From afd99576d73f4ae49635ae6fa16624db9b03d252 Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Mon, 22 Jun 2026 17:31:06 +0200 Subject: cuda: store weight dtype per context --- libffwd/backend_cuda.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) (limited to 'libffwd/backend_cuda.c') 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); } -- cgit v1.3.1