triton-windows 3.2.0.post11__cp39-cp39-win_amd64.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of triton-windows might be problematic. Click here for more details.
- triton/_C/libtriton.pyd +0 -0
- triton/__init__.py +85 -0
- triton/_internal_testing.py +123 -0
- triton/backends/__init__.py +50 -0
- triton/backends/amd/compiler.py +368 -0
- triton/backends/amd/driver.c +211 -0
- triton/backends/amd/driver.py +512 -0
- triton/backends/amd/include/hip/amd_detail/amd_channel_descriptor.h +358 -0
- triton/backends/amd/include/hip/amd_detail/amd_device_functions.h +1031 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_atomic.h +1612 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_bf16.h +1337 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_bfloat16.h +293 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_common.h +32 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_complex.h +174 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_cooperative_groups.h +829 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_fp16.h +1809 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_gl_interop.h +108 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_math_constants.h +124 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_runtime.h +405 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_runtime_pt_api.h +196 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_unsafe_atomics.h +565 -0
- triton/backends/amd/include/hip/amd_detail/amd_hip_vector_types.h +2226 -0
- triton/backends/amd/include/hip/amd_detail/amd_math_functions.h +104 -0
- triton/backends/amd/include/hip/amd_detail/amd_surface_functions.h +244 -0
- triton/backends/amd/include/hip/amd_detail/amd_warp_functions.h +494 -0
- triton/backends/amd/include/hip/amd_detail/concepts.hpp +30 -0
- triton/backends/amd/include/hip/amd_detail/device_library_decls.h +133 -0
- triton/backends/amd/include/hip/amd_detail/functional_grid_launch.hpp +218 -0
- triton/backends/amd/include/hip/amd_detail/grid_launch.h +67 -0
- triton/backends/amd/include/hip/amd_detail/grid_launch.hpp +50 -0
- triton/backends/amd/include/hip/amd_detail/grid_launch_GGL.hpp +26 -0
- triton/backends/amd/include/hip/amd_detail/helpers.hpp +137 -0
- triton/backends/amd/include/hip/amd_detail/hip_api_trace.hpp +1350 -0
- triton/backends/amd/include/hip/amd_detail/hip_assert.h +101 -0
- triton/backends/amd/include/hip/amd_detail/hip_cooperative_groups_helper.h +242 -0
- triton/backends/amd/include/hip/amd_detail/hip_fp16_gcc.h +254 -0
- triton/backends/amd/include/hip/amd_detail/hip_fp16_math_fwd.h +96 -0
- triton/backends/amd/include/hip/amd_detail/hip_ldg.h +100 -0
- triton/backends/amd/include/hip/amd_detail/hip_prof_str.h +10169 -0
- triton/backends/amd/include/hip/amd_detail/hip_runtime_prof.h +77 -0
- triton/backends/amd/include/hip/amd_detail/host_defines.h +180 -0
- triton/backends/amd/include/hip/amd_detail/hsa_helpers.hpp +102 -0
- triton/backends/amd/include/hip/amd_detail/macro_based_grid_launch.hpp +798 -0
- triton/backends/amd/include/hip/amd_detail/math_fwd.h +698 -0
- triton/backends/amd/include/hip/amd_detail/ockl_image.h +177 -0
- triton/backends/amd/include/hip/amd_detail/program_state.hpp +107 -0
- triton/backends/amd/include/hip/amd_detail/texture_fetch_functions.h +491 -0
- triton/backends/amd/include/hip/amd_detail/texture_indirect_functions.h +478 -0
- triton/backends/amd/include/hip/channel_descriptor.h +39 -0
- triton/backends/amd/include/hip/device_functions.h +38 -0
- triton/backends/amd/include/hip/driver_types.h +468 -0
- triton/backends/amd/include/hip/hip_bf16.h +36 -0
- triton/backends/amd/include/hip/hip_bfloat16.h +44 -0
- triton/backends/amd/include/hip/hip_common.h +100 -0
- triton/backends/amd/include/hip/hip_complex.h +38 -0
- triton/backends/amd/include/hip/hip_cooperative_groups.h +46 -0
- triton/backends/amd/include/hip/hip_deprecated.h +95 -0
- triton/backends/amd/include/hip/hip_ext.h +159 -0
- triton/backends/amd/include/hip/hip_fp16.h +36 -0
- triton/backends/amd/include/hip/hip_gl_interop.h +32 -0
- triton/backends/amd/include/hip/hip_hcc.h +24 -0
