liangsu9988 commited on
Commit
ab02a6f
·
verified ·
1 Parent(s): 26784fd

Uploaded using `kernel-builder`.

Browse files
Files changed (24) hide show
  1. build/torch211-cxx11-cu128-x86_64-linux/__init__.py +22 -1
  2. build/torch211-cxx11-cu128-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} +1 -1
  3. build/torch211-cxx11-cu128-x86_64-linux/_ops.py +3 -3
  4. build/torch211-cxx11-cu128-x86_64-linux/metadata.json +5 -5
  5. build/torch211-cxx11-cu130-x86_64-linux/__init__.py +22 -1
  6. build/torch211-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} +1 -1
  7. build/torch211-cxx11-cu130-x86_64-linux/_ops.py +3 -3
  8. build/torch211-cxx11-cu130-x86_64-linux/metadata.json +5 -5
  9. build/torch212-cxx11-cu130-x86_64-linux/__init__.py +22 -1
  10. build/torch212-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} +1 -1
  11. build/torch212-cxx11-cu130-x86_64-linux/_ops.py +3 -3
  12. build/torch212-cxx11-cu130-x86_64-linux/metadata.json +5 -5
  13. build/torch212-cxx11-cu132-x86_64-linux/__init__.py +22 -1
  14. build/torch212-cxx11-cu132-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} +1 -1
  15. build/torch212-cxx11-cu132-x86_64-linux/_ops.py +3 -3
  16. build/torch212-cxx11-cu132-x86_64-linux/metadata.json +5 -5
  17. build/torch213-cxx11-cu130-x86_64-linux/__init__.py +22 -1
  18. build/torch213-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} +1 -1
  19. build/torch213-cxx11-cu130-x86_64-linux/_ops.py +3 -3
  20. build/torch213-cxx11-cu130-x86_64-linux/metadata.json +5 -5
  21. build/torch213-cxx11-cu132-x86_64-linux/__init__.py +22 -1
  22. build/torch213-cxx11-cu132-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} +1 -1
  23. build/torch213-cxx11-cu132-x86_64-linux/_ops.py +3 -3
  24. build/torch213-cxx11-cu132-x86_64-linux/metadata.json +5 -5
build/torch211-cxx11-cu128-x86_64-linux/__init__.py CHANGED
@@ -150,6 +150,19 @@ def _gemv_warpsplit_interleaved_fake(
150
  return None
151
 
152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
153
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
154
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
155
  if a_packed.shape[0] < 512:
@@ -733,8 +746,16 @@ def nvfp4_gemm_m256_bf16(
733
  ) -> torch.Tensor:
734
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
735
  if workspace is None:
 
 
 
 
 
 
 
 
736
  workspace = torch.empty(
737
- nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb),
738
  device=a_packed.device,
739
  dtype=torch.uint8,
740
  )
 
150
  return None
151
 
152
 
153
+ @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_workspace_size"))
154
+ def _m256_workspace_size_fake(a_packed, b_packed, sfa, sfb) -> int:
155
+ del sfa, sfb
156
+ if a_packed.ndim != 2 or b_packed.ndim != 2:
157
+ raise RuntimeError("a_packed and b_packed must be rank-2")
158
+ if a_packed.shape[0] < 512:
159
+ raise RuntimeError("the M256 tier requires M >= 512")
160
+ if a_packed.shape[1] != b_packed.shape[1]:
161
+ raise RuntimeError("packed A and B K dimensions must match")
162
+ # The current persistent M256 CUTLASS schedule has no auxiliary storage.
163
+ return 0
164
+
165
+
166
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
167
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
168
  if a_packed.shape[0] < 512:
 
