blksprs 1.9.2__tar.gz → 1.10__tar.gz

{blksprs-1.9.2 → blksprs-1.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: blksprs
-Version: 1.9.2
+Version: 1.10
 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
@@ -23,14 +23,6 @@ Requires-Dist: build; extra == "build"
 [![GitHub Release](https://img.shields.io/github/v/release/FelixSchoen/blksprs?include_prereleases&label=Latest%20Release)](https://github.com/FelixSchoen/blksprs/releases)
 [![Python Version](https://img.shields.io/badge/Python%20Version-3.11-blue)](https://www.python.org/downloads/release/python-3119/)
 
-## Important Notice
-
-🚨 **Non-Final API** 🚨
-
-Although it already supports a wide variety of functions, this library is still under active development and the API is
-subject to change. For feature requests or bug reports, please open an [issue](https://github.com/FelixSchoen/blksprs/issues).
-We also encourage [pull requests](https://github.com/FelixSchoen/blksprs/pulls).
-
 ## Overview
 
 A lightweight and efficient library for operations on block-sparse matrices in PyTorch using Triton.
@@ -44,7 +36,7 @@ Currently supported operations (includes gradient calculation):
 - 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
+- Splitting and merging of matrices (_currently* only supports splitting and merging along the last dimension_)
 - Conversion to and from sparse form
 - Conversion to different sparsity layouts and different sparsity block sizes
 
@@ -70,13 +62,15 @@ Furthermore, the library provides a set of utility functions
 - for the creation of sparsity layouts based on existing
 dense tensors and for the scatter operation (module ``bs.layouting``),
 - for the application of ``nn.Linear``, ``nn.Dropout``, and ``nn.LayerNorm`` layers to block-sparse tensors,
-- as well as utility functions to apply linear layers,
-ensure correct input dimensionality, and validate input (module ``bs.utils``).
+- as well as utility functions to ensure correct input dimensionality, and validate input (module ``bs.utils``).
+
+_* see the [Roadmap](#roadmap) section for more information_ 
 
 ## Installation
 
-Note that due to the dependency on [Triton](https://github.com/triton-lang/triton) this library is only compatible with
-the Linux platform.
+Note that due to the dependency on [Triton](https://github.com/triton-lang/triton) this library is **only compatible with
+the Linux platform**.
+Keep track of this [issue](https://github.com/triton-lang/triton/issues/1640) for updates.
 
 We recommend installing blksprs from [PyPI](https://pypi.org/project/blksprs/) using pip:
 
@@ -92,6 +86,16 @@ We recommend installing blksprs from [PyPI](https://pypi.org/project/blksprs/) u
 
 See [`CHANGELOG.md`](https://github.com/FelixSchoen/blksprs/blob/main/CHANGELOG.md) for a detailed changelog.
 
+## Roadmap
+
+Note that since this library covers all our current needs it is in a **bugfix-only** state.
+This means that there are no plans to add new features, e.g., support for dimension specification of the ``split`` and ``merge`` operations.
+We will continue to maintain the library and fix any issues that arise.
+Should you find any bugs please open an [issue](https://github.com/FelixSchoen/blksprs/issues).
+We also encourage [pull requests](https://github.com/FelixSchoen/blksprs/pulls).
+
+It might be that this changes with future projects, but as of December 2024, we are content with the current state of the library.
+
 ## Usage
 
 We provide an example below to demonstrate the usage of the library.

{blksprs-1.9.2 → blksprs-1.10}/README.md

@@ -3,14 +3,6 @@
 [![GitHub Release](https://img.shields.io/github/v/release/FelixSchoen/blksprs?include_prereleases&label=Latest%20Release)](https://github.com/FelixSchoen/blksprs/releases)
 [![Python Version](https://img.shields.io/badge/Python%20Version-3.11-blue)](https://www.python.org/downloads/release/python-3119/)
 
-## Important Notice
-
-🚨 **Non-Final API** 🚨
-
-Although it already supports a wide variety of functions, this library is still under active development and the API is
-subject to change. For feature requests or bug reports, please open an [issue](https://github.com/FelixSchoen/blksprs/issues).
-We also encourage [pull requests](https://github.com/FelixSchoen/blksprs/pulls).
-
 ## Overview
 
