liangsu9988 commited on
Commit
ee00eb8
·
verified ·
1 Parent(s): 0821e64

Uploaded using `kernel-builder`.

Browse files
Files changed (21) hide show
  1. benchmarks/benchmark.py +17 -5
  2. build/torch212-cxx11-cu130-x86_64-linux/__init__.py +4 -4
  3. build/torch212-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_7781728.abi3.so → _fp4_gemm_cuda_b708035.abi3.so} +2 -2
  4. build/torch212-cxx11-cu130-x86_64-linux/_ops.py +3 -3
  5. build/torch212-cxx11-cu130-x86_64-linux/fp4_gemm/__init__.py +0 -26
  6. build/torch212-cxx11-cu130-x86_64-linux/metadata.json +7 -7
  7. build/torch212-cxx11-cu132-x86_64-linux/__init__.py +4 -4
  8. build/{torch213-cxx11-cu130-x86_64-linux/_fp4_gemm_cuda_7781728.abi3.so → torch212-cxx11-cu132-x86_64-linux/_fp4_gemm_cuda_b708035.abi3.so} +2 -2
  9. build/torch212-cxx11-cu132-x86_64-linux/_ops.py +3 -3
  10. build/torch212-cxx11-cu132-x86_64-linux/fp4_gemm/__init__.py +0 -26
  11. build/torch212-cxx11-cu132-x86_64-linux/metadata.json +7 -7
  12. build/torch213-cxx11-cu130-x86_64-linux/__init__.py +4 -4
  13. build/{torch212-cxx11-cu132-x86_64-linux/_fp4_gemm_cuda_7781728.abi3.so → torch213-cxx11-cu130-x86_64-linux/_fp4_gemm_cuda_b708035.abi3.so} +2 -2
  14. build/torch213-cxx11-cu130-x86_64-linux/_ops.py +3 -3
  15. build/torch213-cxx11-cu130-x86_64-linux/fp4_gemm/__init__.py +0 -26
  16. build/torch213-cxx11-cu130-x86_64-linux/metadata.json +7 -7
  17. build/torch213-cxx11-cu132-x86_64-linux/__init__.py +4 -4
  18. build/torch213-cxx11-cu132-x86_64-linux/{_fp4_gemm_cuda_7781728.abi3.so → _fp4_gemm_cuda_b708035.abi3.so} +2 -2
  19. build/torch213-cxx11-cu132-x86_64-linux/_ops.py +3 -3
  20. build/torch213-cxx11-cu132-x86_64-linux/fp4_gemm/__init__.py +0 -26
  21. build/torch213-cxx11-cu132-x86_64-linux/metadata.json +7 -7
benchmarks/benchmark.py CHANGED
@@ -6,6 +6,7 @@ from __future__ import annotations
6
  import argparse
7
  import importlib.util
8
  import json
 
9
  import sys
10
  from dataclasses import asdict, dataclass
11
  from pathlib import Path
@@ -79,7 +80,8 @@ def bench_case(helpers, ops, native, name: str, shape: tuple[int, int, int], war
79
  torch_compile_us = measure(compiled_ref, warmup, iters)
80
  stream = torch.cuda.current_stream().cuda_stream
81
  results: list[BenchResult] = []
82
- for variant in (0, 1, 2):
 
83
  out = torch.empty((m, n), device="cuda", dtype=torch.bfloat16)
84
  ops.nvfp4_gemm_bf16(a_packed, b_packed, sfa, sfb, out, 1.0, variant)
85
  torch.cuda.synchronize()
@@ -89,11 +91,14 @@ def bench_case(helpers, ops, native, name: str, shape: tuple[int, int, int], war
89
  warmup,
90
  iters,
91
  )
 
 
 
