blksprs 1.10.1__py3-none-any.whl → 1.11__py3-none-any.whl

blksprs/ops/softmax.py

@@ -3,7 +3,6 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.ops.misc.exp import exp
 from blksprs.ops.misc.row_wise import row_wise_sum, row_wise_max, row_wise_sub
 from blksprs.utils.blksprs_tensor import BlksprsTensor
 from blksprs.utils.tools import get_triton_block_size, stride
@@ -12,7 +11,7 @@ from blksprs.utils.validation import validate_contiguous, validate_dimensions, v
 
 
 def softmax(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int,
-            triton_block_size: int = None) -> BlksprsTensor:
+            triton_block_size: int = None, lut: dict = None) -> BlksprsTensor:
     """Computes the softmax of a block-sparse tensor in compressed form.
 
     Note:
@@ -23,6 +22,7 @@ def softmax(x: BlksprsTensor, 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 result of the softmax operation as a block-sparse tensor in compressed form.
@@ -37,24 +37,38 @@ def softmax(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_lut = torch.nonzero(sparsity_layout).contiguous()
-
-    sparsity_layout_rws, _ = torch.max(sparsity_layout, dim=-1, keepdim=True)
-    sparsity_layout_rws_flat = sparsity_layout_rws.reshape(-1)
-    sparsity_reverse_lut_rws = ((torch.cumsum(sparsity_layout_rws_flat, dim=-1) - 1) *
-                                (sparsity_layout_rws_flat == 1) -
-                                (1 * (sparsity_layout_rws_flat == 0)))
-
-    validate_contiguous(sparsity_layout, sparsity_lut, sparsity_reverse_lut_rws)
+    lut = _BlocksparseSoftmax.build_lut(lut, sparsity_layout)
 
     return BlksprsTensor(_BlocksparseSoftmax.apply(x, sparsity_layout,
-                                                   sparsity_lut,
-                                                   sparsity_reverse_lut_rws,
+                                                   lut["sparsity_lut"],
+                                                   lut["sparsity_reverse_lut_rws"],
                                                    sparsity_block_size, triton_block_size))
 
 
 class _BlocksparseSoftmax(torch.autograd.Function):
 
+    @staticmethod
+    def build_lut(lut: dict, sparsity_layout: Tensor):
+        if lut is None:
+            lut = dict()
+
+        if "sparsity_lut" not in lut:
+            sparsity_lut = torch.nonzero(sparsity_layout).contiguous()
+            lut["sparsity_lut"] = sparsity_lut
+
+
+        if "sparsity_reverse_lut_rws" not in lut:
+            sparsity_layout_rws, _ = torch.max(sparsity_layout, dim=-1, keepdim=True)
+            sparsity_layout_rws_flat = sparsity_layout_rws.reshape(-1)
+            sparsity_reverse_lut_rws = ((torch.cumsum(sparsity_layout_rws_flat, dim=-1) - 1) *
+                                        (sparsity_layout_rws_flat == 1) -
+                                        (1 * (sparsity_layout_rws_flat == 0)))
+            lut["sparsity_reverse_lut_rws"] = sparsity_reverse_lut_rws
+
+        validate_contiguous(sparsity_layout, lut["sparsity_lut"], lut["sparsity_reverse_lut_rws"])
+
+        return lut
+
     @staticmethod
     def forward(ctx, x: Tensor, sparsity_layout: Tensor,
                 sparsity_lut: Tensor,
@@ -72,7 +86,7 @@ class _BlocksparseSoftmax(torch.autograd.Function):
                                                            flag_slice_only=True,
                                                            triton_block_size=triton_block_size)
         x_scaled = row_wise_sub(x, sparsity_layout, x_row_wise_max, sparsity_block_size, triton_block_size)
-        x_exp = exp(x_scaled, sparsity_block_size, triton_block_size=triton_block_size)
+        x_exp = torch.exp(x_scaled)
         x_exp_row_wise_sum, sparsity_layout_rws = row_wise_sum(x_exp, sparsity_layout, sparsity_block_size,
                                                                flag_slice_only=True,
                                                                triton_block_size=triton_block_size)
@@ -182,29 +196,26 @@ class _BlocksparseSoftmax(torch.autograd.Function):
         rev_idx_spa_s_msk = (rev_idx_spa_s_idx >= 0 and rev_idx_spa_s_idx < s_l_s_b * s_l_s_b_s)
         rev_idx_spa_s = tl.load(r_lut_s + rev_idx_spa_s_idx, mask=rev_idx_spa_s_msk).to(tl.int32)
 
