blksprs 1.5__py3-none-any.whl → 1.7__py3-none-any.whl

blksprs/__init__.py

@@ -4,10 +4,11 @@ from blksprs.ops.exp import exp
 from blksprs.ops.matmul import matmul
 from blksprs.ops.softmax import softmax
 from blksprs.ops.transpose import transpose
+from blksprs.ops.partitioning import split, merge
 
 class layout:
     from blksprs.layouting.distribution_layout import build_distribution_layout
-    from blksprs.layouting.sparsity_layout import build_sparsity_layout, build_sparsity_layout_adaption
+    from blksprs.layouting.sparsity_layout import build_sparsity_layout, build_sparsity_layout_adaption, build_sparsity_layout_matmul
 
 class misc:
     from blksprs.misc.broadcast_ops import broadcast_add, broadcast_sub
@@ -18,4 +19,4 @@ class util:
     from blksprs.utils.tools import do_shape_blocksparse, undo_shape_blocksparse, disable_validation
 
 class experimental:
-    from blksprs.experimental.distribution_mdi import gather_mdi
+    from blksprs.experimental.distribution_mdi import gather_mdi, scatter_reduce_mdi

blksprs/experimental/distribution_mdi.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
     validate_sparsity, validate_dtype_int, validate_sparsity_block_size, validate_triton_block_size
 
@@ -51,15 +51,15 @@ class _BlocksparseGatherMDI(torch.autograd.Function):
         output = torch.empty_like(idx_col, dtype=x.dtype)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         s_l_x_b, s_l_x_r, s_l_x_c = sparsity_layout_x.size()
-        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = sparsity_layout_x.stride()
+        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = stride(sparsity_layout_x)
         i_b, i_r, i_c = idx_col.size()
-        i_b_s, i_r_s, i_c_s = idx_col.stride()
+        i_b_s, i_r_s, i_c_s = stride(idx_col)
         s_lut_i_r, s_lut_i_c = sparsity_lut_i.size()
-        s_lut_i_r_s, s_lut_i_c_s = sparsity_lut_i.stride()
+        s_lut_i_r_s, s_lut_i_c_s = stride(sparsity_lut_i)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -224,15 +224,15 @@ class _BlocksparseScatterReduceMDI(torch.autograd.Function):
                              dtype=x.dtype, device=x.device)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         s_lut_x_r, s_lut_x_c = sparsity_lut_x.size()
-        s_lut_x_r_s, s_lut_x_c_s = sparsity_lut_x.stride()
+        s_lut_x_r_s, s_lut_x_c_s = stride(sparsity_lut_x)
         i_b, i_r, i_c = idx_col.size()
-        i_b_s, i_r_s, i_c_s = idx_col.stride()
+        i_b_s, i_r_s, i_c_s = stride(idx_col)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
         s_l_o_b, s_l_o_r, s_l_o_c = sparsity_layout_o.size()
-        s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = sparsity_layout_o.stride()
+        s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_o)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -366,11 +366,11 @@ def build_distribution_layout_mdi(idx_bat: Tensor, idx_row: Tensor, idx_col: Ten
                          dtype=torch.bool, device=idx_col.device)
 
     i_b, i_r, i_c = idx_col.size()
-    i_b_s, i_r_s, i_c_s = idx_col.stride()
+    i_b_s, i_r_s, i_c_s = stride(idx_col)
     s_lut_i_r, s_lut_i_c = sparsity_lut_i.size()
-    s_lut_i_r_s, s_lut_i_c_s = sparsity_lut_i.stride()
+    s_lut_i_r_s, s_lut_i_c_s = stride(sparsity_lut_i)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)

blksprs/layouting/distribution_layout.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_triton_block_size, validate_dimensions, validate_device, \
     validate_contiguous
 
@@ -34,11 +34,11 @@ def build_distribution_layout(indices: Tensor, sparsity_layout_indices: Tensor,
                          dtype=torch.bool, device=indices.device)
 
     i_b, i_r, i_c = indices.size()
-    i_b_s, i_r_s, i_c_s = indices.stride()
+    i_b_s, i_r_s, i_c_s = stride(indices)
     s_lut_i_r, s_lut_i_c = sparsity_lut_i.size()
-    s_lut_i_r_s, s_lut_i_c_s = sparsity_lut_i.stride()
+    s_lut_i_r_s, s_lut_i_c_s = stride(sparsity_lut_i)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)

blksprs/layouting/sparsity_layout.py

@@ -5,7 +5,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_triton_block_size, validate_dimensions, validate_device, \
     validate_contiguous, validate_sparsity, validate_sparsity_block_size
 
@@ -30,9 +30,9 @@ def build_sparsity_layout(x: Tensor, sparsity_block_size: int, triton_block_size
                          dtype=torch.bool, device=x.device)
 
     x_b, x_r, x_c = x.size()
-    x_b_s, x_r_s, x_c_s = x.stride()
+    x_b_s, x_r_s, x_c_s = stride(x)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -120,10 +120,10 @@ def build_sparsity_layout_adaption(x: Tensor, sparsity_layout_from: Tensor,
     output = torch.zeros(o_b, o_r, o_c, dtype=torch.bool, device=x.device)
 
