[CI] Add CI test for flash_attention examples (#558)

* [CI] Add CI test for flash_attention examples

* Update example_gqa_fwd_bshd.py

* Update example_mha_fwd_bshd_wgmma_pipelined.py

* [CI] Added conditional annotations for tests in flash_attention

* [CI] Added conditional annotations for tests in flash_attention

---------
Co-authored-by: Lei Wang <34334180+LeiWang1999@users.noreply.github.com>

examples/flash_attention/example_gqa_bwd.py

 import torch
 import torch.nn.functional as F
 import tilelang
-from tilelang import cached
 from tilelang.autotuner import *
 import tilelang.language as T
 import argparse
 
 
-def flashattn_fwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, block_N, groups=1):
+@tilelang.jit(out_idx=[3, 4])
+def flashattn_fwd(batch, heads, seq_len, dim_qk, dim_v, is_causal, block_M, block_N, groups=1):
     scale = (1.0 / dim_qk)**0.5 * 1.44269504  # log2(e)
     head_kv = heads // groups
     q_shape = [batch, seq_len, heads, dim_qk]
@@ -24,7 +24,7 @@ def flashattn_fwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, bloc
             Output: T.Tensor([batch, seq_len, heads, dim_v], dtype),  # type: ignore
             lse: T.Tensor([batch, heads, seq_len], accum_dtype),  # type: ignore
     ):
-        with T.Kernel(T.ceildiv(seq_len, block_M), heads, batch, threads=128) as (bx, by, bz):
+        with T.Kernel(T.ceildiv(seq_len, block_M), heads, batch, threads=256) as (bx, by, bz):
             Q_shared = T.alloc_shared([block_M, dim_qk], dtype)
             K_shared = T.alloc_shared([block_N, dim_qk], dtype)
             V_shared = T.alloc_shared([block_N, dim_v], dtype)
@@ -44,10 +44,10 @@ def flashattn_fwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, bloc
             T.fill(scores_max, -T.infinity(accum_dtype))
             loop_range = (
                 T.ceildiv(
-                    (bx + 1) * block_M, block_N) if is_casual else T.ceildiv(seq_len, block_N))
+                    (bx + 1) * block_M, block_N) if is_causal else T.ceildiv(seq_len, block_N))
             for k in T.Pipelined(loop_range, num_stages=1):
                 T.copy(K[bz, k * block_N:(k + 1) * block_N, by // groups, :], K_shared)
-                if is_casual:
+                if is_causal:
                     for i, j in T.Parallel(block_M, block_N):
                         acc_s[i, j] = T.if_then_else(bx * block_M + i >= k * block_N + j, 0,
                                                      -T.infinity(acc_s.dtype))
@@ -78,6 +78,7 @@ def flashattn_fwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, bloc
     return flash_fwd
 
 
+@tilelang.jit(out_idx=[2])
 def flashattn_bwd_preprocess(batch, heads, seq_len, dim_v):
     dtype = "float16"
     accum_dtype = "float"
@@ -113,6 +114,7 @@ def make_dq_layout(dQ):
                     lambda b, l, h, d: [b, l // 8, h, d // 8, (d % 2), 4 * (l % 8) + (d % 8) // 2])
 
 
+@tilelang.jit(out_idx=[1])
 def flashattn_bwd_postprocess(batch, heads, seq_len, dim_qk):
     dtype = "float16"
     accum_dtype = "float"
@@ -134,7 +136,8 @@ def flashattn_bwd_postprocess(batch, heads, seq_len, dim_qk):
     return flash_bwd_post
 
 
-def flashattn_bwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, block_N, groups=1):
+@tilelang.jit
+def flashattn_bwd(batch, heads, seq_len, dim_qk, dim_v, is_causal, block_M, block_N, groups=1):
     sm_scale = (1.0 / dim_qk)**0.5
     scale = (1.0 / dim_qk)**0.5 * 1.44269504  # log2(e)
     head_kv = heads // groups
@@ -156,7 +159,7 @@ def flashattn_bwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, bloc
             dK: T.Tensor(k_shape, dtype),  # type: ignore
             dV: T.Tensor(v_shape, dtype),  # type: ignore
     ):
-        with T.Kernel(heads, T.ceildiv(seq_len, block_M), batch, threads=256) as (bx, by, bz):
+        with T.Kernel(heads, T.ceildiv(seq_len, block_M), batch, threads=128) as (bx, by, bz):
             K_shared = T.alloc_shared([block_M, dim_qk], dtype)
             dsT_shared = T.alloc_shared([block_M, block_N], dtype)
             q = T.alloc_shared([block_N, dim_qk], dtype)
@@ -185,7 +188,7 @@ def flashattn_bwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, bloc
             T.copy(V[bz, by * block_M:(by + 1) * block_M, bx // groups, :], 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=1):
                 T.copy(Q[bz, k * block_N:(k + 1) * block_N, bx, :], q)
@@ -194,7 +197,7 @@ def flashattn_bwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, bloc
                 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)
@@ -225,16 +228,16 @@ def flashattn_bwd(batch, heads, seq_len, dim_qk, dim_v, is_casual, block_M, bloc
     return flash_bwd
 
 
+@torch.compile
 class _attention(torch.autograd.Function):
 
     @staticmethod
     def forward(ctx, q, k, v, causal, groups=1):
         BATCH, N_CTX, H, D_HEAD_QK = q.shape
         D_HEAD_V = v.shape[-1]
-        block_M = 64
+        block_M = 128
         block_N = 64
-        mod = cached(flashattn_fwd, [3, 4], BATCH, H, N_CTX, D_HEAD_QK, D_HEAD_V, causal, block_M,
-                     block_N, groups)
+        mod = flashattn_fwd(BATCH, H, N_CTX, D_HEAD_QK, D_HEAD_V, causal, block_M, block_N, groups)
         o, lse = mod(q, k, v)
         ctx.save_for_backward(q, k, v, o, lse)
         ctx.causal = causal
@@ -243,6 +246,9 @@ class _attention(torch.autograd.Function):
     @staticmethod
     def backward(ctx, do):
         q, k, v, o, lse = ctx.saved_tensors
+        BATCH, N_CTX, H, D_HEAD_QK = q.shape
+        HEAD_KV, D_HEAD_V, = v.shape[-2], v.shape[-1]
+        groups = H // HEAD_KV
 
         def maybe_contiguous(x):
             if x.stride(-1) != 1:
@@ -250,14 +256,20 @@ class _attention(torch.autograd.Function):
             return x
 
