blksprs 1.10.2__py3-none-any.whl → 1.11__py3-none-any.whl

blksprs/__init__.py

@@ -13,7 +13,6 @@ class ops:
     class misc:
         from blksprs.ops.misc.row_wise import row_wise_sum, row_wise_max, row_wise_add, row_wise_sub
         from blksprs.ops.misc.broadcast_ops import broadcast_add, broadcast_sub
-        from blksprs.ops.misc.exp import exp
 
 
 class layouting:

blksprs/ops/conversion.py

@@ -19,7 +19,7 @@ def from_blksprs(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size:
 
 
 def to_dense(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int, fill_value: float = 0,
-             triton_block_size: int = None) -> Tensor:
+             triton_block_size: int = None, lut: dict = None) -> Tensor:
     """Converts a block-sparse tensor in compressed form to a block-sparse tensor in regular form based on the given
         sparsity layout.
 
@@ -30,6 +30,7 @@ def to_dense(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int
         fill_value (float): The value to fill the resulting dense tensor with where the block-sparse tensor is not
             present (default ``0``).
         triton_block_size (int): The block size to use for the triton kernel (default ``None``).
+        lut (dict, optional): A dictionary containing the look-up tables for the operation (default ``None``).
 
     Returns:
         Tensor: The block-sparse tensor converted to regular form.
@@ -44,24 +45,35 @@ def to_dense(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int
     validate_sparsity_block_size(sparsity_block_size, x)
     validate_triton_block_size(triton_block_size, sparsity_block_size)
 
-    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)))
-
-    validate_contiguous(sparsity_reverse_lut)
+    lut = _BlocksparseToDense.build_lut(lut, sparsity_layout)
 
     if sparsity_layout.size(1) == 1 and sparsity_layout.size(2) == 1 and torch.all(sparsity_layout):
         return x
 
     return _BlocksparseToDense.apply(x,
-                                     sparsity_layout, sparsity_reverse_lut,
+                                     sparsity_layout, lut["sparsity_reverse_lut"],
                                      sparsity_block_size, fill_value,
                                      triton_block_size)
 
 
 class _BlocksparseToDense(torch.autograd.Function):
 
+    @staticmethod
+    def build_lut(lut: dict, sparsity_layout: Tensor):
+        if lut is None:
+            lut = dict()
+
+        if "sparsity_reverse_lut" not in lut:
+            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)))
+            lut["sparsity_reverse_lut"] = sparsity_reverse_lut
+
+        validate_contiguous(lut["sparsity_reverse_lut"])
+
+        return lut
+
     @staticmethod
     def forward(ctx, x: Tensor,
                 sparsity_layout: Tensor, sparsity_reverse_lut: Tensor,
@@ -160,7 +172,7 @@ def to_blksprs(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int,
 
 
 def to_sparse(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int,
-              triton_block_size: int = None) -> BlksprsTensor:
+              triton_block_size: int = None, lut: dict = None) -> BlksprsTensor:
     """Converts a block-sparse tensor in regular form to a block-sparse tensor in compressed form based on the given
     sparsity layout.
 
@@ -169,6 +181,7 @@ def to_sparse(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int,
         sparsity_layout (Tensor): The sparsity layout of the block-sparse tensor.
         sparsity_block_size (int): The size of the sparsity blocks.
         triton_block_size (int): The block size to use for the triton kernel (default ``None``).
+        lut (dict, optional): A dictionary containing the look-up tables for the operation (default ``None``).
 
     Returns:
         BlksprsTensor: The block-sparse tensor converted to compressed form.
@@ -183,22 +196,36 @@ def to_sparse(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int,
     validate_sparsity_block_size(sparsity_block_size, x)
     validate_triton_block_size(triton_block_size, sparsity_block_size)
 
-    sparsity_lut = torch.nonzero(sparsity_layout).contiguous()
-    n_sparse_blocks = torch.sum(sparsity_layout.to(torch.int)).item()
-
-    validate_contiguous(sparsity_layout, sparsity_lut)
+    lut = _BlocksparseToSparse.build_lut(lut, sparsity_layout)
 
     if sparsity_layout.size(1) == 1 and sparsity_layout.size(2) == 1 and torch.all(sparsity_layout):
         return BlksprsTensor(x)
 
     return BlksprsTensor(_BlocksparseToSparse.apply(x,
-                                                    sparsity_layout, sparsity_lut,
-                                                    sparsity_block_size, n_sparse_blocks,
+                                                    sparsity_layout, lut["sparsity_lut"],
+                                                    sparsity_block_size, lut["n_sparse_blocks"],
                                                     triton_block_size))
 
 
 class _BlocksparseToSparse(torch.autograd.Function):
 
