Version 1.0 Disabled flash attention

asigalov61 · Oct 12, 2023 · 9490b1b · 9490b1b
1 parent f6e958c
commit 9490b1b
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/tegridy-tools/X-Transformer/x_transformer_1_23_2.py b/tegridy-tools/X-Transformer/x_transformer_1_23_2.py
@@ -256,7 +256,7 @@ def flash_attn(
 
         # pytorch 2.0 flash attn: q, k, v, mask, dropout, causal, softmax_scale
 
-        with torch.backends.cuda.sdp_kernel(**config._asdict()):
+        with torch.backends.cuda.sdp_kernel(enable_math=True, enable_mem_efficient=True):
             out = F.scaled_dot_product_attention(
                 q, k, v,
                 attn_mask = mask,