blksprs 1.8.3__py3-none-any.whl → 1.9__py3-none-any.whl

blksprs/layouting/distribution_layout.py

@@ -10,13 +10,14 @@ from blksprs.utils.validation import validate_triton_block_size, validate_dimens
 
 
 def build_distribution_layout(indices: BlksprsTensor, sparsity_layout_indices: Tensor,
-                              size_target: torch.Size,
+                              dim: int, size_target: torch.Size,
                               sparsity_block_size: int, triton_block_size: int = None) -> Tensor:
     """Builds the sparsity layout of either the source of a gather or the target of a scatter operation.
 
     Args:
         indices (BlksprsTensor): The block-sparse indices tensor in compressed form used for the gather or scatter operation.
         sparsity_layout_indices (Tensor): The sparsity layout of the indices block-sparse tensor.
+        dim (int): The dimension along which the operation is conducted.
         size_target (torch.Size): The size of the block-sparse target tensor in regular form.
         sparsity_block_size (int): The size of the sparsity blocks.
         triton_block_size (int, optional): The block size to use for the triton kernel (default ``None``).
@@ -31,6 +32,8 @@ def build_distribution_layout(indices: BlksprsTensor, sparsity_layout_indices: T
 
     sparsity_lut_i = torch.nonzero(sparsity_layout_indices).contiguous()
 
+    adjusted_dim = dim % 3
+
     output = torch.zeros(size_target[0], size_target[1] // sparsity_block_size, size_target[2] // sparsity_block_size,
                          dtype=torch.bool, device=indices.device)
 
@@ -55,6 +58,7 @@ def build_distribution_layout(indices: BlksprsTensor, sparsity_layout_indices: T
       i_b, i_b_s, i_r_s, i_c_s,
       sparsity_lut_i,
       s_lut_i_r, s_lut_i_r_s, s_lut_i_c_s,
+      adjusted_dim,
       output,
       o_b, o_b_s, o_r_s, o_c_s,
       sparsity_block_size,
@@ -68,6 +72,7 @@ def kernel_distribution_layout(i,
                                i_b, i_b_s, i_r_s, i_c_s,
                                s_lut_i,
                                s_lut_i_r, s_lut_i_r_s, s_lut_i_c_s,
+                               dim,
                                o,
                                o_b, o_b_s, o_r_s, o_c_s,
                                sparsity_block_size,
@@ -86,17 +91,30 @@ def kernel_distribution_layout(i,
     spa_row_i_msk = (spa_row_i_idx < s_lut_i_r * s_lut_i_r_s)
     spa_row_i = tl.load(s_lut_i + spa_row_i_idx, mask=spa_row_i_msk)
 
+    spa_col_i_idx = (pid_blk * s_lut_i_r_s + 2 * s_lut_i_c_s)
+    spa_col_i_msk = (spa_col_i_idx < s_lut_i_r * s_lut_i_r_s)
+    spa_col_i = tl.load(s_lut_i + spa_col_i_idx, mask=spa_col_i_msk)
+
     blk_i_idx = (pid_blk * i_b_s +
                  ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * i_r_s)[:, None] +
                  ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * i_c_s)[None, :])
     blk_i_msk = (blk_i_idx < i_b * i_b_s)
     blk_i = tl.load(i + blk_i_idx, mask=blk_i_msk)
 
-    blk_i = blk_i // sparsity_block_size
+    dst_bat_idx = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_bat_i, dtype=tl.int32)
+    dst_row_idx = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_row_i, dtype=tl.int32)
+    dst_col_idx = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_col_i, dtype=tl.int32)
+    if dim == 0:
+        dst_bat_idx = blk_i
+    elif dim == 1:
+        dst_row_idx = blk_i // sparsity_block_size
+    elif dim == 2:
+        dst_col_idx = blk_i // sparsity_block_size
+
     blk_v = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), 1, dtype=tl.int32)
 
