Hugging Face's logo

kernels-community
/
moe

kernel
Model card Files Community
1
danieldk HF Staff commited on
Commit
69e0359
·
1 Parent(s): 441a21f

Build

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. build/torch25-cxx11-cu118-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} +2 -2
  2. build/torch25-cxx11-cu118-x86_64-linux/moe/_ops.py +3 -3
  3. build/torch25-cxx11-cu121-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} +2 -2
  4. build/torch25-cxx11-cu121-x86_64-linux/moe/_ops.py +3 -3
  5. build/torch25-cxx11-cu124-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} +2 -2
  6. build/torch25-cxx11-cu124-x86_64-linux/moe/_ops.py +3 -3
  7. build/torch25-cxx98-cu118-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} +2 -2
  8. build/torch25-cxx98-cu118-x86_64-linux/moe/_ops.py +3 -3
  9. build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  10. build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  11. build/torch25-cxx98-cu121-x86_64-linux/moe/_ops.py +3 -3
  12. build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  13. build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  14. build/torch25-cxx98-cu124-x86_64-linux/moe/_ops.py +3 -3
  15. build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  16. build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  17. build/torch26-cxx11-cu118-x86_64-linux/moe/_ops.py +3 -3
  18. build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  19. build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  20. build/torch26-cxx11-cu124-x86_64-linux/moe/_ops.py +3 -3
  21. build/torch26-cxx11-cu126-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  22. build/torch26-cxx11-cu126-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  23. build/torch26-cxx11-cu126-x86_64-linux/moe/_ops.py +3 -3
  24. build/torch26-cxx98-cu118-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  25. build/torch26-cxx98-cu118-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  26. build/torch26-cxx98-cu118-x86_64-linux/moe/_ops.py +3 -3
  27. build/torch26-cxx98-cu124-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  28. build/torch26-cxx98-cu124-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  29. build/torch26-cxx98-cu124-x86_64-linux/moe/_ops.py +3 -3
  30. build/torch26-cxx98-cu126-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  31. build/torch26-cxx98-cu126-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so +0 -3
  32. build/torch26-cxx98-cu126-x86_64-linux/moe/_ops.py +3 -3
  33. build/torch27-cxx11-cu118-x86_64-linux/moe/__init__.py +80 -0
  34. build/torch27-cxx11-cu118-x86_64-linux/moe/_moe_441a21f.abi3.so +3 -0
  35. build/torch27-cxx11-cu118-x86_64-linux/moe/_ops.py +9 -0
  36. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  37. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  38. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  39. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  40. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  41. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +218 -0
  42. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json +218 -0
  43. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  44. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  45. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  46. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  47. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  48. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  49. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  50. build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
build/torch25-cxx11-cu118-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:17cf96e50dd9d5632a7b5959037c5d11e79fae301989ee6315bc8e3d4d545bde
3
- size 87056376
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8eac11915526a86c44419a9ae870d66487fea4fe417faecd94f5fb80e125747d
3
+ size 67203016
build/torch25-cxx11-cu118-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch25-cxx11-cu121-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:95186380b2815d16e6b8d5cc2d87a93d45dadf31ac197028e37b7d6f65680cf1
3
- size 87250920
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c55bc16e5d34ff5488bc78d950496a23d21fa81c0c7ff93320b9a5cd7834c93f
3
+ size 67344312
build/torch25-cxx11-cu121-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch25-cxx11-cu124-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:53bbb3f5ee8f5cbc0d640e924797f71f2c5b2767ec044470d6200fd154c7887b
3
- size 86965656
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:180cb29050dd618a6ee9cae4b1bf3efc0109a2474a6cc8fc3e0d85f31ff3d992
3
+ size 67190048
build/torch25-cxx11-cu124-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch25-cxx98-cu118-x86_64-linux/moe/{_moe_c558d47_dirty.abi3.so → _moe_441a21f.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d0eedac7fe7e2bcdace617e68444affa54523ae20cf620ad92959476f93f9e9c
3
- size 87052576
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:acbd41fc487d0e2cdec489be22dd04e0979be2e7f79b2400e98bcb59a6e14f82
3
+ size 67195120
build/torch25-cxx98-cu118-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5150a68f433d116840445ee3d308e3e2fa2e66cf625269abef2ab7103e808fa
3
+ size 67336680
build/torch25-cxx98-cu121-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:004cbfe65028ad7b4dd0c2bcffe16bbcb8fe94e3d3771b466808dbd1996fde12
3
- size 87247384
 