746
  ) -> torch.Tensor:
747
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
748
  if workspace is None:
749
+ # Dynamo cannot place a custom operator returning a Python scalar in
750
+ # an FX graph. This is exact for the current persistent M256 schedule;
751
+ # eager setup still queries the native helper as a future-proof check.
752
+ workspace_size = (
753
+ 0
754
+ if torch.compiler.is_compiling()
755
+ else nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb)
756
+ )
757
  workspace = torch.empty(
758
+ workspace_size,
759
  device=a_packed.device,
760
  dtype=torch.uint8,
761
  )
build/torch211-cxx11-cu128-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5d384b35a1df0fce0389696709ac1d7ab6ca045a4e93b6ff389ea3a4666d118a
3
  size 2721416
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d1da1c9c5de85d62d2a39e29178eb3fb6e1fc4d772b04195fbcfd9683b63fcb5
3
  size 2721416
build/torch211-cxx11-cu128-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_e1f10b6
3
- ops = torch.ops._fp4_gemm_cuda_e1f10b6
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_e1f10b6::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_dc57d15
3
+ ops = torch.ops._fp4_gemm_cuda_dc57d15
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_dc57d15::{op_name}"
build/torch211-cxx11-cu128-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_e1f10b6",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
@@ -13,9 +13,9 @@
13
  "digest": {
14
  "algorithm": "sha256",
15
  "files": {
16
- "__init__.py": "x7M8jFG4z1Mx/zhAh30IsOFooDn28i8uyrkkNIjyG+c=",
17
- "_fp4_gemm_cuda_e1f10b6.abi3.so": "XThLNaHfD84DiWlnCawderbKBFpOk7b/OJ6jpGZtEYo=",
18
- "_ops.py": "FGdLd66XdfjhguPY8EEn/zd8T/lGPpQP+r8VJnZaCPE=",
19
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
20
  }
21
  },
@@ -26,7 +26,7 @@
26
  "dirty": false
27
  },
28
  "kernel": {
29
- "sha": "e1f10b68fd3d9f68ecd85895128283aa6768ea6f",
30
  "dirty": false
31
  }
32
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_dc57d15",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
 
13
  "digest": {
14
  "algorithm": "sha256",
15
  "files": {
16
+ "__init__.py": "XycjAPC24qtK6IH+GtF/jrPmWnB6DyNgSny26ep3w1Y=",
17
+ "_fp4_gemm_cuda_dc57d15.abi3.so": "0docnF3oXWLSo54pF46z+24fxNdysEGV+8/ZaDtj/LU=",
18
+ "_ops.py": "TM5HQsh2RfinhrEZ4U0XyWF2hFyrPWpxR79FQ41D5jk=",
19
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
20
  }
21
  },
 
26
  "dirty": false
27
  },
28
  "kernel": {
29
+ "sha": "dc57d15b8bb1d4b52d82012f7f789ae5ae6db17f",
30
  "dirty": false
31
  }
32
  }
build/torch211-cxx11-cu130-x86_64-linux/__init__.py CHANGED
@@ -150,6 +150,19 @@ def _gemv_warpsplit_interleaved_fake(
150
  return None
151
 
152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
153
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
154
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
155
  if a_packed.shape[0] < 512:
@@ -733,8 +746,16 @@ def nvfp4_gemm_m256_bf16(
733
  ) -> torch.Tensor:
734
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
735
  if workspace is None:
 
 
 
 
 
 
 
 
736
  workspace = torch.empty(
737
- nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb),
738
  device=a_packed.device,
739
  dtype=torch.uint8,
740
  )
 
150
  return None
151
 
152
 
153
+ @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_workspace_size"))
154
+ def _m256_workspace_size_fake(a_packed, b_packed, sfa, sfb) -> int:
155
+ del sfa, sfb
156
+ if a_packed.ndim != 2 or b_packed.ndim != 2:
157
+ raise RuntimeError("a_packed and b_packed must be rank-2")
158
+ if a_packed.shape[0] < 512:
159
+ raise RuntimeError("the M256 tier requires M >= 512")
160
+ if a_packed.shape[1] != b_packed.shape[1]:
161
+ raise RuntimeError("packed A and B K dimensions must match")
162
+ # The current persistent M256 CUTLASS schedule has no auxiliary storage.
163
+ return 0
164
+
165
+
166
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
167
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
168
  if a_packed.shape[0] < 512:
 
