blksprs 1.11__py3-none-any.whl → 2.0rc1__py3-none-any.whl

blksprs/ops/softmax.py

@@ -1,17 +1,18 @@
 import torch
 import triton
 from torch import Tensor
+from torch._library import triton_op
+from torch._library.triton import wrap_triton
 from triton import language as tl
 
 from blksprs.ops.misc.row_wise import row_wise_sum, row_wise_max, row_wise_sub
 from blksprs.utils.blksprs_tensor import BlksprsTensor
-from blksprs.utils.tools import get_triton_block_size, stride
+from blksprs.utils.tools import stride, get_autotune_configs
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
-    validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
+    validate_sparsity, validate_sparsity_block_size
 
 
-def softmax(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int,
-            triton_block_size: int = None, lut: dict = None) -> BlksprsTensor:
+def softmax(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int, lut: dict = None) -> BlksprsTensor:
     """Computes the softmax of a block-sparse tensor in compressed form.
 
     Note:
@@ -21,7 +22,6 @@ def softmax(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int,
         x (BlksprsTensor): A block-sparse tensor in compressed form.
         sparsity_layout (Tensor): The sparsity layout of the block-sparse tensor.
         sparsity_block_size (int): The size of the sparsity blocks.
-        triton_block_size (int): The block size to use for the triton kernel (default ``None``).
         lut (dict, optional): A dictionary containing the look-up tables for the operation (default ``None``).
 
     Returns:
@@ -35,244 +35,272 @@ def softmax(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int,
     validate_device(x)
     validate_sparsity(sparsity_block_size, (x, sparsity_layout))
     validate_sparsity_block_size(sparsity_block_size, x)
-    validate_triton_block_size(triton_block_size, sparsity_block_size)
-
-    lut = _BlocksparseSoftmax.build_lut(lut, sparsity_layout)
-
-    return BlksprsTensor(_BlocksparseSoftmax.apply(x, sparsity_layout,
-                                                   lut["sparsity_lut"],
-                                                   lut["sparsity_reverse_lut_rws"],
-                                                   sparsity_block_size, triton_block_size))
-
-
-class _BlocksparseSoftmax(torch.autograd.Function):
-
-    @staticmethod
-    def build_lut(lut: dict, sparsity_layout: Tensor):
-        if lut is None:
-            lut = dict()
-
-        if "sparsity_lut" not in lut:
-            sparsity_lut = torch.nonzero(sparsity_layout).contiguous()
-            lut["sparsity_lut"] = sparsity_lut
-
-
-        if "sparsity_reverse_lut_rws" not in lut:
-            sparsity_layout_rws, _ = torch.max(sparsity_layout, dim=-1, keepdim=True)
-            sparsity_layout_rws_flat = sparsity_layout_rws.reshape(-1)
-            sparsity_reverse_lut_rws = ((torch.cumsum(sparsity_layout_rws_flat, dim=-1) - 1) *
-                                        (sparsity_layout_rws_flat == 1) -
-                                        (1 * (sparsity_layout_rws_flat == 0)))
-            lut["sparsity_reverse_lut_rws"] = sparsity_reverse_lut_rws
-
-        validate_contiguous(sparsity_layout, lut["sparsity_lut"], lut["sparsity_reverse_lut_rws"])
-
-        return lut
-
-    @staticmethod
-    def forward(ctx, x: Tensor, sparsity_layout: Tensor,
-                sparsity_lut: Tensor,
-                sparsity_reverse_lut_rws: Tensor,
-                sparsity_block_size: int, triton_block_size: int) -> Tensor:
-        output = torch.empty_like(x)
-
-        x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = stride(x)
-        s_lut_r, s_lut_c = sparsity_lut.size()
-        s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
-        o_b, o_r, o_c = output.size()
-
-        x_row_wise_max, sparsity_layout_rwm = row_wise_max(x, sparsity_layout, sparsity_block_size,
-                                                           flag_slice_only=True,
-                                                           triton_block_size=triton_block_size)
