failed on h100

.gitmodules

@@ -1,3 +1,3 @@
 [submodule "csrc/sliding_tile_attention/tk"]
-	path = csrc/sliding_tile_attention/tk
+	path = sta_kernel/thunderkitten/tk
 	url = https://github.com/HazyResearch/ThunderKittens.git

sta_kernel/tilelang/sta_tilelang.py

@@ -0,0 +1,252 @@
+# Copyright (c) Tile-AI Corporation.
+# Licensed under the MIT License.
+import math
+import torch
+
+import tilelang
+import tilelang.language as T
+import torch.nn.functional as F
+
+def get_sta_mask(x, canvas_size=(32, 48, 80), tile_size=(4, 8, 8), kernel_size=(1, 1, 1), has_text=False):
+    bsz, num_head, downsample_len, _ = x.shape
+    device = x.device
+    CT, CH, CW = canvas_size
+    TT, TH, TW = tile_size
+    NT, NH, NW = CT // TT, CH // TH, CW // TW
+    KT, KH, KW = kernel_size
+    DT, DH, DW = KT // 2, KH // 2, KW // 2
+    
+    dense_mask = torch.full([bsz, num_head, downsample_len, downsample_len],
+                           False,
+                           dtype=torch.bool,
+                           device=device)
+    indices = torch.arange(downsample_len, device=device)
+    q_t = (indices // (NT * NH * NW)).unsqueeze(-1)
+    q_h = ((indices // (NW)) % NH).unsqueeze(-1)
+    q_w = ((indices // 1) % NW).unsqueeze(-1)
+    
+    q_t = torch.clamp(q_t, DT, NT-DT-1)
+    q_h = torch.clamp(q_h, DH, NH-DH-1)
+    q_w = torch.clamp(q_w, DW, NW-DW-1)
+    
+    k_indices = torch.arange(downsample_len, device=device)
+    k_t = (k_indices // (NT * NH * NW)).unsqueeze(0)
+    k_h = ((k_indices // (NW)) % NH).unsqueeze(0)
+    k_w = ((k_indices // 1) % NW).unsqueeze(0)
+    
+    t_dist = torch.abs(q_t - k_t)
+    h_dist = torch.abs(q_h - k_h)
+    w_dist = torch.abs(q_w - k_w)
+    mask = (t_dist <= DT) & (h_dist <= DH) & (w_dist <= DW)
+    
+    for b in range(bsz):
+        for h in range(num_head):
+            dense_mask[b, h] = mask
+    
+    # for text mask
+    if has_text:
+        text_start = downsample_len - 3
+        dense_mask[:, :, text_start:, :text_start] = True
+        dense_mask[:, :, text_start:, text_start:] = torch.tril(
+            torch.ones(3, 3, device=device, dtype=torch.bool)
+        )
+    return dense_mask
+
+def blocksparse_flashattn(batch, heads, seq_len, dim, downsample_len, is_causal):
+    block_M = 64
+    block_N = 64
+    num_stages = 1
+    threads = 128
+    scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
+    shape = [batch, heads, seq_len, dim]
+    block_mask_shape = [batch, heads, downsample_len, downsample_len]
+
+    dtype = "float16"
+    accum_dtype = "float"
+    block_mask_dtype = "bool"
+
+    def kernel_func(block_M, block_N, num_stages, threads):
+
+        @T.macro
+        def MMA0(
+            K: T.Buffer(shape, dtype),
+            Q_shared: T.Buffer([block_M, dim], dtype),
+            K_shared: T.Buffer([block_N, dim], dtype),
+            acc_s: T.Buffer([block_M, block_N], accum_dtype),
+            k: T.int32,
+            bx: T.int32,
+            by: T.int32,
+            bz: T.int32,
+        ):
+            T.copy(K[bz, by, k * block_N:(k + 1) * block_N, :], K_shared)
+            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))
+            else:
+                T.clear(acc_s)
+            T.gemm(Q_shared, K_shared, acc_s, transpose_B=True, policy=T.GemmWarpPolicy.FullRow)
+
+        @T.macro
+        def MMA1(
+                V: T.Buffer(shape, dtype),
+                V_shared: T.Buffer([block_M, dim], dtype),
+                acc_s_cast: T.Buffer([block_M, block_N], dtype),
+                acc_o: T.Buffer([block_M, dim], accum_dtype),
+                k: T.int32,
+                by: T.int32,