         do, q, k, v, o = [maybe_contiguous(x) for x in (do, q, k, v, o)]
-        block_M = 128
-        block_N = 128
-        mod_prep = cached(flashattn_bwd_preprocess, [2], BATCH, H, N_CTX, D_HEAD_V)
-        mod_post = cached(flashattn_bwd_postprocess, [1], BATCH, H, N_CTX, D_HEAD_QK)
+        block_M = 64
+        block_N = 32
+        mod_prep = flashattn_bwd_preprocess(BATCH, H, N_CTX, D_HEAD_V)
+        mod_post = flashattn_bwd_postprocess(BATCH, H, N_CTX, D_HEAD_QK)
         delta = mod_prep(o, do)
-        mod = cached(flashattn_bwd, [6, 7, 8], BATCH, H, N_CTX, D_HEAD_QK, D_HEAD_V, ctx.causal,
-                     block_M, block_N, groups)
-        dq, dk, dv = mod(q, k, v, do, lse, delta)
+        kernel = flashattn_bwd(BATCH, H, N_CTX, D_HEAD_QK, D_HEAD_V, ctx.causal, block_M, block_N,
+                               groups)
+        shape_q = [BATCH, N_CTX, H, D_HEAD_QK]
+        shape_k = [BATCH, N_CTX, HEAD_KV, D_HEAD_QK]
+        shape_v = [BATCH, N_CTX, HEAD_KV, D_HEAD_V]
+        dq = torch.zeros(shape_q, dtype=torch.float32, device=q.device)
+        dk = torch.zeros(shape_k, dtype=torch.float16, device=q.device)
+        dv = torch.zeros(shape_v, dtype=torch.float16, device=q.device)
+        kernel(q, k, v, do, lse, delta, dq, dk, dv)
         dq = mod_post(dq)
         return dq, dk, dv, None, None
 
@@ -290,22 +302,17 @@ def ref_program(Q, K, V, is_causal, groups=1):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=8, help='Batch size')
-    parser.add_argument('--h', type=int, default=32, help='Number of heads')
-    parser.add_argument('--n_ctx', type=int, default=1024, help='Context size')
-    parser.add_argument('--d_head_qk', type=int, default=192, help='Head dimension for Q/K')
-    parser.add_argument('--d_head_v', type=int, default=128, help='Head dimension for V')
-    parser.add_argument('--casual', type=bool, default=False, help='Casual flag')
-    parser.add_argument('--groups', type=int, default=16, help='groups')
-    args = parser.parse_args()
-    BATCH, H, N_CTX, D_HEAD_QK, D_HEAD_V, groups = args.batch, args.h, args.n_ctx, args.d_head_qk, args.d_head_v, args.groups
-    casual = args.casual
+def main(BATCH: int = 8,
+         H: int = 32,
+         N_CTX: int = 1024,
+         D_HEAD_QK: int = 192,
+         D_HEAD_V: int = 128,
+         groups: int = 16,
+         causal: bool = False):
     flops_per_qk = 2.0 * BATCH * H * N_CTX * N_CTX * D_HEAD_QK
     flops_per_v = 2.0 * BATCH * H * N_CTX * N_CTX * D_HEAD_V
     total_flops = 3 * flops_per_qk + 2 * flops_per_v
-    if casual:
+    if causal:
         total_flops *= 0.5
     Q = (
         torch.empty(BATCH, N_CTX, H, D_HEAD_QK, dtype=torch.half,
@@ -321,20 +328,21 @@ if __name__ == "__main__":
     dO = (
         torch.empty(BATCH, N_CTX, H, D_HEAD_V, dtype=torch.half,
                     device="cuda").normal_().requires_grad_())
-    O = attention(Q, K, V, casual, groups)
+    O = attention(Q, K, V, causal, groups)
     O.backward(dO, retain_graph=True)
     dQ, Q.grad = Q.grad.clone(), None
     dK, K.grad = K.grad.clone(), None
     dV, V.grad = V.grad.clone(), None
 
-    O_ref = ref_program(Q, K, V, casual, groups)
+    O_ref = ref_program(Q, K, V, causal, groups)
     O_ref.backward(dO, retain_graph=True)
     dQ_ref, Q.grad = Q.grad.clone(), None
     dK_ref, K.grad = K.grad.clone(), None
     dV_ref, V.grad = V.grad.clone(), None
 
     assert torch.allclose(O, O_ref, rtol=1e-2, atol=1e-2)
-    assert torch.allclose(dV, dV_ref, rtol=1e-2, atol=1e-2)
+    torch.testing.assert_close(dV, dV_ref, rtol=1e-2, atol=1e-2)
+    # assert torch.allclose(dV, dV_ref, rtol=1e-2, atol=1e-2)
     assert torch.allclose(dK, dK_ref, rtol=1e-2, atol=1e-2)
     assert torch.allclose(dQ, dQ_ref, rtol=1e-2, atol=1e-2)
 
@@ -352,3 +360,16 @@ if __name__ == "__main__":
     latency = do_bench(run1, warmup=500)
     print("tilelang: {:.2f} ms".format(latency))
     print("tilelang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=8, help='Batch size')
+    parser.add_argument('--h', type=int, default=32, help='Number of heads')
+    parser.add_argument('--n_ctx', type=int, default=1024, help='Context size')
+    parser.add_argument('--d_head_qk', type=int, default=192, help='Head dimension for Q/K')
+    parser.add_argument('--d_head_v', type=int, default=128, help='Head dimension for V')
+    parser.add_argument('--causal', type=bool, default=False, help='Causal flag')
+    parser.add_argument('--groups', type=int, default=16, help='groups')
+    args = parser.parse_args()
+    main(args.batch, args.h, args.n_ctx, args.d_head_qk, args.d_head_v, args.groups, args.causal)

examples/flash_attention/example_gqa_fwd_bshd.py

@@ -227,42 +227,52 @@ def ref_program(Q, K, V, is_causal, groups=1):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=1, help='batch size')
-    parser.add_argument('--heads', type=int, default=64, help='heads')
-    parser.add_argument('--seq_len', type=int, default=8192, help='sequence length')
-    parser.add_argument('--dim', type=int, default=128, help='dim')
-    parser.add_argument('--is_causal', action='store_true', help='causal')
-    parser.add_argument('--tune', action='store_true', help='tune configs')
-    parser.add_argument('--groups', type=int, default=16, help='groups')
-    args = parser.parse_args()
-    batch, heads, seq_len, dim, is_causal, groups = args.batch, args.heads, args.seq_len, args.dim, args.is_causal, args.groups
+def main(batch: int = 1,
+         heads: int = 64,
+         seq_len: int = 4096,
+         dim: int = 128,
+         is_causal: bool = False,
+         groups: int = 16,
+         tune: bool = False):
     flops_per_matmul = 2.0 * batch * heads * seq_len * seq_len * dim
     total_flops = 2 * flops_per_matmul
     if is_causal:
         total_flops *= 0.5
 
