Re-enable compilation for Hopper

flash_attn/__init__.py

-__version__ = "2.2.4"
+__version__ = "2.2.4.post1"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,

setup.py

@@ -122,10 +122,10 @@ if not SKIP_CUDA_BUILD:
     # cc_flag.append("arch=compute_75,code=sm_75")
     cc_flag.append("-gencode")
     cc_flag.append("arch=compute_80,code=sm_80")
-    # if CUDA_HOME is not None:
-    #     if bare_metal_version >= Version("11.8"):
-    #         cc_flag.append("-gencode")
-    #         cc_flag.append("arch=compute_90,code=sm_90")
+    if CUDA_HOME is not None:
+        if bare_metal_version >= Version("11.8"):
+            cc_flag.append("-gencode")
+            cc_flag.append("arch=compute_90,code=sm_90")
 
     # HACK: The compiler flag -D_GLIBCXX_USE_CXX11_ABI is set to be the same as
     # torch._C._GLIBCXX_USE_CXX11_ABI

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==2.2.4
+RUN pip install flash-attn==2.2.4.post1
 
 # Install CUDA extensions for fused dense, layer norm
 RUN git clone https://github.com/HazyResearch/flash-attention \
-    && cd flash-attention && git checkout v2.2.4 \
+    && cd flash-attention && git checkout v2.2.4.post1 \
     && cd csrc/layer_norm && pip install . && cd ../../ \
     && cd csrc/fused_dense_lib && pip install . && cd ../../ \
     && cd .. && rm -rf flash-attention