746
  ) -> torch.Tensor:
747
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
748
  if workspace is None:
749
+ # Dynamo cannot place a custom operator returning a Python scalar in
750
+ # an FX graph. This is exact for the current persistent M256 schedule;
751
+ # eager setup still queries the native helper as a future-proof check.
752
+ workspace_size = (
753
+ 0
754
+ if torch.compiler.is_compiling()
755
+ else nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb)
756
+ )
757
  workspace = torch.empty(
758
+ workspace_size,
759
  device=a_packed.device,
760
  dtype=torch.uint8,
761
  )
build/torch211-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b580778bebed486d4865d0e9b5dd09282d2545ea7417f961b0acdb3289874e5a
3
  size 8564792
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:687ec5c797f1519fed83a51f82bc7ecf53059fbe35843981d86fefe4143c3b87
3
  size 8564792
build/torch211-cxx11-cu130-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_e1f10b6
3
- ops = torch.ops._fp4_gemm_cuda_e1f10b6
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_e1f10b6::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_dc57d15
3
+ ops = torch.ops._fp4_gemm_cuda_dc57d15
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_dc57d15::{op_name}"
build/torch211-cxx11-cu130-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_e1f10b6",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
@@ -14,9 +14,9 @@
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
- "__init__.py": "x7M8jFG4z1Mx/zhAh30IsOFooDn28i8uyrkkNIjyG+c=",
18
- "_fp4_gemm_cuda_e1f10b6.abi3.so": "tYB3i+vtSG1IZdDptd0JKC0lRep0F/lhsKzbMomHTlo=",
19
- "_ops.py": "FGdLd66XdfjhguPY8EEn/zd8T/lGPpQP+r8VJnZaCPE=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
@@ -27,7 +27,7 @@
27
  "dirty": false
28
  },
29
  "kernel": {
30
- "sha": "e1f10b68fd3d9f68ecd85895128283aa6768ea6f",
31
  "dirty": false
32
  }
33
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_dc57d15",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
 
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "XycjAPC24qtK6IH+GtF/jrPmWnB6DyNgSny26ep3w1Y=",
18
+ "_fp4_gemm_cuda_dc57d15.abi3.so": "aH7Fx5fxUZ/tg6Ufgrx+z1MFn741hDmB2G/v5BQ8O4c=",
19
+ "_ops.py": "TM5HQsh2RfinhrEZ4U0XyWF2hFyrPWpxR79FQ41D5jk=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
 
27
  "dirty": false
28
  },
29
  "kernel": {
30
+ "sha": "dc57d15b8bb1d4b52d82012f7f789ae5ae6db17f",
31
  "dirty": false
32
  }
33
  }
build/torch212-cxx11-cu130-x86_64-linux/__init__.py CHANGED
@@ -150,6 +150,19 @@ def _gemv_warpsplit_interleaved_fake(
150
  return None
151
 
152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
153
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
154
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
155
  if a_packed.shape[0] < 512:
@@ -733,8 +746,16 @@ def nvfp4_gemm_m256_bf16(
733
  ) -> torch.Tensor:
734
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
735
  if workspace is None:
 
 
 
 
 
 
 
 
736
  workspace = torch.empty(
737
- nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb),
738
  device=a_packed.device,
739
  dtype=torch.uint8,
740
  )
 
150
  return None
151
 
152
 
153
+ @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_workspace_size"))
154
+ def _m256_workspace_size_fake(a_packed, b_packed, sfa, sfb) -> int:
155
+ del sfa, sfb
156
+ if a_packed.ndim != 2 or b_packed.ndim != 2:
157
+ raise RuntimeError("a_packed and b_packed must be rank-2")
158
+ if a_packed.shape[0] < 512:
159
+ raise RuntimeError("the M256 tier requires M >= 512")
160
+ if a_packed.shape[1] != b_packed.shape[1]:
161
+ raise RuntimeError("packed A and B K dimensions must match")
162
+ # The current persistent M256 CUTLASS schedule has no auxiliary storage.
163
+ return 0
164
+
165
+
166
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
167
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
168
  if a_packed.shape[0] < 512:
 
