mps-flash-attn 0.1.13__tar.gz → 0.1.14__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (38) hide show
  1. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/PKG-INFO +1 -1
  2. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/__init__.py +4 -3
  3. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn.egg-info/PKG-INFO +1 -1
  4. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/pyproject.toml +1 -1
  5. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/LICENSE +0 -0
  6. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/README.md +0 -0
  7. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/csrc/mps_flash_attn.mm +0 -0
  8. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/06c421e7a01418cf64aafa07f6b1df0558148583959c596d9a7ce260987f89f0.metallib +0 -0
  9. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/09b9615289be632fdf05444004a0b3b67fb1b70b05a7e0fce8e0ba3a95e3921c.metallib +0 -0
  10. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/0c36461301fb52cbad786d0642b020ad2bfc7229b487ccb5dff44d198423b347.metallib +0 -0
  11. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/2ca9312d1151f792e1a95617db9186928300e3d0ffbe016f0ad53b62ab840bac.metallib +0 -0
  12. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/73254c55475c6b7f7b009f095398994b1f9ae8215beafcf810f100357ccc99b2.metallib +0 -0
  13. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/73254c55475c6b7f7b009f095398994b1f9ae8215beafcf810f100357ccc99b2_1024_1024.bin +0 -0
  14. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/73254c55475c6b7f7b009f095398994b1f9ae8215beafcf810f100357ccc99b2_2048_2048.bin +0 -0
  15. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/73254c55475c6b7f7b009f095398994b1f9ae8215beafcf810f100357ccc99b2_4096_4096.bin +0 -0
  16. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/73254c55475c6b7f7b009f095398994b1f9ae8215beafcf810f100357ccc99b2_512_512.bin +0 -0
  17. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/73254c55475c6b7f7b009f095398994b1f9ae8215beafcf810f100357ccc99b2_8192_8192.bin +0 -0
  18. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/771935bf47d248650e287da82bc82e04bff7c4c52964823e7a12462ccd23408e.metallib +0 -0
  19. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/975aece2b4d3d78035be08a0735a7deacf2e544adee5af81c9c0a3a42a926129.metallib +0 -0
  20. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/a5e2c5c401e3872af0899c1fb3e30b5f52a6070fc49c9dac02982cc1c2f25849.metallib +0 -0
  21. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/ac4573fb201e92310867c59bd569a8ae68f859d60a9352d9d4d5d41c1547c83c.metallib +0 -0
  22. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/adc0f77ff05156bbda8fe78afd9ba8a8d3c890ba8fea0902ae79a6ae8c4f04c3.metallib +0 -0
  23. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/eab4f40de4b0ebd2765b41c25dba7ccab5db4abf6a6eb87d76fff7b5e0829b2f.metallib +0 -0
  24. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/eab4f40de4b0ebd2765b41c25dba7ccab5db4abf6a6eb87d76fff7b5e0829b2f_1024_1024.bin +0 -0
  25. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/eab4f40de4b0ebd2765b41c25dba7ccab5db4abf6a6eb87d76fff7b5e0829b2f_2048_2048.bin +0 -0
  26. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/eab4f40de4b0ebd2765b41c25dba7ccab5db4abf6a6eb87d76fff7b5e0829b2f_4096_4096.bin +0 -0
  27. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/eab4f40de4b0ebd2765b41c25dba7ccab5db4abf6a6eb87d76fff7b5e0829b2f_512_512.bin +0 -0
  28. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/eab4f40de4b0ebd2765b41c25dba7ccab5db4abf6a6eb87d76fff7b5e0829b2f_8192_8192.bin +0 -0
  29. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/f08fe0efd72e055177e068154dae01e08c4d52d3cb883330a04f1431d274aece.metallib +0 -0
  30. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/kernels/manifest.json +0 -0
  31. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/lib/libMFABridge.dylib +0 -0
  32. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn.egg-info/SOURCES.txt +0 -0
  33. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn.egg-info/dependency_links.txt +0 -0
  34. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn.egg-info/requires.txt +0 -0
  35. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn.egg-info/top_level.txt +0 -0
  36. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/setup.cfg +0 -0
  37. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/setup.py +0 -0
  38. {mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/tests/test_attention.py +0 -0
{mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: mps-flash-attn
3
- Version: 0.1.13
3
+ Version: 0.1.14
4
4
  Summary: Flash Attention for PyTorch on Apple Silicon (M1/M2/M3/M4)
5
5
  Author: imperatormk
6
6
  License-Expression: MIT
{mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn/__init__.py RENAMED
@@ -4,7 +4,7 @@ MPS Flash Attention - Flash Attention for PyTorch on Apple Silicon
4
4
  This package provides memory-efficient attention using Metal Flash Attention kernels.
5
5
  """
6
6
 
7
- __version__ = "0.1.13"
7
+ __version__ = "0.1.14"
8
8
 
9
9
  import torch
10
10
  from typing import Optional
@@ -200,12 +200,13 @@ def replace_sdpa():
200
200
  def patched_sdpa(query, key, value, attn_mask=None, dropout_p=0.0,
201
201
  is_causal=False, scale=None):
202
202
  # Use MFA for MPS tensors without dropout
203
- # Only use MFA for seq_len >= 1024 where it outperforms PyTorch's math backend
203
+ # Only use MFA for seq_len >= 1536 where it outperforms PyTorch's math backend
204
204
  # For shorter sequences, PyTorch's simpler matmul+softmax approach is faster
205
+ # Benchmark (BF16, heads=30, head_dim=128): crossover is ~1200-1500
205
206
  if (query.device.type == 'mps' and
206
207
  dropout_p == 0.0 and
207
208
  _HAS_MFA and
208
- query.shape[2] >= 1024):
209
+ query.shape[2] >= 1536):
209
210
  try:
210
211
  # Convert float mask to bool mask if needed
211
212
  # PyTorch SDPA uses additive masks (0 = attend, -inf = mask)
{mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/mps_flash_attn.egg-info/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: mps-flash-attn
3
- Version: 0.1.13
3
+ Version: 0.1.14
4
4
  Summary: Flash Attention for PyTorch on Apple Silicon (M1/M2/M3/M4)
5
5
  Author: imperatormk
6
6
  License-Expression: MIT
{mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/pyproject.toml RENAMED
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
4
4
 
5
5
  [project]
6
6
  name = "mps-flash-attn"
7
- version = "0.1.13"
7
+ version = "0.1.14"
8
8
  description = "Flash Attention for PyTorch on Apple Silicon (M1/M2/M3/M4)"
9
9
  readme = "README.md"
10
10
  license = "MIT"
{mps_flash_attn-0.1.13 → mps_flash_attn-0.1.14}/LICENSE RENAMED
File without changes