-    if (not args.tune):
+    if (not tune):
         program = flashattn(
-            batch, heads, seq_len, dim, is_causal, tune=args.tune, groups=groups)(
-                block_M=128, block_N=128, num_stages=2, threads=128)
-        ref_program = partial(ref_program, is_causal=is_causal, groups=groups)
+            batch, heads, seq_len, dim, is_causal, tune=tune, groups=groups)(
+                block_M=64, block_N=64, num_stages=2, threads=128)
+        ref_program_processed = partial(ref_program, is_causal=is_causal, groups=groups)
         kernel = tilelang.compile(program, out_idx=[3])
         profiler = kernel.get_profiler(tensor_supply_type=tilelang.TensorSupplyType.Normal)
-        profiler.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+        profiler.assert_allclose(ref_program_processed, rtol=0.01, atol=0.01)
         print("All checks pass.")
-        latency = profiler.do_bench(ref_program, warmup=500)
+        latency = profiler.do_bench(ref_program_processed, warmup=500)
         print("Ref: {:.2f} ms".format(latency))
         print("Ref: {:.2f} TFlops".format(total_flops / latency * 1e-9))
         latency = profiler.do_bench(warmup=500)
         print("Tile-lang: {:.2f} ms".format(latency))
         print("Tile-lang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
     else:
-        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=args.tune)
+        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=tune)
         best_latency = best_result.latency
         best_config = best_result.config
         ref_latency = best_result.ref_latency
         print(f"Best latency: {best_latency}")
         print(f"Best TFlops: {total_flops / best_latency * 1e-9}")
         print(f"Best config: {best_config}")
+        print(f"Ref latency: {ref_latency}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=1, help='batch size')
+    parser.add_argument('--heads', type=int, default=64, help='heads')
+    parser.add_argument('--seq_len', type=int, default=4096, help='sequence length')
+    parser.add_argument('--dim', type=int, default=128, help='dim')
+    parser.add_argument('--is_causal', action='store_true', help='causal')
+    parser.add_argument('--tune', action='store_true', help='tune configs')
+    parser.add_argument('--groups', type=int, default=16, help='groups')
+    args = parser.parse_args()
+    main(args.batch, args.heads, args.seq_len, args.dim, args.is_causal, args.groups, args.tune)

examples/flash_attention/example_gqa_fwd_bshd_wgmma_pipelined.py

@@ -199,42 +199,54 @@ def ref_program(Q, K, V, is_causal, groups=1):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=1, help='batch size')
-    parser.add_argument('--heads', type=int, default=64, help='heads')
-    parser.add_argument('--seq_len', type=int, default=8192, help='sequence length')
-    parser.add_argument('--dim', type=int, default=128, help='dim')
-    parser.add_argument('--is_causal', action='store_true', help='causal')
-    parser.add_argument('--tune', action='store_true', help='tune configs')
-    parser.add_argument('--groups', type=int, default=16, help='groups')
-    args = parser.parse_args()
-    batch, heads, seq_len, dim, is_causal, groups = args.batch, args.heads, args.seq_len, args.dim, args.is_causal, args.groups
+def main(
+    batch: int = 1,
+    heads: int = 64,
+    seq_len: int = 4096,
+    dim: int = 128,
+    is_causal: bool = False,
+    groups: int = 16,
+    tune: bool = False,
+):
     flops_per_matmul = 2.0 * batch * heads * seq_len * seq_len * dim
     total_flops = 2 * flops_per_matmul
     if is_causal:
         total_flops *= 0.5
 
-    if (not args.tune):
+    if (not tune):
         program = flashattn(
-            batch, heads, seq_len, dim, is_causal, tune=args.tune, groups=groups)(
+            batch, heads, seq_len, dim, is_causal, tune=tune, groups=groups)(
                 block_M=128, block_N=128, num_stages=2, threads=256)
-        ref_program = partial(ref_program, is_causal=is_causal, groups=groups)
+        ref_program_processed = partial(ref_program, is_causal=is_causal, groups=groups)
         kernel = tilelang.compile(program, out_idx=[3])
         profiler = kernel.get_profiler(tensor_supply_type=tilelang.TensorSupplyType.Normal)
-        profiler.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+        profiler.assert_allclose(ref_program_processed, rtol=0.01, atol=0.01)
         print("All checks pass.")
-        latency = profiler.do_bench(ref_program, warmup=500)
+        latency = profiler.do_bench(ref_program_processed, warmup=500)
         print("Ref: {:.2f} ms".format(latency))
         print("Ref: {:.2f} TFlops".format(total_flops / latency * 1e-9))
         latency = profiler.do_bench(warmup=500)
         print("Tile-lang: {:.2f} ms".format(latency))
         print("Tile-lang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
     else:
-        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=args.tune)
+        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=tune)
         best_latency = best_result.latency
         best_config = best_result.config
         ref_latency = best_result.ref_latency
         print(f"Best latency: {best_latency}")
         print(f"Best TFlops: {total_flops / best_latency * 1e-9}")
         print(f"Best config: {best_config}")
+        print(f"Ref latency: {ref_latency}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=1, help='batch size')
+    parser.add_argument('--heads', type=int, default=64, help='heads')
+    parser.add_argument('--seq_len', type=int, default=4096, help='sequence length')
+    parser.add_argument('--dim', type=int, default=128, help='dim')
+    parser.add_argument('--is_causal', action='store_true', help='causal')
+    parser.add_argument('--tune', action='store_true', help='tune configs')
+    parser.add_argument('--groups', type=int, default=16, help='groups')
+    args = parser.parse_args()
+    main(args.batch, args.heads, args.seq_len, args.dim, args.is_causal, args.groups, args.tune)

examples/flash_attention/example_mha_bwd.py

@@ -6,7 +6,8 @@ import tilelang.language as T
 import argparse
 
 
-def flashattn_fwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
+@tilelang.jit(out_idx=[3, 4])
+def flashattn_fwd(batch, heads, seq_len, dim, is_causal, block_M, block_N):
     scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
     shape = [batch, seq_len, heads, dim]
     dtype = "float16"
@@ -44,10 +45,10 @@ def flashattn_fwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
             #     Q_local[i, j] *= scale
             loop_range = (
                 T.ceildiv(
-                    (bx + 1) * block_M, block_N) if is_casual else T.ceildiv(seq_len, block_N))
+                    (bx + 1) * block_M, block_N) if is_causal else T.ceildiv(seq_len, block_N))
             for k in T.Pipelined(loop_range, num_stages=1):
                 T.copy(K[bz, k * block_N:(k + 1) * block_N, by, :], K_shared)
-                if is_casual:
+                if is_causal:
                     for i, j in T.Parallel(block_M, block_N):
                         acc_s[i, j] = T.if_then_else(bx * block_M + i >= k * block_N + j, 0,
                                                      -T.infinity(acc_s.dtype))
@@ -78,6 +79,7 @@ def flashattn_fwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
     return flash_fwd
 
 
+@tilelang.jit(out_idx=[2])
 def flashattn_bwd_preprocess(batch, heads, seq_len, dim):
     dtype = "float16"
     accum_dtype = "float"
@@ -113,6 +115,7 @@ def make_dq_layout(dQ):
                     lambda b, l, h, d: [b, l // 8, h, d // 8, (d % 2), 4 * (l % 8) + (d % 8) // 2])
 
 
+@tilelang.jit(out_idx=[1])
 def flashattn_bwd_postprocess(batch, heads, seq_len, dim):
     dtype = "float16"
     accum_dtype = "float"
@@ -134,7 +137,8 @@ def flashattn_bwd_postprocess(batch, heads, seq_len, dim):
     return flash_bwd_post
 