- triton/backends/amd/include/hip/hip_math_constants.h +36 -0
- triton/backends/amd/include/hip/hip_profile.h +27 -0
- triton/backends/amd/include/hip/hip_runtime.h +75 -0
- triton/backends/amd/include/hip/hip_runtime_api.h +8919 -0
- triton/backends/amd/include/hip/hip_texture_types.h +29 -0
- triton/backends/amd/include/hip/hip_vector_types.h +41 -0
- triton/backends/amd/include/hip/hip_version.h +17 -0
- triton/backends/amd/include/hip/hiprtc.h +421 -0
- triton/backends/amd/include/hip/library_types.h +78 -0
- triton/backends/amd/include/hip/math_functions.h +42 -0
- triton/backends/amd/include/hip/surface_types.h +63 -0
- triton/backends/amd/include/hip/texture_types.h +194 -0
- triton/backends/amd/include/hsa/Brig.h +1131 -0
- triton/backends/amd/include/hsa/amd_hsa_common.h +91 -0
- triton/backends/amd/include/hsa/amd_hsa_elf.h +436 -0
- triton/backends/amd/include/hsa/amd_hsa_kernel_code.h +269 -0
- triton/backends/amd/include/hsa/amd_hsa_queue.h +109 -0
- triton/backends/amd/include/hsa/amd_hsa_signal.h +80 -0
- triton/backends/amd/include/hsa/hsa.h +5729 -0
- triton/backends/amd/include/hsa/hsa_amd_tool.h +91 -0
- triton/backends/amd/include/hsa/hsa_api_trace.h +566 -0
- triton/backends/amd/include/hsa/hsa_ext_amd.h +3090 -0
- triton/backends/amd/include/hsa/hsa_ext_finalize.h +531 -0
- triton/backends/amd/include/hsa/hsa_ext_image.h +1454 -0
- triton/backends/amd/include/hsa/hsa_ven_amd_aqlprofile.h +488 -0
- triton/backends/amd/include/hsa/hsa_ven_amd_loader.h +667 -0
- triton/backends/amd/include/roctracer/ext/prof_protocol.h +107 -0
- triton/backends/amd/include/roctracer/hip_ostream_ops.h +4435 -0
- triton/backends/amd/include/roctracer/hsa_ostream_ops.h +1467 -0
- triton/backends/amd/include/roctracer/hsa_prof_str.h +3027 -0
- triton/backends/amd/include/roctracer/roctracer.h +779 -0
- triton/backends/amd/include/roctracer/roctracer_ext.h +81 -0
- triton/backends/amd/include/roctracer/roctracer_hcc.h +24 -0
- triton/backends/amd/include/roctracer/roctracer_hip.h +37 -0
- triton/backends/amd/include/roctracer/roctracer_hsa.h +112 -0
- triton/backends/amd/include/roctracer/roctracer_plugin.h +137 -0
- triton/backends/amd/include/roctracer/roctracer_roctx.h +67 -0
- triton/backends/amd/include/roctracer/roctx.h +229 -0
- triton/backends/amd/lib/ockl.bc +0 -0
- triton/backends/amd/lib/ocml.bc +0 -0
- triton/backends/compiler.py +304 -0
- triton/backends/driver.py +48 -0
- triton/backends/nvidia/__init__.py +0 -0
- triton/backends/nvidia/bin/ptxas.exe +0 -0
- triton/backends/nvidia/compiler.py +410 -0
- triton/backends/nvidia/driver.c +451 -0
- triton/backends/nvidia/driver.py +524 -0
- triton/backends/nvidia/include/cuda.h +24359 -0
- triton/backends/nvidia/lib/libdevice.10.bc +0 -0
- triton/backends/nvidia/lib/x64/cuda.lib +0 -0
- triton/compiler/__init__.py +4 -0
- triton/compiler/code_generator.py +1303 -0
- triton/compiler/compiler.py +430 -0
- triton/compiler/errors.py +51 -0
- triton/compiler/make_launcher.py +0 -0
- triton/errors.py +5 -0
- triton/language/__init__.py +294 -0
- triton/language/_utils.py +21 -0
- triton/language/core.py +2694 -0
- triton/language/extra/__init__.py +26 -0
- triton/language/extra/cuda/__init__.py +13 -0
- triton/language/extra/cuda/_experimental_tma.py +108 -0
- triton/language/extra/cuda/libdevice.py +1629 -0
- triton/language/extra/cuda/utils.py +109 -0
- triton/language/extra/hip/__init__.py +3 -0
- triton/language/extra/hip/libdevice.py +475 -0
- triton/language/extra/libdevice.py +786 -0
- triton/language/math.py +250 -0
- triton/language/random.py +207 -0
- triton/language/semantic.py +1796 -0
- triton/language/standard.py +452 -0
- triton/runtime/__init__.py +23 -0
- triton/runtime/autotuner.py +408 -0
- triton/runtime/build.py +111 -0
- triton/runtime/cache.py +295 -0
- triton/runtime/driver.py +60 -0
- triton/runtime/errors.py +26 -0
- triton/runtime/interpreter.py +1235 -0
- triton/runtime/jit.py +951 -0
- triton/testing.py +511 -0
- triton/tools/__init__.py +0 -0
- triton/tools/build_extern.py +365 -0
- triton/tools/compile.c +67 -0
- triton/tools/compile.h +14 -0
- triton/tools/compile.py +155 -0