-        x_scaled = row_wise_sub(x, sparsity_layout, x_row_wise_max, sparsity_block_size, triton_block_size)
-        x_exp = torch.exp(x_scaled)
-        x_exp_row_wise_sum, sparsity_layout_rws = row_wise_sum(x_exp, sparsity_layout, sparsity_block_size,
-                                                               flag_slice_only=True,
-                                                               triton_block_size=triton_block_size)
-
-        s_b, s_r, s_c = x_exp_row_wise_sum.shape
-        s_b_s, s_r_s, s_c_s = stride(x_exp_row_wise_sum)
-        s_l_s_b, s_l_s_r, s_l_s_c = sparsity_layout_rws.shape
-        s_l_s_b_s, s_l_s_r_s, s_l_s_c_s = stride(sparsity_layout_rws)
-
-        if triton_block_size is None:
-            triton_block_size = get_triton_block_size(sparsity_block_size)
-
-        triton_grid = lambda meta: [o_b,
-                                    triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
-                                    triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
-
-        (_BlocksparseSoftmax.kernel_blocksparse_softmax[triton_grid]
-         (x_exp,
-          x_b, x_b_s, x_r_s, x_c_s,
-          sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-          x_exp_row_wise_sum, s_b, s_b_s, s_r_s, s_c_s,
-          s_l_s_b, s_l_s_b_s, s_l_s_r_s,
-          sparsity_reverse_lut_rws,
-          output,
-          triton_block_size))
-
-        # Save for backward pass
-        ctx.save_for_backward(output, sparsity_layout, sparsity_lut)
-        ctx.sparsity_block_size = sparsity_block_size
-        ctx.triton_block_size = triton_block_size
-
-        return output
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        o, sparsity_layout, sparsity_lut = ctx.saved_tensors
-        sparsity_block_size = ctx.sparsity_block_size
-        triton_block_size = ctx.triton_block_size
-
-        s, sparsity_layout_s = row_wise_sum(grad_output * o, sparsity_layout, sparsity_block_size, flag_slice_only=True,
-                                            triton_block_size=triton_block_size)
-
-        sparsity_layout_s_flat = sparsity_layout_s.reshape(-1)
-        sparsity_reverse_lut_s = ((torch.cumsum(sparsity_layout_s_flat, dim=-1) - 1) *
-                                  (sparsity_layout_s_flat == 1) -
-                                  (1 * (sparsity_layout_s_flat == 0)))
-
-        o_b, o_r, o_c = o.size()
-        o_b_s, o_r_s, o_c_s = stride(o)
-        s_lut_r, s_lut_c = sparsity_lut.size()
-        s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
-        s_b, s_r, s_c = s.size()
-        s_b_s, s_r_s, s_c_s = stride(s)
-        s_l_s_b, s_l_s_r, s_l_s_c = sparsity_layout_s.size()
-        s_l_s_b_s, s_l_s_r_s, s_l_s_c_s = stride(sparsity_layout_s)
-
-        grad_x = torch.empty_like(o, dtype=torch.float)
-
-        triton_grid = lambda meta: [o_b,
-                                    triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
-                                    triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
-
-        (_BlocksparseSoftmax.kernel_blocksparse_softmax_grad_x[triton_grid]
-         (grad_output,
-          o_b, o_b_s, o_r_s, o_c_s,
-          o,
-          o_b, o_b_s, o_r_s, o_c_s,
-          sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-          s,
-          s_b, s_b_s, s_r_s, s_c_s,
-          s_l_s_b, s_l_s_b_s, s_l_s_r_s,
-          sparsity_reverse_lut_s,
-          grad_x,
-          o_b, o_b_s, o_r_s, o_c_s,
-          triton_block_size
-          ))
-
-        return grad_x, None, None, None, None, None
-
-    @staticmethod
-    @triton.jit
-    def kernel_blocksparse_softmax(x,
-                                   x_b, x_b_s, x_r_s, x_c_s,
-                                   s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-                                   s, s_b, s_b_s, s_r_s, s_c_s,
-                                   s_l_s_b, s_l_s_b_s, s_l_s_r_s,
-                                   r_lut_s,
-                                   o,
-                                   TRITON_BLOCK_SIZE: tl.constexpr) -> None:
-        # Get triton block indices
-        pid_blk = tl.program_id(axis=0)
