aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-08 22:48:11 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 08:27:03 +0200
commit940303fe2fa3d55746e86fc199d8900f0634568a (patch)
tree161dd1a3d6dcba9c3d776a77dfb2a955f18d4be1
parent26aa937f023652648f201ad849d8dd644ca0261c (diff)
downloadffwd-940303fe2fa3d55746e86fc199d8900f0634568a.tar.gz
ffwd-940303fe2fa3d55746e86fc199d8900f0634568a.zip
cuda: clarify comment gemm_compute
-rw-r--r--pplx_embed_cuda.cu5
1 files changed, 3 insertions, 2 deletions
diff --git a/pplx_embed_cuda.cu b/pplx_embed_cuda.cu
index 4c192be..ea04a10 100644
--- a/pplx_embed_cuda.cu
+++ b/pplx_embed_cuda.cu
@@ -499,8 +499,9 @@ static int launch_check(void)
}
// Compute type for the projection GEMMs. Default is exact F32 (== cublasSgemm).
-// PPLX_CUDA_FAST_GEMM={bf16,tf32,16f} opts into tensor-core matmul that keeps
-// F32 inputs/outputs and F32 accumulation but rounds operands internally.
+// PPLX_CUDA_FAST_GEMM={bf16,tf32,16f} selects a reduced-precision tensor-core
+// matmul that keeps F32 inputs/outputs and F32 accumulation but rounds operands
+// internally.
static cublasComputeType_t gemm_compute(void)
{
static int init = 0;