- triton/tools/disasm.py +144 -0
- triton/tools/experimental_descriptor.py +32 -0
- triton/tools/link.py +322 -0
- triton/windows_utils.py +375 -0
- triton_windows-3.2.0.post11.dist-info/METADATA +39 -0
- triton_windows-3.2.0.post11.dist-info/RECORD +154 -0
- triton_windows-3.2.0.post11.dist-info/WHEEL +5 -0
- triton_windows-3.2.0.post11.dist-info/top_level.txt +12 -0
triton/windows_utils.py
ADDED
|
@@ -0,0 +1,375 @@
|
|
|
1
|
+
import functools
|
|
2
|
+
import os
|
|
3
|
+
import re
|
|
4
|
+
import subprocess
|
|
5
|
+
import sys
|
|
6
|
+
import sysconfig
|
|
7
|
+
import winreg
|
|
8
|
+
from collections.abc import Iterable
|
|
9
|
+
from functools import partial
|
|
10
|
+
from glob import glob
|
|
11
|
+
from pathlib import Path
|
|
12
|
+
from typing import Callable, Optional
|
|
13
|
+
|
|
14
|
+
|
|
15
|
+
def find_in_program_files(rel_path: str) -> Optional[Path]:
|
|
16
|
+
program_files = os.getenv("ProgramFiles(x86)", r"C:\Program Files (x86)")
|
|
17
|
+
path = Path(program_files) / rel_path
|
|
18
|
+
if path.exists():
|
|
19
|
+
return path
|
|
20
|
+
|
|
21
|
+
program_files = os.getenv("ProgramW6432", r"C:\Program Files")
|
|
22
|
+
path = Path(program_files) / rel_path
|
|
23
|
+
if path.exists():
|
|
24
|
+
return path
|
|
25
|
+
|
|
26
|
+
return None
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
def parse_version(s: str, prefix: str = "") -> Optional[tuple[int, ...]]:
|
|
30
|
+
s = s.removeprefix(prefix)
|
|
31
|
+
try:
|
|
32
|
+
return tuple(int(x) for x in s.split("."))
|
|
33
|
+
except ValueError:
|
|
34
|
+
return None
|
|
35
|
+
|
|
36
|
+
|
|
37
|
+
def unparse_version(t: Iterable[int], prefix: str = "") -> str:
|
|
38
|
+
return prefix + ".".join([str(x) for x in t])
|
|
39
|
+
|
|
40
|
+
|
|
41
|
+
def max_version(
|
|
42
|
+
versions: Iterable[str],
|
|
43
|
+
prefix: str = "",
|
|
44
|
+
check: Callable[[str], bool] = lambda x: True,
|
|
45
|
+
) -> Optional[str]:
|
|
46
|
+
versions = [x for x in versions if check(x)]
|
|
47
|
+
versions = [parse_version(x, prefix) for x in versions]
|
|
48
|
+
versions = [x for x in versions if x is not None]
|
|
49
|
+
if not versions:
|
|
50
|
+
return None
|
|
51
|
+
version = unparse_version(max(versions), prefix)
|
|
52
|
+
return version
|
|
53
|
+
|
|
54
|
+
|
|
55
|
+
def check_msvc(msvc_base_path: Path, version: str) -> bool:
|
|
56
|
+
return all(
|
|
57
|
+
x.exists()
|
|
58
|
+
for x in [
|
|
59
|
+
msvc_base_path / version / "include" / "vcruntime.h",
|
|
60
|
+
msvc_base_path / version / "lib" / "x64" / "vcruntime.lib",
|
|
61
|
+
]
|
|
62
|
+
)
|
|
63
|
+
|
|
64
|
+
|
|
65
|
+
def find_msvc_vswhere() -> tuple[Optional[Path], Optional[str]]:
|
|
66
|
+
vswhere_path = find_in_program_files(
|
|
67
|
+
r"Microsoft Visual Studio\Installer\vswhere.exe"
|
|
68
|
+
)
|
|
69
|
+
if vswhere_path is None:
|
|
70
|
+
return None, None
|
|
71
|
+
|
|
72
|
+
command = [
|
|
73
|
+
str(vswhere_path),
|
|
74
|
+
"-prerelease",
|
|
75
|
+
"-products",
|
|
76
|
+
"*",
|
|
77
|
+
"-requires",
|
|
78
|
+
"Microsoft.VisualStudio.Component.VC.Tools.x86.x64",
|
|
79
|
+
"-requires",
|
|
80
|
+
"Microsoft.VisualStudio.Component.Windows10SDK",
|
|
81
|
+
"-latest",
|
|
82
|
+
"-property",
|
|
83
|
+
"installationPath",
|
|
84
|
+
]
|
|
85
|
+
try:
|
|
86
|
+
output = subprocess.check_output(command, text=True).strip()
|
|
87
|
+
except subprocess.CalledProcessError:
|
|
88
|
+
return None, None
|
|
89
|
+
|
|
90
|
+
msvc_base_path = Path(output) / "VC" / "Tools" / "MSVC"
|
|
91
|
+
if not msvc_base_path.exists():
|
|
92
|
+
return None, None
|
|
93
|
+
|
|
94
|
+
version = max_version(
|
|
95
|
+
os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path)
|
|
96
|
+
)
|
|
97
|
+
if version is None:
|
|
98
|
+
return None, None
|
|
99
|
+
|
|
100
|
+
return msvc_base_path, version
|
|
101
|
+
|
|
102
|
+
|
|
103
|
+
def find_msvc_envpath() -> tuple[Optional[Path], Optional[str]]:
|
|
104
|
+
paths = os.getenv("PATH", "").split(os.pathsep)
|
|
105
|
+
for path in paths:
|
|
106
|
+
path = path.replace("/", "\\")
|
|
107
|
+
match = re.compile(r".*\\VC\\Tools\\MSVC\\").match(path)