 
-def flashattn_bwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
+@tilelang.jit
+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]
@@ -153,7 +157,7 @@ def flashattn_bwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
             dK: T.Tensor(shape, dtype),  # type: ignore
             dV: T.Tensor(shape, dtype),  # type: ignore
     ):
-        with T.Kernel(heads, T.ceildiv(seq_len, block_M), batch, threads=256) as (bx, by, bz):
+        with T.Kernel(heads, T.ceildiv(seq_len, block_M), batch, threads=128) as (bx, by, bz):
             K_shared = T.alloc_shared([block_M, dim], dtype)
             dsT_shared = T.alloc_shared([block_M, block_N], dtype)
             # should not store K to local if dim is large
@@ -181,12 +185,11 @@ def flashattn_bwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
                 dv_shared: tilelang.layout.make_swizzled_layout(dv_shared),
                 dk_shared: tilelang.layout.make_swizzled_layout(dk_shared),
             })
-
             T.copy(K[bz, by * block_M:(by + 1) * block_M, bx, :], K_shared)
             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)
@@ -195,7 +198,7 @@ def flashattn_bwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
                 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)
@@ -232,12 +235,7 @@ class _attention(torch.autograd.Function):
         BATCH, N_CTX, H, D_HEAD = q.shape
         block_M = 64
         block_N = 64 if D_HEAD <= 128 else 32
-        kernel = tilelang.compile(
-            flashattn_fwd(BATCH, H, N_CTX, D_HEAD, causal, block_M, block_N),
-            out_idx=[3, 4],
-            target="cuda",
-            execution_backend="cython")
-        o, lse = kernel(q, k, v)
+        o, lse = flashattn_fwd(BATCH, H, N_CTX, D_HEAD, causal, block_M, block_N)(q, k, v)
         ctx.save_for_backward(q, k, v, o, lse)
         ctx.causal = causal
         return o
@@ -245,6 +243,7 @@ class _attention(torch.autograd.Function):
     @staticmethod
     def backward(ctx, do):
         q, k, v, o, lse = ctx.saved_tensors
+        BATCH, N_CTX, H, D_HEAD = q.shape
 
         def maybe_contiguous(x):
             if x.stride(-1) != 1:
@@ -252,25 +251,17 @@ class _attention(torch.autograd.Function):
             return x
 
         do, q, k, v, o = [maybe_contiguous(x) for x in (do, q, k, v, o)]
-        block_M = 128
-        block_N = 128 if D_HEAD <= 64 else 32
-        kernel_prep = tilelang.compile(
-            flashattn_bwd_preprocess(BATCH, H, N_CTX, D_HEAD),
-            out_idx=[2],
-            target="cuda",
-            execution_backend="cython")
-        kernel_post = tilelang.compile(
-            flashattn_bwd_postprocess(BATCH, H, N_CTX, D_HEAD),
-            out_idx=[1],
-            target="cuda",
-            execution_backend="cython")
+        block_M = 64
+        block_N = 64 if D_HEAD <= 64 else 32
+        kernel_prep = flashattn_bwd_preprocess(BATCH, H, N_CTX, D_HEAD)
+        kernel_post = flashattn_bwd_postprocess(BATCH, H, N_CTX, D_HEAD)
         delta = kernel_prep(o, do)
-        kernel = tilelang.compile(
-            flashattn_bwd(BATCH, H, N_CTX, D_HEAD, ctx.causal, block_M, block_N),
-            out_idx=[6, 7, 8],
-            target="cuda",
-            execution_backend="cython")
-        dq, dk, dv = kernel(q, k, v, do, lse, delta)
+        kernel = flashattn_bwd(BATCH, H, N_CTX, D_HEAD, ctx.causal, block_M, block_N)
+        shape = [BATCH, N_CTX, H, D_HEAD]
+        dq = torch.zeros(shape, dtype=torch.float32, device=q.device)
+        dk = torch.empty(shape, dtype=torch.float16, device=q.device)
+        dv = torch.empty(shape, dtype=torch.float16, device=q.device)
+        kernel(q, k, v, do, lse, delta, dq, dk, dv)
         dq = kernel_post(dq)
         return dq, dk, dv, None
 
@@ -292,19 +283,16 @@ def ref_program(Q, K, V, is_causal):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=8, help='Batch size')
-    parser.add_argument('--h', type=int, default=32, help='Number of heads')
-    parser.add_argument('--n_ctx', type=int, default=1024, help='Context size')
-    parser.add_argument('--d_head', type=int, default=64, help='Head dimension')
-    parser.add_argument('--casual', type=bool, default=False, help='Casual flag')
-    args = parser.parse_args()
-    BATCH, H, N_CTX, D_HEAD = args.batch, args.h, args.n_ctx, args.d_head
-    casual = args.casual
+def main(
+    BATCH: int = 8,
+    H: int = 32,
+    N_CTX: int = 1024,
+    D_HEAD: int = 64,
+    causal: bool = False,
+):
     flops_per_matmul = 2.0 * BATCH * H * N_CTX * N_CTX * D_HEAD
     total_flops = 5 * flops_per_matmul
-    if casual:
+    if causal:
         total_flops *= 0.5
     Q = (
         torch.empty(BATCH, N_CTX, H, D_HEAD, dtype=torch.half,
@@ -312,13 +300,13 @@ if __name__ == "__main__":
     K = torch.empty_like(Q).normal_().requires_grad_()
     V = torch.empty_like(Q).normal_().requires_grad_()
     dO = torch.randn_like(Q)
-    O = attention(Q, K, V, casual)
+    O = attention(Q, K, V, causal)
     O.backward(dO, retain_graph=True)
     dQ, Q.grad = Q.grad.clone(), None
     dK, K.grad = K.grad.clone(), None
     dV, V.grad = V.grad.clone(), None
 
-    O_ref = ref_program(Q, K, V, casual)
+    O_ref = ref_program(Q, K, V, causal)
     O_ref.backward(dO, retain_graph=True)
     dQ_ref, Q.grad = Q.grad.clone(), None
     dK_ref, K.grad = K.grad.clone(), None
@@ -343,3 +331,14 @@ if __name__ == "__main__":
     latency = do_bench(run1, warmup=500)
     print("tilelang: {:.2f} ms".format(latency))
     print("tilelang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=8, help='Batch size')
+    parser.add_argument('--h', type=int, default=32, help='Number of heads')
+    parser.add_argument('--n_ctx', type=int, default=1024, help='Context size')
+    parser.add_argument('--d_head', type=int, default=64, help='Head dimension')
+    parser.add_argument('--causal', type=bool, default=False, help='Causal flag')
+    args = parser.parse_args()
+    main(args.batch, args.h, args.n_ctx, args.d_head, args.causal)

examples/flash_attention/example_mha_bwd_wgmma_pipelined.py

 import torch
 import torch.nn.functional as F
 import tilelang
-from tilelang import cached
 from tilelang.autotuner import *
 import tilelang.language as T
 import argparse
 
 
-def flashattn_fwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
+@tilelang.jit(out_idx=[3, 4])
+def flashattn_fwd(batch, heads, seq_len, dim, is_causal, block_M, block_N):
     scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
     shape = [batch, seq_len, heads, dim]
     dtype = "float16"
@@ -45,10 +45,10 @@ def flashattn_fwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
             #     Q_local[i, j] *= scale
             loop_range = (
                 T.ceildiv(
-                    (bx + 1) * block_M, block_N) if is_casual else T.ceildiv(seq_len, block_N))
+                    (bx + 1) * block_M, block_N) if is_causal else T.ceildiv(seq_len, block_N))
             for k in T.Pipelined(loop_range, num_stages=1):
                 T.copy(K[bz, k * block_N:(k + 1) * block_N, by, :], K_shared)
-                if is_casual:
+                if is_causal:
                     for i, j in T.Parallel(block_M, block_N):
                         acc_s[i, j] = T.if_then_else(bx * block_M + i >= k * block_N + j, 0,
                                                      -T.infinity(acc_s.dtype))
@@ -79,6 +79,7 @@ def flashattn_fwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
     return flash_fwd
 