92
  native_function = (
93
  native.fp4_w4a16_gemm_sm120_bf16out
94
- if variant == 0
95
  else native.fp4_w4a16_gemm_sm120_bf16out_widen
96
- if variant == 1
97
  else native.fp4_w4a16_gemm_sm120_bf16out_pingpong
98
  )
99
  native_us = measure(
@@ -140,14 +145,19 @@ def main() -> int:
140
  parser = argparse.ArgumentParser()
141
  parser.add_argument("--backend", choices=["source", "installed"], default="source")
142
  parser.add_argument("--artifact", default=None)
143
- parser.add_argument("--mode", choices=["smoke", "headline"], default="headline")
 
 
144
  parser.add_argument("--warmup", type=int, default=20)
145
  parser.add_argument("--iterations", type=int, default=100)
146
  parser.add_argument("--json-out", default=None)
147
  args = parser.parse_args()
148
 
149
  helpers = load_helpers()
150
- sys.path.insert(0, str(ROOT.parent / "official" / "FlashRT"))
 
 
 
151
  try:
152
  import flash_rt.flash_rt_kernels as native
153
  finally:
@@ -168,6 +178,8 @@ def main() -> int:
168
  }
169
  if args.mode == "smoke":
170
  shapes = {"small_m16_n128_k128": shapes["small_m16_n128_k128"]}
 
 
171
  results: list[BenchResult] = []
172
  for name, shape in shapes.items():
173
  results.extend(
 
6
  import argparse
7
  import importlib.util
8
  import json
9
+ import os
10
  import sys
11
  from dataclasses import asdict, dataclass
12
  from pathlib import Path
 
80
  torch_compile_us = measure(compiled_ref, warmup, iters)
81
  stream = torch.cuda.current_stream().cuda_stream
82
  results: list[BenchResult] = []
83
+ variants = (-1, 0, 1, 2) if torch.cuda.get_device_capability(0) == (11, 0) else (0, 1, 2)
84
+ for variant in variants:
85
  out = torch.empty((m, n), device="cuda", dtype=torch.bfloat16)
86
  ops.nvfp4_gemm_bf16(a_packed, b_packed, sfa, sfb, out, 1.0, variant)
87
  torch.cuda.synchronize()
 
91
  warmup,
92
  iters,
93
  )
94
+ native_variant = variant
95
+ if native_variant < 0:
96
+ native_variant = helpers.select_sm110_variant(shape)
97
  native_function = (
98
  native.fp4_w4a16_gemm_sm120_bf16out
99
+ if native_variant == 0
100
  else native.fp4_w4a16_gemm_sm120_bf16out_widen
101
+ if native_variant == 1
102
  else native.fp4_w4a16_gemm_sm120_bf16out_pingpong
103
  )
104
  native_us = measure(
 
145
  parser = argparse.ArgumentParser()
146
  parser.add_argument("--backend", choices=["source", "installed"], default="source")
147
  parser.add_argument("--artifact", default=None)
148
+ parser.add_argument(
149
+ "--mode", choices=["smoke", "headline", "thor-models"], default="headline"
150
+ )
151
  parser.add_argument("--warmup", type=int, default=20)
152
  parser.add_argument("--iterations", type=int, default=100)
153
  parser.add_argument("--json-out", default=None)
154
  args = parser.parse_args()
155
 
156
  helpers = load_helpers()
157
+ native_root = Path(
158
+ os.environ.get("FLASHRT_NATIVE_ROOT", str(ROOT.parent / "official" / "FlashRT"))
159
+ )
160
+ sys.path.insert(0, str(native_root))
161
  try:
162
  import flash_rt.flash_rt_kernels as native
163
  finally:
 
178
  }
179
  if args.mode == "smoke":
180
  shapes = {"small_m16_n128_k128": shapes["small_m16_n128_k128"]}
181
+ elif args.mode == "thor-models":
182
+ shapes = dict(helpers.SM110_SHAPES)
183
  results: list[BenchResult] = []
184
  for name, shape in shapes.items():
185
  results.extend(
build/torch212-cxx11-cu130-x86_64-linux/__init__.py CHANGED
@@ -31,7 +31,7 @@ def _linear_fake(
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
- variant: int = 0,
35
  ) -> None:
36
  return None
37
 