|
|
108
|
+
if not match:
|
|
109
|
+
continue
|
|
110
|
+
|
|
111
|
+
msvc_base_path = Path(match.group(0))
|
|
112
|
+
if not msvc_base_path.exists():
|
|
113
|
+
continue
|
|
114
|
+
|
|
115
|
+
version = max_version(
|
|
116
|
+
os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path)
|
|
117
|
+
)
|
|
118
|
+
if version is None:
|
|
119
|
+
continue
|
|
120
|
+
|
|
121
|
+
return msvc_base_path, version
|
|
122
|
+
|
|
123
|
+
return None, None
|
|
124
|
+
|
|
125
|
+
|
|
126
|
+
def find_msvc_hardcoded() -> tuple[Optional[Path], Optional[str]]:
|
|
127
|
+
vs_path = find_in_program_files("Microsoft Visual Studio")
|
|
128
|
+
if vs_path is None:
|
|
129
|
+
return None, None
|
|
130
|
+
|
|
131
|
+
paths = glob(str(vs_path / "*" / "*" / "VC" / "Tools" / "MSVC"))
|
|
132
|
+
# First try the highest version
|
|
133
|
+
paths = sorted(paths)[::-1]
|
|
134
|
+
for msvc_base_path in paths:
|
|
135
|
+
msvc_base_path = Path(msvc_base_path)
|
|
136
|
+
version = max_version(
|
|
137
|
+
os.listdir(msvc_base_path), check=partial(check_msvc, msvc_base_path)
|
|
138
|
+
)
|
|
139
|
+
if version is None:
|
|
140
|
+
continue
|
|
141
|
+
return msvc_base_path, version
|
|
142
|
+
|
|
143
|
+
return None, None
|
|
144
|
+
|
|
145
|
+
|
|
146
|
+
def find_msvc() -> tuple[list[str], list[str]]:
|
|
147
|
+
msvc_base_path, version = find_msvc_vswhere()
|
|
148
|
+
if msvc_base_path is None:
|
|
149
|
+
msvc_base_path, version = find_msvc_envpath()
|
|
150
|
+
if msvc_base_path is None:
|
|
151
|
+
msvc_base_path, version = find_msvc_hardcoded()
|
|
152
|
+
if msvc_base_path is None:
|
|
153
|
+
print("WARNING: Failed to find MSVC.")
|
|
154
|
+
return [], []
|
|
155
|
+
|
|
156
|
+
return (
|
|
157
|
+
[str(msvc_base_path / version / "include")],
|
|
158
|
+
[str(msvc_base_path / version / "lib" / "x64")],
|
|
159
|
+
)
|
|
160
|
+
|
|
161
|
+
|
|
162
|
+
def check_winsdk(winsdk_base_path: Path, version: str) -> bool:
|
|
163
|
+
return all(
|
|
164
|
+
x.exists()
|
|
165
|
+
for x in [
|
|
166
|
+
winsdk_base_path / "Include" / version / "ucrt" / "stdlib.h",
|
|
167
|
+
winsdk_base_path / "Lib" / version / "ucrt" / "x64" / "ucrt.lib",
|
|
168
|
+
]
|
|
169
|
+
)
|
|
170
|
+
|
|
171
|
+
|
|
172
|
+
def find_winsdk_registry() -> tuple[Optional[Path], Optional[str]]:
|
|
173
|
+
try:
|
|
174
|
+
reg = winreg.ConnectRegistry(None, winreg.HKEY_LOCAL_MACHINE)
|
|
175
|
+
key = winreg.OpenKeyEx(
|
|
176
|
+
reg, r"SOFTWARE\WOW6432Node\Microsoft\Microsoft SDKs\Windows\v10.0"
|
|
177
|
+
)
|
|
178
|
+
folder = winreg.QueryValueEx(key, "InstallationFolder")[0]
|
|
179
|
+
winreg.CloseKey(key)
|
|
180
|
+
except OSError:
|
|
181
|
+
return None, None
|
|
182
|
+
|
|
183
|
+
winsdk_base_path = Path(folder)
|
|
184
|
+
if not (winsdk_base_path / "Include").exists():
|
|
185
|
+
return None, None
|
|
186
|
+
|
|
187
|
+
version = max_version(
|
|
188
|
+
os.listdir(winsdk_base_path / "Include"),
|
|
189
|
+
check=partial(check_winsdk, winsdk_base_path),
|
|
190
|
+
)
|
|
191
|
+
if version is None:
|
|
192
|
+
return None, None
|
|
193
|
+
|
|
194
|
+
return winsdk_base_path, version
|
|
195
|
+
|
|
196
|
+
|
|
197
|
+
def find_winsdk_hardcoded() -> tuple[Optional[Path], Optional[str]]:
|
|
198
|
+
winsdk_base_path = find_in_program_files(r"Windows Kits\10")
|
|
199
|
+
if winsdk_base_path is None:
|
|
200
|
+
return None, None
|
|
201
|
+
if not (winsdk_base_path / "Include").exists():
|
|
202
|
+
return None, None
|
|
203
|
+
|
|
204
|
+
version = max_version(
|
|
205
|
+
os.listdir(winsdk_base_path / "Include"),
|
|
206
|
+
check=partial(check_winsdk, winsdk_base_path),
|
|
207
|
+
)
|
|
208
|
+
if version is None:
|
|
209
|
+
return None, None
|
|
210
|
+
|
|
211
|
+
return winsdk_base_path, version
|
|
212
|
+
|
|
213
|
+
|
|
214
|
+
def find_winsdk() -> tuple[list[str], list[str]]:
|
|
215
|
+
winsdk_base_path, version = find_winsdk_registry()
|
|
216
|
+
if winsdk_base_path is None:
|
|
217
|
+
winsdk_base_path, version = find_winsdk_hardcoded()
|
|
218
|
+
if winsdk_base_path is None:
|
|
219
|
+
print("WARNING: Failed to find Windows SDK.")