-        if rev_idx_spa_s == -1:
-            tl.device_assert(False)
-            return
-
-        # Load x block
-        blk_x_idx = ((pid_blk * x_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-        blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
-        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
+        if rev_idx_spa_s >= 0:
+            # Load x block
+            blk_x_idx = ((pid_blk * x_b_s) +
+                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
+                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
+            blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
+            blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
 
-        # Load sum block
-        blk_s_idx = (rev_idx_spa_s * s_b_s +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * s_r_s)[:, None] +
-                     (tl.arange(0, 1) * s_c_s)[None, :])
-        blk_s_msk = (blk_s_idx >= 0 and blk_s_idx < s_b * s_b_s)
-        blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
+            # Load sum block
+            blk_s_idx = (rev_idx_spa_s * s_b_s +
+                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * s_r_s)[:, None] +
+                         (tl.arange(0, 1) * s_c_s)[None, :])
+            blk_s_msk = (blk_s_idx >= 0 and blk_s_idx < s_b * s_b_s)
+            blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
 
-        # Compute softmax
-        buf = tl.div_rn(blk_x, blk_s)
+            # Compute softmax
+            buf = tl.div_rn(blk_x, blk_s)
 
-        # Store output
-        tl.store(o + blk_x_idx, buf, mask=blk_x_msk)
+            # Store output
+            tl.store(o + blk_x_idx, buf, mask=blk_x_msk)
 
     @staticmethod
     @triton.jit
@@ -239,32 +250,29 @@ class _BlocksparseSoftmax(torch.autograd.Function):
         rev_idx_spa_s_msk = (rev_idx_spa_s_idx >= 0 and rev_idx_spa_s_idx < s_l_s_b * s_l_s_b_s)
         rev_idx_spa_s = tl.load(r_lut_s + rev_idx_spa_s_idx, mask=rev_idx_spa_s_msk).to(tl.int32)
 
-        if rev_idx_spa_s == -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, :])
-        blk_s_msk = (blk_s_idx >= 0 and blk_s_idx < s_b * s_b_s)
-        blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
-
-        blk_g_idx = ((pid_blk * g_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * g_r_s)[:, None] +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * g_c_s)[None, :])
-        blk_g_msk = (blk_g_idx >= 0 and blk_g_idx < g_b * g_b_s)
-        blk_g = tl.load(g + blk_g_idx, mask=blk_g_msk)
-
-        blk_x_idx = ((pid_blk * x_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-        blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
-        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
-
-        buf = blk_x * (blk_g - blk_s)
-
-        blk_o_idx = ((pid_blk * o_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_c_s)[None, :])
-        blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s)
-        tl.store(o + blk_o_idx, buf, mask=blk_o_msk)
+        if rev_idx_spa_s >= 0:
+            blk_s_idx = (rev_idx_spa_s * s_b_s +
+                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * s_r_s)[:, None] +
+                         (tl.arange(0, 1) * s_c_s)[None, :])
+            blk_s_msk = (blk_s_idx >= 0 and blk_s_idx < s_b * s_b_s)
+            blk_s = tl.load(s + blk_s_idx, mask=blk_s_msk)
+
+            blk_g_idx = ((pid_blk * g_b_s) +
+                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * g_r_s)[:, None] +
+                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * g_c_s)[None, :])
+            blk_g_msk = (blk_g_idx >= 0 and blk_g_idx < g_b * g_b_s)
+            blk_g = tl.load(g + blk_g_idx, mask=blk_g_msk)
+
+            blk_x_idx = ((pid_blk * x_b_s) +
+                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
+                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
+            blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
+            blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
+
+            buf = blk_x * (blk_g - blk_s)
+
+            blk_o_idx = ((pid_blk * o_b_s) +
+                         ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +
+                         ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_c_s)[None, :])
+            blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s)
+            tl.store(o + blk_o_idx, buf, mask=blk_o_msk)

blksprs/ops/transpose.py

@@ -3,14 +3,15 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
+from blksprs.ops.flow import flow_forward_pull
 from blksprs.utils.blksprs_tensor import BlksprsTensor
 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 transpose(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int, triton_block_size: int = None) -> (
-        BlksprsTensor, Tensor):
+def transpose(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: int, triton_block_size: int = None,
+              lut: dict = None) -> (BlksprsTensor, Tensor):
     """Transposes a block-sparse tensor in compressed form.
 
