fix previous typos

Author: Rachmanino

examples/attention_sink/example_gqa_sink_bwd_bhsd.py

@@ -211,7 +211,7 @@ def flashattn_bwd(batch, heads, seq_len, dim, groups, window_size=None):  # None
     block_M, block_N, num_stages, threads = get_bwd_configs()
 
     if window_size is not None:
-        assert window_size % block_N == 0, "window_size must be divisible by block_M"
+        assert window_size % block_N == 0, "window_size must be divisible by block_N"
 
     @T.prim_func
     def flash_bwd(

examples/attention_sink/example_mha_sink_bwd_bhsd.py

@@ -216,7 +216,7 @@ def flashattn_bwd(
     accum_dtype = "float"
 
     if window_size is not None:
-        assert window_size % block_N == 0, "window_size must be divisible by block_M"
+        assert window_size % block_N == 0, "window_size must be divisible by block_N"
 
     @T.prim_func
     def flash_bwd(

examples/flash_attention/example_mha_bwd_wgmma_pipelined.py

@@ -146,7 +146,7 @@ def flash_bwd_post(
 @tilelang.jit(pass_configs={
     tilelang.PassConfigKey.TL_ENABLE_FAST_MATH: True,
 })
-def flashattn_bwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
+def flashattn_bwd(batch, heads, seq_len, dim, is_causal, block_M, block_N):
     sm_scale = (1.0 / dim)**0.5
     scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
     shape = [batch, seq_len, heads, dim]
@@ -198,7 +198,7 @@ def flash_bwd(
             T.copy(V[bz, by * block_M:(by + 1) * block_M, bx, :], V_shared)
             T.clear(dv)
             T.clear(dk)
-            loop_st = T.floordiv(by * block_M, block_N) if is_casual else 0
+            loop_st = T.floordiv(by * block_M, block_N) if is_causal else 0
             loop_ed = T.ceildiv(seq_len, block_N)
             for k in T.Pipelined(loop_st, loop_ed, num_stages=2):
                 T.copy(Q[bz, k * block_N:(k + 1) * block_N, bx, :], q)
@@ -219,7 +219,7 @@ def flash_bwd(
                 T.copy(lse[bz, bx, k * block_N:(k + 1) * block_N], lse_shared)
                 for i, j in T.Parallel(block_M, block_N):
                     qkT[i, j] = T.exp2(qkT[i, j] * scale - lse_shared[j])
-                if is_casual:
+                if is_causal:
                     for i, j in T.Parallel(block_M, block_N):
                         qkT[i, j] = T.if_then_else(by * block_M + i <= k * block_N + j, qkT[i, j],
                                                    0)