From 338ee4bf525175044df39add6eb559e5cc76edf6 Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Tue, 23 Jun 2026 01:07:04 +0200 Subject: cuda: add bert ffn16 intermediate path --- tests/check_batch_parity.py | 2 +- tests/check_cuda_fast_gemm_parity.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) (limited to 'tests') diff --git a/tests/check_batch_parity.py b/tests/check_batch_parity.py index 3b02a34..fa94540 100755 --- a/tests/check_batch_parity.py +++ b/tests/check_batch_parity.py @@ -140,7 +140,7 @@ def main(): choices=(0, 8)) ap.add_argument("--mlx-quant-group-size", "--mlx-quantize-group-size", dest="mlx_quantize_group_size", type=int, default=64) - ap.add_argument("--cuda-weight-dtype", choices=("f32", "bf16"), default=None, + ap.add_argument("--cuda-weight-dtype", choices=("f32", "bf16", "f16"), default=None, help="force CUDA weight storage dtype (default: snapshot dtype)") ap.add_argument("texts", nargs="*") args = ap.parse_args() diff --git a/tests/check_cuda_fast_gemm_parity.py b/tests/check_cuda_fast_gemm_parity.py index 57adf72..57dc310 100755 --- a/tests/check_cuda_fast_gemm_parity.py +++ b/tests/check_cuda_fast_gemm_parity.py @@ -131,7 +131,7 @@ def main() -> int: ap.add_argument("--model-dir", required=True) ap.add_argument("--mode", choices=("f32", "tf32", "bf16", "16f"), default="tf32") - ap.add_argument("--weights", choices=("f32", "bf16"), default="f32") + ap.add_argument("--weights", choices=("f32", "bf16", "f16"), default="f32") ap.add_argument("--batch-size", type=int, default=8) ap.add_argument("--max-diff", type=float, default=None) ap.add_argument("--min-cosine", type=float, default=0.999) -- cgit v1.3.1