-    blk_o_idx = ((spa_bat_i * o_b_s) +
-                 (spa_row_i * o_r_s) +
-                 (blk_i * o_c_s))
+    blk_o_idx = ((dst_bat_idx * o_b_s) +
+                 (dst_row_idx * o_r_s) +
+                 (dst_col_idx * o_c_s))
     blk_o_msk = (blk_o_idx < o_b * o_b_s)
     tl.store(o + blk_o_idx, blk_v, mask=blk_o_msk)

blksprs/ops/distribution.py

@@ -3,19 +3,23 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
+from blksprs.ops.conversion import to_dense
 from blksprs.utils.blksprs_tensor import BlksprsTensor
 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
 
 
-def gather(src: BlksprsTensor, sparsity_layout_src: Tensor, idx: BlksprsTensor, sparsity_layout_idx: Tensor,
+def gather(src: BlksprsTensor, sparsity_layout_src: Tensor,
+           dim: int,
+           idx: BlksprsTensor, sparsity_layout_idx: Tensor,
            sparsity_block_size: int, triton_block_size: int = None) -> BlksprsTensor:
     """Applies a gather operation on a block-sparse tensor in compressed form.
 
     Args:
         src (BlksprsTensor): The source block-sparse tensor in compressed form to gather from.
         sparsity_layout_src (Tensor): The sparsity layout of the source block-sparse tensor.
+        dim (int): The dimension along which to gather.
         idx (BlksprsTensor): The block-sparse indices tensor in compressed form specifying how to gather from the source tensor.
         sparsity_layout_idx (Tensor): The sparsity layout of the indices block-sparse tensor.
         sparsity_block_size (int): The size of the sparsity blocks.
@@ -46,16 +50,18 @@ def gather(src: BlksprsTensor, sparsity_layout_src: Tensor, idx: BlksprsTensor,
     validate_contiguous(sparsity_layout_src, sparsity_reverse_lut_x,
                         sparsity_layout_idx, sparsity_lut_i)
 
+    adjusted_dim = dim % 3
+
     return BlksprsTensor(_BlocksparseGather.apply(src, sparsity_layout_src, sparsity_reverse_lut_x,
-                                    idx, sparsity_layout_idx, sparsity_lut_i,
-                                    sparsity_block_size, triton_block_size))
+                                                  adjusted_dim, idx, sparsity_layout_idx, sparsity_lut_i,
+                                                  sparsity_block_size, triton_block_size))
 
 
 class _BlocksparseGather(torch.autograd.Function):
 
     @staticmethod
     def forward(ctx, x: Tensor, sparsity_layout_x: Tensor, sparsity_reverse_lut_x: Tensor,
-                i: Tensor, sparsity_layout_i: Tensor, sparsity_lut_i: Tensor,
+                dim: int, i: Tensor, sparsity_layout_i: Tensor, sparsity_lut_i: Tensor,
                 sparsity_block_size: int, triton_block_size: int = None) -> Tensor:
         output = torch.empty_like(i, dtype=x.dtype)
 
@@ -82,6 +88,7 @@ class _BlocksparseGather(torch.autograd.Function):
           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_reverse_lut_x,
+          dim,
           i,
           i_b, i_b_s, i_r_s, i_c_s,
           output,
@@ -91,6 +98,7 @@ class _BlocksparseGather(torch.autograd.Function):
           triton_block_size))
 
         ctx.save_for_backward(sparsity_layout_x, i, sparsity_layout_i)
+        ctx.dim = dim
         ctx.sparsity_block_size = sparsity_block_size
         ctx.triton_block_size = triton_block_size
 
@@ -99,15 +107,15 @@ class _BlocksparseGather(torch.autograd.Function):
     @staticmethod
     def backward(ctx, grad_output):
         sparsity_layout_x, i, sparsity_layout_i = ctx.saved_tensors
+        dim = ctx.dim
         sparsity_block_size = ctx.sparsity_block_size
         triton_block_size = ctx.triton_block_size
 
         return scatter_reduce(grad_output, sparsity_layout_i,
-                              i,
-                              sparsity_layout_x,
-                              sparsity_block_size,
+                              dim, i,
+                              sparsity_layout_x, sparsity_block_size,
                               reduce_op="sum",
-                              triton_block_size=triton_block_size), None, None, None, None, None, None, None
+                              triton_block_size=triton_block_size), None, None, None, None, None, None, None, None
 
