From 038a59e2cd0bdf484b76fa2cc1f403a5cf65b150 Mon Sep 17 00:00:00 2001 From: Mohammad Nasirifar Date: Thu, 12 Jun 2025 07:06:13 -0700 Subject: [PATCH] Initialize flash attn flag (#38768) _flash_supports_window_size is used further down in this file and relied on by e.g. [ring-flash-attention](https://github.com/zhuzilin/ring-flash-attention/blob/123f924/ring_flash_attn/adapters/hf_adapter.py#L9-L11). Even though it is an unexported name, it still makes sense to keep the state of `globals()` in this file consistent. --- src/transformers/modeling_flash_attention_utils.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/src/transformers/modeling_flash_attention_utils.py b/src/transformers/modeling_flash_attention_utils.py index f01db89f5b1..6ab4cb748aa 100644 --- a/src/transformers/modeling_flash_attention_utils.py +++ b/src/transformers/modeling_flash_attention_utils.py @@ -46,6 +46,9 @@ if is_torch_npu_available(): from .integrations.npu_flash_attention import npu_flash_attn_varlen_func as flash_attn_varlen_func +_flash_supports_window_size = False + + if flash_attn_func: _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)