 
+@tilelang.jit(out_idx=[2])
 def flashattn_bwd_preprocess(batch, heads, seq_len, dim):
     dtype = "float16"
     accum_dtype = "float"
@@ -114,6 +115,7 @@ def make_dq_layout(dQ):
                     lambda b, l, h, d: [b, l // 8, h, d // 8, (d % 2), 4 * (l % 8) + (d % 8) // 2])
 
 
+@tilelang.jit(out_idx=[1])
 def flashattn_bwd_postprocess(batch, heads, seq_len, dim):
     dtype = "float16"
     accum_dtype = "float"
@@ -135,6 +137,7 @@ def flashattn_bwd_postprocess(batch, heads, seq_len, dim):
     return flash_bwd_post
 
 
+@tilelang.jit
 def flashattn_bwd(batch, heads, seq_len, dim, is_casual, block_M, block_N):
     sm_scale = (1.0 / dim)**0.5
     scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
@@ -244,7 +247,7 @@ class _attention(torch.autograd.Function):
         BATCH, N_CTX, H, D_HEAD = q.shape
         block_M = 64
         block_N = 64 if D_HEAD <= 128 else 32
-        mod = cached(flashattn_fwd, [3, 4], BATCH, H, N_CTX, D_HEAD, causal, block_M, block_N)
+        mod = flashattn_fwd(BATCH, H, N_CTX, D_HEAD, causal, block_M, block_N)
         o, lse = mod(q, k, v)
         ctx.save_for_backward(q, k, v, o, lse)
         ctx.causal = causal
@@ -253,6 +256,7 @@ class _attention(torch.autograd.Function):
     @staticmethod
     def backward(ctx, do):
         q, k, v, o, lse = ctx.saved_tensors
+        BATCH, N_CTX, H, D_HEAD = q.shape
 
         def maybe_contiguous(x):
             if x.stride(-1) != 1:
@@ -260,14 +264,17 @@ class _attention(torch.autograd.Function):
             return x
 
         do, q, k, v, o = [maybe_contiguous(x) for x in (do, q, k, v, o)]
-        block_M = 128
-        block_N = 128 if D_HEAD <= 64 else 32
-        mod_prep = cached(flashattn_bwd_preprocess, [2], BATCH, H, N_CTX, D_HEAD)
-        mod_post = cached(flashattn_bwd_postprocess, [1], BATCH, H, N_CTX, D_HEAD)
+        block_M = 64
+        block_N = 64 if D_HEAD <= 64 else 32
+        mod_prep = flashattn_bwd_preprocess(BATCH, H, N_CTX, D_HEAD)
+        mod_post = flashattn_bwd_postprocess(BATCH, H, N_CTX, D_HEAD)
         delta = mod_prep(o, do)
-        mod = cached(flashattn_bwd, [6, 7, 8], BATCH, H, N_CTX, D_HEAD, ctx.causal, block_M,
-                     block_N)
-        dq, dk, dv = mod(q, k, v, do, lse, delta)
+        mod = flashattn_bwd(BATCH, H, N_CTX, D_HEAD, ctx.causal, block_M, block_N)
+        shape = [BATCH, N_CTX, H, D_HEAD]
+        dq = torch.zeros(shape, dtype=torch.float32, device=q.device)
+        dk = torch.empty(shape, dtype=torch.float16, device=q.device)
+        dv = torch.empty(shape, dtype=torch.float16, device=q.device)
+        mod(q, k, v, do, lse, delta, dq, dk, dv)
         dq = mod_post(dq)
         return dq, dk, dv, None
 
@@ -289,19 +296,16 @@ def ref_program(Q, K, V, is_causal):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=8, help='Batch size')
-    parser.add_argument('--h', type=int, default=32, help='Number of heads')
-    parser.add_argument('--n_ctx', type=int, default=1024, help='Context size')
-    parser.add_argument('--d_head', type=int, default=64, help='Head dimension')
-    parser.add_argument('--casual', type=bool, default=False, help='Casual flag')
-    args = parser.parse_args()
-    BATCH, H, N_CTX, D_HEAD = args.batch, args.h, args.n_ctx, args.d_head
-    casual = args.casual
+def main(
+    BATCH: int = 8,
+    H: int = 32,
+    N_CTX: int = 1024,
+    D_HEAD: int = 64,
+    causal: bool = False,
+):
     flops_per_matmul = 2.0 * BATCH * H * N_CTX * N_CTX * D_HEAD
     total_flops = 5 * flops_per_matmul
-    if casual:
+    if causal:
         total_flops *= 0.5
     Q = (
         torch.empty(BATCH, N_CTX, H, D_HEAD, dtype=torch.half,
@@ -309,13 +313,13 @@ if __name__ == "__main__":
     K = torch.empty_like(Q).normal_().requires_grad_()
     V = torch.empty_like(Q).normal_().requires_grad_()
     dO = torch.randn_like(Q)
-    O = attention(Q, K, V, casual)
+    O = attention(Q, K, V, causal)
     O.backward(dO, retain_graph=True)
     dQ, Q.grad = Q.grad.clone(), None
     dK, K.grad = K.grad.clone(), None
     dV, V.grad = V.grad.clone(), None
 
-    O_ref = ref_program(Q, K, V, casual)
+    O_ref = ref_program(Q, K, V, causal)
     O_ref.backward(dO, retain_graph=True)
     dQ_ref, Q.grad = Q.grad.clone(), None
     dK_ref, K.grad = K.grad.clone(), None
@@ -340,3 +344,14 @@ if __name__ == "__main__":
     latency = do_bench(run1, warmup=500)
     print("tilelang: {:.2f} ms".format(latency))
     print("tilelang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=8, help='Batch size')
+    parser.add_argument('--h', type=int, default=32, help='Number of heads')
+    parser.add_argument('--n_ctx', type=int, default=1024, help='Context size')
+    parser.add_argument('--d_head', type=int, default=64, help='Head dimension')
+    parser.add_argument('--causal', type=bool, default=False, help='Causal flag')
+    args = parser.parse_args()
+    main(args.batch, args.h, args.n_ctx, args.d_head, args.causal)

examples/flash_attention/example_mha_fwd_bhsd.py

@@ -182,43 +182,55 @@ def ref_program(Q, K, V, is_causal):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=1, help='batch size')
-    parser.add_argument('--heads', type=int, default=1, help='heads')
-    parser.add_argument('--seq_q', type=int, default=256, help='query sequence length')
-    parser.add_argument('--seq_kv', type=int, default=256, help='key/value sequence length')
-    parser.add_argument('--dim', type=int, default=64, help='dim')
-    parser.add_argument('--is_causal', action='store_true', help='causal')
-    parser.add_argument('--tune', action='store_true', help='tune configs')
-    args = parser.parse_args()
-    batch, heads, seq_q, seq_kv, dim, is_causal = args.batch, args.heads, args.seq_q, args.seq_kv, args.dim, args.is_causal
+def main(
+    batch: int = 1,
+    heads: int = 1,
+    seq_q: int = 256,
+    seq_kv: int = 256,
+    dim: int = 64,
+    is_causal: bool = False,
+    tune: bool = False,
+):
     flops_per_matmul = 2.0 * batch * heads * seq_q * seq_kv * dim
     total_flops = 2 * flops_per_matmul
     if is_causal:
         total_flops *= 0.5
 