-        pid_row = tl.program_id(axis=1)
-        pid_col = tl.program_id(axis=2)
-
-        # Get position of current sparsity block consisting of its batch and row index
-        spa_bat_idx = (pid_blk * s_lut_r_s + 0 * s_lut_c_s)
-        spa_bat_msk = (spa_bat_idx >= 0 and spa_bat_idx < s_lut_r * s_lut_r_s)
-        spa_bat = tl.load(s_lut + spa_bat_idx, mask=spa_bat_msk)
-
-        spa_row_idx = (pid_blk * s_lut_r_s + 1 * s_lut_c_s)
-        spa_row_msk = (spa_row_idx >= 0 and spa_row_idx < s_lut_r * s_lut_r_s)
-        spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
-
-        # Get reverse sparsity indices for s
-        rev_idx_spa_s_idx = (spa_bat * s_l_s_b_s +
-                             spa_row * s_l_s_r_s)
-        rev_idx_spa_s_msk = (rev_idx_spa_s_idx >= 0 and rev_idx_spa_s_idx < s_l_s_b * s_l_s_b_s)
-        rev_idx_spa_s = tl.load(r_lut_s + rev_idx_spa_s_idx, mask=rev_idx_spa_s_msk).to(tl.int32)
-
-        if rev_idx_spa_s >= 0:
-            # Load x block
-            blk_x_idx = ((pid_blk * x_b_s) +
-                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-            blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
-            blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
-
-            # Load sum block
-            blk_s_idx = (rev_idx_spa_s * s_b_s +
-                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * s_r_s)[:, None] +
-                         (tl.arange(0, 1) * s_c_s)[None, :])
-            blk_s_msk = (blk_s_idx >= 0 and blk_s_idx < s_b * s_b_s)
-            blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
-
-            # Compute softmax
-            buf = tl.div_rn(blk_x, blk_s)
-
-            # Store output
-            tl.store(o + blk_x_idx, buf, mask=blk_x_msk)
-
-    @staticmethod
-    @triton.jit
-    def kernel_blocksparse_softmax_grad_x(g,
-                                          g_b, g_b_s, g_r_s, g_c_s,
-                                          x,
-                                          x_b, x_b_s, x_r_s, x_c_s,
-                                          s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-                                          s,
-                                          s_b, s_b_s, s_r_s, s_c_s,
-                                          s_l_s_b, s_l_s_b_s, s_l_s_r_s,
-                                          r_lut_s,
-                                          o,
-                                          o_b, o_b_s, o_r_s, o_c_s,
-                                          TRITON_BLOCK_SIZE: tl.constexpr) -> None:
-        # Get triton block indices
-        pid_blk = tl.program_id(axis=0)
-        pid_row = tl.program_id(axis=1)
-        pid_col = tl.program_id(axis=2)
-
-        # Get position of current sparsity block consisting of its batch and row index
-        spa_bat_idx = (pid_blk * s_lut_r_s + 0 * s_lut_c_s)
-        spa_bat_msk = (spa_bat_idx >= 0 and spa_bat_idx < s_lut_r * s_lut_r_s)
-        spa_bat = tl.load(s_lut + spa_bat_idx, mask=spa_bat_msk)
-
-        spa_row_idx = (pid_blk * s_lut_r_s + 1 * s_lut_c_s)
-        spa_row_msk = (spa_row_idx >= 0 and spa_row_idx < s_lut_r * s_lut_r_s)
-        spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
-
-        rev_idx_spa_s_idx = (spa_bat * s_l_s_b_s +
-                             spa_row * s_l_s_r_s)
-        rev_idx_spa_s_msk = (rev_idx_spa_s_idx >= 0 and rev_idx_spa_s_idx < s_l_s_b * s_l_s_b_s)
-        rev_idx_spa_s = tl.load(r_lut_s + rev_idx_spa_s_idx, mask=rev_idx_spa_s_msk).to(tl.int32)
-
-        if rev_idx_spa_s >= 0:
-            blk_s_idx = (rev_idx_spa_s * s_b_s +
-                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * s_r_s)[:, None] +
-                         (tl.arange(0, 1) * s_c_s)[None, :])
-            blk_s_msk = (blk_s_idx >= 0 and blk_s_idx < s_b * s_b_s)
-            blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
-
-            blk_g_idx = ((pid_blk * g_b_s) +
-                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * g_r_s)[:, None] +
-                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * g_c_s)[None, :])