     @staticmethod
     @triton.jit
@@ -115,6 +123,7 @@ class _BlocksparseGather(torch.autograd.Function):
                                   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,
                                   r_lut_x,
+                                  dim,
                                   i,
                                   i_b, i_b_s, i_r_s, i_c_s,
                                   o,
@@ -136,6 +145,10 @@ class _BlocksparseGather(torch.autograd.Function):
         spa_row_o_msk = (spa_row_o_idx < s_lut_o_r * s_lut_o_r_s)
         spa_row_o = tl.load(s_lut_o + spa_row_o_idx, mask=spa_row_o_msk)
 
+        spa_col_o_idx = (pid_blk * s_lut_o_r_s + 2 * s_lut_o_c_s)
+        spa_col_o_msk = (spa_col_o_idx < s_lut_o_r * s_lut_o_r_s)
+        spa_col_o = tl.load(s_lut_o + spa_col_o_idx, mask=spa_col_o_msk)
+
         # Load index values
         blk_i_idx = ((pid_blk * i_b_s) +
                      ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * i_r_s)[:, None] +
@@ -143,33 +156,50 @@ class _BlocksparseGather(torch.autograd.Function):
         blk_i_msk = (blk_i_idx < i_b * i_b_s)
         blk_i = tl.load(i + blk_i_idx, mask=blk_i_msk).to(tl.int32)
 
-        # Get positions of sparsity blocks
+        # Get indices of sparsity blocks and positions within the blocks
         pos_spa_blk_x = blk_i // sparsity_block_size
-        pos_spa_col_x = blk_i % sparsity_block_size
+        pos_spa_int_x = blk_i % sparsity_block_size
+
+        rev_dst_bat_x = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_bat_o, dtype=tl.int32)
+        rev_dst_row_x = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_row_o, dtype=tl.int32)
+        rev_dst_col_x = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_col_o, dtype=tl.int32)
+        dst_row_x = (((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None]
+                     .broadcast_to((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE)))
+        dst_col_x = (((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :]
+                     .broadcast_to((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE)))
+        if dim == 0:
+            rev_dst_bat_x = blk_i
+        elif dim == 1:
+            rev_dst_row_x = pos_spa_blk_x
+            dst_row_x = pos_spa_int_x * x_r_s
+        elif dim == 2:
+            rev_dst_col_x = pos_spa_blk_x
+            dst_col_x = pos_spa_int_x * x_c_s
 
         # Load reverse sparsity indices for x
-        rev_idx_spa_x_idx = ((spa_bat_o * s_l_x_b_s) +
-                             (spa_row_o * s_l_x_r_s) +
-                             (pos_spa_blk_x * s_l_x_c_s))
+        rev_idx_spa_x_idx = ((rev_dst_bat_x * s_l_x_b_s) +
+                             (rev_dst_row_x * s_l_x_r_s) +
+                             (rev_dst_col_x * s_l_x_c_s))
         rev_idx_spa_x_msk = (rev_idx_spa_x_idx < s_l_x_b * s_l_x_b_s)
         rev_idx_spa_x = tl.load(r_lut_x + rev_idx_spa_x_idx, mask=rev_idx_spa_x_msk).to(tl.int32)
 
         # Load x values
         blk_x_idx = ((rev_idx_spa_x * x_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                     (pos_spa_col_x * x_c_s))
-        blk_x_msk = (blk_x_idx < x_b * x_b_s)
+                     dst_row_x +
+                     dst_col_x)
+        blk_x_msk = ((blk_x_idx < x_b * x_b_s) & rev_idx_spa_x_msk != -1)
         blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
 
