skip compiler bug on navi

ROCm · Dec 4, 2024 · 35c8925 · 35c8925
1 parent 76a48f4
commit 35c8925
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/tests/test_flash_attn_triton_amd.py b/tests/test_flash_attn_triton_amd.py
@@ -18,7 +18,7 @@
 from flash_attn.bert_padding import pad_input, unpad_input
 from flash_attn.flash_attn_interface import _get_block_size_n
 from flash_attn.layers.rotary import apply_rotary_emb
-from flash_attn.flash_attn_triton_amd.utils import DEBUG
+from flash_attn.flash_attn_triton_amd.utils import DEBUG, is_rdna
 
 # Test ROCM Triton Backend
 USE_TRITON_ROCM = os.getenv("FLASH_ATTENTION_TRITON_AMD_ENABLE", "FALSE") == "TRUE"
@@ -1570,6 +1570,10 @@ def test_flash_attn_varlen_output(
 )
 # @pytest.mark.parametrize('seqlen_q,seqlen_k', [(256, 128)])
 def test_flash_attn_causal(seqlen_q, seqlen_k, swap_sq_sk, d, local, dtype):
+    if USE_TRITON_ROCM:
+        if is_rdna():
+            if seqlen_q == 1 and seqlen_k == 239 and d == 256:
+                pytest.skip("This config doesnot work on RDNA Devices.")
     if (
         max(seqlen_q, seqlen_k) >= 2048
         and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30