746
  ) -> torch.Tensor:
747
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
748
  if workspace is None:
749
+ # Dynamo cannot place a custom operator returning a Python scalar in
750
+ # an FX graph. This is exact for the current persistent M256 schedule;
751
+ # eager setup still queries the native helper as a future-proof check.
752
+ workspace_size = (
753
+ 0
754
+ if torch.compiler.is_compiling()
755
+ else nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb)
756
+ )
757
  workspace = torch.empty(
758
+ workspace_size,
759
  device=a_packed.device,
760
  dtype=torch.uint8,
761
  )
build/torch212-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f950eb405f45cdfa822aa2ee646f87ddcda0c0c6de4d26a81b51947dd50bf3d7
3
  size 8566680
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0f26894cd8000a3c464811b255f07732758d8d53521c2b84c305e9cc07d96619
3
  size 8566680
build/torch212-cxx11-cu130-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_e1f10b6
3
- ops = torch.ops._fp4_gemm_cuda_e1f10b6
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_e1f10b6::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_dc57d15
3
+ ops = torch.ops._fp4_gemm_cuda_dc57d15
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_dc57d15::{op_name}"
build/torch212-cxx11-cu130-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_e1f10b6",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
@@ -14,9 +14,9 @@
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
- "__init__.py": "x7M8jFG4z1Mx/zhAh30IsOFooDn28i8uyrkkNIjyG+c=",
18
- "_fp4_gemm_cuda_e1f10b6.abi3.so": "+VDrQF9FzfqCKqLuZG+H3c2gwMbeTSaoG1GUfdUL89c=",
19
- "_ops.py": "FGdLd66XdfjhguPY8EEn/zd8T/lGPpQP+r8VJnZaCPE=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
@@ -27,7 +27,7 @@
27
  "dirty": false
28
  },
29
  "kernel": {
30
- "sha": "e1f10b68fd3d9f68ecd85895128283aa6768ea6f",
31
  "dirty": false
32
  }
33
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_dc57d15",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
 
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "XycjAPC24qtK6IH+GtF/jrPmWnB6DyNgSny26ep3w1Y=",
18
+ "_fp4_gemm_cuda_dc57d15.abi3.so": "DyaJTNgACjxGSBGyVfB3MnWNjVNSHCuEwwXpzAfZZhk=",
19
+ "_ops.py": "TM5HQsh2RfinhrEZ4U0XyWF2hFyrPWpxR79FQ41D5jk=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
 
27
  "dirty": false
28
  },
29
  "kernel": {
30
+ "sha": "dc57d15b8bb1d4b52d82012f7f789ae5ae6db17f",
31
  "dirty": false
32
  }
33
  }
build/torch212-cxx11-cu132-x86_64-linux/__init__.py CHANGED
@@ -150,6 +150,19 @@ def _gemv_warpsplit_interleaved_fake(
150
  return None
151
 
152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
153
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
154
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
155
  if a_packed.shape[0] < 512:
@@ -733,8 +746,16 @@ def nvfp4_gemm_m256_bf16(
733
  ) -> torch.Tensor:
734
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
735
  if workspace is None:
 
 
 
 
 
 
 
 
736
  workspace = torch.empty(
737
- nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb),
738
  device=a_packed.device,
739
  dtype=torch.uint8,
740
  )
 
150
  return None
151
 
152
 
153
+ @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_workspace_size"))
154
+ def _m256_workspace_size_fake(a_packed, b_packed, sfa, sfb) -> int:
155
+ del sfa, sfb
156
+ if a_packed.ndim != 2 or b_packed.ndim != 2:
157
+ raise RuntimeError("a_packed and b_packed must be rank-2")
158
+ if a_packed.shape[0] < 512:
159
+ raise RuntimeError("the M256 tier requires M >= 512")
160
+ if a_packed.shape[1] != b_packed.shape[1]:
161
+ raise RuntimeError("packed A and B K dimensions must match")
162
+ # The current persistent M256 CUTLASS schedule has no auxiliary storage.
163
+ return 0
164
+
165
+
166
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
167
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
168
  if a_packed.shape[0] < 512:
 