-            blk_g_msk = (blk_g_idx >= 0 and blk_g_idx < g_b * g_b_s)
-            blk_g = tl.load(g + blk_g_idx, mask=blk_g_msk)
-
-            blk_x_idx = ((pid_blk * x_b_s) +
-                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-            blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
-            blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
-
-            buf = blk_x * (blk_g - blk_s)
-
-            blk_o_idx = ((pid_blk * o_b_s) +
-                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +
-                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_c_s)[None, :])
-            blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s)
-            tl.store(o + blk_o_idx, buf, mask=blk_o_msk)
+
+    lut = softmax_build_lut(lut, sparsity_layout)
+
+    return BlksprsTensor(softmax_forward(x, sparsity_layout,
+                                         lut["sparsity_lut"],
+                                         lut["sparsity_reverse_lut_rws"],
+                                         sparsity_block_size))
+
+
+@triton_op("blksprs::softmax", mutates_args={})
+def softmax_forward(x: Tensor, sparsity_layout: Tensor,
+                    sparsity_lut: Tensor,
+                    sparsity_reverse_lut_rws: Tensor,
+                    sparsity_block_size: int) -> Tensor:
+    output = torch.empty_like(x)
+
+    x_b, x_r, x_c = x.size()
+    x_b_s, x_r_s, x_c_s = stride(x)
+    s_lut_r, s_lut_c = sparsity_lut.size()
+    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
+    o_b, o_r, o_c = output.size()
+
+    x_row_wise_max, sparsity_layout_rwm = row_wise_max(x, sparsity_layout, sparsity_block_size,
+                                                       flag_slice_only=True)
+    x_scaled = row_wise_sub(x, sparsity_layout, x_row_wise_max, sparsity_block_size)
+    x_exp = torch.exp(x_scaled)
+    x_exp_row_wise_sum, sparsity_layout_rws = row_wise_sum(x_exp, sparsity_layout, sparsity_block_size,
+                                                           flag_slice_only=True)
+
+    s_b, s_r, s_c = x_exp_row_wise_sum.shape
+    s_b_s, s_r_s, s_c_s = stride(x_exp_row_wise_sum)
+    s_l_s_b, s_l_s_r, s_l_s_c = sparsity_layout_rws.shape
+    s_l_s_b_s, s_l_s_r_s, s_l_s_c_s = stride(sparsity_layout_rws)
+
+    triton_grid = lambda meta: [o_b,
+                                triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
+                                triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
+
+    (wrap_triton(softmax_kernel)[triton_grid]
+     (x_exp,
+      x_b, x_b_s, x_r_s, x_c_s,
+      sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+      x_exp_row_wise_sum, s_b, s_b_s, s_r_s, s_c_s,
+      s_l_s_b, s_l_s_b_s, s_l_s_r_s,
+      sparsity_reverse_lut_rws,
+      output,
+      sparsity_block_size))
+
+    return output
+
+
+def softmax_backward(ctx, grad_output):
+    o, sparsity_layout, sparsity_lut = ctx.saved_tensors
+    sparsity_block_size = ctx.sparsity_block_size
+
+    s, sparsity_layout_s = row_wise_sum(grad_output * o, sparsity_layout, sparsity_block_size, flag_slice_only=True)
+
+    sparsity_layout_s_flat = sparsity_layout_s.reshape(-1)
+    sparsity_reverse_lut_s = ((torch.cumsum(sparsity_layout_s_flat, dim=-1) - 1) *
+                              (sparsity_layout_s_flat == 1) -
+                              (1 * (sparsity_layout_s_flat == 0)))
+
+    o_b, o_r, o_c = o.size()
+    o_b_s, o_r_s, o_c_s = stride(o)
+    s_lut_r, s_lut_c = sparsity_lut.size()
+    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
+    s_b, s_r, s_c = s.size()
+    s_b_s, s_r_s, s_c_s = stride(s)
+    s_l_s_b, s_l_s_r, s_l_s_c = sparsity_layout_s.size()
+    s_l_s_b_s, s_l_s_r_s, s_l_s_c_s = stride(sparsity_layout_s)
+
+    grad_x = torch.empty_like(o, dtype=torch.float)
+
+    triton_grid = lambda meta: [o_b,
+                                triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
+                                triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
+
+    (wrap_triton(softmax_kernel_grad)[triton_grid]
+     (grad_output,