 
 
 
build/torch25-cxx98-cu121-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:23c4badd5ef623b42dce429e7e0ef12f25cec48334c7968d29bf7e3ccc65ca6a
3
+ size 67178320
build/torch25-cxx98-cu124-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:3785d7743d33c834d0a3ac99dcb8add3a951314290dd9f7d4eb87045c7424f90
3
- size 86958024
 
 
 
 
build/torch25-cxx98-cu124-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:34f65dbaacb9153df07d5b72e6cd3a1e784529256ff7756c746469d5a8b342ab
3
+ size 67202944
build/torch26-cxx11-cu118-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:6fc2df0cc4b652b7c1a2a34191bedc3a3190cac9e74b44403ce3f0c763bdf072
3
- size 87060400
 
 
 
 
build/torch26-cxx11-cu118-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:32c101d7deb2f3a2cf582794546d56c3ebd80cc818615816c1a2f878e369d0c0
3
+ size 67190096
build/torch26-cxx11-cu124-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:511359ec7e5a3c4ea2beda0a31b005a0210aadf3d2247b87e990eefa0f8ff8ba
3
- size 86965712
 
 
 
 
build/torch26-cxx11-cu124-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch26-cxx11-cu126-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ee80c4772427bb7eef2f4b7129771b67fdc42343be176145f8d5d10407695003
3
+ size 67579576
build/torch26-cxx11-cu126-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:b546237970e07f6125675971da472a7cacfaa81384f6ddfa4eb36f4da644c0fc
3
- size 87424816
 
 
 
 
build/torch26-cxx11-cu126-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch26-cxx98-cu118-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0703d488723b4f28431c1a74679b367d0b8c7db3671c574f5e07c675bb583f2
3
+ size 67186880
build/torch26-cxx98-cu118-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:889767dfddb5c30f49394395f34b6e25c2f097b32826e02bc73e7381a8ef8b24
3
- size 87048456
 
 
 
 
build/torch26-cxx98-cu118-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch26-cxx98-cu124-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e89acab57b21cd80022701dd8c457efa09f50ede4d2762ed7e4b35d42232ac7f
3
+ size 67178288
build/torch26-cxx98-cu124-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:a8b46694ef9faa8dc3bc01413602d4c5a7ce10fdd7e34c6773e778b7917a9066
3
- size 86949808
 
 
 
 
build/torch26-cxx98-cu124-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch26-cxx98-cu126-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d8f063ad5ea46c358960b9feb878d77b81a130d90fa633d5f97dcebde6742472
3
+ size 67559576
build/torch26-cxx98-cu126-x86_64-linux/moe/_moe_c558d47_dirty.abi3.so DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:6023b925f8f17a3e8dca1f673bdb6348c661e94423f86114f14fef1bf5f69947
3
- size 87417112
 
 
 
 
build/torch26-cxx98-cu126-x86_64-linux/moe/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _moe_c558d47_dirty
3
- ops = torch.ops._moe_c558d47_dirty
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_moe_c558d47_dirty::{op_name}"
 
