aboutsummaryrefslogtreecommitdiff
path: root/pplx_embed_cuda.cu
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-10 07:30:34 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 08:27:03 +0200
commitc14151290df20e1a2ebc9fb06a77b075ca4ad959 (patch)
treed4b53dd131f0e5f5f8f5dd60727cc13a76c42534 /pplx_embed_cuda.cu
parentdcc54e7a267f88ed456209c63f58c7130274e8da (diff)
downloadffwd-c14151290df20e1a2ebc9fb06a77b075ca4ad959.tar.gz
ffwd-c14151290df20e1a2ebc9fb06a77b075ca4ad959.zip
cli: rename flags, add --stream, split serving into pplx-embed-server
Diffstat (limited to 'pplx_embed_cuda.cu')
-rw-r--r--pplx_embed_cuda.cu4
1 files changed, 2 insertions, 2 deletions
diff --git a/pplx_embed_cuda.cu b/pplx_embed_cuda.cu
index bda1be2..a00fc84 100644
--- a/pplx_embed_cuda.cu
+++ b/pplx_embed_cuda.cu
@@ -537,14 +537,14 @@ static int launch_check(void)
}
// Compute type for CUDA GEMMs. Default is exact F32 (== cublasSgemm).
-// PPLX_CUDA_FAST_GEMM={bf16,tf32,16f} selects a reduced-precision tensor-core
+// PPLX_CUDA_GEMM_MODE={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 env_init = 0;
if (!g_gemm_compute_set && !env_init) {
- const char *e = getenv("PPLX_CUDA_FAST_GEMM");
+ const char *e = getenv("PPLX_CUDA_GEMM_MODE");
if (e && parse_gemm_compute(e, &g_gemm_compute) == 0)
g_gemm_compute_set = 1;
env_init = 1;