 A lightweight and efficient library for operations on block-sparse matrices in PyTorch using Triton.
@@ -24,7 +16,7 @@ Currently supported operations (includes gradient calculation):
 - 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
+- Splitting and merging of matrices (_currently* only supports splitting and merging along the last dimension_)
 - Conversion to and from sparse form
 - Conversion to different sparsity layouts and different sparsity block sizes
 
@@ -50,13 +42,15 @@ Furthermore, the library provides a set of utility functions
 - for the creation of sparsity layouts based on existing
 dense tensors and for the scatter operation (module ``bs.layouting``),
 - for the application of ``nn.Linear``, ``nn.Dropout``, and ``nn.LayerNorm`` layers to block-sparse tensors,
-- as well as utility functions to apply linear layers,
-ensure correct input dimensionality, and validate input (module ``bs.utils``).
+- as well as utility functions to ensure correct input dimensionality, and validate input (module ``bs.utils``).
+
+_* see the [Roadmap](#roadmap) section for more information_ 
 
 ## Installation
 
-Note that due to the dependency on [Triton](https://github.com/triton-lang/triton) this library is only compatible with
-the Linux platform.
+Note that due to the dependency on [Triton](https://github.com/triton-lang/triton) this library is **only compatible with
+the Linux platform**.
+Keep track of this [issue](https://github.com/triton-lang/triton/issues/1640) for updates.
 
 We recommend installing blksprs from [PyPI](https://pypi.org/project/blksprs/) using pip:
 
@@ -72,6 +66,16 @@ We recommend installing blksprs from [PyPI](https://pypi.org/project/blksprs/) u
 
 See [`CHANGELOG.md`](https://github.com/FelixSchoen/blksprs/blob/main/CHANGELOG.md) for a detailed changelog.
 
+## Roadmap
+
+Note that since this library covers all our current needs it is in a **bugfix-only** state.
+This means that there are no plans to add new features, e.g., support for dimension specification of the ``split`` and ``merge`` operations.
+We will continue to maintain the library and fix any issues that arise.
+Should you find any bugs please open an [issue](https://github.com/FelixSchoen/blksprs/issues).
+We also encourage [pull requests](https://github.com/FelixSchoen/blksprs/pulls).
+
+It might be that this changes with future projects, but as of December 2024, we are content with the current state of the library.
+
 ## Usage
 
 We provide an example below to demonstrate the usage of the library.

{blksprs-1.9.2 → blksprs-1.10}/blksprs/__init__.py

@@ -15,9 +15,6 @@ class ops:
         from blksprs.ops.misc.broadcast_ops import broadcast_add, broadcast_sub
         from blksprs.ops.misc.exp import exp
 
-    class experimental:
-        from blksprs.ops.experimental.distribution_mdi import gather_mdi, scatter_reduce_mdi
-
 
 class layouting:
     from blksprs.layouting.distribution_layout import build_distribution_layout
@@ -25,9 +22,6 @@ class layouting:
         build_sparsity_layout_matmul, build_sparsity_layout_matmul_fast
     from blksprs.utils.layout_utils import build_full_sparsity_layout
 
-    class experimental:
-        from blksprs.ops.experimental.distribution_mdi import build_distribution_layout_mdi
-
 
 class utils:
     from blksprs.utils.processing import apply_torch_linear, apply_torch_normalisation, apply_torch_dropout, \