     x_b, x_r, x_c = x.size()
-    x_b_s, x_r_s, x_c_s = x.stride()
+    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 = sparsity_lut.stride()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
+    o_b_s, o_r_s, o_c_s = stride(output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size_from)
@@ -188,3 +188,39 @@ def kernel_sparsity_layout_adaption(x,
                        // sparsity_block_size_to) * o_c_s))
         blk_o_msk = (blk_o_idx < o_b * o_b_s)
         tl.store(o + blk_o_idx, 1, mask=blk_o_msk)
+
+
+def build_sparsity_layout_matmul(sparsity_layout_x: Tensor, sparsity_layout_y: Tensor):
+    """Builds the precise sparsity layout of the result of a matrix multiplication between the two input tensors.
+
+    Args:
+        sparsity_layout_x (Tensor): The sparsity layout of the first block-sparse tensor.
+        sparsity_layout_y (Tensor): The sparsity layout of the second block-sparse tensor.
+
+    Returns:
+        Tensor: The precise sparsity layout of the result of a matrix multiplication between the two input tensors.
+
+    """
+    return torch.matmul(sparsity_layout_x.to(torch.float), sparsity_layout_y.to(torch.float)).to(torch.bool)
+
+
+def build_sparsity_layout_matmul_fast(sparsity_layout_x: Tensor, sparsity_layout_y: Tensor):
+    """Builds the approximate sparsity layout of the result of a matrix multiplication between the two input tensors.
+
+    Note:
+        This function is faster than the ``build_sparsity_layout_matmul`` function due to the fact that it only checks
+            whether at least one of the blocks in either of the vectors participating in the matmul is non-sparse. The
+            resulting sparsity layout may thus overestimate the actual sparsity of the result.
+
+    Args:
+        sparsity_layout_x (Tensor): The sparsity layout of the first block-sparse tensor.
+        sparsity_layout_y (Tensor): The sparsity layout of the second block-sparse tensor.
+
+    Returns:
+        Tensor: The approximate sparsity layout of the result of a matrix multiplication between the two input tensors.
+
+    """
+    sparsity_layout_x_slice = torch.max(sparsity_layout_x, dim=-1).values.unsqueeze(-1)
+    sparsity_layout_y_slice = torch.max(sparsity_layout_y, dim=-2).values.unsqueeze(1)
+
+    return torch.logical_or(sparsity_layout_x_slice, sparsity_layout_y_slice)

blksprs/misc/broadcast_ops.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_device, \
     validate_sparsity_block_size, validate_triton_block_size
 
@@ -44,13 +44,13 @@ def broadcast_add(x: Tensor, y: Tensor, sparsity_layout_output: Tensor,
     output = torch.zeros(n_sparse_blocks, sparsity_block_size, sparsity_block_size, dtype=x.dtype, device=x.device)
 
     x_b, x_c = x.size()
-    x_b_s, x_c_s = x.stride()
+    x_b_s, x_c_s = stride(x)
     y_b, y_c = y.size()
-    y_b_s, y_c_s = y.stride()
+    y_b_s, y_c_s = stride(y)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
     s_lut_o_r, s_lut_o_c = sparsity_lut_o.size()
-    s_lut_o_r_s, s_lut_o_c_s = sparsity_lut_o.stride()
+    s_lut_o_r_s, s_lut_o_c_s = stride(sparsity_lut_o)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)

blksprs/misc/repeat_interleave.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_device, \
     validate_sparsity_block_size, validate_triton_block_size, validate_dimensions
 
@@ -35,7 +35,7 @@ def repeat_interleave(x: Tensor, sparsity_layout: Tensor, repeats: int,
     validate_sparsity_block_size(sparsity_block_size, x)
     validate_triton_block_size(triton_block_size, sparsity_block_size)
 
-    sparsity_layout_output = torch.repeat_interleave(sparsity_layout, 3, dim=0).contiguous()
+    sparsity_layout_output = torch.repeat_interleave(sparsity_layout, repeats, dim=0).contiguous()
 
     sparsity_lut = torch.nonzero(sparsity_layout).contiguous()
 
@@ -52,13 +52,13 @@ def repeat_interleave(x: Tensor, sparsity_layout: Tensor, repeats: int,
                          dtype=x.dtype, device=x.device)
 
     x_b, x_r, x_c = x.size()
-    x_b_s, x_r_s, x_c_s = x.stride()
+    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 = sparsity_lut.stride()
+    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
     s_l_o_b, s_l_o_r, s_l_o_c = sparsity_layout_output.size()
-    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = sparsity_layout_output.stride()
+    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)

blksprs/misc/row_wise.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_dimensions, validate_contiguous, validate_device, validate_sparsity, \
     validate_sparsity_block_size, validate_triton_block_size
 
@@ -60,13 +60,13 @@ def row_wise_sum(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int,
                          device=x.device)
 
     x_b, x_r, x_c = x.size()
-    x_b_s, x_r_s, x_c_s = x.stride()
+    x_b_s, x_r_s, x_c_s = stride(x)
     s_lut_x_r, s_lut_x_c = sparsity_lut.size()
-    s_lut_x_r_s, s_lut_x_c_s = sparsity_lut.stride()
+    s_lut_x_r_s, s_lut_x_c_s = stride(sparsity_lut)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
     s_l_o_b, s_l_o_r, s_l_o_c = sparsity_layout_output.size()
-    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = sparsity_layout_output.stride()
+    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -183,13 +183,13 @@ def row_wise_max(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int,
                         device=x.device)
 
     x_b, x_r, x_c = x.size()
-    x_b_s, x_r_s, x_c_s = x.stride()
+    x_b_s, x_r_s, x_c_s = stride(x)
     s_lut_x_r, s_lut_x_c = sparsity_lut.size()
-    s_lut_x_r_s, s_lut_x_c_s = sparsity_lut.stride()
+    s_lut_x_r_s, s_lut_x_c_s = stride(sparsity_lut)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
     s_l_o_b, s_l_o_r, s_l_o_c = sparsity_layout_output.size()
-    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = sparsity_layout_output.stride()
+    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -290,15 +290,15 @@ def row_wise_add(x: Tensor, sparsity_layout_x: Tensor, y: Tensor,
     output = torch.empty_like(x)
 
     x_b, x_r, x_c = x.size()
-    x_b_s, x_r_s, x_c_s = x.stride()
+    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 = sparsity_lut.stride()
+    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
     y_b, y_r, y_c = y.size()
-    y_b_s, y_r_s, y_c_s = y.stride()
+    y_b_s, y_r_s, y_c_s = stride(y)
     s_l_y_b, s_l_y_r, s_l_y_c = sparsity_layout_rwm.size()
-    s_l_y_b_s, s_l_y_r_s, s_l_y_c_s = sparsity_layout_rwm.stride()
+    s_l_y_b_s, s_l_y_r_s, s_l_y_c_s = stride(sparsity_layout_rwm)
     o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = output.stride()
+    o_b_s, o_r_s, o_c_s = stride(output)
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -361,7 +361,8 @@ def kernel_blocksparse_row_wise_add(x,
     rev_idx_spa_s = tl.load(r_lut_y + rev_idx_spa_s_idx, mask=rev_idx_spa_s_msk).to(tl.int32)
 
     if rev_idx_spa_s == -1:
-        assert False, "Invalid sparsity block"
+        tl.device_assert(False)
+        return
 
