aboutsummaryrefslogtreecommitdiff
path: root/libffwd/backend_cuda.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-22 17:31:06 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-22 17:31:06 +0200
commitafd99576d73f4ae49635ae6fa16624db9b03d252 (patch)
tree0ca8b99c7aaca6197fe6c82eeb0e42996ae44e7b /libffwd/backend_cuda.c
parent74fc389fb382cc79f67b67591569144b3fa393bd (diff)
downloadffwd-afd99576d73f4ae49635ae6fa16624db9b03d252.tar.gz
ffwd-afd99576d73f4ae49635ae6fa16624db9b03d252.zip
cuda: store weight dtype per context
Diffstat (limited to 'libffwd/backend_cuda.c')
-rw-r--r--libffwd/backend_cuda.c6
1 files changed, 3 insertions, 3 deletions
diff --git a/libffwd/backend_cuda.c b/libffwd/backend_cuda.c
index db39345..2635fb2 100644
--- a/libffwd/backend_cuda.c
+++ b/libffwd/backend_cuda.c
@@ -23,8 +23,8 @@ void ffwd_cli_help(FILE *f) {
fputs(" --gpu-gemm-mode MODE\n"
" GPU GEMM compute mode: f32, tf32, bf16, or 16f (default: f32)\n"
" --gpu-weight-dtype DTYPE\n"
- " GPU weight storage: f32 or bf16 (default: f32). bf16 halves\n"
- " weight memory and uses BF16 tensor cores\n",
+ " GPU weight storage: f32 or bf16 (default: model dtype).\n"
+ " bf16 halves weight memory and uses BF16 tensor cores\n",
f);
}
@@ -32,7 +32,7 @@ void ffwd_server_help(FILE *f) {
fputs(" --gpu-gemm-mode MODE GPU GEMM compute: f32, tf32, bf16, or 16f\n"
" (default: f32, exact)\n"
" --gpu-weight-dtype DTYPE GPU weight storage: f32 or bf16 (default:\n"
- " f32); bf16 halves weight memory\n",
+ " model dtype); bf16 halves weight memory\n",
f);
}