blksprs 1.6.1__py3-none-any.whl → 1.8__py3-none-any.whl

blksprs/ops/repeat.py

@@ -0,0 +1,322 @@
+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):
+    """Repeats a block-spare tensor in compressed form according to the given repeats.
+    
+    Repeats is a 3-tuple of integers, where each integer represents the number of times the tensor should be repeated in
+        the first, second and third dimension respectively.
+        
+    Note:
+        An output sparsity layout can be provided, in which case only the indicated blocks are filled. This may result
+        in blocks not being present in the output that were present in the input if the output sparsity layout indicates
+        them to be sparse.
+    
+    Args:
+        x (Tensor): A block-sparse tensor in compressed form.
+        sparsity_layout_x (Tensor): The sparsity layout of the block-sparse tensor.
+        repeats (tuple[int, int, int]): The number of times the tensor should be repeated in the first, second and
+            third dimension respectively.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        sparsity_layout_output (Tensor): The desired sparsity layout of the output tensor (default ``None``).
+        triton_block_size (int): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: A block-sparse tensor in compressed form containing the repeated values.
+        Tensor: The sparsity layout of the resulting output 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
+
+
+def repeat_interleave(x: Tensor, sparsity_layout_x: Tensor, repeats: int,
+                      sparsity_block_size: int, sparsity_layout_output: Tensor = None,
+                      triton_block_size: int = None) -> (
+        Tensor, Tensor):
+    """Repeats and interleaves the block-sparse tensor in compressed form.
+
+    Repeats each matrix contained in the tensors by ``repeats`` amount and places them consecutively in the output
+        tensor.
+
+    Note:
+        In similar fashion to the regular ``repeat`` an output sparsity layout can be provided. In this case only
+        non-sparse blocks will be filled.
+
+    Args:
+        x (Tensor): A block-sparse tensor in compressed form.
+        sparsity_layout_x (Tensor): The sparsity layout of the block-sparse tensor.
+        repeats (int): The number of times to repeat the matrices.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        sparsity_layout_output (Tensor): The desired sparsity layout of the output tensor (default ``None``).
+        triton_block_size (int): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: A block-sparse tensor in compressed form containing the repeated and interleaved matrices.
+        Tensor: The sparsity layout of the resulting output 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 = torch.repeat_interleave(sparsity_layout_x, repeats, dim=0).contiguous()
+
+    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_interleave(repeats, dim=0)
+                            .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)
+
+    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

@@ -3,9 +3,9 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.ops.exp import exp
+from blksprs.misc.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
 
@@ -63,13 +63,13 @@ class _BlocksparseTranspose(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_b, s_l_r, s_l_c = sparsity_layout_o.size()
-        s_l_b_s, s_l_r_s, s_l_c_s = sparsity_layout_o.stride()
+        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)
@@ -140,7 +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

@@ -3,6 +3,7 @@ from torch import Tensor
 
 VALIDATION = True
 
+
 def validate_dimensions(*tensors: Tensor, dims=3) -> None:
     if _check_skip_validation():
         return
@@ -71,10 +72,25 @@ def validate_sparsity(sparsity_block_size: int, *tensor_sparsity_layout_tuples:
             raise ValueError("Mismatch between sparsity layout and blocks")
 
 
+def validate_sparsity_dense(sparsity_block_size: int, *tensor_sparsity_layout_tuples: tuple[Tensor, Tensor]) -> None:
+    if _check_skip_validation():
+        return
+
+    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) // sparsity_block_size == sparsity_layout.size(-1) and
+                tensor.size(-2) // sparsity_block_size == sparsity_layout.size(-2)):
+            raise ValueError("Tensor not conforming to sparsity layout")
+
+
 def _validate_sparsity_layout_values(sparsity_layout: Tensor):
     if not torch.all(torch.logical_or(sparsity_layout == 0, sparsity_layout == 1)):
         raise ValueError("Sparsity layout values must be either 0 or 1")
 