     # Load x block
     blk_x_idx = ((pid_blk * x_b_s) +

blksprs/ops/conversion.py

@@ -6,7 +6,7 @@ from torch import Tensor
 from triton import language as tl
 
 from blksprs.layouting.sparsity_layout import build_sparsity_layout_adaption
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
     validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
 
@@ -65,11 +65,11 @@ class _BlocksparseToDense(torch.autograd.Function):
                             dtype=x.dtype, device=x.device)
 
         x_b, x_r, x_c = x.shape
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         s_l_b, s_l_r, s_l_c = sparsity_layout.size()
-        s_l_b_s, s_l_r_s, s_l_c_s = sparsity_layout.stride()
+        s_l_b_s, s_l_r_s, s_l_c_s = stride(sparsity_layout)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -190,11 +190,11 @@ class _BlocksparseToSparse(torch.autograd.Function):
                              dtype=x.dtype, device=x.device)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        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 = sparsity_lut.stride()
+        s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -347,13 +347,13 @@ class _BlocksparseAdaptLayout(torch.autograd.Function):
                              dtype=x.dtype, device=x.device)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         s_l_x_b, s_l_x_r, s_l_x_c = sparsity_layout_from.size()
-        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = sparsity_layout_from.stride()
+        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = stride(sparsity_layout_from)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
         s_lut_o_r, s_lut_o_c = sparsity_lut_to.size()
-        s_lut_o_r_s, s_lut_o_c_s = sparsity_lut_to.stride()
+        s_lut_o_r_s, s_lut_o_c_s = stride(sparsity_lut_to)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(min_sparsity_block_size)

blksprs/ops/distribution.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
     validate_sparsity, validate_dtype_int, validate_sparsity_block_size, validate_triton_block_size
 
@@ -59,15 +59,15 @@ class _BlocksparseGather(torch.autograd.Function):
         output = torch.empty_like(i, dtype=x.dtype)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         s_l_x_b, s_l_x_r, s_l_x_c = sparsity_layout_x.size()
-        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = sparsity_layout_x.stride()
+        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = stride(sparsity_layout_x)
         i_b, i_r, i_c = i.size()
-        i_b_s, i_r_s, i_c_s = i.stride()
+        i_b_s, i_r_s, i_c_s = stride(i)
         s_lut_i_r, s_lut_i_c = sparsity_lut_i.size()
-        s_lut_i_r_s, s_lut_i_c_s = sparsity_lut_i.stride()
+        s_lut_i_r_s, s_lut_i_c_s = stride(sparsity_lut_i)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -248,15 +248,15 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
                              dtype=x.dtype, device=x.device)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         s_lut_x_r, s_lut_x_c = sparsity_lut_x.size()
-        s_lut_x_r_s, s_lut_x_c_s = sparsity_lut_x.stride()
+        s_lut_x_r_s, s_lut_x_c_s = stride(sparsity_lut_x)
         i_b, i_r, i_c = i.size()
-        i_b_s, i_r_s, i_c_s = i.stride()
+        i_b_s, i_r_s, i_c_s = stride(i)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
         s_l_o_b, s_l_o_r, s_l_o_c = sparsity_layout_o.size()
-        s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = sparsity_layout_o.stride()
+        s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_o)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)

blksprs/ops/exp.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
     validate_sparsity_block_size, validate_triton_block_size
 
@@ -43,9 +43,9 @@ class _BlocksparseExp(torch.autograd.Function):
         output = torch.empty_like(x)
 
         x_b, x_r, x_c = x.shape
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         o_b, o_r, o_c = output.shape
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)

blksprs/ops/matmul.py

@@ -4,7 +4,7 @@ from torch import Tensor
 from triton import language as tl
 
 from blksprs.ops.transpose import transpose
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
     validate_sparsity, validate_sparsity_block_size, validate_triton_block_size, validate_dtype_float
 
@@ -82,17 +82,17 @@ class _BlocksparseMatmulSSS(torch.autograd.Function):
                              dtype=x.dtype, device=x.device)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
         s_l_x_b, s_l_x_r, s_l_x_c = sparsity_layout_x.size()
-        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = sparsity_layout_x.stride()
+        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = stride(sparsity_layout_x)
         y_b, y_r, y_c = y.size()
-        y_b_s, y_r_s, y_c_s = y.stride()
+        y_b_s, y_r_s, y_c_s = stride(y)
         s_l_y_b, s_l_y_r, s_l_y_c = sparsity_layout_y.size()
-        s_l_y_b_s, s_l_y_r_s, s_l_y_c_s = sparsity_layout_y.stride()
+        s_l_y_b_s, s_l_y_r_s, s_l_y_c_s = stride(sparsity_layout_y)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
         s_lut_o_r, s_lut_o_c = sparsity_lut_o.size()
-        s_lut_o_r_s, s_lut_o_c_s = sparsity_lut_o.stride()
+        s_lut_o_r_s, s_lut_o_c_s = stride(sparsity_lut_o)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)