746
  ) -> torch.Tensor:
747
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
748
  if workspace is None:
749
+ # Dynamo cannot place a custom operator returning a Python scalar in
750
+ # an FX graph. This is exact for the current persistent M256 schedule;
751
+ # eager setup still queries the native helper as a future-proof check.
752
+ workspace_size = (
753
+ 0
754
+ if torch.compiler.is_compiling()
755
+ else nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb)
756
+ )
757
  workspace = torch.empty(
758
+ workspace_size,
759
  device=a_packed.device,
760
  dtype=torch.uint8,
761
  )
build/torch212-cxx11-cu132-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:20b51c0058864befdf526936c452585e317d43341059d3a924ca6797250ae23e
3
  size 8550344
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3b1fc2155f02132014a4ed86a37eb5c945d705651eeda4cf4fc294c4e4d5a442
3
  size 8550344
build/torch212-cxx11-cu132-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_e1f10b6
3
- ops = torch.ops._fp4_gemm_cuda_e1f10b6
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_e1f10b6::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_dc57d15
3
+ ops = torch.ops._fp4_gemm_cuda_dc57d15
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_dc57d15::{op_name}"
build/torch212-cxx11-cu132-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_e1f10b6",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
@@ -14,9 +14,9 @@
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
- "__init__.py": "x7M8jFG4z1Mx/zhAh30IsOFooDn28i8uyrkkNIjyG+c=",
18
- "_fp4_gemm_cuda_e1f10b6.abi3.so": "ILUcAFiGS+/fUmk2xFJYXjF9QzQQWdOpJMpnlyUK4j4=",
19
- "_ops.py": "FGdLd66XdfjhguPY8EEn/zd8T/lGPpQP+r8VJnZaCPE=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
@@ -27,7 +27,7 @@
27
  "dirty": false
28
  },
29
  "kernel": {
30
- "sha": "e1f10b68fd3d9f68ecd85895128283aa6768ea6f",
31
  "dirty": false
32
  }
33
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_dc57d15",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
 
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "XycjAPC24qtK6IH+GtF/jrPmWnB6DyNgSny26ep3w1Y=",
18
+ "_fp4_gemm_cuda_dc57d15.abi3.so": "Ox/CFV8CEyAUpO2Go361yUXXBWUe7aTPT8KUxOTVpEI=",
19
+ "_ops.py": "TM5HQsh2RfinhrEZ4U0XyWF2hFyrPWpxR79FQ41D5jk=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
 
27
  "dirty": false
28
  },
29
  "kernel": {
30
+ "sha": "dc57d15b8bb1d4b52d82012f7f789ae5ae6db17f",
31
  "dirty": false
32
  }
33
  }
build/torch213-cxx11-cu130-x86_64-linux/__init__.py CHANGED
@@ -150,6 +150,19 @@ def _gemv_warpsplit_interleaved_fake(
150
  return None
151
 
152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
153
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
154
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
155
  if a_packed.shape[0] < 512:
@@ -733,8 +746,16 @@ def nvfp4_gemm_m256_bf16(
733
  ) -> torch.Tensor:
734
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
735
  if workspace is None:
 
 
 
 
 
 
 
 
736
  workspace = torch.empty(
737
- nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb),
738
  device=a_packed.device,
739
  dtype=torch.uint8,
740
  )
 