|
|
220
|
+
return [], []
|
|
221
|
+
|
|
222
|
+
return (
|
|
223
|
+
[
|
|
224
|
+
str(winsdk_base_path / "Include" / version / "shared"),
|
|
225
|
+
str(winsdk_base_path / "Include" / version / "ucrt"),
|
|
226
|
+
str(winsdk_base_path / "Include" / version / "um"),
|
|
227
|
+
],
|
|
228
|
+
[
|
|
229
|
+
str(winsdk_base_path / "Lib" / version / "ucrt" / "x64"),
|
|
230
|
+
str(winsdk_base_path / "Lib" / version / "um" / "x64"),
|
|
231
|
+
],
|
|
232
|
+
)
|
|
233
|
+
|
|
234
|
+
|
|
235
|
+
@functools.cache
|
|
236
|
+
def find_msvc_winsdk() -> tuple[list[str], list[str]]:
|
|
237
|
+
msvc_inc_dirs, msvc_lib_dirs = find_msvc()
|
|
238
|
+
winsdk_inc_dirs, winsdk_lib_dirs = find_winsdk()
|
|
239
|
+
return msvc_inc_dirs + winsdk_inc_dirs, msvc_lib_dirs + winsdk_lib_dirs
|
|
240
|
+
|
|
241
|
+
|
|
242
|
+
@functools.cache
|
|
243
|
+
def find_python() -> list[str]:
|
|
244
|
+
for python_base_path in [
|
|
245
|
+
sys.exec_prefix,
|
|
246
|
+
sys.base_exec_prefix,
|
|
247
|
+
os.path.dirname(sys.executable),
|
|
248
|
+
]:
|
|
249
|
+
python_lib_dir = Path(python_base_path) / "libs"
|
|
250
|
+
if (python_lib_dir / "python3.lib").exists():
|
|
251
|
+
return [str(python_lib_dir)]
|
|
252
|
+
|
|
253
|
+
print("WARNING: Failed to find Python libs.")
|
|
254
|
+
return []
|
|
255
|
+
|
|
256
|
+
|
|
257
|
+
def find_cuda_bundled() -> tuple[Optional[str], list[str], list[str]]:
|
|
258
|
+
cuda_base_path = (
|
|
259
|
+
Path(sysconfig.get_paths()["platlib"]) / "triton" / "backends" / "nvidia"
|
|
260
|
+
)
|
|
261
|
+
if check_cuda_system_wide(cuda_base_path):
|
|
262
|
+
return (
|
|
263
|
+
str(cuda_base_path / "bin"),
|
|
264
|
+
[str(cuda_base_path / "include")],
|
|
265
|
+
[str(cuda_base_path / "lib" / "x64")],
|
|
266
|
+
)
|
|
267
|
+
|
|
268
|
+
return None, [], []
|
|
269
|
+
|
|
270
|
+
|
|
271
|
+
def check_cuda_pip(nvidia_base_path: Path) -> bool:
|
|
272
|
+
return all(
|
|
273
|
+
x.exists()
|
|
274
|
+
for x in [
|
|
275
|
+
nvidia_base_path / "cuda_nvcc" / "bin" / "ptxas.exe",
|
|
276
|
+
nvidia_base_path / "cuda_runtime" / "include" / "cuda.h",
|
|
277
|
+
nvidia_base_path / "cuda_runtime" / "lib" / "x64" / "cuda.lib",
|
|
278
|
+
]
|
|
279
|
+
)
|
|
280
|
+
|
|
281
|
+
|
|
282
|
+
def find_cuda_pip() -> tuple[Optional[str], list[str], list[str]]:
|
|
283
|
+
nvidia_base_path = Path(sysconfig.get_paths()["platlib"]) / "nvidia"
|
|
284
|
+
if check_cuda_pip(nvidia_base_path):
|
|
285
|
+
return (
|
|
286
|
+
str(nvidia_base_path / "cuda_nvcc" / "bin"),
|
|
287
|
+
[str(nvidia_base_path / "cuda_runtime" / "include")],
|
|
288
|
+
[str(nvidia_base_path / "cuda_runtime" / "lib" / "x64")],
|
|
289
|
+
)
|
|
290
|
+
|
|
291
|
+
return None, [], []
|