+    @staticmethod
+    def build_lut(lut: dict, sparsity_layout: Tensor):
+        if lut is None:
+            lut = dict()
+
+        if "sparsity_lut" not in lut:
+            sparsity_lut = torch.nonzero(sparsity_layout).contiguous()
+            lut["sparsity_lut"] = sparsity_lut
+
+        if "n_sparse_blocks" not in lut:
+            n_sparse_blocks = torch.sum(sparsity_layout.to(torch.int)).item()
+            lut["n_sparse_blocks"] = n_sparse_blocks
+
+        validate_contiguous(sparsity_layout, lut["sparsity_lut"])
+
+        return lut
+
     @staticmethod
     def forward(ctx, x: Tensor,
                 sparsity_layout: Tensor, sparsity_lut: Tensor,

blksprs/ops/distribution.py

@@ -13,7 +13,7 @@ from blksprs.utils.validation import validate_contiguous, validate_dimensions, v
 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:
+           sparsity_block_size: int, triton_block_size: int = None, lut: dict = None) -> BlksprsTensor:
     """Applies a gather operation on a block-sparse tensor in compressed form.
 
     Args:
@@ -24,6 +24,7 @@ def gather(src: BlksprsTensor, sparsity_layout_src: Tensor,
         sparsity_layout_idx (Tensor): The sparsity layout of the indices block-sparse tensor.
         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``).
+        lut (dict, optional): A dictionary containing the look-up tables for the operation (default ``None``).
 
     Returns:
         BlksprsTensor: The result of the gather operation as a block-sparse tensor in compressed form.
@@ -40,25 +41,38 @@ def gather(src: BlksprsTensor, sparsity_layout_src: Tensor,
     validate_sparsity_block_size(sparsity_block_size, src, idx)
     validate_triton_block_size(triton_block_size, sparsity_block_size)
 
-    sparsity_layout_x_flat = sparsity_layout_src.reshape(-1)
-    sparsity_reverse_lut_x = ((torch.cumsum(sparsity_layout_x_flat, dim=-1) - 1) *
-                              (sparsity_layout_x_flat == 1) -
-                              (1 * (sparsity_layout_x_flat == 0)))
-
-    sparsity_lut_i = torch.nonzero(sparsity_layout_idx).contiguous()
-
-    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,
-                                                  adjusted_dim, idx, sparsity_layout_idx, sparsity_lut_i,
+    lut = _BlocksparseGather.build_lut(lut, sparsity_layout_src, sparsity_layout_idx)
+
+    return BlksprsTensor(_BlocksparseGather.apply(src, sparsity_layout_src, lut["sparsity_reverse_lut_x"],
+                                                  adjusted_dim, idx, sparsity_layout_idx, lut["sparsity_lut_i"],
                                                   sparsity_block_size, triton_block_size))
 
 
 class _BlocksparseGather(torch.autograd.Function):
 
+    @staticmethod
+    def build_lut(lut: dict, sparsity_layout_src: Tensor, sparsity_layout_idx: Tensor):
+        if lut is None:
+            lut = dict()
+
+        if "sparsity_reverse_lut_x" not in lut:
+            sparsity_layout_x_flat = sparsity_layout_src.reshape(-1)
+            sparsity_reverse_lut_x = ((torch.cumsum(sparsity_layout_x_flat, dim=-1) - 1) *
+                                      (sparsity_layout_x_flat == 1) -
+                                      (1 * (sparsity_layout_x_flat == 0)))
+            lut["sparsity_reverse_lut_x"] = sparsity_reverse_lut_x
+
+        if "sparsity_lut_i" not in lut:
+            sparsity_lut_i = torch.nonzero(sparsity_layout_idx).contiguous()
+            lut["sparsity_lut_i"] = sparsity_lut_i
+
+        validate_contiguous(sparsity_layout_src, lut["sparsity_reverse_lut_x"],
+                            sparsity_layout_idx, lut["sparsity_lut_i"])
+
+        return lut
+
     @staticmethod
     def forward(ctx, x: Tensor, sparsity_layout_x: Tensor, sparsity_reverse_lut_x: Tensor,
                 dim: int, i: Tensor, sparsity_layout_i: Tensor, sparsity_lut_i: Tensor,
@@ -202,7 +216,7 @@ 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:
+            sparsity_block_size: int, triton_block_size: int = None, lut: dict = None) -> BlksprsTensor:
     """Wrapper for ``scatter_reduce`` with ``reduce_op="none"``.
 
