From d7b133dc1f013e38b1c1820bcd64e3b855c2ea7c Mon Sep 17 00:00:00 2001 From: Sunny Liu Date: Wed, 22 Jan 2025 20:33:13 -0500 Subject: [PATCH] llama sdpa patching WIP - static class function import --- src/axolotl/monkeypatch/llama_patch_multipack.py | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/src/axolotl/monkeypatch/llama_patch_multipack.py b/src/axolotl/monkeypatch/llama_patch_multipack.py index f4779d69d..9f03f551d 100644 --- a/src/axolotl/monkeypatch/llama_patch_multipack.py +++ b/src/axolotl/monkeypatch/llama_patch_multipack.py @@ -28,13 +28,11 @@ def hijack_llama_prepare_4d_mask(): # ) def llama_patched_prepare_4d_causal_attention_mask_with_cache_position( - attention_mask: Optional[torch.Tensor], - *args, + attention_mask: Optional[torch.Tensor], *args, **kwargs ): dtype = torch.bfloat16 if is_torch_bf16_gpu_available() else torch.float32 return LlamaModel._prepare_4d_causal_attention_mask_with_cache_position( - mask_2d_to_4d(attention_mask, dtype=dtype), - *args, + mask_2d_to_4d(attention_mask, dtype=dtype), *args, **kwargs ) LlamaModel._prepare_4d_causal_attention_mask_with_cache_position = ( # pylint: disable=protected-access