@@ -44,7 +44,7 @@ def _legacy_linear_fake(
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
- variant: int = 0,
48
  ) -> None:
49
  return None
50
 
@@ -117,7 +117,7 @@ def nvfp4_gemm_bf16(
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
- variant: int = 0,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
@@ -132,7 +132,7 @@ def fp4_w4a16_linear_bf16(
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
- variant: int = 0,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
 
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
+ variant: int = -1,
35
  ) -> None:
36
  return None
37
 
 
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
+ variant: int = -1,
48
  ) -> None:
49
  return None
50
 
 
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
+ variant: int = -1,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
 
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
+ variant: int = -1,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
build/torch212-cxx11-cu130-x86_64-linux/{_fp4_gemm_cuda_7781728.abi3.so → _fp4_gemm_cuda_b708035.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:613e52c9f5c1c0e3bc1ab0baa45cefcb88af74b53393cbb280bbc1d0128ecebe
3
- size 1731384
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e4c05290522d8245fcf823d1ea679d2e54a1b6e5dc757e57e3d4ad9bf550d124
3
+ size 2213120
build/torch212-cxx11-cu130-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_7781728
3
- ops = torch.ops._fp4_gemm_cuda_7781728
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_7781728::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_b708035
3
+ ops = torch.ops._fp4_gemm_cuda_b708035
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_b708035::{op_name}"
build/torch212-cxx11-cu130-x86_64-linux/fp4_gemm/__init__.py DELETED
@@ -1,26 +0,0 @@
1
- import ctypes
2
- import importlib.util
3
- import sys
4
- from pathlib import Path
5
- from types import ModuleType
6
-
7
-
8
- def _import_from_path(file_path: Path) -> ModuleType:
9
- # We cannot use the module name as-is, after adding it to `sys.modules`,
10
- # it would also be used for other imports. So, we make a module name that
11
- # depends on the path for it to be unique using the hex-encoded hash of
12
- # the path.
13
- path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
- module_name = path_hash
15
- spec = importlib.util.spec_from_file_location(module_name, file_path)
16
- if spec is None:
17
- raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
- module = importlib.util.module_from_spec(spec)
19
- if module is None:
20
- raise ImportError(f"Cannot load module {module_name} from spec")
21
- sys.modules[module_name] = module
22
- spec.loader.exec_module(module) # type: ignore
23
- return module
24
-
25
-
26
- globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
build/torch212-cxx11-cu130-x86_64-linux/metadata.json CHANGED
@@ -1,32 +1,32 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_7781728",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
 
10
  "12.0a"
11
  ]
12
  },
13
  "digest": {
14
  "algorithm": "sha256",
15
  "files": {
16
- "__init__.py": "FD6tCd6u2NRYfLKq1Uz7OYi+tZV5o9GDAb+KN3GX2oU=",
17
- "_fp4_gemm_cuda_7781728.abi3.so": "YT5SyfXBwOO8GrC6pFzvy4ivdLUzk8uygLvB0BKOzr4=",
18
- "_ops.py": "4wRBLd5HIJqgJskLFCRBmKJpvQfYeE9AmKTqIPQtPnw=",
19
- "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
- "sha": "19aaa6421e674e9fecc352bbae6eab81d19a6bf4",
26
  "dirty": false
27
  },
28
  "kernel": {
29
- "sha": "77817286c522d8f7abf8d2cd873e1f84a79357b9",
30
  "dirty": false
31
  }
32
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_b708035",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
10
+ "11.0a",
11
  "12.0a"
12
  ]
13
  },
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "EJ8HPhVwOKhifotSgdwCmB6MRbtnP1S0QfUGOVascy0=",
18
+ "_fp4_gemm_cuda_b708035.abi3.so": "5MBSkFItgkX8+CPR6medLlShtuXcdX5X49Stm/VQ0SQ=",
19
+ "_ops.py": "1J2lcfXUXZPX8+zk7ydYdEIScP8Kc745y6jTWOHAFW4="
 
