diff --git a/csrc/flash_attn/fmha_api.cpp b/csrc/flash_attn/fmha_api.cpp
index be2f501..73d9cdb 100644
--- a/csrc/flash_attn/fmha_api.cpp
+++ b/csrc/flash_attn/fmha_api.cpp
@@ -413,8 +413,8 @@ mha_bwd(const at::Tensor &dout,  // total_q x num_heads, x head_size
     const int total_k = k.size(TOTAL_DIM);
     TORCH_CHECK(batch_size > 0);
     TORCH_CHECK((head_size % 8 == 0) && (head_size <= 128));
-    if (head_size > 64) {  // TODO: eventually we should support SM86 and SM70 with d=128 as well
-        TORCH_CHECK(is_sm80 || is_sm90);
+    if (head_size > 64) {
+        TORCH_CHECK(is_sm80 || is_sm90, "FlashAttention backward for head dim > 64 requires A100 or H100 GPUs as the implementation needs a large amount of shared memory.");
     }
 
     CHECK_SHAPE(q, total_q, num_heads, head_size);
diff --git a/setup.py b/setup.py
index ebf16b3..ca7f209 100644
--- a/setup.py
+++ b/setup.py
@@ -162,7 +162,7 @@ ext_modules.append(
 
 setup(
     name="flash_attn",
-    version="1.0.3.post0",
+    version="1.0.4",
     packages=find_packages(
         exclude=("build", "csrc", "include", "tests", "dist", "docs", "benchmarks", "flash_attn.egg-info",)
     ),
diff --git a/training/Dockerfile b/training/Dockerfile
index e915ee3..7e113ee 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==1.0.3.post0
+RUN pip install flash-attn==1.0.4
 
 # Install CUDA extensions for cross-entropy, fused dense, layer norm
 RUN git clone https://github.com/HazyResearch/flash-attention \
-    && cd flash-attention && git checkout v1.0.3.post0 \
+    && cd flash-attention && git checkout v1.0.4 \
     && cd csrc/fused_softmax && pip install . && cd ../../ \
     && cd csrc/rotary && pip install . && cd ../../ \
     && cd csrc/xentropy && pip install . && cd ../../ \