{blksprs-1.9.2 → blksprs-1.10}/blksprs/layouting/distribution_layout.py

@@ -84,21 +84,21 @@ def kernel_distribution_layout(i,
 
     # Get position of current sparsity block consisting of its batch, row, and column index
     spa_bat_i_idx = (pid_blk * s_lut_i_r_s + 0 * s_lut_i_c_s)
-    spa_bat_i_msk = (spa_bat_i_idx < s_lut_i_r * s_lut_i_r_s)
+    spa_bat_i_msk = (spa_bat_i_idx >= 0 and spa_bat_i_idx < s_lut_i_r * s_lut_i_r_s)
     spa_bat_i = tl.load(s_lut_i + spa_bat_i_idx, mask=spa_bat_i_msk)
 
     spa_row_i_idx = (pid_blk * s_lut_i_r_s + 1 * s_lut_i_c_s)
-    spa_row_i_msk = (spa_row_i_idx < s_lut_i_r * s_lut_i_r_s)
+    spa_row_i_msk = (spa_row_i_idx >= 0 and 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_msk = (spa_col_i_idx >= 0 and 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_msk = (blk_i_idx >= 0 and blk_i_idx < i_b * i_b_s)
     blk_i = tl.load(i + blk_i_idx, mask=blk_i_msk)
 
     dst_bat_idx = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), spa_bat_i, dtype=tl.int32)
@@ -111,10 +111,10 @@ def kernel_distribution_layout(i,
     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_v = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), 1, dtype=tl.int1)
 
     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)
+    blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s)
     tl.store(o + blk_o_idx, blk_v, mask=blk_o_msk)

{blksprs-1.9.2 → blksprs-1.10}/blksprs/layouting/sparsity_layout.py

@@ -71,7 +71,7 @@ def kernel_sparsity_layout(x,
     blk_x_idx = (pid_bat * 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_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)
 
     # Store sparsity layout value
@@ -79,7 +79,7 @@ def kernel_sparsity_layout(x,
         blk_o_idx = (pid_bat * o_b_s +
                      (((pid_row * TRITON_BLOCK_SIZE) // sparsity_block_size) * o_r_s +
                       ((pid_col * TRITON_BLOCK_SIZE) // sparsity_block_size) * o_c_s))
-        blk_o_msk = (blk_o_idx < o_b * o_b_s)
+        blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s)
         tl.store(o + blk_o_idx, 1, mask=blk_o_msk)
 
 
@@ -162,22 +162,22 @@ def kernel_sparsity_layout_adaption(x,
 
     # 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_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 < s_lut_r * s_lut_r_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 < s_lut_r * s_lut_r_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)
 
     # 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] +
                  ((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_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)
 
     # Store sparsity layout value
@@ -187,7 +187,7 @@ def kernel_sparsity_layout_adaption(x,
                        // sparsity_block_size_to) * o_r_s) +
                      (((spa_col * sparsity_block_size_from + pid_col * TRITON_BLOCK_SIZE)
                        // sparsity_block_size_to) * o_c_s))
-        blk_o_msk = (blk_o_idx < o_b * o_b_s)
+        blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s)
         tl.store(o + blk_o_idx, 1, mask=blk_o_msk)
 
 

{blksprs-1.9.2 → blksprs-1.10}/blksprs/ops/conversion.py

@@ -1,5 +1,3 @@
-from typing import Any
-
 import torch
 import triton
 from torch import Tensor
@@ -133,7 +131,7 @@ class _BlocksparseToDense(torch.autograd.Function):
 
         # Get reverse sparsity index for current block
         rev_idx_spa_idx = (pid_blk * s_l_b_s + spa_row * s_l_r_s + spa_col * s_l_c_s)
-        rev_idx_spa_msk = (rev_idx_spa_idx < s_l_b * s_l_b_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(sparsity_reverse_lut + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
 
         # If block is present commence operations
@@ -143,13 +141,13 @@ class _BlocksparseToDense(torch.autograd.Function):
                          tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
                        (((pid_col % (sparsity_block_size // TRITON_BLOCK_SIZE)) * TRITON_BLOCK_SIZE +
                          tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-            blk_msk = (blk_idx < x_b * x_b_s)
+            blk_msk = (blk_idx >= 0 and blk_idx < x_b * x_b_s)
             blk = tl.load(x + blk_idx, mask=blk_msk)
 
             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, :])
-            o_msk = (o_idx < o_b * o_b_s)
+            o_msk = (o_idx >= 0 and o_idx < o_b * o_b_s)
             tl.store(o + o_idx, blk, o_msk)
 
