diff options
| author | Johannes Gäßler <johannesg@5d6.de> | 2023-12-23 09:16:33 +0100 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2023-12-23 09:16:33 +0100 |
| commit | e0a4002273907b2c414b6b5442d99e08bfe2df35 (patch) | |
| tree | 6c9ecc8f515c9f0ececc609150be3cb467043498 | |
| parent | 7082d24cec35e9ce9147535a2224dfc67ee0a78c (diff) | |
CUDA: fixed row rounding for 0 tensor splits (#4594)
| -rw-r--r-- | ggml-cuda.cu | 8 |
1 files changed, 6 insertions, 2 deletions
diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 7c2a834e..490081ca 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -7937,12 +7937,16 @@ static void ggml_cuda_op_mul_mat( if (id != 0) { row_low[id] = ne01*g_tensor_split[id]; - row_low[id] -= row_low[id] % rounding; + if (row_low[id] < ne01) { + row_low[id] -= row_low[id] % rounding; + } } if (id != g_device_count - 1) { row_high[id] = ne01*g_tensor_split[id + 1]; - row_high[id] -= row_high[id] % rounding; + if (row_high[id] < ne01) { + row_high[id] -= row_high[id] % rounding; + } } } } |