150
  return None
151
 
152
 
153
+ @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_workspace_size"))
154
+ def _m256_workspace_size_fake(a_packed, b_packed, sfa, sfb) -> int:
155
+ del sfa, sfb
156
+ if a_packed.ndim != 2 or b_packed.ndim != 2:
157
+ raise RuntimeError("a_packed and b_packed must be rank-2")
158
+ if a_packed.shape[0] < 512:
159
+ raise RuntimeError("the M256 tier requires M >= 512")
160
+ if a_packed.shape[1] != b_packed.shape[1]:
161
+ raise RuntimeError("packed A and B K dimensions must match")
162
+ # The current persistent M256 CUTLASS schedule has no auxiliary storage.
163
+ return 0
164
+
165
+
166
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
167
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
168
  if a_packed.shape[0] < 512:
 
746
  ) -> torch.Tensor:
747
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
748
  if workspace is None:
749
+ # Dynamo cannot place a custom operator returning a Python scalar in
750
+ # an FX graph. This is exact for the current persistent M256 schedule;
751
+ # eager setup still queries the native helper as a future-proof check.
752
+ workspace_size = (
753
+ 0
754
+ if torch.compiler.is_compiling()
755
+ else nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb)
756
+ )
757
  workspace = torch.empty(
758
+ workspace_size,
759
  device=a_packed.device,
760
  dtype=torch.uint8,
761
  )
build/torch213-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f37120b35de71e695142d6f77a82f68e97a0403647fd6b0a7f251a9e9d564101
3
  size 8566520
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a2e973560e69449ee79707984b7703d19e8f8ef9bba3902513887c57b64682c0
3
  size 8566520
build/torch213-cxx11-cu130-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_e1f10b6
3
- ops = torch.ops._fp4_gemm_cuda_e1f10b6
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_e1f10b6::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_dc57d15
3
+ ops = torch.ops._fp4_gemm_cuda_dc57d15
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_dc57d15::{op_name}"
build/torch213-cxx11-cu130-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_e1f10b6",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
@@ -14,9 +14,9 @@
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
- "__init__.py": "x7M8jFG4z1Mx/zhAh30IsOFooDn28i8uyrkkNIjyG+c=",
18
- "_fp4_gemm_cuda_e1f10b6.abi3.so": "83Egs13nHmlRQtb3eoL2jpegQDZH/WsKfyUanp1WQQE=",
19
- "_ops.py": "FGdLd66XdfjhguPY8EEn/zd8T/lGPpQP+r8VJnZaCPE=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
@@ -27,7 +27,7 @@
27
  "dirty": false
28
  },
29
  "kernel": {
30
- "sha": "e1f10b68fd3d9f68ecd85895128283aa6768ea6f",
31
  "dirty": false
32
  }
33
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_dc57d15",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
 
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "XycjAPC24qtK6IH+GtF/jrPmWnB6DyNgSny26ep3w1Y=",
18
+ "_fp4_gemm_cuda_dc57d15.abi3.so": "oulzVg5pRJ7nlweYS3cD0Z6Pjvm7o5AlE4h8V7ZGgsA=",
19
+ "_ops.py": "TM5HQsh2RfinhrEZ4U0XyWF2hFyrPWpxR79FQ41D5jk=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
 
27
  "dirty": false
28
  },
29
  "kernel": {
30
+ "sha": "dc57d15b8bb1d4b52d82012f7f789ae5ae6db17f",
31
  "dirty": false
32
  }
33
  }
build/torch213-cxx11-cu132-x86_64-linux/__init__.py CHANGED
@@ -150,6 +150,19 @@ def _gemv_warpsplit_interleaved_fake(
150
  return None
151
 
152
 
 
 
 
 
 
 
 
 
 
 
 
 
 
153
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
154
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
155
  if a_packed.shape[0] < 512:
@@ -733,8 +746,16 @@ def nvfp4_gemm_m256_bf16(
733
  ) -> torch.Tensor:
734
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
735
  if workspace is None:
 
 
 
 
 
 
 
 
736
  workspace = torch.empty(
737
- nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb),
738
  device=a_packed.device,
739
  dtype=torch.uint8,
740
  )
 