 
@@ -260,15 +258,15 @@ class _BlocksparseToSparse(torch.autograd.Function):
 
         # 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_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 < s_lut_r * s_lut_r_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 < s_lut_r * s_lut_r_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)
 
         # Load block from dense tensor
@@ -277,14 +275,14 @@ class _BlocksparseToSparse(torch.autograd.Function):
                        tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
                      ((spa_col * sparsity_block_size + pid_col * TRITON_BLOCK_SIZE +
                        tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-        blk_d_msk = (blk_d_idx < x_b * x_b_s)
+        blk_d_msk = (blk_d_idx >= 0 and blk_d_idx < x_b * x_b_s)
         blk_d = tl.load(x + blk_d_idx, mask=blk_d_msk)
 
         # Store block in sparse tensor
         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 < (pid_blk + 1) * o_b_s)
+        blk_o_msk = (blk_o_idx >= 0 and blk_o_idx < (pid_blk + 1) * o_b_s)
         tl.store(o + blk_o_idx, blk_d, mask=blk_o_msk)
 
 
@@ -424,15 +422,15 @@ class _BlocksparseAdaptLayout(torch.autograd.Function):
 
         # Get position of current sparsity block consisting of its batch, row, and column index
         spa_bat_o_idx = (pid_blk * s_lut_o_r_s + 0 * s_lut_o_c_s)
-        spa_bat_o_msk = (spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
+        spa_bat_o_msk = (spa_bat_o_idx >= 0 and spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
         spa_bat_o = tl.load(s_lut_o + spa_bat_o_idx, mask=spa_bat_o_msk)
 
         spa_row_o_idx = (pid_blk * s_lut_o_r_s + 1 * s_lut_o_c_s)
-        spa_row_o_msk = (spa_row_o_idx < s_lut_o_r * s_lut_o_r_s)
+        spa_row_o_msk = (spa_row_o_idx >= 0 and 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_msk = (spa_col_o_idx >= 0 and 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)
 
         # Get equivalent sparsity block in from layout
@@ -444,7 +442,7 @@ class _BlocksparseAdaptLayout(torch.autograd.Function):
         rev_idx_spa_x_idx = (spa_bat_x * s_l_x_b_s +
                              spa_row_x * s_l_x_r_s +
                              spa_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_msk = (rev_idx_spa_x_idx >= 0 and 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 block is present commence operations
@@ -459,12 +457,12 @@ class _BlocksparseAdaptLayout(torch.autograd.Function):
             blk_x_idx = ((rev_idx_spa_x * x_b_s) +
                          ((shift_row_x * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
                          ((shift_col_x * 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_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)
 
             # 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 >= 0 and blk_o_idx < o_b * o_b_s)
             tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)

{blksprs-1.9.2 → blksprs-1.10}/blksprs/ops/distribution.py

@@ -138,22 +138,22 @@ class _BlocksparseGather(torch.autograd.Function):
 
         # Get position of current sparsity block consisting of its batch, row, and column index
         spa_bat_o_idx = (pid_blk * s_lut_o_r_s + 0 * s_lut_o_c_s)
-        spa_bat_o_msk = (spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
+        spa_bat_o_msk = (spa_bat_o_idx >= 0 and spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
         spa_bat_o = tl.load(s_lut_o + spa_bat_o_idx, mask=spa_bat_o_msk)
 
         spa_row_o_idx = (pid_blk * s_lut_o_r_s + 1 * s_lut_o_c_s)
-        spa_row_o_msk = (spa_row_o_idx < s_lut_o_r * s_lut_o_r_s)
+        spa_row_o_msk = (spa_row_o_idx >= 0 and 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_msk = (spa_col_o_idx >= 0 and 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] +
                      ((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_msk = (blk_i_idx >= 0 and blk_i_idx < i_b * i_b_s)
         blk_i = tl.load(i + blk_i_idx, mask=blk_i_msk).to(tl.int32)
 
