diff options
| author | Johannes Gäßler <johannesg@5d6.de> | 2024-06-01 23:26:10 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-06-01 23:26:10 +0200 |
| commit | e141ce624af57bdffbaf57014a044eb1d9689230 (patch) | |
| tree | f41f3a1b08c3ce608de244e0e45ff1ed699d2ce9 /ggml-cuda | |
| parent | 2e666832e6ac78194edf030bd1c295e21bdb022c (diff) | |
Fix FlashAttention debug test, FP32 assert (#7684)
Diffstat (limited to 'ggml-cuda')
| -rw-r--r-- | ggml-cuda/fattn-vec-f32.cuh | 4 |
1 files changed, 0 insertions, 4 deletions
diff --git a/ggml-cuda/fattn-vec-f32.cuh b/ggml-cuda/fattn-vec-f32.cuh index ce23a4eb..ddf0c837 100644 --- a/ggml-cuda/fattn-vec-f32.cuh +++ b/ggml-cuda/fattn-vec-f32.cuh @@ -278,14 +278,10 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, template <int D, ggml_type type_K, ggml_type type_V> void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - ggml_tensor * KQV = dst; ggml_tensor * Q = dst->src[0]; ggml_tensor * K = dst->src[1]; ggml_tensor * V = dst->src[2]; - const int32_t precision = KQV->op_params[2]; - GGML_ASSERT(precision == GGML_PREC_DEFAULT); - GGML_ASSERT(K->type == type_K); GGML_ASSERT(V->type == type_V); |