1
  import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_moe_441a21f::{op_name}"
build/torch27-cxx11-cu118-x86_64-linux/moe/__init__.py ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ from . import layers
4
+ from ._ops import ops
5
+ from .fp8_utils import per_token_group_quant_fp8, w8a8_block_fp8_matmul
6
+ from .fused_marlin_moe import fused_marlin_moe
7
+ from .fused_moe import fused_experts, fused_moe, fused_topk, grouped_topk
8
+ from .scalar_type import ScalarType, scalar_types
9
+
10
+
11
+ def gptq_marlin_moe_repack(
12
+ b_q_weight: torch.Tensor,
13
+ perm: torch.Tensor,
14
+ size_k: int,
15
+ size_n: int,
16
+ num_bits: int,
17
+ ) -> torch.Tensor:
18
+ num_experts = b_q_weight.shape[0]
19
+ assert size_k % 16 == 0
20
+ output = torch.empty(
21
+ (num_experts, size_k // 16, size_n * (num_bits // 2)),
22
+ device=b_q_weight.device,
23
+ dtype=b_q_weight.dtype,
24
+ )
25
+ for e in range(num_experts):
26
+ output[e] = ops.gptq_marlin_repack(
27
+ b_q_weight[e], perm[e], size_k, size_n, num_bits
28
+ )
29
+ return output
30
+
31
+
32
+ def awq_marlin_moe_repack(
33
+ b_q_weight: torch.Tensor,
34
+ perm: torch.Tensor,
35
+ size_k: int,
36
+ size_n: int,
37
+ num_bits: int,
38
+ ) -> torch.Tensor:
39
+ num_experts = b_q_weight.shape[0]
40
+ assert size_k % 16 == 0
41
+ output = torch.empty(
42
+ (num_experts, size_k // 16, size_n * (num_bits // 2)),
43
+ device=b_q_weight.device,
44
+ dtype=b_q_weight.dtype,
45
+ )
46
+ for e in range(num_experts):
47
+ output[e] = ops.awq_marlin_repack(b_q_weight[e], size_k, size_n, num_bits)
48
+ return output
49
+
50
+
51
+ def moe_sum(input: torch.Tensor, output: torch.Tensor):
52
+ ops.moe_sum(input, output)
53
+
54
+
55
+ def topk_softmax(
56
+ topk_weights: torch.Tensor,
57
+ topk_ids: torch.Tensor,
58
+ token_expert_indicies: torch.Tensor,
59
+ gating_output: float,
60
+ ) -> None:
61
+ ops.topk_softmax(topk_weights, topk_ids, token_expert_indicies, gating_output)
62
+
63
+
64
+ __all__ = [
65
+ "ScalarType",
66
+ "awq_marlin_moe_repack",
67
+ "fused_experts",
68
+ "fused_marlin_moe",
69
+ "fused_moe",
70
+ "fused_topk",
71
+ "gptq_marlin_moe_repack",
72
+ "grouped_topk",
73
+ "layers",
74
+ "moe_align_block_size",
75
+ "moe_sum",
76
+ "per_token_group_quant_fp8",
77
+ "scalar_types",
78
+ "topk_softmax",
79
+ "w8a8_block_fp8_matmul",
80
+ ]
build/torch27-cxx11-cu118-x86_64-linux/moe/_moe_441a21f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6ff28f76e584e020e7e59f0d0cf3309415d32941a4d03153ae771c258470004
3
+ size 67203152
build/torch27-cxx11-cu118-x86_64-linux/moe/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _moe_441a21f
3
+ ops = torch.ops._moe_441a21f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_moe_441a21f::{op_name}"
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 64,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 64,
46
+ "GROUP_SIZE_M": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 2
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 64,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 16,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 32,
95
+ "num_warps": 4,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 32,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ }
146
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 32,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 16,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 64,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 32,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 32,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 16,
63
+ "num_warps": 4,
64
+ "num_stages": 5
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 256,
77
+ "BLOCK_SIZE_K": 64,
78
+ "GROUP_SIZE_M": 16,
79
+ "num_warps": 4,
80
+ "num_stages": 5
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 64,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 16,
95
+ "num_warps": 8,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 8,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 64,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 8,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 64,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 64,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 64,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 256,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 4,
152
+ "num_stages": 4
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 64,
156
+ "BLOCK_SIZE_N": 256,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 32,
159
+ "num_warps": 4,
160
+ "num_stages": 4
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 256,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 4,
168
+ "num_stages": 4
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 64,
172
+ "BLOCK_SIZE_N": 256,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 32,
175
+ "num_warps": 4,
176
+ "num_stages": 4
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 64,
180
+ "BLOCK_SIZE_N": 256,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 4
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 64,
188
+ "BLOCK_SIZE_N": 256,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 4
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 64,
196
+ "BLOCK_SIZE_N": 256,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 4,
200
+ "num_stages": 4
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 64,
204
+ "BLOCK_SIZE_N": 256,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 32,
207
+ "num_warps": 4,
208
+ "num_stages": 4
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 64,
212
+ "BLOCK_SIZE_N": 256,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 4
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 5
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 5
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 32,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 5
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 64,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 5
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 64,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 5
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 64,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 128,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 4,