|
292
|
+
|
|
293
|
+
|
|
294
|
+
def check_cuda_conda(cuda_base_path: Path) -> bool:
|
|
295
|
+
return all(
|
|
296
|
+
x.exists()
|
|
297
|
+
for x in [
|
|
298
|
+
cuda_base_path / "bin" / "ptxas.exe",
|
|
299
|
+
cuda_base_path / "include" / "cuda.h",
|
|
300
|
+
cuda_base_path / "lib" / "cuda.lib",
|
|
301
|
+
]
|
|
302
|
+
)
|
|
303
|
+
|
|
304
|
+
|
|
305
|
+
def find_cuda_conda() -> tuple[Optional[str], list[str], list[str]]:
|
|
306
|
+
cuda_base_path = Path(sys.exec_prefix) / "Library"
|
|
307
|
+
if check_cuda_conda(cuda_base_path):
|
|
308
|
+
return (
|
|
309
|
+
str(cuda_base_path / "bin"),
|
|
310
|
+
[str(cuda_base_path / "include")],
|
|
311
|
+
[str(cuda_base_path / "lib")],
|
|
312
|
+
)
|
|
313
|
+
|
|
314
|
+
return None, [], []
|
|
315
|
+
|
|
316
|
+
|
|
317
|
+
def check_cuda_system_wide(cuda_base_path: Path) -> bool:
|
|
318
|
+
return all(
|
|
319
|
+
x.exists()
|
|
320
|
+
for x in [
|
|
321
|
+
cuda_base_path / "bin" / "ptxas.exe",
|
|
322
|
+
cuda_base_path / "include" / "cuda.h",
|
|
323
|
+
cuda_base_path / "lib" / "x64" / "cuda.lib",
|
|
324
|
+
]
|
|
325
|
+
)
|
|
326
|
+
|
|
327
|
+
|
|
328
|
+
def find_cuda_env() -> Optional[Path]:
|
|
329
|
+
for cuda_base_path in ["CUDA_PATH", "CUDA_HOME"]:
|
|
330
|
+
cuda_base_path = os.getenv(cuda_base_path)
|
|
331
|
+
if cuda_base_path is None:
|
|
332
|
+
continue
|
|
333
|
+
|
|
334
|
+
cuda_base_path = Path(cuda_base_path)
|
|
335
|
+
if check_cuda_system_wide(cuda_base_path):
|
|
336
|
+
return cuda_base_path
|
|
337
|
+
|
|
338
|
+
return None
|
|
339
|
+
|
|
340
|
+
|
|
341
|
+
def find_cuda_hardcoded() -> Optional[Path]:
|
|
342
|
+
parent = find_in_program_files(r"NVIDIA GPU Computing Toolkit\CUDA")
|
|
343
|
+
if parent is None:
|
|
344
|
+
return None
|
|
345
|
+
|
|
346
|
+
paths = glob(str(parent / "v12*"))
|
|
347
|
+
# First try the highest version
|
|
348
|
+
paths = sorted(paths)[::-1]
|
|
349
|
+
for path in paths:
|
|
350
|
+
cuda_base_path = Path(path)
|
|
351
|
+
if check_cuda_system_wide(cuda_base_path):
|
|
352
|
+
return cuda_base_path
|
|
353
|
+
|
|
354
|
+
return None
|
|
355
|
+
|
|
356
|
+
|
|
357
|
+
@functools.cache
|
|
358
|
+
def find_cuda() -> tuple[Optional[str], list[str], list[str]]:
|
|
359
|
+
for f in [find_cuda_bundled, find_cuda_pip, find_cuda_conda]:
|
|
360
|
+
cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs = f()
|
|
361
|
+
if cuda_bin_path:
|
|
362
|
+
return cuda_bin_path, cuda_inc_dirs, cuda_lib_dirs
|
|
363
|
+
|
|
364
|
+
cuda_base_path = find_cuda_env()
|
|
365
|
+
if cuda_base_path is None:
|
|
366
|
+
cuda_base_path = find_cuda_hardcoded()
|
|
367
|
+
if cuda_base_path is None:
|
|
368
|
+
print("WARNING: Failed to find CUDA.")