150
  return None
151
 
152
 
153
+ @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_workspace_size"))
154
+ def _m256_workspace_size_fake(a_packed, b_packed, sfa, sfb) -> int:
155
+ del sfa, sfb
156
+ if a_packed.ndim != 2 or b_packed.ndim != 2:
157
+ raise RuntimeError("a_packed and b_packed must be rank-2")
158
+ if a_packed.shape[0] < 512:
159
+ raise RuntimeError("the M256 tier requires M >= 512")
160
+ if a_packed.shape[1] != b_packed.shape[1]:
161
+ raise RuntimeError("packed A and B K dimensions must match")
162
+ # The current persistent M256 CUTLASS schedule has no auxiliary storage.
163
+ return 0
164
+
165
+
166
  @torch.library.register_fake(add_op_namespace_prefix("nvfp4_gemm_m256_bf16"))
167
  def _m256_fake(a_packed, b_packed, sfa, sfb, workspace, out, alpha: float = 1.0) -> None:
168
  if a_packed.shape[0] < 512:
 
746
  ) -> torch.Tensor:
747
  """SM120 M>=512 NVFP4 GEMM with a caller-owned graph-stable workspace."""
748
  if workspace is None:
749
+ # Dynamo cannot place a custom operator returning a Python scalar in
750
+ # an FX graph. This is exact for the current persistent M256 schedule;
751
+ # eager setup still queries the native helper as a future-proof check.
752
+ workspace_size = (
753
+ 0
754
+ if torch.compiler.is_compiling()
755
+ else nvfp4_gemm_m256_workspace_size(a_packed, b_packed, sfa, sfb)
756
+ )
757
  workspace = torch.empty(
758
+ workspace_size,
759
  device=a_packed.device,
760
  dtype=torch.uint8,
761
  )
build/torch213-cxx11-cu132-x86_64-linux/{_fp4_gemm_cuda_e1f10b6.abi3.so → _fp4_gemm_cuda_dc57d15.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a519e2bfab7a53c2a6564c2c932c91f8377213082a22471e45bf42a35c8c9cc8
3
  size 8550192
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c5c466015e9fda36acc0d9426f2c9281bb9e8af175afb1686edd37348a4f975
3
  size 8550192
build/torch213-cxx11-cu132-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_e1f10b6
3
- ops = torch.ops._fp4_gemm_cuda_e1f10b6
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_e1f10b6::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_dc57d15
3
+ ops = torch.ops._fp4_gemm_cuda_dc57d15
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_dc57d15::{op_name}"
build/torch213-cxx11-cu132-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_e1f10b6",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
@@ -14,9 +14,9 @@
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
- "__init__.py": "x7M8jFG4z1Mx/zhAh30IsOFooDn28i8uyrkkNIjyG+c=",
18
- "_fp4_gemm_cuda_e1f10b6.abi3.so": "pRniv6t6U8KmVkwskyyR+DdyEwgqIkceRb9Co1yMnMg=",
19
- "_ops.py": "FGdLd66XdfjhguPY8EEn/zd8T/lGPpQP+r8VJnZaCPE=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
@@ -27,7 +27,7 @@
27
  "dirty": false
28
  },
29
  "kernel": {
30
- "sha": "e1f10b68fd3d9f68ecd85895128283aa6768ea6f",
31
  "dirty": false
32
  }
33
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_dc57d15",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
 
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "XycjAPC24qtK6IH+GtF/jrPmWnB6DyNgSny26ep3w1Y=",
18
+ "_fp4_gemm_cuda_dc57d15.abi3.so": "PFxGYBXp/aNqzA2UJvLJKBu56K8XWvsWhu3Tc0ik+XU=",
19
+ "_ops.py": "TM5HQsh2RfinhrEZ4U0XyWF2hFyrPWpxR79FQ41D5jk=",
20
  "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
21
  }
22
  },
 
27
  "dirty": false
28
  },
29
  "kernel": {
30
+ "sha": "dc57d15b8bb1d4b52d82012f7f789ae5ae6db17f",
31
  "dirty": false
32
  }
33
  }