         # Store output
         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)
+        blk_o_msk = ((blk_o_idx < o_b * o_b_s) &  rev_idx_spa_x_msk != -1)
         tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)
 
 
 def scatter(src: BlksprsTensor, sparsity_layout_src: Tensor,
+            dim: int,
             idx: BlksprsTensor,
             sparsity_layout_tgt: Tensor,
             sparsity_block_size: int, triton_block_size: int = None) -> BlksprsTensor:
@@ -184,6 +214,7 @@ def scatter(src: BlksprsTensor, sparsity_layout_src: Tensor,
 
 
 def scatter_reduce(src: BlksprsTensor, sparsity_layout_src: Tensor,
+                   dim: int,
                    idx: BlksprsTensor,
                    sparsity_layout_tgt: Tensor,
                    sparsity_block_size: int,
@@ -193,6 +224,7 @@ def scatter_reduce(src: BlksprsTensor, sparsity_layout_src: Tensor,
     Args:
         src (BlksprsTensor): The source block-sparse tensor in compressed form to scatter from.
         sparsity_layout_src (Tensor): The sparsity layout of the source block-sparse tensor.
+        dim (int): The dimension along which to scatter.
         idx (BlksprsTensor): The block-sparse indices tensor in compressed form specifying how to scatter to the target tensor.
         sparsity_layout_tgt (Tensor): The sparsity layout of the target block-sparse tensor.
         sparsity_block_size (int): The size of the sparsity blocks.
@@ -230,18 +262,20 @@ def scatter_reduce(src: BlksprsTensor, sparsity_layout_src: Tensor,
     validate_contiguous(sparsity_layout_src, sparsity_lut_x,
                         sparsity_layout_tgt, sparsity_reverse_lut_o)
 
+    adjusted_dim = dim % 3
+
     return BlksprsTensor(_BlocksparseScatterReduce.apply(src, sparsity_layout_src, sparsity_lut_x,
-                                           idx,
-                                           sparsity_layout_tgt, sparsity_reverse_lut_o,
-                                           sparsity_block_size, n_sparse_blocks,
-                                           reduce_op, triton_block_size))
+                                                         adjusted_dim, idx,
+                                                         sparsity_layout_tgt, sparsity_reverse_lut_o,
+                                                         sparsity_block_size, n_sparse_blocks,
+                                                         reduce_op, triton_block_size))
 
 
 class _BlocksparseScatterReduce(torch.autograd.Function):
 
     @staticmethod
     def forward(ctx, x: Tensor, sparsity_layout_x: Tensor, sparsity_lut_x: Tensor,
-                i: Tensor,
+                dim: int, i: Tensor,
                 sparsity_layout_o: Tensor, sparsity_reverse_lut_o: Tensor,
                 sparsity_block_size: int, n_sparse_blocks: int,
                 reduce_op: str, triton_block_size: int) -> Tensor:
@@ -274,10 +308,11 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
          (x,
           x_b, x_b_s, x_r_s, x_c_s,
           sparsity_lut_x, s_lut_x_r, s_lut_x_r_s, s_lut_x_c_s,
+          dim,
           i,
           i_b, i_b_s, i_r_s, i_c_s,
           output,
-          o_b, o_b_s, o_r_s, o_c_s,
+          o_b, o_b_s,
           s_l_o_b, s_l_o_b_s, s_l_o_r_s, s_l_o_c_s,
           sparsity_reverse_lut_o,
           reduce_op_ind,
@@ -285,6 +320,7 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
           triton_block_size))
 
         ctx.save_for_backward(sparsity_layout_x, i, sparsity_layout_o)
+        ctx.dim = dim
         ctx.sparsity_block_size = sparsity_block_size
         ctx.reduce_op = reduce_op
         ctx.triton_block_size = triton_block_size
@@ -294,13 +330,14 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
     @staticmethod
     def backward(ctx, grad_output):
         sparsity_layout_x, i, sparsity_layout_o = ctx.saved_tensors
+        dim = ctx.dim
         sparsity_block_size = ctx.sparsity_block_size
         reduce_op = ctx.reduce_op
         triton_block_size = ctx.triton_block_size
 