         # Get indices of sparsity blocks and positions within the blocks
@@ -180,21 +180,21 @@ class _BlocksparseGather(torch.autograd.Function):
         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_msk = (rev_idx_spa_x_idx >= 0 and 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) +
                      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_msk = ((blk_x_idx >= 0 and blk_x_idx < x_b * x_b_s) and 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) &  rev_idx_spa_x_msk != -1)
+        blk_o_msk = ((blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s) and rev_idx_spa_x_msk != -1)
         tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)
 
 
@@ -364,29 +364,29 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
 
         # Get position of current sparsity block consisting of its batch, row, and column index
         spa_bat_x_idx = (pid_blk * s_lut_x_r_s + 0 * s_lut_x_c_s)
-        spa_bat_x_msk = (spa_bat_x_idx < s_lut_x_r * s_lut_x_r_s)
+        spa_bat_x_msk = (spa_bat_x_idx >= 0 and spa_bat_x_idx < s_lut_x_r * s_lut_x_r_s)
         spa_bat_x = tl.load(s_lut_x + spa_bat_x_idx, mask=spa_bat_x_msk)
 
         spa_row_x_idx = (pid_blk * s_lut_x_r_s + 1 * s_lut_x_c_s)
-        spa_row_x_msk = (spa_row_x_idx < s_lut_x_r * s_lut_x_r_s)
+        spa_row_x_msk = (spa_row_x_idx >= 0 and 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_msk = (spa_col_x_idx >= 0 and 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] +
                      ((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_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 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] +
                      ((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_msk = (blk_i_idx >= 0 and blk_i_idx < i_b * i_b_s)
         blk_i = tl.load(i + blk_i_idx, mask=blk_i_msk).to(tl.int32)
 
         # Get indices of sparsity blocks and positions within the blocks
@@ -413,14 +413,14 @@ class _BlocksparseScatterReduce(torch.autograd.Function):
         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_msk = (rev_idx_spa_o_idx >= 0 and 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) +
                      dst_row_o +
                      dst_col_o)
-        blk_o_msk = ((blk_o_idx < o_b * o_b_s) & rev_idx_spa_o_msk != -1)
+        blk_o_msk = ((blk_o_idx >= 0 and blk_o_idx < o_b * o_b_s) and rev_idx_spa_o_msk != -1)
 
         if reduce_op_ind == 0:
             tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)

{blksprs-1.9.2 → blksprs-1.10}/blksprs/ops/flow.py

@@ -22,22 +22,22 @@ def kernel_blocksparse_flow_pull(x,
 
     # 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_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 < s_lut_r * s_lut_r_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 < s_lut_r * s_lut_r_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_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_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:
@@ -47,13 +47,13 @@ def kernel_blocksparse_flow_pull(x,
     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_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 < o_b * o_b_s)
+    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)
 
 
@@ -73,22 +73,22 @@ def kernel_blocksparse_flow_push(x,
 
     # 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_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 < s_lut_r * s_lut_r_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 < s_lut_r * s_lut_r_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_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_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:
@@ -98,13 +98,13 @@ def kernel_blocksparse_flow_push(x,
     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_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 < o_b * o_b_s)
+    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)
 
 

{blksprs-1.9.2 → blksprs-1.10}/blksprs/ops/matmul.py

@@ -164,15 +164,15 @@ class _BlocksparseMatmulSSS(torch.autograd.Function):
 
         # Get position of current sparsity block consisting of its batch, row, and column index
         spa_bat_o_idx = (pid_blk * s_lut_o_r_s + 0 * s_lut_o_c_s)
-        spa_bat_o_msk = (spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
+        spa_bat_o_msk = (spa_bat_o_idx >= 0 and spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
         spa_bat_o = tl.load(s_lut_o + spa_bat_o_idx, mask=spa_bat_o_msk)
 
         spa_row_o_idx = (pid_blk * s_lut_o_r_s + 1 * s_lut_o_c_s)
-        spa_row_o_msk = (spa_row_o_idx < s_lut_o_r * s_lut_o_r_s)
+        spa_row_o_msk = (spa_row_o_idx >= 0 and 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_msk = (spa_col_o_idx >= 0 and 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)
 