blksprs/ops/partitioning.py

@@ -0,0 +1,155 @@
+import torch
+from torch import Tensor
+
+from blksprs.ops.repeat import forward_flow
+
+from blksprs.utils.validation import validate_dimensions, validate_contiguous, validate_device, \
+    validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
+
+
+def split(x: Tensor, sparsity_layout: Tensor, partitions: int,
+          sparsity_block_size: int, triton_block_size: int = None) -> (Tensor, Tensor):
+    """Splits a block-sparse tensor in compressed form along the last dimension into partitions.
+
+    Args:
+        x (Tensor): A block-sparse tensor in compressed form.
+        sparsity_layout (Tensor): The sparsity layout of the block-sparse tensor.
+        partitions (int): The number of partitions to split the block-sparse tensor into.
+        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``).
+
+    Returns:
+        Tensor: The block-sparse tensor split into partitions in compressed form.
+        Tensor: The sparsity layout of the output tensor.
+
+    """
+    x = x.contiguous()
+
+    validate_dimensions(x)
+    validate_contiguous(x)
+    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)
+
+    sparsity_layout_output = (sparsity_layout
+                              .reshape(sparsity_layout.size(0), sparsity_layout.size(1), partitions,
+                                       sparsity_layout.size(2) // partitions)
+                              .permute(0, 2, 1, 3)
+                              .reshape(sparsity_layout.size(0) * partitions, sparsity_layout.size(1),
+                                       sparsity_layout.size(2) // partitions).contiguous())
+
+    sparsity_lut = torch.nonzero(sparsity_layout_output).contiguous()
+
+    sparsity_layout_flat = sparsity_layout.reshape(-1)
+    sparsity_reverse_lut = (((torch.cumsum(sparsity_layout_flat, dim=-1) - 1) *
+                             (sparsity_layout_flat == 1) -
+                             (1 * (sparsity_layout_flat == 0)))
+                            .reshape(sparsity_layout.size(0), sparsity_layout.size(1), partitions,
+                                     sparsity_layout.size(2) // partitions)
+                            .permute(0, 2, 1, 3).reshape(-1).contiguous())
+
+    n_sparse_blocks = torch.sum(sparsity_layout_output.to(torch.int)).item()
+
+    validate_contiguous(sparsity_layout_output, sparsity_lut, sparsity_reverse_lut)
+
+    return _BlocksparseSplit.apply(x, sparsity_layout_output, sparsity_lut, sparsity_reverse_lut, partitions,
+                                   sparsity_block_size, n_sparse_blocks, triton_block_size), sparsity_layout_output
+
+
+class _BlocksparseSplit(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x: Tensor, sparsity_layout_o: Tensor, sparsity_lut: Tensor, sparsity_reverse_lut: Tensor,
+                num_partitions: int, sparsity_block_size: int, n_sparse_blocks: int, triton_block_size: int) -> Tensor:
+        ctx.save_for_backward(sparsity_layout_o)
+        ctx.num_partitions = num_partitions
+
+        return forward_flow(ctx, x, sparsity_layout_o, sparsity_lut, sparsity_reverse_lut, sparsity_block_size,
+                            n_sparse_blocks, triton_block_size)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        sparsity_layout = ctx.saved_tensors[0]
+        num_partitions = ctx.num_partitions
+        sparsity_block_size = ctx.sparsity_block_size
+        triton_block_size = ctx.triton_block_size
+
+        return merge(grad_output, sparsity_layout, num_partitions,
+                     sparsity_block_size, triton_block_size)[0], None, None, None, None, None, None, None
+
+
+def merge(x: Tensor, sparsity_layout: Tensor, partitions: int,
+          sparsity_block_size: int, triton_block_size: int = None) -> (Tensor, Tensor):
+    """Merges the specified partitions of a block-sparse tensor in compressed form along the last dimension.
+
+    Args:
+        x (Tensor): A block-sparse tensor in compressed form.
+        sparsity_layout (Tensor): The sparsity layout of the block-sparse tensor.
+        partitions (int): The number of partitions to be merged.
+        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``).
+
+    Returns:
+        Tensor: The merged block-sparse tensor in compressed form.
+        Tensor: The sparsity layout of the output tensor.
+
+    """
+    x = x.contiguous()
+
+    validate_dimensions(x)
+    validate_contiguous(x)
+    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)
+
+    sparsity_layout_output = (sparsity_layout.reshape(sparsity_layout.size(0) // partitions, partitions,
+                                                      sparsity_layout.size(1), sparsity_layout.size(2))
+                              .permute(0, 2, 1, 3)
+                              .reshape(sparsity_layout.size(0) // partitions,
+                                       sparsity_layout.size(1), sparsity_layout.size(2) * partitions).contiguous())
+
+    sparsity_lut = torch.nonzero(sparsity_layout_output).contiguous()
+
+    sparsity_layout_flat = sparsity_layout.reshape(-1)
+    sparsity_reverse_lut = (((torch.cumsum(sparsity_layout_flat, dim=-1) - 1) *
+                             (sparsity_layout_flat == 1) -
+                             (1 * (sparsity_layout_flat == 0)))
+                            .reshape(sparsity_layout.size(0) // partitions, partitions,
+                                     sparsity_layout.size(1), sparsity_layout.size(2))
+                            .permute(0, 2, 1, 3)
+                            .reshape(sparsity_layout.size(0) // partitions,
+                                     sparsity_layout.size(1), sparsity_layout.size(2) * partitions)
+                            .reshape(-1).contiguous())
+
+    n_sparse_blocks = torch.sum(sparsity_layout_output.to(torch.int)).item()
+
+    validate_contiguous(sparsity_layout_output, sparsity_lut, sparsity_reverse_lut)
+
+    return _BlocksparseMerge.apply(x, sparsity_layout_output, sparsity_lut, sparsity_reverse_lut, partitions,
+                                   sparsity_block_size, n_sparse_blocks, triton_block_size), sparsity_layout_output
+
+
+class _BlocksparseMerge(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x: Tensor, sparsity_layout_o: Tensor, sparsity_lut: Tensor, sparsity_reverse_lut: Tensor,
+                num_partitions: int, sparsity_block_size: int, n_sparse_blocks: int, triton_block_size: int) -> Tensor:
+        ctx.save_for_backward(sparsity_layout_o)
+        ctx.num_partitions = num_partitions
+
+        return forward_flow(ctx, x, sparsity_layout_o, sparsity_lut, sparsity_reverse_lut, sparsity_block_size,
+                            n_sparse_blocks, triton_block_size)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        sparsity_layout = ctx.saved_tensors[0]
+        num_partitions = ctx.num_partitions
+        sparsity_block_size = ctx.sparsity_block_size
+        triton_block_size = ctx.triton_block_size
+
+        return split(grad_output, sparsity_layout, num_partitions,
+                     sparsity_block_size, triton_block_size)[0], None, None, None, None, None, None, None
+
+