|
|
369
|
+
return None, [], []
|
|
370
|
+
|
|
371
|
+
return (
|
|
372
|
+
str(cuda_base_path / "bin"),
|
|
373
|
+
[str(cuda_base_path / "include")],
|
|
374
|
+
[str(cuda_base_path / "lib" / "x64")],
|
|
375
|
+
)
|
|
@@ -0,0 +1,39 @@
|
|
|
1
|
+
Metadata-Version: 2.2
|
|
2
|
+
Name: triton-windows
|
|
3
|
+
Version: 3.2.0.post11
|
|
4
|
+
Summary: A language and compiler for custom Deep Learning operations
|
|
5
|
+
Home-page: https://github.com/woct0rdho/triton-windows
|
|
6
|
+
Author: Philippe Tillet, Dian Wu
|
|
7
|
+
Author-email: phil@openai.com, woctordho@outlook.com
|
|
8
|
+
Keywords: Compiler,Deep Learning
|
|
9
|
+
Classifier: Development Status :: 4 - Beta
|
|
10
|
+
Classifier: Intended Audience :: Developers
|
|
11
|
+
Classifier: Topic :: Software Development :: Build Tools
|
|
12
|
+
Classifier: License :: OSI Approved :: MIT License
|
|
13
|
+
Classifier: Programming Language :: Python :: 3.9
|
|
14
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
15
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
16
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
17
|
+
Classifier: Programming Language :: Python :: 3.13
|
|
18
|
+
Provides-Extra: build
|
|
19
|
+
Requires-Dist: cmake>=3.20; extra == "build"
|
|
20
|
+
Requires-Dist: lit; extra == "build"
|
|
21
|
+
Provides-Extra: tests
|
|
22
|
+
Requires-Dist: autopep8; extra == "tests"
|
|
23
|
+
Requires-Dist: flake8; extra == "tests"
|
|
24
|
+
Requires-Dist: isort; extra == "tests"
|
|
25
|
+
Requires-Dist: numpy; extra == "tests"
|
|
26
|
+
Requires-Dist: pytest; extra == "tests"
|
|
27
|
+
Requires-Dist: scipy>=1.7.1; extra == "tests"
|
|
28
|
+
Requires-Dist: llnl-hatchet; extra == "tests"
|
|
29
|
+
Provides-Extra: tutorials
|
|
30
|
+
Requires-Dist: matplotlib; extra == "tutorials"
|
|
31
|
+
Requires-Dist: pandas; extra == "tutorials"
|
|
32
|
+
Requires-Dist: tabulate; extra == "tutorials"
|
|
33
|
+
Dynamic: author
|
|
34
|
+
Dynamic: author-email
|
|
35
|
+
Dynamic: classifier
|
|
36
|
+
Dynamic: home-page
|
|
37
|
+
Dynamic: keywords
|
|
38
|
+
Dynamic: provides-extra
|
|
39
|
+
Dynamic: summary
|
|
@@ -0,0 +1,154 @@
|
|
|
1
|
+
triton/__init__.py,sha256=gG2MzgdBYh-TQo8LqSbSnEAgc3HdtZnb5ab4m8gO_eU,1896
|
|
2
|
+
triton/_internal_testing.py,sha256=4pzyc_36u05khUveZ1TWL3MQ-7EVMJ1C2c1dRd8BMsw,4269
|
|
3
|
+
triton/errors.py,sha256=8WfnuRKLG578mgY6cBA3ECruVMf9ULEKFNgRcJ6IhWM,89
|
|
4
|
+
triton/testing.py,sha256=fX3pn9bjC3Z-z5qzSKW56C_2WF8h3mHLy5RJqpZ-HsA,19382
|
|
5
|
+
triton/windows_utils.py,sha256=OStlF-dMd-rM7CbTknEp3A7q_iU8a5Ou4U-XjjfFwsI,10898
|
|
6
|
+
triton/_C/libtriton.pyd,sha256=Qu1WZReuSYGQPJcB3-ckwOGnGaS3jdSoNHi8GZ1mUCc,87242752
|
|
7
|
+
triton/backends/__init__.py,sha256=opAo_vgEMt3tLO_bYFrYGksnIu0qohbmyuu_s3-rNAs,1595
|
|
8
|
+
triton/backends/compiler.py,sha256=JZiiEbB9Wws3tjU6KXrydKtlOQI7Suk-mTYPlafa0Qk,11388
|
|
9
|
+
triton/backends/driver.py,sha256=QX_6P1Go9ajdlHZi4Hv3nCtdHyDA6o8_lM3NMnlH1mk,1386
|
|
10
|
+
triton/backends/amd/compiler.py,sha256=0FnS5mBLsPB9FQzNr3I9BPd3-KEB1Do29hDsnFGu25k,16203
|
|
11
|
+
triton/backends/amd/driver.c,sha256=obiiiPndny5NyhUcJ8iyrVHrXU1ruLpLGd_LgaKQEbU,8459
|
|
12
|
+
triton/backends/amd/driver.py,sha256=5hrhChpNC8Npm8kRSAde9lB2wxyNkZtZrz7trHbdu7E,18596