+
 def validate_sparsity_block_size(sparsity_block_size: int, *tensors):
     if _check_skip_validation():
         return
@@ -86,6 +102,7 @@ def validate_sparsity_block_size(sparsity_block_size: int, *tensors):
         if not (tensor.size(-1) % sparsity_block_size == 0 and tensor.size(-2) % sparsity_block_size == 0):
             raise ValueError("Tensor sizes must be divisible by sparsity block size")
 
+
 def validate_triton_block_size(triton_block_size: int, sparsity_block_size: int):
     if _check_skip_validation():
         return
@@ -99,9 +116,11 @@ def validate_triton_block_size(triton_block_size: int, sparsity_block_size: int)
     if triton_block_size > sparsity_block_size:
         raise ValueError("Triton block size cannot be larger than sparsity block size")
 
+
 def _check_skip_validation():
     return not VALIDATION
 
+
 def _set_skip_validation(skip_validation: bool):
     global VALIDATION
-    VALIDATION = not skip_validation
+    VALIDATION = not skip_validation

{blksprs-1.6.1.dist-info → blksprs-1.8.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: blksprs
-Version: 1.6.1
+Version: 1.8
 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
@@ -28,12 +28,13 @@ A lightweight and efficient library for operations on block-sparse matrices in P
 
 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_)
+- Matrix multiplication
 - Softmax
 - Transpose
 - Gather
 - Scatter (_supports either no reduction or summation, gradients are only available for summation_)
+- Repeat (_supports target sparsity layout_)
+- Repeat Interleave (_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
@@ -50,8 +51,14 @@ These include, e.g.,
 Note that in order to correctly apply element-wise operations between two sparse tensors their sparsity layouts have to
 match.
 
+Further helpful operations (included in the ``bs.misc`` module) that do **not** support gradient calculation include:
+
+- Row-wise sum, max, addition, and subtraction
+- Broadcast addition and subtraction between slices
+
 Furthermore, the library provides a set of utility functions for the creation of sparsity layouts based on existing
-dense tensors.
+dense tensors and for the scatter operation (module ``bs.layout``), as well as utility functions to ensure correct input
+dimensionality (module ``bs.util``).
 
 ## Installation
 
@@ -64,7 +71,7 @@ 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-1.8.dist-info/RECORD

@@ -0,0 +1,21 @@
+blksprs/__init__.py,sha256=qDqoB-X5vo5_3PlrN54sp59XR5hg6EanIsADS67QnH0,1058
+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/exp.py,sha256=cdF0s93Q9iucIXuEE3howsB0N6D60xgvem7C-a-yiGI,3704
+blksprs/misc/partitioning.py,sha256=nBRZzfi3XYAhDLEBzYflQkvGa3MIZ-qNeIlrZ16k44g,7533
+blksprs/misc/row_wise.py,sha256=0vDJA8uCocmebSIPIbFeND5_PQIE10pUj3DBOQXlTvE,16888
+blksprs/ops/conversion.py,sha256=9xVdCrj38m1cMh43LQs-GrXZ5pNRjhQyKx6paaw3C6A,21898
+blksprs/ops/distribution.py,sha256=V3TK5SlNT_JdGHNaDNl-U4U5vwAYsgkAOg4eTmYxbuA,16877
+blksprs/ops/matmul.py,sha256=uqVe6Dz2aaCbCglM1uS2eRHVKh7PQcuecaIBWFubPEw,11256
+blksprs/ops/repeat.py,sha256=OSsa2rj6BHL3Kedfu3wr0D82mn4HmbJ1l7XEmT-6ehg,14423
+blksprs/ops/softmax.py,sha256=5nAgeT68nucgOugjtCy1aBIMa7Kyk1KNN-j8fgmeVuk,11996
+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=WzihRPibXYzss3PMkhDt5_d3Q3NHA_d1TzTz3CoGPGg,4136
+blksprs-1.8.dist-info/METADATA,sha256=koey4w8ynY84Z0dM5u9y_P831rtR0w-Z-dBcje4O6ko,8007
+blksprs-1.8.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+blksprs-1.8.dist-info/top_level.txt,sha256=qyp0IHeY3H2GQA97i4hk_To5rRBS2YcE1HRPSLy04fk,8
+blksprs-1.8.dist-info/RECORD,,