         if reduce_op == "sum":
-            return gather(grad_output, sparsity_layout_o, i, sparsity_layout_x, sparsity_block_size,
-                          triton_block_size=triton_block_size), None, None, None, None, None, None, None, None, None
+            return gather(grad_output, sparsity_layout_o, dim, i, sparsity_layout_x, sparsity_block_size,
+                          triton_block_size=triton_block_size), None, None, None, None, None, None, None, None, None, None
         else:
             raise ValueError(f"Reduction operation '{reduce_op}' does not support backward pass")
 
@@ -309,10 +346,11 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
     def kernel_blocksparse_scatter(x,
                                    x_b, x_b_s, x_r_s, x_c_s,
                                    s_lut_x, s_lut_x_r, s_lut_x_r_s, s_lut_x_c_s,
+                                   dim,
                                    i,
                                    i_b, i_b_s, i_r_s, i_c_s,
                                    o,
-                                   o_b, o_b_s, o_r_s, o_c_s,
+                                   o_b, o_b_s,
                                    s_l_o_b, s_l_o_b_s, s_l_o_r_s, s_l_o_c_s,
                                    r_lut_o,
                                    reduce_op_ind,
@@ -332,6 +370,10 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
         spa_row_x_msk = (spa_row_x_idx < s_lut_x_r * s_lut_x_r_s)
         spa_row_x = tl.load(s_lut_x + spa_row_x_idx, mask=spa_row_x_msk)
 
+        spa_col_x_idx = (pid_blk * s_lut_x_r_s + 2 * s_lut_x_c_s)
+        spa_col_x_msk = (spa_col_x_idx < s_lut_x_r * s_lut_x_r_s)
+        spa_col_x = tl.load(s_lut_x + spa_col_x_idx, mask=spa_col_x_msk)
+
         # Load x values
         blk_x_idx = ((pid_blk * x_b_s) +
                      ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
@@ -346,22 +388,38 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
         blk_i_msk = (blk_i_idx < i_b * i_b_s)
         blk_i = tl.load(i + blk_i_idx, mask=blk_i_msk).to(tl.int32)
 
-        # Get positions of sparsity blocks
-        pos_spa_blk_o = blk_i // sparsity_block_size
-        pos_spa_col_o = blk_i % sparsity_block_size
+        # Get indices of sparsity blocks and positions within the blocks
+        pos_spa_blk_x = blk_i // sparsity_block_size
+        pos_spa_int_x = blk_i % sparsity_block_size
+
+        rev_dst_bat_o = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_bat_x, dtype=tl.int32)
+        rev_dst_row_o = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_row_x, dtype=tl.int32)
+        rev_dst_col_o = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_col_x, dtype=tl.int32)
+        dst_row_o = (((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None]
+                     .broadcast_to((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE)))
+        dst_col_o = (((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :]
+                     .broadcast_to((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE)))
+        if dim == 0:
+            rev_dst_bat_o = blk_i
+        elif dim == 1:
+            rev_dst_row_o = pos_spa_blk_x
+            dst_row_o = pos_spa_int_x * x_r_s
+        elif dim == 2:
+            rev_dst_col_o = pos_spa_blk_x
+            dst_col_o = pos_spa_int_x * x_c_s
 
         # Load reverse sparsity indices for o
-        rev_idx_spa_o_idx = ((spa_bat_x * s_l_o_b_s) +
-                             (spa_row_x * s_l_o_r_s) +
-                             (pos_spa_blk_o * s_l_o_c_s))
+        rev_idx_spa_o_idx = ((rev_dst_bat_o * s_l_o_b_s) +
+                             (rev_dst_row_o * s_l_o_r_s) +
+                             (rev_dst_col_o * s_l_o_c_s))
         rev_idx_spa_o_msk = (rev_idx_spa_o_idx < s_l_o_b * s_l_o_b_s)
         rev_idx_spa_o = tl.load(r_lut_o + rev_idx_spa_o_idx, mask=rev_idx_spa_o_msk).to(tl.int32)
 