-    if (not args.tune):
+    if (not tune):
         program = flashattn(
-            batch, heads, seq_q, seq_kv, dim, is_causal, tune=args.tune)(
+            batch, heads, seq_q, seq_kv, dim, is_causal, tune=tune)(
                 block_M=64, block_N=64, num_stages=1, threads=128)
-        ref_program = partial(ref_program, is_causal=is_causal)
+        ref_program_processed = partial(ref_program, is_causal=is_causal)
         kernel = tilelang.compile(program, out_idx=[3])
 
         profiler = kernel.get_profiler()
-        profiler.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+        profiler.assert_allclose(ref_program_processed, rtol=0.01, atol=0.01)
         print("All checks pass.")
-        latency = profiler.do_bench(ref_program, warmup=500)
+        latency = profiler.do_bench(ref_program_processed, warmup=500)
         print("Ref: {:.2f} ms".format(latency))
         print("Ref: {:.2f} TFlops".format(total_flops / latency * 1e-9))
         latency = profiler.do_bench(warmup=500)
         print("Tile-lang: {:.2f} ms".format(latency))
         print("Tile-lang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
     else:
-        best_result = flashattn(batch, heads, seq_q, seq_kv, dim, is_causal, tune=args.tune)
+        best_result = flashattn(batch, heads, seq_q, seq_kv, dim, is_causal, tune=tune)
         best_latency = best_result.latency
         best_config = best_result.config
         ref_latency = best_result.ref_latency
         print(f"Best latency: {best_latency}")
         print(f"Best TFlops: {total_flops / best_latency * 1e-9}")
         print(f"Best config: {best_config}")
+        print(f"Ref latency: {ref_latency}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=1, help='batch size')
+    parser.add_argument('--heads', type=int, default=1, help='heads')
+    parser.add_argument('--seq_q', type=int, default=256, help='query sequence length')
+    parser.add_argument('--seq_kv', type=int, default=256, help='key/value sequence length')
+    parser.add_argument('--dim', type=int, default=64, help='dim')
+    parser.add_argument('--is_causal', action='store_true', help='causal')
+    parser.add_argument('--tune', action='store_true', help='tune configs')
+    args = parser.parse_args()
+    main(args.batch, args.heads, args.seq_q, args.seq_kv, args.dim, args.is_causal, args.tune)

examples/flash_attention/example_mha_fwd_bhsd_wgmma_pipelined.py

@@ -187,43 +187,55 @@ def ref_program(Q, K, V, is_causal):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=8, help='batch size')
-    parser.add_argument('--heads', type=int, default=32, help='heads')
-    parser.add_argument('--seq_q', type=int, default=4096, help='query sequence length')
-    parser.add_argument('--seq_kv', type=int, default=4096, help='key/value sequence length')
-    parser.add_argument('--dim', type=int, default=128, help='dim')
-    parser.add_argument('--is_causal', action='store_true', help='causal')
-    parser.add_argument('--tune', action='store_true', help='tune configs')
-    args = parser.parse_args()
-    batch, heads, seq_q, seq_kv, dim, is_causal = args.batch, args.heads, args.seq_q, args.seq_kv, args.dim, args.is_causal
+def main(
+    batch: int = 8,
+    heads: int = 32,
+    seq_q: int = 4096,
+    seq_kv: int = 4096,
+    dim: int = 128,
+    is_causal: bool = False,
+    tune: bool = False,
+):
     flops_per_matmul = 2.0 * batch * heads * seq_q * seq_kv * dim
     total_flops = 2 * flops_per_matmul
     if is_causal:
         total_flops *= 0.5
 
-    if (not args.tune):
+    if (not tune):
         program = flashattn(
-            batch, heads, seq_q, seq_kv, dim, is_causal, tune=args.tune)(
+            batch, heads, seq_q, seq_kv, dim, is_causal, tune=tune)(
                 block_M=128, block_N=128, num_stages=2, threads=256)
-        ref_program = partial(ref_program, is_causal=is_causal)
+        ref_program_processed = partial(ref_program, is_causal=is_causal)
         kernel = tilelang.compile(program, out_idx=[3])
 
         profiler = kernel.get_profiler()
-        profiler.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+        profiler.assert_allclose(ref_program_processed, rtol=0.01, atol=0.01)
         print("All checks pass.")
-        latency = profiler.do_bench(ref_program, warmup=500)
+        latency = profiler.do_bench(ref_program_processed, warmup=500)
         print("Ref: {:.2f} ms".format(latency))
         print("Ref: {:.2f} TFlops".format(total_flops / latency * 1e-9))
         latency = profiler.do_bench(warmup=500)
         print("Tile-lang: {:.2f} ms".format(latency))
         print("Tile-lang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
     else:
-        best_result = flashattn(batch, heads, seq_q, seq_kv, dim, is_causal, tune=args.tune)
+        best_result = flashattn(batch, heads, seq_q, seq_kv, dim, is_causal, tune=tune)
         best_latency = best_result.latency
         best_config = best_result.config
         ref_latency = best_result.ref_latency
         print(f"Best latency: {best_latency}")
         print(f"Best TFlops: {total_flops / best_latency * 1e-9}")
         print(f"Best config: {best_config}")
+        print(f"Ref latency: {ref_latency}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=8, help='batch size')
+    parser.add_argument('--heads', type=int, default=32, help='heads')
+    parser.add_argument('--seq_q', type=int, default=4096, help='query sequence length')
+    parser.add_argument('--seq_kv', type=int, default=4096, help='key/value sequence length')
+    parser.add_argument('--dim', type=int, default=128, help='dim')
+    parser.add_argument('--is_causal', action='store_true', help='causal')
+    parser.add_argument('--tune', action='store_true', help='tune configs')
+    args = parser.parse_args()
+    main(args.batch, args.heads, args.seq_q, args.seq_kv, args.dim, args.is_causal, args.tune)

examples/flash_attention/example_mha_fwd_bshd.py

@@ -177,41 +177,52 @@ def ref_program(Q, K, V, is_causal):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=8, help='batch size')
-    parser.add_argument('--heads', type=int, default=32, help='heads')
-    parser.add_argument('--seq_len', type=int, default=4096, help='sequence length')
-    parser.add_argument('--dim', type=int, default=128, help='dim')
-    parser.add_argument('--is_causal', action='store_true', help='causal')
-    parser.add_argument('--tune', action='store_true', help='tune configs')
-    args = parser.parse_args()
-    batch, heads, seq_len, dim, is_causal = args.batch, args.heads, args.seq_len, args.dim, args.is_causal
+def main(
+    batch: int = 8,
+    heads: int = 32,
+    seq_len: int = 4096,
+    dim: int = 128,
+    is_causal: bool = False,
+    tune: bool = False,
+):
     flops_per_matmul = 2.0 * batch * heads * seq_len * seq_len * dim
     total_flops = 2 * flops_per_matmul
     if is_causal:
         total_flops *= 0.5
 