     Note:
@@ -21,6 +22,7 @@ def transpose(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: in
         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 transposed block-sparse tensor in compressed form.
@@ -28,6 +30,7 @@ def transpose(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: in
 
     """
     x = x.contiguous()
+    x_t = x.transpose(-1, -2).contiguous()
 
     validate_dimensions(x)
     validate_contiguous(x)
@@ -36,66 +39,53 @@ def transpose(x: BlksprsTensor, sparsity_layout: Tensor, sparsity_block_size: in
     validate_sparsity_block_size(sparsity_block_size, x)
     validate_triton_block_size(triton_block_size, sparsity_block_size)
 
-    sparsity_layout_t = sparsity_layout.transpose(-1, -2).contiguous()
+    lut = _BlocksparseTranspose.build_lut(lut, sparsity_layout)
 
-    sparsity_lut = torch.nonzero(sparsity_layout_t).contiguous()
+    return BlksprsTensor(
+        _BlocksparseTranspose.apply(x_t, lut["sparsity_layout_t"], lut["sparsity_lut"], lut["sparsity_reverse_lut"],
+                                    sparsity_block_size,
+                                    lut["n_sparse_blocks"], triton_block_size)), lut["sparsity_layout_t"]
 
-    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()).transpose(-1, -2).contiguous().reshape(-1))
 
-    n_sparse_blocks = torch.sum(sparsity_layout.to(torch.int)).item()
+class _BlocksparseTranspose(torch.autograd.Function):
 
-    validate_contiguous(sparsity_layout_t, sparsity_lut, sparsity_reverse_lut)
+    @staticmethod
+    def build_lut(lut: dict, sparsity_layout: Tensor):
+        if lut is None:
+            lut = dict()
 
-    return BlksprsTensor(
-        _BlocksparseTranspose.apply(x, sparsity_layout_t, sparsity_lut, sparsity_reverse_lut, sparsity_block_size,
-                                    n_sparse_blocks, triton_block_size)), sparsity_layout_t
+        if "sparsity_layout_t" not in lut:
+            sparsity_layout_t = sparsity_layout.transpose(-1, -2).contiguous()
+            lut["sparsity_layout_t"] = sparsity_layout_t
 
+        if "sparsity_lut" not in lut:
+            sparsity_lut = torch.nonzero(lut["sparsity_layout_t"]).contiguous()
+            lut["sparsity_lut"] = sparsity_lut
 
-class _BlocksparseTranspose(torch.autograd.Function):
+        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)))
+                                    .reshape(sparsity_layout.size()).transpose(-1, -2).contiguous().reshape(-1))
+            lut["sparsity_reverse_lut"] = sparsity_reverse_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(lut["sparsity_layout_t"], lut["sparsity_lut"], lut["sparsity_reverse_lut"])
+
+        return lut
 
     @staticmethod
     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 = 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 = 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: [o_b,
-                                    triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
-                                    triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
-
-        (_BlocksparseTranspose.kernel_blocksparse_transpose[triton_grid]
-         (x,
-          x_b, x_b_s, x_r_s, x_c_s,
-          s_l_b, s_l_b_s, s_l_r_s, s_l_c_s,
-          sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-          sparsity_reverse_lut,
-          output,
-          o_b, o_b_s,
-          triton_block_size))
-
-        # Save for backward pass
         ctx.save_for_backward(sparsity_layout_o)
-        ctx.sparsity_block_size = sparsity_block_size
-        ctx.triton_block_size = triton_block_size
 