+                bz: T.int32,
+        ):
+            T.copy(V[bz, by, k * block_N:(k + 1) * block_N, :], V_shared)
+            T.gemm(acc_s_cast, V_shared, acc_o, policy=T.GemmWarpPolicy.FullRow)
+
+        @T.macro
+        def Softmax(
+                acc_s: T.Buffer([block_M, block_N], accum_dtype),
+                acc_s_cast: T.Buffer([block_M, block_N], dtype),
+                scores_max: T.Buffer([block_M], accum_dtype),
+                scores_max_prev: T.Buffer([block_M], accum_dtype),
+                scores_scale: T.Buffer([block_M], accum_dtype),
+                scores_sum: T.Buffer([block_M], accum_dtype),
+                logsum: T.Buffer([block_M], accum_dtype),
+        ):
+            T.copy(scores_max, scores_max_prev)
+            T.fill(scores_max, -T.infinity(accum_dtype))
+            T.reduce_max(acc_s, scores_max, dim=1, clear=False)
+            # To do causal softmax, we need to set the scores_max to 0 if it is -inf
+            # This process is called Check_inf in FlashAttention3 code, and it only need to be done
+            # in the first ceil_div(kBlockM, kBlockN) steps.
+            # for i in T.Parallel(block_M):
+            #     scores_max[i] = T.if_then_else(scores_max[i] == -T.infinity(accum_dtype), 0, scores_max[i])
+            for i in T.Parallel(block_M):
+                scores_scale[i] = T.exp2(scores_max_prev[i] * scale - scores_max[i] * scale)
+            for i, j in T.Parallel(block_M, block_N):
+                # Instead of computing exp(x - max), we compute exp2(x * log_2(e) -
+                # max * log_2(e)) This allows the compiler to use the ffma
+                # instruction instead of fadd and fmul separately.
+                acc_s[i, j] = T.exp2(acc_s[i, j] * scale - scores_max[i] * scale)
+            T.reduce_sum(acc_s, scores_sum, dim=1)
+            for i in T.Parallel(block_M):
+                logsum[i] = logsum[i] * scores_scale[i] + scores_sum[i]
+            T.copy(acc_s, acc_s_cast)
+
+        @T.macro
+        def Rescale(
+                acc_o: T.Buffer([block_M, dim], accum_dtype),
+                scores_scale: T.Buffer([block_M], accum_dtype),
+        ):
+            for i, j in T.Parallel(block_M, dim):
+                acc_o[i, j] *= scores_scale[i]
+
+        @T.prim_func
+        def main(
+                Q: T.Buffer(shape, dtype),
+                K: T.Buffer(shape, dtype),
+                V: T.Buffer(shape, dtype),
+                BlockSparseMask: T.Buffer(block_mask_shape, block_mask_dtype),
+                Output: T.Buffer(shape, dtype),
+        ):
+            with T.Kernel(
+                    T.ceildiv(seq_len, block_M), heads, batch, threads=threads) as (bx, by, bz):
+                Q_shared = T.alloc_shared([block_M, dim], dtype)
+                K_shared = T.alloc_shared([block_N, dim], dtype)
+                V_shared = T.alloc_shared([block_N, dim], dtype)
+                O_shared = T.alloc_shared([block_M, dim], dtype)
+                acc_s = T.alloc_fragment([block_M, block_N], accum_dtype)
+                acc_s_cast = T.alloc_fragment([block_M, block_N], dtype)
+                acc_o = T.alloc_fragment([block_M, dim], accum_dtype)
+                scores_max = T.alloc_fragment([block_M], accum_dtype)
+                scores_max_prev = T.alloc_fragment([block_M], accum_dtype)
+                scores_scale = T.alloc_fragment([block_M], accum_dtype)
+                scores_sum = T.alloc_fragment([block_M], accum_dtype)
+                logsum = T.alloc_fragment([block_M], accum_dtype)
+                block_mask = T.alloc_local([downsample_len], block_mask_dtype)
