1d_dw builds its im2col as f32 when the kernel is bf16, then
multiplies the two, so a depthwise convolution over bf16 weights asks
for kernel_mul_mv_f32_bf16, which was never instantiated. The base, the
_4 and the _short families are filled in next to their bf16 neighbours,
inside the same runtime guard, so a device without bf16 support is
unaffected.
This commit is contained in:
Pascal
2026-09-23 20:46:47 +03:00
committed by Georgi Gerganov
parent f48aebe2f1
commit 8917ea0432
1 changed files with 3 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/kernels/mul_mv.metal
+3
View File
@@ -1043,6 +1043,7 @@ template [[host_name("kernel_mul_mv_f16_f16")]] kernel mul_mv_t_t kernel_mul_m
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mv_bf16_f32")]] kernel mul_mv_t_t kernel_mul_mv_t_t<bfloat, float>;
template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t_t kernel_mul_mv_t_t<bfloat, bfloat>;
template [[host_name("kernel_mul_mv_f32_bf16")]] kernel mul_mv_t_t kernel_mul_mv_t_t<float, bfloat>;
#endif
template<typename T0, typename T04, typename T1, typename T14, short NR0, typename args_t>
@@ -1167,6 +1168,7 @@ template [[host_name("kernel_mul_mv_f16_f16_4")]] kernel mul_mv_t_t_4 kernel_m
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mv_bf16_f32_4")]] kernel mul_mv_t_t_4 kernel_mul_mv_t_t_4<bfloat, bfloat4, float, float4>;
template [[host_name("kernel_mul_mv_bf16_bf16_4")]] kernel mul_mv_t_t_4 kernel_mul_mv_t_t_4<bfloat, bfloat4, bfloat, bfloat4>;
template [[host_name("kernel_mul_mv_f32_bf16_4")]] kernel mul_mv_t_t_4 kernel_mul_mv_t_t_4<float, float4, bfloat, bfloat4>;
#endif
template<typename T0, typename T1, typename args_t>
@@ -1232,6 +1234,7 @@ template [[host_name("kernel_mul_mv_f16_f16_short")]] kernel mul_mv_t_t_short_t
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mv_bf16_f32_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short<bfloat, float>;
template [[host_name("kernel_mul_mv_bf16_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short<bfloat, bfloat>;
template [[host_name("kernel_mul_mv_f32_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short<float, bfloat>;
#endif
template<int nr0, typename args_t>