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
+ "sha": "d720fa90fb9cd92d1bc60a9dc5c55bef2aafabb8",
26
  "dirty": false
27
  },
28
  "kernel": {
29
+ "sha": "b708035936a53871c033c5c4f49f1635bd74ad57",
30
  "dirty": false
31
  }
32
  }
build/torch212-cxx11-cu132-x86_64-linux/__init__.py CHANGED
@@ -31,7 +31,7 @@ def _linear_fake(
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
- variant: int = 0,
35
  ) -> None:
36
  return None
37
 
@@ -44,7 +44,7 @@ def _legacy_linear_fake(
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
- variant: int = 0,
48
  ) -> None:
49
  return None
50
 
@@ -117,7 +117,7 @@ def nvfp4_gemm_bf16(
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
- variant: int = 0,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
@@ -132,7 +132,7 @@ def fp4_w4a16_linear_bf16(
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
- variant: int = 0,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
 
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
+ variant: int = -1,
35
  ) -> None:
36
  return None
37
 
 
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
+ variant: int = -1,
48
  ) -> None:
49
  return None
50
 
 
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
+ variant: int = -1,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
 
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
+ variant: int = -1,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
build/{torch213-cxx11-cu130-x86_64-linux/_fp4_gemm_cuda_7781728.abi3.so → torch212-cxx11-cu132-x86_64-linux/_fp4_gemm_cuda_b708035.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8094f1a754b442779a3422dca4cb8626602bb38a5e5a3804cff765d0247da470
3
- size 1727128
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a620c09ef02f710ab141393d479ed09ca13321619b3ee8a73d5be605c62272e4
3
+ size 2217168
build/torch212-cxx11-cu132-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_7781728
3
- ops = torch.ops._fp4_gemm_cuda_7781728
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_7781728::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_b708035
3
+ ops = torch.ops._fp4_gemm_cuda_b708035
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_b708035::{op_name}"
build/torch212-cxx11-cu132-x86_64-linux/fp4_gemm/__init__.py DELETED
@@ -1,26 +0,0 @@
1
- import ctypes
2
- import importlib.util
3
- import sys
4
- from pathlib import Path
5
- from types import ModuleType
6
-
7
-
8
- def _import_from_path(file_path: Path) -> ModuleType:
9
- # We cannot use the module name as-is, after adding it to `sys.modules`,
10
- # it would also be used for other imports. So, we make a module name that
11
- # depends on the path for it to be unique using the hex-encoded hash of
12
- # the path.
13
- path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
- module_name = path_hash
15
- spec = importlib.util.spec_from_file_location(module_name, file_path)
16
- if spec is None:
17
- raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
- module = importlib.util.module_from_spec(spec)
19
- if module is None:
20
- raise ImportError(f"Cannot load module {module_name} from spec")
21
- sys.modules[module_name] = module
22
- spec.loader.exec_module(module) # type: ignore
23
- return module
24
-
25
-
26
- globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
build/torch212-cxx11-cu132-x86_64-linux/metadata.json CHANGED
@@ -1,32 +1,32 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_7781728",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
 