     """
@@ -219,7 +233,7 @@ def scatter_reduce(src: BlksprsTensor, sparsity_layout_src: Tensor,
                    idx: BlksprsTensor,
                    sparsity_layout_tgt: Tensor,
                    sparsity_block_size: int,
-                   reduce_op: str = "sum", triton_block_size: int = None) -> BlksprsTensor:
+                   reduce_op: str = "sum", triton_block_size: int = None, lut: dict = None) -> BlksprsTensor:
     """Applies a scatter operation on a block-sparse tensor in compressed form.
 
     Args:
@@ -232,6 +246,7 @@ def scatter_reduce(src: BlksprsTensor, sparsity_layout_src: Tensor,
         reduce_op (str, optional): The reduction operation to apply during the scatter operation (default ``"sum"``).
             Supported operations are ``"none"`` and ``"sum"``.
         triton_block_size (int, optional): The block size to use for the triton kernel (default ``None``).
+        lut (dict, optional): A dictionary containing the look-up tables for the operation (default ``None``).
 
     Returns:
         BlksprsTensor: The result of the scatter operation as a block-sparse tensor in compressed form.
@@ -251,29 +266,44 @@ def scatter_reduce(src: BlksprsTensor, sparsity_layout_src: Tensor,
     if reduce_op not in ["none", "sum"]:
         raise ValueError(f"Reduction operation '{reduce_op}' is not supported")
 
-    sparsity_lut_x = torch.nonzero(sparsity_layout_src).contiguous()
-
-    sparsity_layout_o_flat = sparsity_layout_tgt.reshape(-1)
-    sparsity_reverse_lut_o = ((torch.cumsum(sparsity_layout_o_flat, dim=-1) - 1) *
-                              (sparsity_layout_o_flat == 1) -
-                              (1 * (sparsity_layout_o_flat == 0)))
-
-    n_sparse_blocks = torch.sum(sparsity_layout_tgt.to(torch.int)).item()
-
-    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,
+    lut = _BlocksparseScatterReduce.build_lut(lut, sparsity_layout_src, sparsity_layout_tgt)
+
+    return BlksprsTensor(_BlocksparseScatterReduce.apply(src, sparsity_layout_src, lut["sparsity_lut_x"],
                                                          adjusted_dim, idx,
-                                                         sparsity_layout_tgt, sparsity_reverse_lut_o,
-                                                         sparsity_block_size, n_sparse_blocks,
+                                                         sparsity_layout_tgt, lut["sparsity_reverse_lut_o"],
+                                                         sparsity_block_size, lut["n_sparse_blocks"],
                                                          reduce_op, triton_block_size))
 
 
 class _BlocksparseScatterReduce(torch.autograd.Function):
 
+    @staticmethod
+    def build_lut(lut: dict, sparsity_layout_src: Tensor, sparsity_layout_tgt: Tensor):
+        if lut is None:
+            lut = dict()
+
+        if "sparsity_lut_x" not in lut:
+            sparsity_lut_x = torch.nonzero(sparsity_layout_src).contiguous()
+            lut["sparsity_lut_x"] = sparsity_lut_x
+
+        if "sparsity_reverse_lut_o" not in lut:
+            sparsity_layout_o_flat = sparsity_layout_tgt.reshape(-1)
+            sparsity_reverse_lut_o = ((torch.cumsum(sparsity_layout_o_flat, dim=-1) - 1) *
+                                      (sparsity_layout_o_flat == 1) -
+                                      (1 * (sparsity_layout_o_flat == 0)))
+            lut["sparsity_reverse_lut_o"] = sparsity_reverse_lut_o
+
+        if "n_sparse_blocks" not in lut:
+            n_sparse_blocks = torch.sum(sparsity_layout_tgt.to(torch.int)).item()
+            lut["n_sparse_blocks"] = n_sparse_blocks
+
+        validate_contiguous(sparsity_layout_src, lut["sparsity_lut_x"],
+                            sparsity_layout_tgt, lut["sparsity_reverse_lut_o"])
+
+        return lut
+
     @staticmethod
     def forward(ctx, x: Tensor, sparsity_layout_x: Tensor, sparsity_lut_x: Tensor,
                 dim: int, i: Tensor,