-        return output
+        return flow_forward_pull(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):
@@ -105,56 +95,3 @@ class _BlocksparseTranspose(torch.autograd.Function):
 
         return transpose(grad_output, sparsity_layout, sparsity_block_size, triton_block_size)[
             0], None, None, None, None, None, None
-
-    @staticmethod
-    @triton.jit
-    def kernel_blocksparse_transpose(x,
-                                     x_b, x_b_s, x_r_s, x_c_s,
-                                     s_l_b, s_l_b_s, s_l_r_s, s_l_c_s,
-                                     s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-                                     r_lut,
-                                     o,
-                                     o_b, o_b_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 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 >= 0 and spa_bat_idx < s_lut_r * s_lut_r_s)
-        spa_bat = tl.load(s_lut + spa_bat_idx, mask=spa_bat_msk)
-
-        spa_row_idx = (pid_blk * s_lut_r_s + 1 * s_lut_c_s)
-        spa_row_msk = (spa_row_idx >= 0 and spa_row_idx < s_lut_r * s_lut_r_s)
-        spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
-
-        spa_col_idx = (pid_blk * s_lut_r_s + 2 * s_lut_c_s)
-        spa_col_msk = (spa_col_idx >= 0 and 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_b_s +
-                           spa_row * s_l_r_s +
-                           spa_col * s_l_c_s)
-        rev_idx_spa_msk = (rev_idx_spa_idx >= 0 and rev_idx_spa_idx < s_l_b * s_l_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 >= 0 and blk_x_idx < x_b * x_b_s)
-        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
-
-        blk_x_t = tl.trans(blk_x)
-
-        blk_o_idx = (pid_blk * o_b_s +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_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_t, mask=blk_o_msk)

blksprs/utils/tools.py

@@ -1,3 +1,4 @@
+import torch
 from torch import Tensor, Size
 
 
@@ -20,4 +21,9 @@ def get_triton_block_size(sparsity_block_size: int, limit: int = 128):
 
 
 def stride(x: Tensor):
-    return x.view(x.shape).stride()
+    if x.dim() == 2:
+        return x.size(1), 1
+    elif x.dim() == 3:
+        return x.size(1) * x.size(2), x.size(2), 1
+    else:
+        raise NotImplementedError

{blksprs-1.10.1.dist-info → blksprs-1.11.dist-info}/METADATA

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: blksprs
-Version: 1.10.1
+Version: 1.11
 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-1.11.dist-info/RECORD

@@ -0,0 +1,23 @@
+blksprs/__init__.py,sha256=AJYVfR40nOfE5F3waHPVSuajwYDcoGkiEQc8HhQbUBU,1721
+blksprs/layouting/distribution_layout.py,sha256=xDGY5-J7uSD8oenlf8bEJ2amMiQG3NBf2klTTydbTJE,5140
+blksprs/layouting/sparsity_layout.py,sha256=IVtHc_nN3ZM2y4GFcys70PqDWmWc7tkHlVGlToErANk,9894
+blksprs/ops/conversion.py,sha256=QFtZ-nmY2JAWutheiO07vatXqz3eSZBP5Ym_U2Q1oWk,23299
+blksprs/ops/distribution.py,sha256=nHTuE7Tq0Q404VN8bWNC2sEwmmdAtgZI6I7auRICdps,21749
+blksprs/ops/flow.py,sha256=7tOXfTBKOAixYmDa_VXg7TwviLV5ZQMHQjtbyOjqA00,7879
+blksprs/ops/matmul.py,sha256=eVj_BGj78bJkXYuvw4KctMfcfveQBt5OdYmeXzdpO88,12631
+blksprs/ops/partitioning.py,sha256=qMv9w3yFWXwXIhIppdcJ_JMsoZ25HCH38vb6GRneoLM,10416
+blksprs/ops/repeat.py,sha256=i824ijprfYpCaEjiSG5FTUZz7wMS5ksVy_-vY7ZX8Fg,9729
+blksprs/ops/softmax.py,sha256=_mGkA2jHN8cXwtWXYswobEPyM7UC0JyzRszoE4ZYs7w,13063
+blksprs/ops/transpose.py,sha256=O1XhGIGiVkhOSKcBD0HrYaeK6HmpvEEzLb7zJl7xsIM,4246
+blksprs/ops/misc/broadcast_ops.py,sha256=pv0nssSDOdDbQFttpqUIs2ZXShqfm2RYCfJH-C5x3H0,5544
+blksprs/ops/misc/row_wise.py,sha256=DnV5-xEJUbqZlK2fETwHiPQDUMwT-lkc0VUhBlnJ5Y0,17458
+blksprs/utils/benchmarking.py,sha256=4pLVlnPW_2EM-NT3n4SClaRznVYEljztLbJcccz8kZE,1360
+blksprs/utils/blksprs_tensor.py,sha256=pfoz59aJixj_fIoFx76ySiygwRQUemmgjMKepZ2c4j0,244
+blksprs/utils/layout_utils.py,sha256=49ZdPS_gMn_IrWty3FARbi2rda5a8g5DmAEL8LOrC30,670
+blksprs/utils/processing.py,sha256=WLuMJQ8v-YovXwcDjhlDn3N31WMZXrtyeeyKSgq_zn4,3642
+blksprs/utils/tools.py,sha256=k2OfEplbQiAwVjP84zZf7SNB8FzvMtOFBL9sC98OCbI,683
+blksprs/utils/validation.py,sha256=CbxBbeQWJo8wox5eMoVzaTlP9FVBwt3-gxUOmi3EUgw,4213
+blksprs-1.11.dist-info/METADATA,sha256=NUEiHexWiFNbMxQI2TUEzMw9iGBhxqflhWr2xCgOw28,9105
+blksprs-1.11.dist-info/WHEEL,sha256=52BFRY2Up02UkjOa29eZOS2VxUrpPORXg1pkohGGUS8,91
+blksprs-1.11.dist-info/top_level.txt,sha256=qyp0IHeY3H2GQA97i4hk_To5rRBS2YcE1HRPSLy04fk,8
+blksprs-1.11.dist-info/RECORD,,