         # Store output
         blk_o_idx = ((rev_idx_spa_o * o_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +
-                     (pos_spa_col_o * o_c_s))
-        blk_o_msk = (blk_o_idx < o_b * o_b_s)
+                     dst_row_o +
+                     dst_col_o)
+        blk_o_msk = ((blk_o_idx < o_b * o_b_s) & rev_idx_spa_o_msk != -1)
 
         if reduce_op_ind == 0:
             tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)

blksprs/ops/experimental/distribution_mdi.py

@@ -153,6 +153,10 @@ class _BlocksparseGatherMDI(torch.autograd.Function):
         rev_idx_spa_x_msk = (rev_idx_spa_x_idx < s_l_x_b * s_l_x_b_s)
         rev_idx_spa_x = tl.load(r_lut_x + rev_idx_spa_x_idx, mask=rev_idx_spa_x_msk).to(tl.int32)
 
+        if rev_idx_spa_x == -1:
+            tl.device_assert(False)
+            return
+
         # Load x values
         blk_x_idx = ((rev_idx_spa_x * x_b_s) +
                      ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
@@ -342,6 +346,10 @@ class _BlocksparseScatterReduceMDI(torch.autograd.Function):
         rev_idx_spa_o_msk = (rev_idx_spa_o_idx < s_l_o_b * s_l_o_b_s)
         rev_idx_spa_o = tl.load(r_lut_o + rev_idx_spa_o_idx, mask=rev_idx_spa_o_msk).to(tl.int32)
 
+        if rev_idx_spa_o == -1:
+            tl.device_assert(False)
+            return
+
         # Store output
         blk_o_idx = ((rev_idx_spa_o * o_b_s) +
                      ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +

blksprs/ops/misc/row_wise.py

@@ -117,6 +117,10 @@ def kernel_blocksparse_row_wise_sum(x,
     rev_idx_spa_msk = (rev_idx_spa_idx < s_l_o_b * s_l_o_b_s)
     rev_idx_spa = tl.load(r_lut_o + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
 
+    if rev_idx_spa == -1:
+        tl.device_assert(False)
+        return
+
     blk_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, :])
@@ -240,6 +244,10 @@ def kernel_blocksparse_row_wise_max(x,
     rev_idx_spa_msk = (rev_idx_spa_idx < s_l_o_b * s_l_o_b_s)
     rev_idx_spa = tl.load(r_lut_o + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
 
+    if rev_idx_spa == -1:
+        tl.device_assert(False)
+        return
+
     blk_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, :])

blksprs/ops/softmax.py

@@ -238,6 +238,10 @@ class _BlocksparseSoftmax(torch.autograd.Function):
         rev_idx_spa_s_msk = (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 == -1:
+            tl.device_assert(False)
+            return
+
         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, :])

{blksprs-1.8.3.dist-info → blksprs-1.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: blksprs
-Version: 1.8.3
+Version: 1.9
 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