152
+ "num_stages": 3
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 128,
156
+ "BLOCK_SIZE_N": 128,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 16,
159
+ "num_warps": 4,
160
+ "num_stages": 3
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 128,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 4,
168
+ "num_stages": 3
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 128,
172
+ "BLOCK_SIZE_N": 128,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 16,
175
+ "num_warps": 4,
176
+ "num_stages": 3
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 128,
180
+ "BLOCK_SIZE_N": 128,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 3
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 128,
188
+ "BLOCK_SIZE_N": 128,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 3
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 128,
196
+ "BLOCK_SIZE_N": 128,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 4,
200
+ "num_stages": 3
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 128,
204
+ "BLOCK_SIZE_N": 128,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 16,
207
+ "num_warps": 4,
208
+ "num_stages": 3
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 128,
212
+ "BLOCK_SIZE_N": 128,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 3
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 32,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 32,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 128,
84
+ "BLOCK_SIZE_N": 32,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 16,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 256,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 4,
152
+ "num_stages": 4
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 64,
156
+ "BLOCK_SIZE_N": 256,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 32,
159
+ "num_warps": 4,
160
+ "num_stages": 3
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 256,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 32,
167
+ "num_warps": 4,
168
+ "num_stages": 4
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 64,
172
+ "BLOCK_SIZE_N": 256,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 16,
175
+ "num_warps": 4,
176
+ "num_stages": 4
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 64,
180
+ "BLOCK_SIZE_N": 256,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 4
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 64,
188
+ "BLOCK_SIZE_N": 256,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 4
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 64,
196
+ "BLOCK_SIZE_N": 256,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 32,
199
+ "num_warps": 4,
200
+ "num_stages": 4
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 64,
204
+ "BLOCK_SIZE_N": 256,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 16,
207
+ "num_warps": 4,
208
+ "num_stages": 4
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 64,
212
+ "BLOCK_SIZE_N": 256,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 4
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 8,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 8,
24
+ "num_stages": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 8,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 8,
40
+ "num_stages": 5
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 8,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 32,
61
+ "BLOCK_SIZE_K": 256,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 32,
69
+ "BLOCK_SIZE_K": 256,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 64,
77
+ "BLOCK_SIZE_K": 256,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 256,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 256,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 8,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 256,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 256,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 8,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 256,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 8,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 256,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 8,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 256,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 8,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 256,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 8,
144
+ "num_stages": 3
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 256,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 128,
150
+ "GROUP_SIZE_M": 1,
151
+ "num_warps": 8,
152
+ "num_stages": 3
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 256,
156
+ "BLOCK_SIZE_N": 128,
157
+ "BLOCK_SIZE_K": 128,
158
+ "GROUP_SIZE_M": 1,
159
+ "num_warps": 8,
160
+ "num_stages": 3
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 256,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 128,
166
+ "GROUP_SIZE_M": 1,
167
+ "num_warps": 8,
168
+ "num_stages": 3
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 256,
172
+ "BLOCK_SIZE_N": 128,
173
+ "BLOCK_SIZE_K": 128,
174
+ "GROUP_SIZE_M": 1,
175
+ "num_warps": 8,
176
+ "num_stages": 3
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 256,
180
+ "BLOCK_SIZE_N": 128,
181
+ "BLOCK_SIZE_K": 128,
182
+ "GROUP_SIZE_M": 1,
183
+ "num_warps": 8,
184
+ "num_stages": 3
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 256,
188
+ "BLOCK_SIZE_N": 128,
189
+ "BLOCK_SIZE_K": 128,
190
+ "GROUP_SIZE_M": 1,
191
+ "num_warps": 8,
192
+ "num_stages": 3
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 256,
196
+ "BLOCK_SIZE_N": 128,
197
+ "BLOCK_SIZE_K": 128,
198
+ "GROUP_SIZE_M": 1,
199
+ "num_warps": 8,
200
+ "num_stages": 3
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 256,
204
+ "BLOCK_SIZE_N": 128,
205
+ "BLOCK_SIZE_K": 128,
206
+ "GROUP_SIZE_M": 1,
207
+ "num_warps": 8,
208
+ "num_stages": 3
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 256,
212
+ "BLOCK_SIZE_N": 128,
213
+ "BLOCK_SIZE_K": 128,
214
+ "GROUP_SIZE_M": 1,
215
+ "num_warps": 8,
216
+ "num_stages": 3
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 5
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 8,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 5