         # Setup buffer
@@ -192,12 +192,12 @@ class _BlocksparseMatmulSSS(torch.autograd.Function):
             rev_idx_spa_x_idx = (spa_bat_o * s_l_x_b_s +
                                  spa_row_o * s_l_x_r_s +
                                  i_seg_spa * 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_msk = (rev_idx_spa_x_idx >= 0 and 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)
 
             # Get reverse sparsity indices for y
             rev_idx_spa_y_idx = (spa_bat_o * s_l_y_b_s + i_seg_spa * s_l_y_r_s + spa_col_o * s_l_y_c_s)
-            rev_idx_spa_y_msk = (rev_idx_spa_y_idx < s_l_y_b * s_l_y_b_s)
+            rev_idx_spa_y_msk = (rev_idx_spa_y_idx >= 0 and rev_idx_spa_y_idx < s_l_y_b * s_l_y_b_s)
             rev_idx_spa_y = tl.load(r_lut_y + rev_idx_spa_y_idx, mask=rev_idx_spa_y_msk).to(tl.int32)
 
             # If both blocks are present commence calculation
@@ -206,14 +206,14 @@ class _BlocksparseMatmulSSS(torch.autograd.Function):
                              ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
                              ((i_seg_tri_mod * 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_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_y_idx = ((rev_idx_spa_y * y_b_s) +
                              ((i_seg_tri_mod * TRITON_BLOCK_SIZE +
                                tl.arange(0, TRITON_BLOCK_SIZE)) * y_r_s)[:, None] +
                              ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * y_c_s)[None, :])
-                blk_y_msk = (blk_y_idx < y_b * y_b_s)
+                blk_y_msk = (blk_y_idx >= 0 and blk_y_idx < y_b * y_b_s)
                 blk_y = tl.load(y + blk_y_idx, mask=blk_y_msk)
 
                 # Perform matrix multiplication
@@ -223,5 +223,5 @@ class _BlocksparseMatmulSSS(torch.autograd.Function):
         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 >= 0 and blk_o_idx < o_b * o_b_s)
         tl.store(o + blk_o_idx, buf, mask=blk_o_msk)

{blksprs-1.9.2 → blksprs-1.10}/blksprs/ops/misc/broadcast_ops.py

@@ -99,29 +99,29 @@ def kernel_broadcast_addition(x,
 
     # Get position of current sparsity block consisting of its batch, row, and column index
     spa_bat_o_idx = (pid_blk * s_lut_o_r_s + 0 * s_lut_o_c_s)
-    spa_bat_o_msk = (spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
+    spa_bat_o_msk = (spa_bat_o_idx >= 0 and spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
     spa_bat_o = tl.load(s_lut_o + spa_bat_o_idx, mask=spa_bat_o_msk)
 
     spa_row_o_idx = (pid_blk * s_lut_o_r_s + 1 * s_lut_o_c_s)
-    spa_row_o_msk = (spa_row_o_idx < s_lut_o_r * s_lut_o_r_s)
+    spa_row_o_msk = (spa_row_o_idx >= 0 and 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_msk = (spa_col_o_idx >= 0 and 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 x block
     blk_x_idx = (spa_bat_o * x_b_s +
                  ((spa_row_o * sparsity_block_size + pid_row * 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_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 y block
     blk_y_idx = (spa_bat_o * y_b_s +
                  ((spa_col_o * sparsity_block_size + pid_col * TRITON_BLOCK_SIZE +
                    tl.arange(0, TRITON_BLOCK_SIZE)) * y_c_s)[None, :])
-    blk_y_msk = (blk_y_idx < y_b * y_b_s)
+    blk_y_msk = (blk_y_idx >= 0 and blk_y_idx < y_b * y_b_s)
     blk_y = tl.load(y + blk_y_idx, mask=blk_y_msk)
 
     # Compute sum
@@ -132,5 +132,5 @@ def kernel_broadcast_addition(x,
     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 >= 0 and blk_o_idx < o_b * o_b_s)
     tl.store(o + blk_o_idx, buf, mask=blk_o_msk)