{blksprs-1.8.3.dist-info → blksprs-1.9.dist-info}/RECORD

@@ -1,23 +1,23 @@
 blksprs/__init__.py,sha256=YMrERuEf1hTv5vVdOvPEzh9rESn4uqOB7WHB12Qs5lU,1836
-blksprs/layouting/distribution_layout.py,sha256=wmj1SwWyY_fhbvMmh6AXrR77LoSp6xLwUWCCyO9i5lk,4239
+blksprs/layouting/distribution_layout.py,sha256=9f_Bx2YQF4LTH95C0S7OuB9eeOuh73NcE0Z7Wrtug38,5034
 blksprs/layouting/sparsity_layout.py,sha256=-sScIn4hhG35j9BXytrojEzp8jnFkMargJjtivPV1fc,9755
 blksprs/ops/conversion.py,sha256=ol-iV45wDzp9G1dJEkY53EdrvnmHzcl7QQmPJ-xqQTs,22410
-blksprs/ops/distribution.py,sha256=fXZV6UegCVpIwzh-A825OSYClHWu5k0UMYdO2UGDUpM,17067
+blksprs/ops/distribution.py,sha256=OWTH_dfO43uIMY6S44wpvRoIBuKzaTy1f57BOEf7EYA,19925
 blksprs/ops/matmul.py,sha256=yh2ZnO0ZltT1AgadiFP0vX28YJ4n74xO-I_5vFUmOmA,11452
 blksprs/ops/partitioning.py,sha256=K0ExR2a3W62d_9xxCJzsdJDLgtbxTI6P8loOOBdhPzE,7674
 blksprs/ops/repeat.py,sha256=IvSIRbuyFn0b57LObymLgup0LqlWQ3ndIw-QuiYQcaU,14564
-blksprs/ops/softmax.py,sha256=CDQT2KnwkJ4hGIgT0EUp6P92uiYpCdJQ9zxcdgSAAJA,12102
+blksprs/ops/softmax.py,sha256=V-1vqRefjjwSp6JPwKxVxh5pTng9gOdtgGlXHDPbpYM,12190
 blksprs/ops/transpose.py,sha256=jxzFFffrj4S_9tiCrwwUMdz6EA98o1dziWXjlqb64a4,6859
-blksprs/ops/experimental/distribution_mdi.py,sha256=HaRUu6LTWATzjuHWgddIUE-0fgY-O87STpJO4JY7k_8,20357
+blksprs/ops/experimental/distribution_mdi.py,sha256=F_0tl4Gn-9JZs_TZfDtZqO_RPFl7sejqQNF8UNIoCbs,20533
 blksprs/ops/misc/broadcast_ops.py,sha256=cPtRJa3pkZfY1QG51CJ-zDn4SK-CRpX5LEXoKGGMvRU,5418
 blksprs/ops/misc/exp.py,sha256=FnSFosBfJHuiEbD0MD-i4axLghRn4a0f8KvHXrKBB6M,3802
-blksprs/ops/misc/row_wise.py,sha256=SvJuNww-_QoVKTyTjMvjmzHlBuUlTKamkuq_rKzwAqs,17081
+blksprs/ops/misc/row_wise.py,sha256=U4Kk0-P4oOuMNjMHXxP2gP9njMIeMfz8RZrzItNIF94,17229
 blksprs/utils/benchmarking.py,sha256=4pLVlnPW_2EM-NT3n4SClaRznVYEljztLbJcccz8kZE,1360
 blksprs/utils/blksprs_tensor.py,sha256=VjplBgDhnf9sxf-1R5feA0xp5FDCDdaeZmCeoIRdCnc,151
 blksprs/utils/processing.py,sha256=hYsFxEbQKcbqU4WtZWusPnWMHg8ZAZF1SKZJYjez9aU,2060
 blksprs/utils/tools.py,sha256=r7Y4C37vfSWUyQTGwa8NyRqgovmsq9hMufkenqYHOxo,539
 blksprs/utils/validation.py,sha256=IZxH2HZpePmv7lRqLsSwV_6FwsdnTXv9q4j98vCMSsQ,4195
-blksprs-1.8.3.dist-info/METADATA,sha256=DZkJ_HeetF1V6-_F6GeG0uXT-QmttMFOq4ao8fiSMgQ,8458
-blksprs-1.8.3.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-blksprs-1.8.3.dist-info/top_level.txt,sha256=qyp0IHeY3H2GQA97i4hk_To5rRBS2YcE1HRPSLy04fk,8
-blksprs-1.8.3.dist-info/RECORD,,
+blksprs-1.9.dist-info/METADATA,sha256=9mMjmvJ2_Rz0uyiY9S8SKTRcs6YW5Jk1w6PRobh6Q3c,8456
+blksprs-1.9.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+blksprs-1.9.dist-info/top_level.txt,sha256=qyp0IHeY3H2GQA97i4hk_To5rRBS2YcE1HRPSLy04fk,8
+blksprs-1.9.dist-info/RECORD,,