{blksprs-1.10.1.dist-info → blksprs-1.11.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (75.6.0)
+Generator: setuptools (76.0.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

blksprs/ops/misc/exp.py

@@ -1,104 +0,0 @@
-import torch
-import triton
-from torch import Tensor
-from triton import language as tl
-
-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_block_size, validate_triton_block_size
-
-
-def exp(x: BlksprsTensor, sparsity_block_size: int, triton_block_size: int = None) -> BlksprsTensor:
-    """Applies the element-wise exponential function to a block-sparse tensor.
-
-    Note:
-        This operation does not consider sparse blocks, i.e., these will not be set to ``e^0``.
-        Consider this when converting back to tensors in regular form.
-
-    Args:
-        x (BlksprsTensor): A block-sparse tensor in compressed form.
-        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:
-        BlksprsTensor: The exponential function applied to all elements of the input tensor as a block-sparse tensor in
-            compressed form.
-
-    """
-    x = x.contiguous()
-
-    validate_dimensions(x)
-    validate_contiguous(x)
-    validate_device(x)
-    validate_sparsity_block_size(sparsity_block_size, x)
-    validate_triton_block_size(triton_block_size, sparsity_block_size)
-
-    return BlksprsTensor(_BlocksparseExp.apply(x, sparsity_block_size, triton_block_size))
-
-
-class _BlocksparseExp(torch.autograd.Function):
-
-    @staticmethod
-    def forward(ctx, x: Tensor, sparsity_block_size: int, triton_block_size: int) -> Tensor:
-        output = torch.empty_like(x)
-
-        x_b, x_r, x_c = x.shape
-        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 = stride(output)
-
-        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"])]
-
-        (_BlocksparseExp.kernel_blocksparse_exp[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,
-          triton_block_size))
-
-        ctx.save_for_backward(output)
-
-        return output
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        o = ctx.saved_tensors[0]
-
-        grad_x = torch.mul(grad_output, o)
-
-        return grad_x, None, None
-
-    @staticmethod
-    @triton.jit
-    def kernel_blocksparse_exp(x,
-                               x_b, x_b_s, x_r_s, x_c_s,
-                               o,
-                               o_b, o_b_s, o_r_s, o_c_s,
-                               TRITON_BLOCK_SIZE: tl.constexpr) -> None:
-        # Get triton block indices
-        pid_blk = tl.program_id(axis=0)
-        pid_row = tl.program_id(axis=1)
-        pid_col = tl.program_id(axis=2)
-
-        # Load block
-        blk_x_idx = ((pid_blk * x_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-        blk_x_msk = (blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s)
-        blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
-
-        # Compute exp
-        buf = tl.exp(blk_x)
-
-        # Store block
-        blk_o_idx = ((pid_blk * o_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_c_s)[None, :])
-        blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s)
-        tl.store(o + blk_o_idx, buf, mask=blk_o_msk)