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 5
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 16,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 16,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 8,
96
+ "num_stages": 5
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 8,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 32,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 8,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 64,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 32,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 128,
148
+ "BLOCK_SIZE_N": 256,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 8,
152
+ "num_stages": 4
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 128,
156
+ "BLOCK_SIZE_N": 256,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 16,
159
+ "num_warps": 8,
160
+ "num_stages": 3
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 128,
164
+ "BLOCK_SIZE_N": 256,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 8,
168
+ "num_stages": 3
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 128,
172
+ "BLOCK_SIZE_N": 256,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 16,
175
+ "num_warps": 8,
176
+ "num_stages": 3
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 128,
180
+ "BLOCK_SIZE_N": 256,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 8,
184
+ "num_stages": 4
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 128,
188
+ "BLOCK_SIZE_N": 256,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 8,
192
+ "num_stages": 3
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 128,
196
+ "BLOCK_SIZE_N": 256,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 8,
200
+ "num_stages": 3
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 128,
204
+ "BLOCK_SIZE_N": 256,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 16,
207
+ "num_warps": 8,
208
+ "num_stages": 3
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 128,
212
+ "BLOCK_SIZE_N": 256,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 8,
216
+ "num_stages": 3
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 16,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 8,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 64,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 64,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 32,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 256,
108
+ "BLOCK_SIZE_N": 32,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 256,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 32,
151
+ "num_warps": 4,
152
+ "num_stages": 4
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 64,
156
+ "BLOCK_SIZE_N": 256,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 32,
159
+ "num_warps": 4,
160
+ "num_stages": 4
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 256,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 4,
168
+ "num_stages": 4
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 64,
172
+ "BLOCK_SIZE_N": 256,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 16,
175
+ "num_warps": 4,
176
+ "num_stages": 4
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 64,
180
+ "BLOCK_SIZE_N": 256,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 4
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 64,
188
+ "BLOCK_SIZE_N": 256,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 4
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 64,
196
+ "BLOCK_SIZE_N": 256,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 4,
200
+ "num_stages": 4
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 64,
204
+ "BLOCK_SIZE_N": 256,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 16,
207
+ "num_warps": 4,
208
+ "num_stages": 4
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 64,
212
+ "BLOCK_SIZE_N": 256,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 4
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 32,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 4,
32
+ "num_stages": 5
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 64,
39
+ "num_warps": 4,
40
+ "num_stages": 5
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 32,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 32,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 5
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 64,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 5
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 64,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 64,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 128,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 4,
152
+ "num_stages": 3
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 128,
156
+ "BLOCK_SIZE_N": 128,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 16,
159
+ "num_warps": 4,
160
+ "num_stages": 3
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 128,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 4,
168
+ "num_stages": 3
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 128,
172
+ "BLOCK_SIZE_N": 128,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 16,
175
+ "num_warps": 4,
176
+ "num_stages": 3
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 128,
180
+ "BLOCK_SIZE_N": 128,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 3
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 128,
188
+ "BLOCK_SIZE_N": 128,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 3
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 128,
196
+ "BLOCK_SIZE_N": 128,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 4,
200
+ "num_stages": 3
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 128,
204
+ "BLOCK_SIZE_N": 128,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 16,
207
+ "num_warps": 4,
208
+ "num_stages": 3
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 128,
212
+ "BLOCK_SIZE_N": 128,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 3
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 16,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 32,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 32,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 128,
84