blksprs/ops/repeat.py

@@ -0,0 +1,241 @@
+import torch
+import triton
+from triton import language as tl
+from torch import Tensor
+
+from blksprs.utils.tools import get_triton_block_size, stride
+from blksprs.utils.validation import validate_dimensions, validate_contiguous, validate_device, \
+    validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
+
+
+def repeat(x: Tensor, sparsity_layout_x: Tensor, repeats: tuple[int, int, int],
+           sparsity_block_size: int, sparsity_layout_output: Tensor = None, triton_block_size: int = None) -> (
+        Tensor, Tensor):
+    x = x.contiguous()
+
+    validate_dimensions(x)
+    validate_contiguous(x)
+    validate_device(x)
+    validate_sparsity(sparsity_block_size, (x, sparsity_layout_x))
+    validate_sparsity_block_size(sparsity_block_size, x)
+    validate_triton_block_size(triton_block_size, sparsity_block_size)
+
+    sparsity_layout_o = sparsity_layout_x.repeat(repeats[0], repeats[1], repeats[2])
+
+    if sparsity_layout_output is not None:
+        sparsity_layout_o = torch.logical_and(sparsity_layout_o, sparsity_layout_output)
+
+    sparsity_lut = torch.nonzero(sparsity_layout_o).contiguous()
+
+    sparsity_layout_flat = sparsity_layout_x.reshape(-1)
+    sparsity_reverse_lut = (((torch.cumsum(sparsity_layout_flat, dim=-1) - 1) *
+                             (sparsity_layout_flat == 1) -
+                             (1 * (sparsity_layout_flat == 0)))
+                            .reshape(sparsity_layout_x.size())
+                            .repeat(repeats[0], repeats[1], repeats[2])
+                            .reshape(-1).contiguous())
+
+    n_sparse_blocks = torch.sum(sparsity_layout_o.to(torch.int)).item()
+
+    validate_contiguous(sparsity_layout_o, sparsity_lut, sparsity_reverse_lut)
+
+    return _BlocksparseRepeat.apply(x, sparsity_layout_x, sparsity_layout_o, sparsity_lut, sparsity_reverse_lut,
+                                    sparsity_block_size, n_sparse_blocks, triton_block_size), sparsity_layout_o
+
+
+class _BlocksparseRepeat(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x: Tensor, sparsity_layout_x: Tensor, sparsity_layout_o: Tensor, sparsity_lut: Tensor,
+                sparsity_reverse_lut: Tensor,
+                sparsity_block_size: int, n_sparse_blocks: int,
+                triton_block_size: int) -> Tensor:
+        ctx.save_for_backward(sparsity_layout_x, sparsity_layout_o, sparsity_lut, sparsity_reverse_lut)
+        ctx.x_size = x.size()
+        ctx.x_stride = stride(x)
+
+        return forward_flow(ctx, x, sparsity_layout_o, sparsity_lut, sparsity_reverse_lut, sparsity_block_size,
+                            n_sparse_blocks, triton_block_size)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        sparsity_layout_x, sparsity_layout_o, sparsity_lut, sparsity_reverse_lut = ctx.saved_tensors
+        x_size = ctx.x_size
+        x_stride = ctx.x_stride
+        sparsity_block_size = ctx.sparsity_block_size
+        triton_block_size = ctx.triton_block_size
+
+        n_sparse_blocks = torch.sum(sparsity_layout_x.to(torch.int)).item()
+
+        output = torch.zeros(size=(n_sparse_blocks, sparsity_block_size, sparsity_block_size),
+                             dtype=grad_output.dtype, device=grad_output.device)
+
+        x_b, x_r, x_c = grad_output.size()
+        x_b_s, x_r_s, x_c_s = stride(grad_output)
+        s_l_x_b, s_l_x_r, s_l_x_c = sparsity_layout_o.size()
+        s_l_x_b_s, s_l_x_r_s, s_l_x_c_s = stride(sparsity_layout_o)
+        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 = x_size
+        o_b_s, o_r_s, o_c_s = x_stride
+
+        if triton_block_size is None:
+            triton_block_size = get_triton_block_size(sparsity_block_size)
+
+        triton_grid = lambda meta: [x_b,
+                                    triton.cdiv(x_r, meta["TRITON_BLOCK_SIZE"]),
+                                    triton.cdiv(x_c, meta["TRITON_BLOCK_SIZE"])]
+
+        (kernel_blocksparse_flow_push[triton_grid]
+         (grad_output,
+          x_b, x_b_s, x_r_s, x_c_s,
+          s_l_x_b, s_l_x_b_s, s_l_x_r_s, s_l_x_c_s,
+          sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+          sparsity_reverse_lut,
+          output,
+          o_b, o_b_s, o_r_s, o_c_s,
+          triton_block_size))
+
+        return output, None, None, None, None, None, None, None
+
+
+@triton.jit
+def kernel_blocksparse_flow_pull(x,
+                                 x_b, x_b_s, x_r_s, x_c_s,
+                                 o,
+                                 o_b, o_b_s, o_r_s, o_c_s,
+                                 s_l_o_b, s_l_o_b_s, s_l_o_r_s, s_l_o_c_s,
+                                 s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+                                 r_lut,
+                                 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 sparsity index of current output block consisting of its batch, row, and column index
+    spa_bat_idx = (pid_blk * s_lut_r_s + 0 * s_lut_c_s)
+    spa_bat_msk = (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 < s_lut_r * s_lut_r_s)
+    spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
+
+    spa_col_idx = (pid_blk * s_lut_r_s + 2 * s_lut_c_s)
+    spa_col_msk = (spa_col_idx < s_lut_r * s_lut_r_s)
+    spa_col = tl.load(s_lut + spa_col_idx, mask=spa_col_msk)
+
+    # Get reverse sparsity index
+    rev_idx_spa_idx = (spa_bat * s_l_o_b_s +
+                       spa_row * s_l_o_r_s +
+                       spa_col * s_l_o_c_s)
+    rev_idx_spa_msk = (rev_idx_spa_idx < s_l_o_b * s_l_o_b_s)
+    rev_idx_spa = tl.load(r_lut + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
+
+    if rev_idx_spa == -1:
+        tl.device_assert(False)
+        return
+
+    blk_x_idx = (rev_idx_spa * 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 < x_b * x_b_s)
+    blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
+
+    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 < o_b * o_b_s)
+    tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)
+
+
+@triton.jit
+def kernel_blocksparse_flow_push(x,
+                                 x_b, x_b_s, x_r_s, x_c_s,
+                                 s_l_x_b, s_l_x_b_s, s_l_x_r_s, s_l_x_c_s,
+                                 s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+                                 r_lut,
+                                 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 sparsity index of current input block consisting of its batch, row, and column index
+    spa_bat_idx = (pid_blk * s_lut_r_s + 0 * s_lut_c_s)
+    spa_bat_msk = (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 < s_lut_r * s_lut_r_s)
+    spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
+
+    spa_col_idx = (pid_blk * s_lut_r_s + 2 * s_lut_c_s)
+    spa_col_msk = (spa_col_idx < s_lut_r * s_lut_r_s)
+    spa_col = tl.load(s_lut + spa_col_idx, mask=spa_col_msk)
+
+    # Get reverse sparsity index
+    rev_idx_spa_idx = (spa_bat * s_l_x_b_s +
+                       spa_row * s_l_x_r_s +
+                       spa_col * s_l_x_c_s)
+    rev_idx_spa_msk = (rev_idx_spa_idx < s_l_x_b * s_l_x_b_s)
+    rev_idx_spa = tl.load(r_lut + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
+
+    if rev_idx_spa == -1:
+        tl.device_assert(False)
+        return
+
+    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 < x_b * x_b_s)
+    blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
+
+    blk_o_idx = (rev_idx_spa * 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 < o_b * o_b_s)
+    tl.atomic_add(o + blk_o_idx, blk_x, mask=blk_o_msk)
+
+
+def forward_flow(ctx, x: Tensor, sparsity_layout_o: Tensor, sparsity_lut: Tensor, sparsity_reverse_lut: Tensor,
+                 sparsity_block_size: int, n_sparse_blocks: int, triton_block_size: int) -> Tensor:
+    output = torch.empty(size=(n_sparse_blocks, sparsity_block_size, sparsity_block_size),
+                         dtype=x.dtype, device=x.device)
+    output = torch.zeros_like(output)
+
+    x_b, x_r, x_c = x.size()
+    x_b_s, x_r_s, x_c_s = stride(x)
+    o_b, o_r, o_c = output.size()
+    o_b_s, o_r_s, o_c_s = stride(output)
+    s_l_o_b, s_l_o_r, s_l_o_c = sparsity_layout_o.size()
+    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_o)
+    s_lut_r, s_lut_c = sparsity_lut.size()
+    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
+    asdf = torch.tensor(sparsity_lut).stride()
+
+    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"])]
+
+    (kernel_blocksparse_flow_pull[triton_grid]
+     (x,
+      x_b, x_b_s, x_r_s, x_c_s,
+      output,
+      o_b, o_b_s, o_r_s, o_c_s,
+      s_l_o_b, s_l_o_b_s, s_l_o_r_s, s_l_o_c_s,
+      sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
+      sparsity_reverse_lut,
+      triton_block_size))
+
+    # Save for backward pass
+    ctx.sparsity_block_size = sparsity_block_size
+    ctx.triton_block_size = triton_block_size
+
+    return output