-    if (not args.tune):
+    if (not tune):
         program = flashattn(
-            batch, heads, seq_len, dim, is_causal, tune=args.tune)(
+            batch, heads, seq_len, dim, is_causal, tune=tune)(
                 block_M=128, block_N=128, num_stages=1, threads=128)
-        ref_program = partial(ref_program, is_causal=is_causal)
+        ref_program_processed = partial(ref_program, is_causal=is_causal)
         kernel = tilelang.compile(program, out_idx=[3])
         profiler = kernel.get_profiler()
-        profiler.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+        profiler.assert_allclose(ref_program_processed, rtol=0.01, atol=0.01)
         print("All checks pass.")
-        latency = profiler.do_bench(ref_program, warmup=500)
+        latency = profiler.do_bench(ref_program_processed, warmup=500)
         print("Ref: {:.2f} ms".format(latency))
         print("Ref: {:.2f} TFlops".format(total_flops / latency * 1e-9))
         latency = profiler.do_bench(warmup=500)
         print("Tile-lang: {:.2f} ms".format(latency))
         print("Tile-lang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
     else:
-        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=args.tune)
+        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=tune)
         best_latency = best_result.latency
         best_config = best_result.config
         ref_latency = best_result.ref_latency
         print(f"Best latency: {best_latency}")
         print(f"Best TFlops: {total_flops / best_latency * 1e-9}")
         print(f"Best config: {best_config}")
+        print(f"Ref latency: {ref_latency}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=8, help='batch size')
+    parser.add_argument('--heads', type=int, default=32, help='heads')
+    parser.add_argument('--seq_len', type=int, default=4096, help='sequence length')
+    parser.add_argument('--dim', type=int, default=128, help='dim')
+    parser.add_argument('--is_causal', action='store_true', help='causal')
+    parser.add_argument('--tune', action='store_true', help='tune configs')
+    args = parser.parse_args()
+    main(args.batch, args.heads, args.seq_len, args.dim, args.is_causal, args.tune)

examples/flash_attention/example_mha_fwd_bshd_wgmma_pipelined.py

@@ -182,41 +182,52 @@ def ref_program(Q, K, V, is_causal):
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=8, help='batch size')
-    parser.add_argument('--heads', type=int, default=32, help='heads')
-    parser.add_argument('--seq_len', type=int, default=4096, help='sequence length')
-    parser.add_argument('--dim', type=int, default=128, help='dim')
-    parser.add_argument('--is_causal', action='store_true', help='causal')
-    parser.add_argument('--tune', action='store_true', help='tune configs')
-    args = parser.parse_args()
-    batch, heads, seq_len, dim, is_causal = args.batch, args.heads, args.seq_len, args.dim, args.is_causal
+def main(
+    batch: int = 8,
+    heads: int = 32,
+    seq_len: int = 4096,
+    dim: int = 128,
+    is_causal: bool = False,
+    tune: bool = False,
+):
     flops_per_matmul = 2.0 * batch * heads * seq_len * seq_len * dim
     total_flops = 2 * flops_per_matmul
     if is_causal:
         total_flops *= 0.5
 
-    if (not args.tune):
+    if (not tune):
         program = flashattn(
-            batch, heads, seq_len, dim, is_causal, tune=args.tune)(
+            batch, heads, seq_len, dim, is_causal, tune=tune)(
                 block_M=128, block_N=128, num_stages=2, threads=256)
-        ref_program = partial(ref_program, is_causal=is_causal)
+        ref_program_processed = partial(ref_program, is_causal=is_causal)
         kernel = tilelang.compile(program, out_idx=[3])
         profiler = kernel.get_profiler(tensor_supply_type=tilelang.TensorSupplyType.Normal)
-        profiler.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+        profiler.assert_allclose(ref_program_processed, rtol=0.01, atol=0.01)
         print("All checks pass.")
-        latency = profiler.do_bench(ref_program, warmup=500)
+        latency = profiler.do_bench(ref_program_processed, warmup=500)
         print("Ref: {:.2f} ms".format(latency))
         print("Ref: {:.2f} TFlops".format(total_flops / latency * 1e-9))
         latency = profiler.do_bench(warmup=500)
         print("Tile-lang: {:.2f} ms".format(latency))
         print("Tile-lang: {:.2f} TFlops".format(total_flops / latency * 1e-9))
     else:
-        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=args.tune)
+        best_result = flashattn(batch, heads, seq_len, dim, is_causal, tune=tune)
         best_latency = best_result.latency
         best_config = best_result.config
         ref_latency = best_result.ref_latency
         print(f"Best latency: {best_latency}")
         print(f"Best TFlops: {total_flops / best_latency * 1e-9}")
         print(f"Best config: {best_config}")
+        print(f"Ref latency: {ref_latency}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=8, help='batch size')
+    parser.add_argument('--heads', type=int, default=32, help='heads')
+    parser.add_argument('--seq_len', type=int, default=4096, help='sequence length')
+    parser.add_argument('--dim', type=int, default=128, help='dim')
+    parser.add_argument('--is_causal', action='store_true', help='causal')
+    parser.add_argument('--tune', action='store_true', help='tune configs')
+    args = parser.parse_args()
+    main(args.batch, args.heads, args.seq_len, args.dim, args.is_causal, args.tune)

examples/flash_attention/example_mha_fwd_varlen.py

@@ -46,7 +46,7 @@ def generate_qkv(q,
     assert v.shape == (batch_size, seqlen_k, nheads_k, d)
 
     if query_padding_mask is not None:
-        q_unpad, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, query_padding_mask)
+        q_unpad, indices_q, cu_seqlens_q, max_seqlen_q, _ = unpad_input(q, query_padding_mask)
         output_pad_fn = lambda output_unpad: pad_input(output_unpad, indices_q, batch_size, seqlen_q
                                                       )
     else:
@@ -58,8 +58,8 @@ def generate_qkv(q,
             output_unpad, "(b s) h d -> b s h d", b=batch_size)
 
     if key_padding_mask is not None:
-        k_unpad, indices_k, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)
-        v_unpad, _, _, _ = unpad_input(v, key_padding_mask)
+        k_unpad, indices_k, cu_seqlens_k, max_seqlen_k, _ = unpad_input(k, key_padding_mask)
+        v_unpad, _, _, _, _ = unpad_input(v, key_padding_mask)
     else:
         k_unpad = rearrange(k, "b s h d -> (b s) h d")
         v_unpad = rearrange(v, "b s h d -> (b s) h d")
@@ -197,6 +197,7 @@ def attention_ref(
     dtype_og = q.dtype
     if upcast:
         q, k, v = q.float(), k.float(), v.float()
+    dim = q.shape[-1]
     scale = (1.0 / dim)**0.5  # log2(e)
     k = repeat(k, "b s h d -> b s (h g) d", g=q.shape[2] // k.shape[2])
     v = repeat(v, "b s h d -> b s (h g) d", g=q.shape[2] // v.shape[2])
@@ -358,15 +359,7 @@ def flashattn(batch_size, UQ, UKV, heads, dim, is_causal):
     return kernel_func(block_M, block_N, num_stages, threads)
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--batch', type=int, default=2, help='batch size')
-    parser.add_argument('--heads', type=int, default=16, help='heads')
-    parser.add_argument('--seq_len', type=int, default=256, help='sequence length')
-    parser.add_argument('--dim', type=int, default=32, help='dim')
-
-    args = parser.parse_args()
-    batch, heads, seq_len, dim = args.batch, args.heads, args.seq_len, args.dim
+def main(batch: int = 2, heads: int = 16, seq_len: int = 256, dim: int = 32):
     flops_per_matmul = 2.0 * batch * heads * seq_len * seq_len * dim
     total_flops = 2 * flops_per_matmul
 
