Fix test with alibi and cache_leftpad

tests/test_flash_attn.py

@@ -27,7 +27,7 @@ is_sm90 = torch.cuda.get_device_capability("cuda") == (9, 0)
 
 
 def attn_bias_from_alibi_slopes(
-    slopes, seqlen_q, seqlen_k, query_padding_mask=None, key_padding_mask=None, causal=False
+    slopes, seqlen_q, seqlen_k, query_padding_mask=None, key_padding_mask=None, causal=False, key_leftpad=None
 ):
     batch, nheads = slopes.shape
     device = slopes.device
@@ -37,6 +37,10 @@ def attn_bias_from_alibi_slopes(
     else:
         row_idx = rearrange(torch.arange(seqlen_q, device=device, dtype=torch.long), "s -> s 1")
         col_idx = torch.arange(seqlen_k, device=device, dtype=torch.long)
+        if key_leftpad is not None:
+            key_leftpad = rearrange(key_leftpad, "b -> b 1 1 1")
+            col_idx = repeat(col_idx, "s -> b 1 1 s", b=key_leftpad.shape[0])
+            col_idx = torch.where(col_idx >= key_leftpad, col_idx - key_leftpad, 2**32)
         sk = (
             seqlen_k
             if key_padding_mask is None
@@ -1993,7 +1997,7 @@ def test_flash_attn_kvcache(
     if alibi:
         alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
         attn_bias = attn_bias_from_alibi_slopes(
-            alibi_slopes, seqlen_q, seqlen_k, None, key_padding_mask, causal=causal
+            alibi_slopes, seqlen_q, seqlen_k, None, key_padding_mask, causal=causal, key_leftpad=cache_leftpad
         )
     else:
         alibi_slopes, attn_bias = None, None