blksprs-1.10.1.dist-info/RECORD

@@ -1,24 +0,0 @@
-blksprs/__init__.py,sha256=wnpk-20jXq7xV0xa-WpHfPQuauI2gEZz9sH-0blKxP0,1766
-blksprs/layouting/distribution_layout.py,sha256=xDGY5-J7uSD8oenlf8bEJ2amMiQG3NBf2klTTydbTJE,5140
-blksprs/layouting/sparsity_layout.py,sha256=IVtHc_nN3ZM2y4GFcys70PqDWmWc7tkHlVGlToErANk,9894
-blksprs/ops/conversion.py,sha256=NK5uXMepPJ9yYh0vnxKwx5_Ffj_bAvhqPVogf_7PY0g,22248
-blksprs/ops/distribution.py,sha256=qK5t5XgQSJxXPced8RohprqCtUMMTaEP2pFm3KU1c8o,20267
-blksprs/ops/flow.py,sha256=SWHDQ5zx0cjnPR0CcAcRNZdSusSAHSU840SwDNUr24g,6437
-blksprs/ops/matmul.py,sha256=LAQyPNwWVmBMRnAex3msLSPD_aG5SblLCMiutJWqvus,11632
-blksprs/ops/partitioning.py,sha256=ugKnpvH36ND7qeJQp56M74qqfACkzcTVuXebzw__28Y,8286
-blksprs/ops/repeat.py,sha256=RCa-dITomA5v12K5Oxa5_ReA361zS7WHPNNHxSp9PGw,8578
-blksprs/ops/softmax.py,sha256=i8NJhvPRYya94AzpN6qiki6_G9KfDrtPifhWd7wbYzk,12496
-blksprs/ops/transpose.py,sha256=oAtUu7QzQnNAH3lvRs_MIvIKpBu9h74f9Sk07AxKnDM,6991
-blksprs/ops/misc/broadcast_ops.py,sha256=pv0nssSDOdDbQFttpqUIs2ZXShqfm2RYCfJH-C5x3H0,5544
-blksprs/ops/misc/exp.py,sha256=ygfw7oD6ALdPwNQX_HelKgO8I3-LCgIXH_x0gWzkUN8,3840
-blksprs/ops/misc/row_wise.py,sha256=DnV5-xEJUbqZlK2fETwHiPQDUMwT-lkc0VUhBlnJ5Y0,17458
-blksprs/utils/benchmarking.py,sha256=4pLVlnPW_2EM-NT3n4SClaRznVYEljztLbJcccz8kZE,1360
-blksprs/utils/blksprs_tensor.py,sha256=pfoz59aJixj_fIoFx76ySiygwRQUemmgjMKepZ2c4j0,244
-blksprs/utils/layout_utils.py,sha256=49ZdPS_gMn_IrWty3FARbi2rda5a8g5DmAEL8LOrC30,670
-blksprs/utils/processing.py,sha256=WLuMJQ8v-YovXwcDjhlDn3N31WMZXrtyeeyKSgq_zn4,3642
-blksprs/utils/tools.py,sha256=r7Y4C37vfSWUyQTGwa8NyRqgovmsq9hMufkenqYHOxo,539
-blksprs/utils/validation.py,sha256=CbxBbeQWJo8wox5eMoVzaTlP9FVBwt3-gxUOmi3EUgw,4213
-blksprs-1.10.1.dist-info/METADATA,sha256=5in6lYCZo1bd8urYR0wkTxIiTTAIAANukLpKeZfGasY,9107
-blksprs-1.10.1.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-blksprs-1.10.1.dist-info/top_level.txt,sha256=qyp0IHeY3H2GQA97i4hk_To5rRBS2YcE1HRPSLy04fk,8
-blksprs-1.10.1.dist-info/RECORD,,