10
  "12.0a"
11
  ]
12
  },
13
  "digest": {
14
  "algorithm": "sha256",
15
  "files": {
16
- "__init__.py": "FD6tCd6u2NRYfLKq1Uz7OYi+tZV5o9GDAb+KN3GX2oU=",
17
- "_fp4_gemm_cuda_7781728.abi3.so": "p+zkZzz86MQv5Yr8EeLAXSPohAj6bv97A9+IUnl9Q1A=",
18
- "_ops.py": "4wRBLd5HIJqgJskLFCRBmKJpvQfYeE9AmKTqIPQtPnw=",
19
- "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
- "sha": "19aaa6421e674e9fecc352bbae6eab81d19a6bf4",
26
  "dirty": false
27
  },
28
  "kernel": {
29
- "sha": "77817286c522d8f7abf8d2cd873e1f84a79357b9",
30
  "dirty": false
31
  }
32
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_b708035",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
10
+ "11.0a",
11
  "12.0a"
12
  ]
13
  },
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "EJ8HPhVwOKhifotSgdwCmB6MRbtnP1S0QfUGOVascy0=",
18
+ "_fp4_gemm_cuda_b708035.abi3.so": "piDAnvAvcQqxQTk9R57QnKEzIWGbPuinPVvmBcYicuQ=",
19
+ "_ops.py": "1J2lcfXUXZPX8+zk7ydYdEIScP8Kc745y6jTWOHAFW4="
 
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
+ "sha": "d720fa90fb9cd92d1bc60a9dc5c55bef2aafabb8",
26
  "dirty": false
27
  },
28
  "kernel": {
29
+ "sha": "b708035936a53871c033c5c4f49f1635bd74ad57",
30
  "dirty": false
31
  }
32
  }
build/torch213-cxx11-cu130-x86_64-linux/__init__.py CHANGED
@@ -31,7 +31,7 @@ def _linear_fake(
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
- variant: int = 0,
35
  ) -> None:
36
  return None
37
 
@@ -44,7 +44,7 @@ def _legacy_linear_fake(
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
- variant: int = 0,
48
  ) -> None:
49
  return None
50
 
@@ -117,7 +117,7 @@ def nvfp4_gemm_bf16(
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
- variant: int = 0,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
@@ -132,7 +132,7 @@ def fp4_w4a16_linear_bf16(
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
- variant: int = 0,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
 
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
+ variant: int = -1,
35
  ) -> None:
36
  return None
37
 
 
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
+ variant: int = -1,
48
  ) -> None:
49
  return None
50
 
 
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
+ variant: int = -1,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
 
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
+ variant: int = -1,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
build/{torch212-cxx11-cu132-x86_64-linux/_fp4_gemm_cuda_7781728.abi3.so → torch213-cxx11-cu130-x86_64-linux/_fp4_gemm_cuda_b708035.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a7ece4673cfce8c42fe58afc11e2c05d23e88408fa6eff7b03df8852797d4350
3
- size 1731432
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0e47aa04752fe715e5979aa938632b78f333c421aa738defc7457fd90e31e4f0
3
+ size 2212960
build/torch213-cxx11-cu130-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_7781728
3
- ops = torch.ops._fp4_gemm_cuda_7781728
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_7781728::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_b708035
3
+ ops = torch.ops._fp4_gemm_cuda_b708035
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_b708035::{op_name}"
build/torch213-cxx11-cu130-x86_64-linux/fp4_gemm/__init__.py DELETED
@@ -1,26 +0,0 @@
1
- import ctypes
2
- import importlib.util
3
- import sys
4
- from pathlib import Path
5
- from types import ModuleType
6
-
7
-
8
- def _import_from_path(file_path: Path) -> ModuleType:
9
- # We cannot use the module name as-is, after adding it to `sys.modules`,
10
- # it would also be used for other imports. So, we make a module name that
11
- # depends on the path for it to be unique using the hex-encoded hash of
12
- # the path.
13
- path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
- module_name = path_hash
15
- spec = importlib.util.spec_from_file_location(module_name, file_path)
16
- if spec is None:
17
- raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
- module = importlib.util.module_from_spec(spec)
19
- if module is None:
20
- raise ImportError(f"Cannot load module {module_name} from spec")
21
- sys.modules[module_name] = module
22
- spec.loader.exec_module(module) # type: ignore
23
- return module
24
-
25
-
26
- globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
build/torch213-cxx11-cu130-x86_64-linux/metadata.json CHANGED
@@ -1,32 +1,32 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_7781728",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
 