blksprs/ops/softmax.py

@@ -5,7 +5,7 @@ from triton import language as tl
 
 from blksprs.ops.exp import exp
 from blksprs.misc.row_wise import row_wise_sum, row_wise_max, row_wise_sub
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
     validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
 
@@ -61,9 +61,9 @@ class _BlocksparseSoftmax(torch.autograd.Function):
         output = torch.empty_like(x)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
+        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 = sparsity_lut.stride()
+        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,
@@ -76,9 +76,9 @@ class _BlocksparseSoftmax(torch.autograd.Function):
                                                                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 = x_exp_row_wise_sum.stride()
+        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 = sparsity_layout_rws.stride()
+        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)
@@ -119,13 +119,13 @@ class _BlocksparseSoftmax(torch.autograd.Function):
                                   (1 * (sparsity_layout_s_flat == 0)))
 
         o_b, o_r, o_c = o.size()
-        o_b_s, o_r_s, o_c_s = o.stride()
+        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 = sparsity_lut.stride()
+        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 = s.stride()
+        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 = sparsity_layout_s.stride()
+        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)
 
@@ -181,7 +181,8 @@ class _BlocksparseSoftmax(torch.autograd.Function):
         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 == -1:
-            assert False, "Invalid sparsity block"
+            tl.device_assert(False)
+            return
 