@@ -408,7 +401,7 @@ if __name__ == "__main__":
     UKV = k_unpad.shape[0]  # unpadded query key length
 
     program = flashattn(batch, UQ, UKV, heads, dim, causal)
-    kernel = tilelang.compile(program, out_idx=-1, execution_backend="cython")
+    kernel = tilelang.compile(program, [6])
     print(kernel.get_kernel_source())
 
     out_unpad = kernel(q_unpad, k_unpad, v_unpad, cu_seqlens_q, cu_seqlens_k, max_seqlen_q)
@@ -437,3 +430,14 @@ if __name__ == "__main__":
     fla_out = output_pad_fn(fla_out_unpad)
     torch.testing.assert_close(out, out_ref, rtol=1e-2, atol=1e-2)
     print("Assert Equal Passed")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--batch', type=int, default=2, help='batch size')
+    parser.add_argument('--heads', type=int, default=16, help='heads')
+    parser.add_argument('--seq_len', type=int, default=256, help='sequence length')
+    parser.add_argument('--dim', type=int, default=32, help='dim')
+
+    args = parser.parse_args()
+    main(args.batch, args.heads, args.seq_len, args.dim)

examples/flash_attention/example_mha_inference.py

@@ -294,7 +294,7 @@ def flash_split_ref(Q, K, V, causal):
                                              3), gacc_o.to(torch.float16).permute(1, 2, 3, 0, 4)
 
 
-if __name__ == "__main__":
+def main():
     BATCH, H, Q_CTX, KV_CTX, D_HEAD = 1, 32, 128, 8192, 128
     causal = False
     flops_per_matmul = 2.0 * BATCH * H * Q_CTX * KV_CTX * D_HEAD
@@ -304,15 +304,19 @@ if __name__ == "__main__":
     BLOCK_M = 128
     BLOCK_N = 64  # if D_HEAD <= 128 else 32
     program = flashattn(BATCH, H, Q_CTX, KV_CTX, D_HEAD, causal, BLOCK_M, BLOCK_N)
-    ref_program = partial(ref_program, causal=causal)
+    ref_program_processed = partial(ref_program, causal=causal)
     kernel = tilelang.compile(program, out_idx=[5])
     profiler = kernel.get_profiler(tilelang.TensorSupplyType.Normal)
-    profiler.assert_allclose(ref_program, rtol=0.01, atol=0.01)
+    profiler.assert_allclose(ref_program_processed, rtol=0.01, atol=0.01)
     print("All checks passed!")
 
-    latency = profiler.do_bench(ref_program, warmup=500)
+    latency = profiler.do_bench(ref_program_processed, warmup=500)
     print("{:.2f} ms".format(latency))
     print("{:.2f} TFlops".format(total_flops / latency * 1e-9))
-    latency = profiler.do_bench(n_warmup=10, n_repeat=10, profiler="tvm")
+    latency = profiler.do_bench(n_warmup=10, n_repeat=10)
     print("{:.4f} ms".format(latency))
     print("{:.2f} TFlops".format(total_flops / latency * 1e-9))
+
+
+if __name__ == "__main__":
+    main()

examples/flash_attention/test_example_flash_attention.py

+import tilelang.testing
+
+import example_gqa_bwd
+import example_mha_bwd
+import example_mha_fwd_bhsd_wgmma_pipelined
+import example_gqa_fwd_bshd
+import example_mha_fwd_bshd
+import example_gqa_fwd_bshd_wgmma_pipelined
+import example_mha_fwd_bshd_wgmma_pipelined
+import example_mha_fwd_varlen
+import example_mha_bwd_wgmma_pipelined
+import example_mha_inference
+import example_mha_fwd_bhsd
+
+
+@tilelang.testing.requires_cuda
+def test_example_gqa_bwd():
+    example_gqa_bwd.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_mha_bwd():
+    example_mha_bwd.main()
+
+
+@tilelang.testing.requires_cuda
+@tilelang.testing.requires_cuda_compute_version_ge(9, 0)
+def test_example_mha_bwd_wgmma_pipelined():
+    example_mha_bwd_wgmma_pipelined.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_gqa_fwd_bshd_wgmma_pipelined():
+    example_gqa_fwd_bshd_wgmma_pipelined.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_gqa_fwd_bshd():
+    example_gqa_fwd_bshd.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_mha_fwd_bhsd_wgmma_pipelined():
+    example_mha_fwd_bhsd_wgmma_pipelined.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_mha_fwd_bhsd():
+    example_mha_fwd_bhsd.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_mha_fwd_bshd_wgmma_pipelined():
+    example_mha_fwd_bshd_wgmma_pipelined.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_mha_fwd_bshd():
+    example_mha_fwd_bshd.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_mha_fwd_varlen():
+    example_mha_fwd_varlen.main()
+
+
+@tilelang.testing.requires_cuda
+def test_example_mha_inference():
+    example_mha_inference.main()
+
+
+if __name__ == "__main__":
+    tilelang.testing.main()

tilelang/engine/phase.py

@@ -46,6 +46,13 @@ def allow_vectorize(pass_ctx: Optional[PassContext] = None) -> bool:
     return not disable_vectorize
 
 
+def allow_global_thread_synchronization(pass_ctx: Optional[PassContext] = None) -> bool:
+    if pass_ctx is None:
+        pass_ctx = tilelang.transform.get_pass_context()
+    enable_global_thread_sync = pass_ctx.config.get("tir.detect_global_barrier", False)
+    return enable_global_thread_sync
+
+
 def LowerAndLegalize(mod: IRModule, target: Target) -> IRModule:
     # Bind the target device information to the module
     mod = tir.transform.BindTarget(target)(mod)
@@ -133,11 +140,11 @@ def OptimizeForTarget(mod: IRModule, target: Target) -> IRModule:
     mod = tir.transform.InferFragment()(mod)
     mod = tir.transform.LowerThreadAllreduce()(mod)
     mod = tilelang.transform.LowerHopperIntrin()(mod)
-    mod = tilelang.transform.ThreadSync("global")(mod)
 
     # Global Barrier Synchronization must be applied before
     # SplitHostDevice pass, as the global barrier
-    mod = tilelang.transform.ThreadSync("global")(mod)
+    if allow_global_thread_synchronization():
+        mod = tilelang.transform.ThreadSync("global")(mod)
     mod = tilelang.transform.AnnotateDeviceRegions()(mod)
     mod = tir.transform.SplitHostDevice()(mod)