10
  "12.0a"
11
  ]
12
  },
13
  "digest": {
14
  "algorithm": "sha256",
15
  "files": {
16
- "__init__.py": "FD6tCd6u2NRYfLKq1Uz7OYi+tZV5o9GDAb+KN3GX2oU=",
17
- "_fp4_gemm_cuda_7781728.abi3.so": "gJTxp1S0QneaNCLcpMuGJmArs4peWjgEz/dl0CR9pHA=",
18
- "_ops.py": "4wRBLd5HIJqgJskLFCRBmKJpvQfYeE9AmKTqIPQtPnw=",
19
- "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
- "sha": "19aaa6421e674e9fecc352bbae6eab81d19a6bf4",
26
  "dirty": false
27
  },
28
  "kernel": {
29
- "sha": "77817286c522d8f7abf8d2cd873e1f84a79357b9",
30
  "dirty": false
31
  }
32
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_b708035",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
10
+ "11.0a",
11
  "12.0a"
12
  ]
13
  },
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "EJ8HPhVwOKhifotSgdwCmB6MRbtnP1S0QfUGOVascy0=",
18
+ "_fp4_gemm_cuda_b708035.abi3.so": "DkeqBHUv5xXll5qpOGMrePMzxCGqc43vx0V/2Q4x5PA=",
19
+ "_ops.py": "1J2lcfXUXZPX8+zk7ydYdEIScP8Kc745y6jTWOHAFW4="
 
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
+ "sha": "d720fa90fb9cd92d1bc60a9dc5c55bef2aafabb8",
26
  "dirty": false
27
  },
28
  "kernel": {
29
+ "sha": "b708035936a53871c033c5c4f49f1635bd74ad57",
30
  "dirty": false
31
  }
32
  }
build/torch213-cxx11-cu132-x86_64-linux/__init__.py CHANGED
@@ -31,7 +31,7 @@ def _linear_fake(
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
- variant: int = 0,
35
  ) -> None:
36
  return None
37
 
@@ -44,7 +44,7 @@ def _legacy_linear_fake(
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
- variant: int = 0,
48
  ) -> None:
49
  return None
50
 
@@ -117,7 +117,7 @@ def nvfp4_gemm_bf16(
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
- variant: int = 0,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
@@ -132,7 +132,7 @@ def fp4_w4a16_linear_bf16(
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
- variant: int = 0,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
 
31
  sfb: torch.Tensor,
32
  out: torch.Tensor,
33
  alpha: float = 1.0,
34
+ variant: int = -1,
35
  ) -> None:
36
  return None
37
 
 
44
  sfb: torch.Tensor,
45
  out: torch.Tensor,
46
  alpha: float = 1.0,
47
+ variant: int = -1,
48
  ) -> None:
49
  return None
50
 
 
117
  sfb: torch.Tensor,
118
  alpha: float = 1.0,
119
  out: torch.Tensor | None = None,
120
+ variant: int = -1,
121
  ) -> torch.Tensor:
122
  if out is None:
123
  out = torch.empty((a_packed.shape[0], b_packed.shape[0]), device=a_packed.device, dtype=torch.bfloat16)
 
132
  sfb: torch.Tensor,
133
  alpha: float = 1.0,
134
  out: torch.Tensor | None = None,
135
+ variant: int = -1,
136
  ) -> torch.Tensor:
137
  """Compatibility alias for :func:`nvfp4_gemm_bf16`."""
138
  return nvfp4_gemm_bf16(
build/torch213-cxx11-cu132-x86_64-linux/{_fp4_gemm_cuda_7781728.abi3.so → _fp4_gemm_cuda_b708035.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c9a7111f4e0f3db4a1766b796c906f968cf3b7253b2e3ba2ef2228263d19f5fd
3
- size 1731272
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05ba221f1c0eea9f69c70907e6c5b3ccd5b8a427b051244aeeb6c4a06aee0d2c
3
+ size 2217008
build/torch213-cxx11-cu132-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _fp4_gemm_cuda_7781728
3
- ops = torch.ops._fp4_gemm_cuda_7781728
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_fp4_gemm_cuda_7781728::{op_name}"
 
