tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-27 11:43:35 -04:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: disable fp32->fp16 nnpa conversions for now

Browse Source 
there are some conversion failures in nnpa that requires the eyes of an
ibm stsm. will create a separate pr to introduce the fp32->fp16 change.

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-06-21 16:58:43 +08:00
parent 4f017d718a
commit 27131e5f34
2 changed files with 0 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -3142,18 +3142,6 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
_mm_storel_epi64((__m128i *)(y + i), y_vec);
}
#endif
#if defined(__NNPA__)
for (; i + 7 < n; i += 8) {
float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
uint16x8_t v_xd = vec_round_from_fp32(v_xh, v_xl, 0);
uint16x8_t v_xdi = vec_convert_to_fp16(v_xd, 0);
vec_xst(v_xdi, 0, (ggml_fp16_t *)(y + i));
printf("breakpoint");
}
#endif
for (; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(x[i]);
}