         # Load x block
         blk_x_idx = ((pid_blk * x_b_s) +

blksprs/ops/transpose.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_dimensions, validate_contiguous, validate_device, \
     validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
 
@@ -56,20 +56,20 @@ def transpose(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int, trit
 class _BlocksparseTranspose(torch.autograd.Function):
 
     @staticmethod
-    def forward(ctx, x: Tensor,
-                sparsity_layout: Tensor, sparsity_lut: Tensor, sparsity_reverse_lut: Tensor, sparsity_block_size: int,
-                n_sparse_blocks: int, triton_block_size: int) -> (Tensor, Tensor):
+    def forward(ctx, x: Tensor, sparsity_layout_o: Tensor, sparsity_lut: Tensor, sparsity_reverse_lut: Tensor,
+                sparsity_block_size: int,
+                n_sparse_blocks: int, triton_block_size: int) -> Tensor:
         output = torch.empty(size=(n_sparse_blocks, sparsity_block_size, sparsity_block_size),
                              dtype=x.dtype, device=x.device)
 
         x_b, x_r, x_c = x.size()
-        x_b_s, x_r_s, x_c_s = x.stride()
-        s_l_b, s_l_r, s_l_c = sparsity_layout.size()
-        s_l_b_s, s_l_r_s, s_l_c_s = sparsity_layout.stride()
+        x_b_s, x_r_s, x_c_s = stride(x)
+        s_l_b, s_l_r, s_l_c = sparsity_layout_o.size()
+        s_l_b_s, s_l_r_s, s_l_c_s = stride(sparsity_layout_o)
         s_lut_r, s_lut_c = sparsity_lut.shape
-        s_lut_r_s, s_lut_c_s = sparsity_lut.stride()
+        s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
         o_b, o_r, o_c = output.size()
-        o_b_s, o_r_s, o_c_s = output.stride()
+        o_b_s, o_r_s, o_c_s = stride(output)
 
         if triton_block_size is None:
             triton_block_size = get_triton_block_size(sparsity_block_size)
@@ -89,8 +89,7 @@ class _BlocksparseTranspose(torch.autograd.Function):
           triton_block_size))
 
         # Save for backward pass
-        ctx.save_for_backward(sparsity_layout)
-        ctx.sparsity_layout = sparsity_layout
+        ctx.save_for_backward(sparsity_layout_o)
         ctx.sparsity_block_size = sparsity_block_size
         ctx.triton_block_size = triton_block_size
 
@@ -141,7 +140,8 @@ class _BlocksparseTranspose(torch.autograd.Function):
         rev_idx_spa = tl.load(r_lut + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
 
         if rev_idx_spa == -1:
-            assert False, "Invalid sparsity block"
+            tl.device_assert(False)
+            return
 
         blk_x_idx = (rev_idx_spa * x_b_s +
                      ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +

blksprs/utils/tools.py

@@ -23,3 +23,6 @@ def get_triton_block_size(sparsity_block_size: int, limit: int = 128):
 
 def disable_validation():
     _set_skip_validation(True)
+
+def stride(x: Tensor):
+    return x.view(x.shape).stride()

blksprs/utils/validation.py

@@ -63,6 +63,8 @@ def validate_sparsity(sparsity_block_size: int, *tensor_sparsity_layout_tuples:
     for (tensor, sparsity_layout) in tensor_sparsity_layout_tuples:
         _validate_sparsity_layout_values(sparsity_layout)
 
+        if not sparsity_layout.dim() == 3:
+            raise ValueError("Sparsity layout must have exactly 3 dimensions")
         if not (tensor.size(-1) == tensor.size(-2) == sparsity_block_size):
             raise ValueError("Blocks not conforming to sparsity block size")
         if not tensor.size(0) == torch.sum(sparsity_layout.reshape(-1)):

{blksprs-1.5.dist-info → blksprs-1.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: blksprs
-Version: 1.5
+Version: 1.7
 Summary: A lightweight library for operations on blocksparse matrices in PyTorch.
 Author-email: Felix Schön <schoen@kr.tuwien.ac.at>
 Project-URL: Homepage, https://github.com/FelixSchoen/blksprs
@@ -31,9 +31,11 @@ Currently supported operations (includes gradient calculation):
 - Sparse matrix multiplication (_supports any combination of sparse and dense matrices due to support
   for `sparse = sparse @ sparse` matmul_)
 - Softmax
-- Transposition
+- Transpose
 - Gather
 - Scatter (_supports either no reduction or summation, gradients are only available for summation_)
+- Repeat (_supports target sparsity layout_)
+- Splitting and merging of matrices along the last dimension
 - Conversion to and from sparse form
 - Conversion to different sparsity layouts and different sparsity block sizes
 
@@ -63,7 +65,7 @@ We recommend installing blksprs from [PyPI](https://pypi.org/project/blksprs/) u
 
 ### Dependencies
 
-- [PyTorch](https://pytorch.org/) (built with v2.4.0)
+- [PyTorch](https://pytorch.org/) (built with v2.5.0)
 - _[Triton](https://github.com/triton-lang/triton) (included with PyTorch)_
 