1
  import torch
2
+ from . import _fp4_gemm_cuda_b708035
3
+ ops = torch.ops._fp4_gemm_cuda_b708035
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_fp4_gemm_cuda_b708035::{op_name}"
build/torch213-cxx11-cu132-x86_64-linux/fp4_gemm/__init__.py DELETED
@@ -1,26 +0,0 @@
1
- import ctypes
2
- import importlib.util
3
- import sys
4
- from pathlib import Path
5
- from types import ModuleType
6
-
7
-
8
- def _import_from_path(file_path: Path) -> ModuleType:
9
- # We cannot use the module name as-is, after adding it to `sys.modules`,
10
- # it would also be used for other imports. So, we make a module name that
11
- # depends on the path for it to be unique using the hex-encoded hash of
12
- # the path.
13
- path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
- module_name = path_hash
15
- spec = importlib.util.spec_from_file_location(module_name, file_path)
16
- if spec is None:
17
- raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
- module = importlib.util.module_from_spec(spec)
19
- if module is None:
20
- raise ImportError(f"Cannot load module {module_name} from spec")
21
- sys.modules[module_name] = module
22
- spec.loader.exec_module(module) # type: ignore
23
- return module
24
-
25
-
26
- globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
build/torch213-cxx11-cu132-x86_64-linux/metadata.json CHANGED
@@ -1,32 +1,32 @@
1
  {
2
  "name": "fp4-gemm",
3
- "id": "_fp4_gemm_cuda_7781728",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
 
10
  "12.0a"
11
  ]
12
  },
13
  "digest": {
14
  "algorithm": "sha256",
15
  "files": {
16
- "__init__.py": "FD6tCd6u2NRYfLKq1Uz7OYi+tZV5o9GDAb+KN3GX2oU=",
17
- "_fp4_gemm_cuda_7781728.abi3.so": "yacRH04PPbShdmt5bJBvlozztyU7Ljui7yIoJj0Z9f0=",
18
- "_ops.py": "4wRBLd5HIJqgJskLFCRBmKJpvQfYeE9AmKTqIPQtPnw=",
19
- "fp4_gemm/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY="
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
- "sha": "19aaa6421e674e9fecc352bbae6eab81d19a6bf4",
26
  "dirty": false
27
  },
28
  "kernel": {
29
- "sha": "77817286c522d8f7abf8d2cd873e1f84a79357b9",
30
  "dirty": false
31
  }
32
  }
 
1
  {
2
  "name": "fp4-gemm",
3
+ "id": "_fp4_gemm_cuda_b708035",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "python-depends": [],
7
  "backend": {
8
  "type": "cuda",
9
  "archs": [
10
+ "11.0a",
11
  "12.0a"
12
  ]
13
  },
14
  "digest": {
15
  "algorithm": "sha256",
16
  "files": {
17
+ "__init__.py": "EJ8HPhVwOKhifotSgdwCmB6MRbtnP1S0QfUGOVascy0=",
18
+ "_fp4_gemm_cuda_b708035.abi3.so": "BboiHxwO6p9pxwkH5sWzzNW4pCewUSRK7rbEoGruDSw=",
19
+ "_ops.py": "1J2lcfXUXZPX8+zk7ydYdEIScP8Kc745y6jTWOHAFW4="
 
20
  }
21
  },
22
  "provenance": {
23
  "kernel-builder": {
24
  "version": "0.17.0-dev0",
25
+ "sha": "d720fa90fb9cd92d1bc60a9dc5c55bef2aafabb8",
26
  "dirty": false
27
  },
28
  "kernel": {
29
+ "sha": "b708035936a53871c033c5c4f49f1635bd74ad57",
30
  "dirty": false
31
  }
32
  }