Set block size to 64 x 64 for kvcache to avoid nvcc segfaults

Dao-AILab · Sep 17, 2023 · c984208 · c984208
1 parent 8c8b4d3
commit c984208
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 14 deletions.
diff --git a/csrc/flash_attn/src/flash_fwd_launch_template.h b/csrc/flash_attn/src/flash_fwd_launch_template.h
@@ -115,18 +115,12 @@ void run_flash_splitkv_fwd(Flash_fwd_params &params, cudaStream_t stream) {
 
 template<typename T, int Headdim>
 void run_mha_fwd_splitkv_dispatch(Flash_fwd_params &params, cudaStream_t stream) {
-    auto dprops = at::cuda::getCurrentDeviceProperties();
-    bool is_sm8x = dprops->major == 8 && dprops->minor > 0;
     constexpr int kBlockM = 64;  // Fixed for all head dimensions
-    if (!is_sm8x) {   // A100, H100
-        // TD [2023-08-28]: nvcc segfaults for headdim 96 with block size 64 x 256,
-        // and for headdim 192 with block size 64 x 128.
-        constexpr int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 160 ? 128 : 64);
-        run_flash_splitkv_fwd<Flash_fwd_kernel_traits<Headdim, kBlockM, kBlockN, 4, false, false, T>>(params, stream);
-    } else {  // Only 99KB of smem, so we have to set kBlockN smaller for Headdim 160 and above
-        constexpr int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 128 ? 128 : 64);
-        run_flash_splitkv_fwd<Flash_fwd_kernel_traits<Headdim, kBlockM, kBlockN, 4, false, false, T>>(params, stream);
-    }
+    // TD [2023-08-28]: nvcc segfaults for headdim 96 with block size 64 x 256,
+    // and for headdim 192 with block size 64 x 128.
+    // Also for headdim 160 with block size 64 x 128 after the rotary addition.
+    constexpr int kBlockN = Headdim <= 64 ? 256 : (Headdim <= 128 ? 128 : 64);
+    run_flash_splitkv_fwd<Flash_fwd_kernel_traits<Headdim, kBlockM, kBlockN, 4, false, false, T>>(params, stream);
 }
 
 template<typename T>

diff --git a/flash_attn/__init__.py b/flash_attn/__init__.py
@@ -1,4 +1,4 @@
-__version__ = "2.2.3"
+__version__ = "2.2.3.post1"
 
 from flash_attn.flash_attn_interface import (
     flash_attn_func,

diff --git 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==2.2.3
+RUN pip install flash-attn==2.2.3.post1
 
 # Install CUDA extensions for cross-entropy, fused dense, layer norm
 RUN git clone https://github.com/HazyResearch/flash-attention \
-    && cd flash-attention && git checkout v2.2.3 \
+    && cd flash-attention && git checkout v2.2.3.post1 \
     && cd csrc/fused_softmax && pip install . && cd ../../ \
     && cd csrc/rotary && pip install . && cd ../../ \
     && cd csrc/layer_norm && pip install . && cd ../../ \