diff --git a/configs/requirements_training.txt b/configs/requirements_training.txt
index 910ef161..0692586b 100644
--- a/configs/requirements_training.txt
+++ b/configs/requirements_training.txt
@@ -31,7 +31,7 @@ omegaconf>=2.4.0
 # Install the one matching the target GPU/CUDA toolchain:
 #   pip install flash-attn           # FA2 (Hopper H100/Ampere A100)
 #   pip install "flash-attn-4[cu13]" # FA4 (Blackwell B200, CUDA 13)
-flash-attn>=2.7.0
+flash-attn>=2.8.3.post2
 
 # Security: Pin transitive dependencies to secure versions
 urllib3>=2.8.0  # Fixes GHSA-gm62-xv2j-4w53, GHSA-2xpw-w6gg-jr37 (CVE-2025-66418, CVE-2025-66471) — 2.6.2 is the minimum that enforces the decompression bomb limit