From e141ce624af57bdffbaf57014a044eb1d9689230 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Sat, 1 Jun 2024 23:26:10 +0200 Subject: Fix FlashAttention debug test, FP32 assert (#7684) --- ggml-cuda/fattn-vec-f32.cuh | 4 ---- 1 file changed, 4 deletions(-) (limited to 'ggml-cuda') 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 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); -- cgit v1.2.3