 ## Changelog

blksprs-1.7.dist-info/RECORD

@@ -0,0 +1,22 @@
+blksprs/__init__.py,sha256=FpvHMo1W6XvuiA1PMDp2_EJz-Xwc15cHz7WeIYXQJC4,1019
+blksprs/experimental/distribution_mdi.py,sha256=jE_SbB0SMGPcCoeM0699cceWAikBkBpGM_8Oo5A7Ets,20083
+blksprs/layouting/distribution_layout.py,sha256=nCR3fCF6eNpi97DI6MMLF2hq_9Lwyo6_aUCIZiJfiX4,4170
+blksprs/layouting/sparsity_layout.py,sha256=JNwbJ4L-418zCiCbt-vLfZ0xU7ReP0zr_tLHs_dytRA,9686
+blksprs/misc/broadcast_ops.py,sha256=0RLnLMYV7GAPI2YL8RotcxjIUSBZKGxdVcsGaJFeL_I,5327
+blksprs/misc/repeat_interleave.py,sha256=P5gfsZXuemLiAijUZfFkBFgMjlU9rlPEzai1xeGOFnw,5678
+blksprs/misc/row_wise.py,sha256=0vDJA8uCocmebSIPIbFeND5_PQIE10pUj3DBOQXlTvE,16888
+blksprs/ops/conversion.py,sha256=iyKIlkWGrK6q55KNRM8N6rY1k4b9k8QUkUl158yZUDA,21330
+blksprs/ops/distribution.py,sha256=V3TK5SlNT_JdGHNaDNl-U4U5vwAYsgkAOg4eTmYxbuA,16877
+blksprs/ops/exp.py,sha256=cdF0s93Q9iucIXuEE3howsB0N6D60xgvem7C-a-yiGI,3704
+blksprs/ops/matmul.py,sha256=uqVe6Dz2aaCbCglM1uS2eRHVKh7PQcuecaIBWFubPEw,11256
+blksprs/ops/partitioning.py,sha256=nBRZzfi3XYAhDLEBzYflQkvGa3MIZ-qNeIlrZ16k44g,7533
+blksprs/ops/repeat.py,sha256=6Wa6GG9Cx6rJXuFpvmOe5hHwYd3l9UYMosKEDsbh9XI,10408
+blksprs/ops/softmax.py,sha256=2dMLbkHNH18jSJmkgOJvZOKwWHhuUogAVCWv2Bwc3oQ,11995
+blksprs/ops/transpose.py,sha256=67pDdCEb7r-Xifupl82fBKAYsxKcCUDy--cPPfduRvU,6761
+blksprs/utils/benchmarking.py,sha256=4pLVlnPW_2EM-NT3n4SClaRznVYEljztLbJcccz8kZE,1360
+blksprs/utils/tools.py,sha256=S3836Zuc-BMigv-5mLTjRznCzuaF6oYW-Ir9zzUnr3o,655
+blksprs/utils/validation.py,sha256=h2oki3xC5qLWZR4-W5QIna-wVSXvRehQEH-ynrOciVE,3467
+blksprs-1.7.dist-info/METADATA,sha256=raZ3ycSMUEAW71bwm-807d_dse44qKdSkWMhH4GI2Qg,7709
+blksprs-1.7.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+blksprs-1.7.dist-info/top_level.txt,sha256=qyp0IHeY3H2GQA97i4hk_To5rRBS2YcE1HRPSLy04fk,8
+blksprs-1.7.dist-info/RECORD,,

blksprs-1.5.dist-info/RECORD

@@ -1,20 +0,0 @@
-blksprs/__init__.py,sha256=OY9ofdbzBGsvY6hx0oLCrSszlJFdMns9x7gKE0asFI0,919
-blksprs/experimental/distribution_mdi.py,sha256=shu-3Nt7nkaLIb4O2kSajC8Lh7IWFXO9rsjzP14ASYA,20088
-blksprs/layouting/distribution_layout.py,sha256=Zv-b2t5VOvW6-ejdX42kUV7X1yYsvDCY_PXFE_wKwi0,4165
-blksprs/layouting/sparsity_layout.py,sha256=vZL8r5LkMwILYYqTYPZcN_NYFJuVFIB6mmBkdtRyXmI,7893
-blksprs/misc/broadcast_ops.py,sha256=ahm7_lI12bJ6VTKRuSkwEeaEYWRY-BeMIOhtei35zpQ,5323
-blksprs/misc/repeat_interleave.py,sha256=KJeapmxbpA7zGFfa5hUhCGrk4aFmhOhlMw-hbTh9PLI,5668
-blksprs/misc/row_wise.py,sha256=1UtjLplrGx1FkxhzQ2hjSBBY11ToLQs0JiLaXKRAkL4,16893
-blksprs/ops/conversion.py,sha256=vuiNwrwyuGI6H4PKrS_UHI7OKWJwNZd2i3LSjf6RetU,21332
-blksprs/ops/distribution.py,sha256=KhtHRVcv4_woyNlldAjIWF-7021-KX-xyIcN6rE-UgE,16879
-blksprs/ops/exp.py,sha256=CVWVq_emO2CnS_xk6Unx67P7EI7IL26dwtsmBJZOLzQ,3698
-blksprs/ops/matmul.py,sha256=743XeD5M4iUv28sYf7q6mVXDd4jZpV04JAx8bF7hWkw,11254
-blksprs/ops/softmax.py,sha256=cs1utM6UCzHhdJpf-ZysBr6CwbjI-5aQG0ahYY37Zy0,11991
-blksprs/ops/transpose.py,sha256=Ru4YKyg796WT6OnDSTCYG45tMmdgvju3hMFzkwsJnO8,6801
-blksprs/utils/benchmarking.py,sha256=4pLVlnPW_2EM-NT3n4SClaRznVYEljztLbJcccz8kZE,1360
-blksprs/utils/tools.py,sha256=JAuwsLISr_hcvxIgUVvKz5ZPf9M5ycquplsBU5dVfDc,596
-blksprs/utils/validation.py,sha256=rP6yr-C2ghXfJEERry_pfvVJ0g0VyqV4sL4HkBRlJg8,3345
-blksprs-1.5.dist-info/METADATA,sha256=dql0_6s1Vfdnx6sLFusayZWSeU9uxvfAjBDdLPk43so,7607
-blksprs-1.5.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
-blksprs-1.5.dist-info/top_level.txt,sha256=qyp0IHeY3H2GQA97i4hk_To5rRBS2YcE1HRPSLy04fk,8
-blksprs-1.5.dist-info/RECORD,,