+
+                T.copy(Q[bz, by, bx * block_M:(bx + 1) * block_M, :], Q_shared)
+                T.fill(acc_o, 0)
+                T.fill(logsum, 0)
+                T.fill(scores_max, -T.infinity(accum_dtype))
+
+                for vj in T.serial(downsample_len):
+                    block_mask[vj] = BlockSparseMask[bz, by, bx, vj]
+
+                loop_range = (
+                    T.min(T.ceildiv(seq_len, block_N), T.ceildiv(
+                        (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=num_stages):
+                    if block_mask[k] != 0:
+                        MMA0(K, Q_shared, K_shared, acc_s, k, bx, by, bz)
+                        Softmax(acc_s, acc_s_cast, scores_max, scores_max_prev, scores_scale,
+                                scores_sum, logsum)
+                        Rescale(acc_o, scores_scale)
+                        MMA1(V, V_shared, acc_s_cast, acc_o, k, by, bz)
+                for i, j in T.Parallel(block_M, dim):
+                    acc_o[i, j] /= logsum[i]
+                T.copy(acc_o, O_shared)
+                T.copy(O_shared, Output[bz, by, bx * block_M:(bx + 1) * block_M, :])
+
+        return main
+
+    return kernel_func(block_M, block_N, num_stages, threads)
+
+
+def test_sta_attention():
+    # Config
+    BATCH, N_HEADS, SEQ_LEN, D_HEAD = 1, 24, 2048, 128
+    torch.manual_seed(0)
+
+    # Create inputs
+    q = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.float16)
+    k = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.float16)
+    v = torch.randn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, device='cuda', dtype=torch.float16)
+
+    sm_scale = 1.0 / (D_HEAD**0.5)
+
+    # Create sparse mask (downsampled to block level)
+    canvas_size = (32, 48, 80)
+    tile_size = (4, 8, 8)
+    delta_size = (2, 2, 2)
+    BLOCK = tile_size[0] * tile_size[1] * tile_size[2]
+    downsample_factor = BLOCK
+    downsample_len = math.ceil(SEQ_LEN / downsample_factor)
+    x_ds = torch.randn([BATCH, N_HEADS, downsample_len, downsample_len],
+                       device='cuda',
+                       dtype=torch.bfloat16)
+    block_mask = get_sta_mask(x_ds, canvas_size, tile_size, delta_size, has_text=True)
+    # print mask density
+    print("mask density", block_mask.sum() / block_mask.numel())
+    print("block_mask", block_mask)
+
+    # Run Triton kernel
+    program = blocksparse_flashattn(BATCH, N_HEADS, SEQ_LEN, D_HEAD, downsample_len, is_causal=True)
+    kernel = tilelang.compile(program, out_idx=[4])
+    
+    cuda_source = kernel.get_kernel_source()
+    print("Generated CUDA kernel:\n", cuda_source)
+
+    tilelang_output = kernel(q, k, v, block_mask)
+
+    if True:
+        # Compute reference
+        # Expand block mask to full attention matrix
+        full_mask = torch.kron(block_mask.float(), torch.ones(BLOCK, BLOCK, device='cuda'))
+        full_mask = full_mask[..., :SEQ_LEN, :SEQ_LEN].bool()
+        full_mask = full_mask & torch.tril(torch.ones_like(full_mask))  # Apply causal
+
+        # PyTorch reference implementation
+        attn = torch.einsum('bhsd,bhtd->bhst', q, k) * sm_scale
+        attn = attn.masked_fill(~full_mask, float('-inf'))
+        attn = F.softmax(attn, dim=-1)
+        ref_output = torch.einsum('bhst,bhtd->bhsd', attn, v)
+
+        print("ref_output", ref_output)
+        print("tilelang_output", tilelang_output)
+
+        # Verify accuracy
+        torch.testing.assert_close(tilelang_output, ref_output, atol=1e-2, rtol=1e-2)
+        print("Pass topk sparse attention test with qlen == klen")
+
+
+if __name__ == "__main__":
+    test_sta_attention()