diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 70195b4..6765bee 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -151,7 +151,7 @@ jobs:
           export PATH=/usr/local/nvidia/bin:/usr/local/nvidia/lib64:$PATH
           export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
           # Currently for this setting the runner goes OOM if we pass --threads 4 to nvcc
-          if [[ ${MATRIX_CUDA_VERSION} =~ "12." && ${MATRIX_TORCH_VERSION} == "2.1" ]]; then
+          if [[ ( ${MATRIX_CUDA_VERSION} == "121" || ${MATRIX_CUDA_VERSION} == "122" ) && ${MATRIX_TORCH_VERSION} == "2.1" ]]; then
             export FLASH_ATTENTION_FORCE_SINGLE_THREAD="TRUE"
           fi
           # Limit MAX_JOBS otherwise the github runner goes OOM
diff --git a/flash_attn/__init__.py b/flash_attn/__init__.py
index 5efd449..6695f00 100644
--- a/flash_attn/__init__.py
+++ b/flash_attn/__init__.py
@@ -1,4 +1,4 @@
-__version__ = "2.1.2.post2"
+__version__ = "2.1.2.post3"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,
diff --git a/training/Dockerfile b/training/Dockerfile
index 133703e..009770d 100644
--- a/training/Dockerfile
+++ b/training/Dockerfile
@@ -85,11 +85,11 @@ RUN pip install transformers==4.25.1 datasets==2.8.0 pytorch-lightning==1.8.6 tr
 RUN pip install git+https://github.com/mlcommons/logging.git@2.1.0
 
 # Install FlashAttention
-RUN pip install flash-attn==2.1.2.post2
+RUN pip install flash-attn==2.1.2.post3
 
 # Install CUDA extensions for cross-entropy, fused dense, layer norm
 RUN git clone https://github.com/HazyResearch/flash-attention \
-    && cd flash-attention && git checkout v2.1.2.post2 \
+    && cd flash-attention && git checkout v2.1.2.post3 \
     && cd csrc/fused_softmax && pip install . && cd ../../ \
     && cd csrc/rotary && pip install . && cd ../../ \
     && cd csrc/xentropy && pip install . && cd ../../ \