+      o_b, o_b_s, o_r_s, o_c_s,
+      o,
+      o_b, o_b_s, o_r_s, o_c_s,
+      sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+      s,
+      s_b, s_b_s, s_r_s, s_c_s,
+      s_l_s_b, s_l_s_b_s, s_l_s_r_s,
+      sparsity_reverse_lut_s,
+      grad_x,
+      o_b, o_b_s, o_r_s, o_c_s,
+      sparsity_block_size))
+
+    return grad_x, None, None, None, None, None
+
+
+@triton.autotune(
+    configs=get_autotune_configs(),
+    key=[]
+)
+@triton.jit
+def softmax_kernel(x,
+                   x_b, x_b_s, x_r_s, x_c_s,
+                   s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+                   s, s_b, s_b_s, s_r_s, s_c_s,
+                   s_l_s_b, s_l_s_b_s, s_l_s_r_s,
+                   r_lut_s,
+                   o,
+                   sparsity_block_size,
+                   TRITON_BLOCK_SIZE: tl.constexpr) -> None:
+    # Get triton block indices
+    pid_blk = tl.program_id(axis=0)
+    pid_row = tl.program_id(axis=1)
+    pid_col = tl.program_id(axis=2)
+
+    # Get valid triton block size
+    val_tbs = min(sparsity_block_size, TRITON_BLOCK_SIZE)
+
+    # Get position of current sparsity block consisting of its batch and row index
+    spa_bat_idx = (pid_blk * s_lut_r_s + 0 * s_lut_c_s)
+    spa_bat_msk = (spa_bat_idx >= 0 and spa_bat_idx < s_lut_r * s_lut_r_s)
+    spa_bat = tl.load(s_lut + spa_bat_idx, mask=spa_bat_msk)
+
+    spa_row_idx = (pid_blk * s_lut_r_s + 1 * s_lut_c_s)
+    spa_row_msk = (spa_row_idx >= 0 and spa_row_idx < s_lut_r * s_lut_r_s)
+    spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
+
+    # Get reverse sparsity indices for s
+    rev_idx_spa_s_idx = (spa_bat * s_l_s_b_s +
+                         spa_row * s_l_s_r_s)
+    rev_idx_spa_s_msk = (rev_idx_spa_s_idx >= 0 and rev_idx_spa_s_idx < s_l_s_b * s_l_s_b_s)
+    rev_idx_spa_s = tl.load(r_lut_s + rev_idx_spa_s_idx, mask=rev_idx_spa_s_msk).to(tl.int32)
+
+    if rev_idx_spa_s >= 0:
+        # Load x block
+        blk_x_idx = ((pid_blk * x_b_s) +
+                     ((pid_row * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
+                     ((pid_col * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
+        blk_x_msk = ((blk_x_idx >= 0 and
+                      blk_x_idx < x_b * x_b_s) and
+                     (tl.arange(0, TRITON_BLOCK_SIZE)[:, None] < val_tbs and
+                      tl.arange(0, TRITON_BLOCK_SIZE)[None, :] < val_tbs))
+        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
+
+        # Load sum block
+        blk_s_idx = (rev_idx_spa_s * s_b_s +
+                     ((pid_row * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * s_r_s)[:, None] +
+                     (tl.arange(0, 1) * s_c_s)[None, :])
+        blk_s_msk = ((blk_s_idx >= 0 and
+                      blk_s_idx < s_b * s_b_s) and
+                     (tl.arange(0, TRITON_BLOCK_SIZE)[:, None] < val_tbs and
+                      tl.arange(0, 1)[None, :] < val_tbs))
+        blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
+
+        # Compute softmax
+        buf = tl.div_rn(blk_x, blk_s)
+
+        # Store output
+        tl.store(o + blk_x_idx, buf, mask=blk_x_msk)
+
+
+@triton.autotune(
+    configs=get_autotune_configs(),
+    key=[]
+)
+@triton.jit
+def softmax_kernel_grad(g,
+                        g_b, g_b_s, g_r_s, g_c_s,
+                        x,
+                        x_b, x_b_s, x_r_s, x_c_s,
+                        s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+                        s,
+                        s_b, s_b_s, s_r_s, s_c_s,
+                        s_l_s_b, s_l_s_b_s, s_l_s_r_s,
+                        r_lut_s,
+                        o,
+                        o_b, o_b_s, o_r_s, o_c_s,
+                        sparsity_block_size,
+                        TRITON_BLOCK_SIZE: tl.constexpr) -> None:
+    # Get triton block indices
+    pid_blk = tl.program_id(axis=0)
+    pid_row = tl.program_id(axis=1)
+    pid_col = tl.program_id(axis=2)
+
+    # Get valid triton block size
+    val_tbs = min(sparsity_block_size, TRITON_BLOCK_SIZE)
+
+    # Get position of current sparsity block consisting of its batch and row index
