ClaudeBot/llama.cpp
0
0
Fork 1
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-17 20:31:47 +02:00
Code Issues Packages Projects Releases Wiki Activity

spacemit : fix wrong transpose function for int16 data (#25161)

Browse Source 
The `sizeof(int16_t)` branch in `permute_transpose_impl` calls
`rvv_transposed_s32_mn_to_nm` instead of `rvv_transposed_s16_mn_to_nm`.
This is a copy-paste bug from the `sizeof(int32_t)` branch above it.

The s32 function uses 32-bit segment load/stores (`vssseg8e32.v`) on 16-bit
data, reading 2x bytes per element and producing completely wrong
transposition results -- 14 out of 16 positions are corrupted for a 4x4
int16 matrix.

The correct function `rvv_transposed_s16_mn_to_nm` already exists (line 390)
and is used elsewhere in flash attention (line 1488).
This commit is contained in:
I3eg1nner
2026-09-16 14:19:47 +03:00
committed by GitHub
parent 60199339bc
commit f266648fa9
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp
+1 -1
View File
@@ -639,7 +639,7 @@ static void permute_transpose_impl(const ggml_tensor * src0,
}
} else if (n_src_stride == sizeof(int16_t)) {
for (int64_t bi = ith; bi < batch; bi += nth) {
rvv_transposed_s32_mn_to_nm((int8_t *) ((char *) dst->data + bi * batch_stride), n_dst_stride,
rvv_transposed_s16_mn_to_nm((int8_t *) ((char *) dst->data + bi * batch_stride), n_dst_stride,
(int8_t *) ((char *) src0->data + bi * batch_stride), m_src_stride, m, n);
}
} else {