mark failing

flash_attn/flash_attn_triton_amd/test.py

@@ -918,26 +918,26 @@ def test_op_prefill_fp8(Z, HQ, HK, N_CTX_Q, N_CTX_K, D_HEAD, causal, dropout_p,
         (1, 1, 1, 2, 2, 2),  # small enough to debug
         (1, 1, 1, 128, 128, 32),  # only one block
         (1, 1, 1, 127, 127, 32),  # only one block but with masking
-        (1, 1, 1, 129, 129, 1),  # two blocks with 2nd block small enough to debug
+        (1, 1, 1, 129, 129, 1),  # two blocks with 2nd block small enough to debug # fails similar to the normal case
         (1, 1, 1, 350, 350, 1),  # two blocks with 2nd block small enough to debug
         (1, 1, 1, 350, 350, 68),  # generic masking on q, k and head
         (4, 1, 1, 512, 512, 128),  # batch > 1
-        (4, 8, 2, 512, 512, 128),  # GQA
-        (4, 8, 2, 512, 512, 68),   # non-power-of-2 head_dim
+        (4, 8, 2, 512, 512, 128),  # GQA  # fails
+        (4, 8, 2, 512, 512, 68),   # non-power-of-2 head_dim # fails
         (4, 8, 2, 500, 500, 68),  # comprehensive case for seqlen q == k
         # seqlen q > k
         (1, 1, 1, 64, 32, 8),
         (1, 1, 1, 192, 128, 32),
-        (4, 8, 2, 1024, 512, 68), 
+        (4, 8, 2, 1024, 512, 68),  # fails
         (1, 1, 1, 729, 516, 68),  
-        (16, 16, 4, 2753, 1528, 68),  # a comprehensive seqlen_q > seqlen_k
+        (16, 16, 4, 2753, 1528, 68),  # a comprehensive seqlen_q > seqlen_k # fails
         # seqlen q < k
         (1, 1, 1, 32, 64, 8),
         (1, 1, 1, 128, 192, 32),
-        (4, 8, 2, 512, 1024, 68),
+        (4, 8, 2, 512, 1024, 68), # fails
         (1, 1, 1, 200, 413, 1),
         (1, 1, 1, 782, 1546, 1),
-        (16, 16, 4, 1528, 2753, 68),  # a comprehensive seqlen_q < seqlen_k
+        (16, 16, 4, 1528, 2753, 68),  # a comprehensive seqlen_q < seqlen_k # fails
         # misc
         (1, 1, 1, 1, 1, 1),
         (1, 1, 1, 2, 4, 16),