+    spa_bat_idx = (pid_blk * s_lut_r_s + 0 * s_lut_c_s)
+    spa_bat_msk = (spa_bat_idx >= 0 and spa_bat_idx < s_lut_r * s_lut_r_s)
+    spa_bat = tl.load(s_lut + spa_bat_idx, mask=spa_bat_msk)
+
+    spa_row_idx = (pid_blk * s_lut_r_s + 1 * s_lut_c_s)
+    spa_row_msk = (spa_row_idx >= 0 and spa_row_idx < s_lut_r * s_lut_r_s)
+    spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
+
+    rev_idx_spa_s_idx = (spa_bat * s_l_s_b_s +
+                         spa_row * s_l_s_r_s)
+    rev_idx_spa_s_msk = (rev_idx_spa_s_idx >= 0 and rev_idx_spa_s_idx < s_l_s_b * s_l_s_b_s)
+    rev_idx_spa_s = tl.load(r_lut_s + rev_idx_spa_s_idx, mask=rev_idx_spa_s_msk).to(tl.int32)
+
+    if rev_idx_spa_s >= 0:
+        blk_s_idx = (rev_idx_spa_s * s_b_s +
+                     ((pid_row * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * s_r_s)[:, None] +
+                     (tl.arange(0, 1) * s_c_s)[None, :])
+        blk_s_msk = ((blk_s_idx >= 0 and
+                      blk_s_idx < s_b * s_b_s) and
+                     (tl.arange(0, TRITON_BLOCK_SIZE)[:, None] < val_tbs and
+                      tl.arange(0, 1)[None, :] < val_tbs))
+        blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
+
+        blk_g_idx = ((pid_blk * g_b_s) +
+                     ((pid_row * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * g_r_s)[:, None] +
+                     ((pid_col * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * g_c_s)[None, :])
+        blk_g_msk = ((blk_g_idx >= 0 and
+                      blk_g_idx < g_b * g_b_s) and
+                     (tl.arange(0, TRITON_BLOCK_SIZE)[:, None] < val_tbs and
+                      tl.arange(0, TRITON_BLOCK_SIZE)[None, :] < val_tbs))
+        blk_g = tl.load(g + blk_g_idx, mask=blk_g_msk)
+
+        blk_x_idx = ((pid_blk * x_b_s) +
+                     ((pid_row * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
+                     ((pid_col * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
+        blk_x_msk = ((blk_x_idx >= 0 and
+                      blk_x_idx < x_b * x_b_s) and
+                     (tl.arange(0, TRITON_BLOCK_SIZE)[:, None] < val_tbs and
+                      tl.arange(0, TRITON_BLOCK_SIZE)[None, :] < val_tbs))
+        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
+
+        buf = blk_x * (blk_g - blk_s)
+
+        blk_o_idx = ((pid_blk * o_b_s) +
+                     ((pid_row * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +
+                     ((pid_col * val_tbs + tl.arange(0, TRITON_BLOCK_SIZE)) * o_c_s)[None, :])
+        blk_o_msk = ((blk_o_idx >= 0 and
+                      blk_o_idx < o_b * o_b_s) and
+                     (tl.arange(0, TRITON_BLOCK_SIZE)[:, None] < val_tbs and
+                      tl.arange(0, TRITON_BLOCK_SIZE)[None, :] < val_tbs))
+        tl.store(o + blk_o_idx, buf, mask=blk_o_msk)
+
+
+def softmax_build_lut(lut: dict, sparsity_layout: Tensor):
+    if lut is None:
+        lut = dict()
+
+    if "sparsity_lut" not in lut:
+        sparsity_lut = torch.nonzero(sparsity_layout).contiguous()
+        lut["sparsity_lut"] = sparsity_lut
+
+    if "sparsity_reverse_lut_rws" not in lut:
+        sparsity_layout_rws, _ = torch.max(sparsity_layout, dim=-1, keepdim=True)
+        sparsity_layout_rws_flat = sparsity_layout_rws.reshape(-1)
+        sparsity_reverse_lut_rws = ((torch.cumsum(sparsity_layout_rws_flat, dim=-1) - 1) *
+                                    (sparsity_layout_rws_flat == 1) -
+                                    (1 * (sparsity_layout_rws_flat == 0)))
+        lut["sparsity_reverse_lut_rws"] = sparsity_reverse_lut_rws
+
+    validate_contiguous(sparsity_layout, lut["sparsity_lut"], lut["sparsity_reverse_lut_rws"])
+
+    return lut
+
+
+# noinspection PyUnusedLocal
+def softmax_setup_context(ctx, inputs, output):
+    (_, sparsity_layout, sparsity_lut, _, sparsity_block_size) = inputs
+
+    ctx.save_for_backward(output, sparsity_layout, sparsity_lut)
+    ctx.sparsity_block_size = sparsity_block_size
+
+
+softmax_forward.register_autograd(softmax_backward, setup_context=softmax_setup_context)