blksprs/ops/flow.py

@@ -40,21 +40,18 @@ def kernel_blocksparse_flow_pull(x,
     rev_idx_spa_msk = (rev_idx_spa_idx >= 0 and 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
+    if rev_idx_spa >= 0:
+        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 >= 0 and blk_x_idx < x_b * x_b_s)
+        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
 
-    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 >= 0 and 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 >= 0 and blk_o_idx < o_b * o_b_s)
-    tl.store(o + blk_o_idx, blk_x, mask=blk_o_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 >= 0 and blk_o_idx < o_b * o_b_s)
+        tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)
 
 
 @triton.jit
@@ -91,25 +88,22 @@ def kernel_blocksparse_flow_push(x,
     rev_idx_spa_msk = (rev_idx_spa_idx >= 0 and 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 >= 0 and blk_x_idx < x_b * x_b_s)
-    blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
+    if rev_idx_spa >= 0:
+        blk_x_idx = (pid_blk * x_b_s +
+                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
+                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
+        blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
+        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
 
-    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 >= 0 and blk_o_idx < o_b * o_b_s)
-    tl.atomic_add(o + blk_o_idx, blk_x, mask=blk_o_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 >= 0 and blk_o_idx < o_b * o_b_s)
+        tl.atomic_add(o + blk_o_idx, blk_x, mask=blk_o_msk)
 
 
-def flow_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:
+def flow_forward_pull(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)
 
@@ -144,3 +138,42 @@ def flow_forward(ctx, x: Tensor, sparsity_layout_o: Tensor, sparsity_lut: Tensor
     ctx.triton_block_size = triton_block_size
 
     return output
+
+
+def flow_forward_push(ctx, x: Tensor, sparsity_layout_x: Tensor, sparsity_lut: Tensor, sparsity_reverse_lut: Tensor,
+                      sparsity_block_size: int, n_sparse_blocks: int, triton_block_size: int) -> Tensor:
+    output = torch.zeros(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 = 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 = stride(sparsity_layout_x)
+    s_lut_r, s_lut_c = sparsity_lut.size()
+    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
+    o_b, o_r, o_c = output.size()
+    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)
+
+    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]
+     (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,
+      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))
+
+    # Save for backward pass
+    if ctx is not None:
+        ctx.sparsity_block_size = sparsity_block_size
+        ctx.triton_block_size = triton_block_size
+
+    return output

blksprs/ops/matmul.py

@@ -13,7 +13,7 @@ from blksprs.utils.validation import validate_contiguous, validate_dimensions, v
 def matmul(x: BlksprsTensor, sparsity_layout_x: Tensor,
            y: BlksprsTensor, sparsity_layout_y: Tensor,
            sparsity_layout_output: Tensor,
-           sparsity_block_size: int, triton_block_size: int = None) -> BlksprsTensor:
+           sparsity_block_size: int, triton_block_size: int = None, lut: dict = None) -> BlksprsTensor:
     """Performs matrix multiplication between two block-sparse tensors.
 
     The sparsity layout of the output tensor is used to only calculate blocks that will be present in the output.
@@ -26,6 +26,7 @@ def matmul(x: BlksprsTensor, sparsity_layout_x: Tensor,
         sparsity_layout_output (Tensor): The sparsity layout of the output tensor.
         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``).
+        lut (dict, optional): A dictionary containing the look-up tables for the operation (default ``None``).
 