+ "BLOCK_SIZE_N": 32,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 4,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 32,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 64,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 32,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 256,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 32,
151
+ "num_warps": 4,
152
+ "num_stages": 4
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 64,
156
+ "BLOCK_SIZE_N": 256,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 32,
159
+ "num_warps": 4,
160
+ "num_stages": 3
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 256,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 4,
168
+ "num_stages": 4
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 64,
172
+ "BLOCK_SIZE_N": 256,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 16,
175
+ "num_warps": 4,
176
+ "num_stages": 4
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 64,
180
+ "BLOCK_SIZE_N": 256,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 4
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 64,
188
+ "BLOCK_SIZE_N": 256,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 4
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 64,
196
+ "BLOCK_SIZE_N": 256,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 4,
200
+ "num_stages": 4
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 64,
204
+ "BLOCK_SIZE_N": 256,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 16,
207
+ "num_warps": 4,
208
+ "num_stages": 4
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 64,
212
+ "BLOCK_SIZE_N": 256,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 4
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 64,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 64,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 32,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 64,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 4
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 32,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 64,
78
+ "GROUP_SIZE_M": 64,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 64,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 256,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 32,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 128,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 4,
152
+ "num_stages": 3
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 128,
156
+ "BLOCK_SIZE_N": 128,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 16,
159
+ "num_warps": 4,
160
+ "num_stages": 3
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 128,
164
+ "BLOCK_SIZE_N": 128,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 4,
168
+ "num_stages": 3
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 128,
172
+ "BLOCK_SIZE_N": 128,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 16,
175
+ "num_warps": 4,
176
+ "num_stages": 3
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 128,
180
+ "BLOCK_SIZE_N": 128,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 3
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 128,
188
+ "BLOCK_SIZE_N": 128,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 3
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 128,
196
+ "BLOCK_SIZE_N": 128,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 4,
200
+ "num_stages": 3
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 128,
204
+ "BLOCK_SIZE_N": 128,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 16,
207
+ "num_warps": 4,
208
+ "num_stages": 3
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 128,
212
+ "BLOCK_SIZE_N": 128,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 3
217
+ }
218
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 64,
22
+ "GROUP_SIZE_M": 16,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 8,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 256,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 4
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 256,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 64,
77
+ "BLOCK_SIZE_K": 256,
78
+ "GROUP_SIZE_M": 64,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 32,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 32,
95
+ "num_warps": 4,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 8,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 256,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 4
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 256,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 32,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 256,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 32,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 8,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 256,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 64,
22
+ "GROUP_SIZE_M": 64,
23
+ "num_warps": 8,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 256,
29
+ "BLOCK_SIZE_K": 64,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 8,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 64,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 64,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 256,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 4
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 64,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 32,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 32,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }
build/torch27-cxx11-cu118-x86_64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 64,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 8,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 8,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 32,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 8,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 256,
62
+ "GROUP_SIZE_M": 32,
63
+ "num_warps": 8,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 256,
70
+ "GROUP_SIZE_M": 32,
71
+ "num_warps": 8,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 256,
78
+ "GROUP_SIZE_M": 32,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 32,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 16,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ }
146
+ }