     Returns:
         BlksprsTensor: The result of the matrix multiplication as a block-sparse tensor in compressed form.
@@ -44,35 +45,52 @@ def matmul(x: BlksprsTensor, sparsity_layout_x: Tensor,
     validate_sparsity_block_size(sparsity_block_size, x, y)
     validate_triton_block_size(triton_block_size, sparsity_block_size)
 
-    sparsity_layout_x_flat = sparsity_layout_x.reshape(-1)
-    sparsity_reverse_lut_x = ((torch.cumsum(sparsity_layout_x_flat, dim=-1) - 1) *
-                              (sparsity_layout_x_flat == 1) -
-                              (1 * (sparsity_layout_x_flat == 0)))
-
-    sparsity_layout_y_flat = sparsity_layout_y.reshape(-1)
-    sparsity_reverse_lut_y = ((torch.cumsum(sparsity_layout_y_flat, dim=-1) - 1) *
-                              (sparsity_layout_y_flat == 1) -
-                              (1 * (sparsity_layout_y_flat == 0)))
-
-    sparsity_lut_o = torch.nonzero(sparsity_layout_output).contiguous()
-
-    n_sparse_blocks = torch.sum(sparsity_layout_output.to(torch.int)).item()
-
-    validate_contiguous(sparsity_layout_x, sparsity_reverse_lut_x,
-                        sparsity_layout_y, sparsity_reverse_lut_y,
-                        sparsity_layout_output, sparsity_lut_o)
+    lut = _BlocksparseMatmulSSS.build_lut(lut, sparsity_layout_x, sparsity_layout_y, sparsity_layout_output)
 
     return BlksprsTensor(_BlocksparseMatmulSSS.apply(x, y,
-                                                     sparsity_layout_x, sparsity_reverse_lut_x,
-                                                     sparsity_layout_y, sparsity_reverse_lut_y,
-                                                     sparsity_layout_output, sparsity_lut_o,
+                                                     sparsity_layout_x, lut["sparsity_reverse_lut_x"],
+                                                     sparsity_layout_y, lut["sparsity_reverse_lut_y"],
+                                                     sparsity_layout_output, lut["sparsity_lut_o"],
                                                      sparsity_block_size,
-                                                     n_sparse_blocks,
+                                                     lut["n_sparse_blocks"],
                                                      triton_block_size))
 
 
 class _BlocksparseMatmulSSS(torch.autograd.Function):
 
+    @staticmethod
+    def build_lut(lut: dict, sparsity_layout_x: Tensor, sparsity_layout_y: Tensor, sparsity_layout_output: Tensor):
+        if lut is None:
+            lut = dict()
+
+        if "sparsity_reverse_lut_x" not in lut:
+            sparsity_layout_x_flat = sparsity_layout_x.reshape(-1)
+            sparsity_reverse_lut_x = ((torch.cumsum(sparsity_layout_x_flat, dim=-1) - 1) *
+                                      (sparsity_layout_x_flat == 1) -
+                                      (1 * (sparsity_layout_x_flat == 0)))
+            lut["sparsity_reverse_lut_x"] = sparsity_reverse_lut_x
+
+        if "sparsity_reverse_lut_y" not in lut:
+            sparsity_layout_y_flat = sparsity_layout_y.reshape(-1)
+            sparsity_reverse_lut_y = ((torch.cumsum(sparsity_layout_y_flat, dim=-1) - 1) *
+                                      (sparsity_layout_y_flat == 1) -
+                                      (1 * (sparsity_layout_y_flat == 0)))
+            lut["sparsity_reverse_lut_y"] = sparsity_reverse_lut_y
+
+        if "sparsity_lut_o" not in lut:
+            sparsity_lut_o = torch.nonzero(sparsity_layout_output).contiguous()
+            lut["sparsity_lut_o"] = sparsity_lut_o
+
+        if "n_sparse_blocks" not in lut:
+            n_sparse_blocks = torch.sum(sparsity_layout_output.to(torch.int)).item()
+            lut["n_sparse_blocks"] = n_sparse_blocks
+
+        validate_contiguous(sparsity_layout_x, lut["sparsity_reverse_lut_x"],
+                            sparsity_layout_y, lut["sparsity_reverse_lut_y"],
+                            sparsity_layout_output, lut["sparsity_lut_o"])
+
+        return lut
+
     @staticmethod
     def forward(ctx, x: Tensor, y: Tensor,
                 sparsity_layout_x: Tensor, sparsity_reverse_lut_x: Tensor,