|
|
13
|
+
triton/backends/amd/include/hip/channel_descriptor.h,sha256=gTYe7SzIg-m3ThOQY2vr5Rh6-uWvUP_d37v8F4T2Q14,1773
|
|
14
|
+
triton/backends/amd/include/hip/device_functions.h,sha256=vkybrdk6wyZP-T1I5PRjtfcMqGYXDeBpB5jhYj358GU,1589
|
|
15
|
+
triton/backends/amd/include/hip/driver_types.h,sha256=m1HI80HC80qkTeco2Jd07woL_jTy48lz9JiDCV_8zsg,18985
|
|
16
|
+
triton/backends/amd/include/hip/hip_bf16.h,sha256=lLw6K5ltb6AqSuINYTq8flxxsDkBP8Y2zbqmUjBcG9c,1571
|
|
17
|
+
triton/backends/amd/include/hip/hip_bfloat16.h,sha256=Nqoy9VjfjglVx2_NJcp8hyT1sJUukXRWj8XMlidv1yA,1755
|
|
18
|
+
triton/backends/amd/include/hip/hip_common.h,sha256=q5aPhG3DHW0iUJ7ayS5lfM_ZnZQNbMmLmfdHlOwbPdA,3450
|
|
19
|
+
triton/backends/amd/include/hip/hip_complex.h,sha256=TmdzQP5oVPfhBVARJYcR5eyv9HInmKMFuFoQ_1ECk_I,1594
|
|
20
|
+
triton/backends/amd/include/hip/hip_cooperative_groups.h,sha256=gMLvaYQ3b-f1vcoMtEwtkN0hO5__zNfP5p5oBKmv_SE,1878
|
|
21
|
+
triton/backends/amd/include/hip/hip_deprecated.h,sha256=gFLuCuKn7R_xCfum_i_Q-vi3Lg8NWHKphKZKze8DwEo,6340
|
|
22
|
+
triton/backends/amd/include/hip/hip_ext.h,sha256=jK1Qc-SXgUyRTj8bBa9ZP__95Qgd2-W1mwnJo6Qpnoo,8560
|
|
23
|
+
triton/backends/amd/include/hip/hip_fp16.h,sha256=vKJh-zgDWUW7NyXxtv2ho6aVLXX8BIPfzCigEQ5d6I4,1523
|
|
24
|
+
triton/backends/amd/include/hip/hip_gl_interop.h,sha256=-GwkSFMBneM8akFE7pqlhi0k-Ft2uz5674wGoiaU43Q,1438
|
|
25
|
+
triton/backends/amd/include/hip/hip_hcc.h,sha256=RYrArDlnTEP89xKbzIpW17_bsBY5moCitq00PL-4oWI,1307
|
|
26
|
+
triton/backends/amd/include/hip/hip_math_constants.h,sha256=8bSfve5E7cDuvNAUkFUeQwSLg3iJJHuqhuD4FmHNxEM,1588
|
|
27
|
+
triton/backends/amd/include/hip/hip_profile.h,sha256=sjsNuduu5Jd6s7sJndZvZLlE0RZ0wN1rTVwv5nR7If0,1304
|
|
28
|
+
triton/backends/amd/include/hip/hip_runtime.h,sha256=uy90l8Nep6xNUzeGcHMoDv84BT3hMpieTV-5ijkpL5A,3058
|
|
29
|
+
triton/backends/amd/include/hip/hip_runtime_api.h,sha256=fzb_xktisCVcp2pWG-ZKhIG-YVQzDjGyPt4wvA4iayM,386498
|
|
30
|
+
triton/backends/amd/include/hip/hip_texture_types.h,sha256=AhkvjG4cDjf_ZFLg5SsSTfBnXG614PBK1XVPa7irZbk,1237
|
|
31
|
+
triton/backends/amd/include/hip/hip_vector_types.h,sha256=6FcBMBkP3ZN1Enalpa9hV0VopxdBJvbUCuaxISgzbTY,1630
|
|
32
|
+
triton/backends/amd/include/hip/hip_version.h,sha256=J3vgzfZH0UkK8RYvyHVj1PbUNSZH1JPtlcmXxLBgwVk,407
|
|
33
|
+
triton/backends/amd/include/hip/hiprtc.h,sha256=npK6f2ZkYIe5blJIGuofuTG0PrSMS2mkFBUqrdOp0A0,15631
|
|
34
|
+
triton/backends/amd/include/hip/library_types.h,sha256=tPOJTQedPH5qC9meawLgKpnbFrQC2WKlfo6s0rhKoZc,2370
|
|
35
|
+
triton/backends/amd/include/hip/math_functions.h,sha256=frzdJ4veBG8n9ALO4EmRrdOiDguR6FP6ygLnvOnVVSM,1815
|
|
36
|
+
triton/backends/amd/include/hip/surface_types.h,sha256=uQHjITphDM7k4pnuEoDEupMUxBobzvhJpSy0unpegh4,1959
|
|
37
|
+
triton/backends/amd/include/hip/texture_types.h,sha256=CtmdykZfDikhnrVfdJk3w2VK5X3Af_6rEKzU-VgLu24,6687
|
|
38
|
+
triton/backends/amd/include/hip/amd_detail/amd_channel_descriptor.h,sha256=_2myGIdBTE0plFbGKOSx8HUqGZd0UBHo-YvKe2xkpbU,11708
|
|
39
|
+
triton/backends/amd/include/hip/amd_detail/amd_device_functions.h,sha256=zfYTHJE_M_y2Y2ssP8ZH_EOczMBg4Iq2guglaKcI5js,31425
|
|
40
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_atomic.h,sha256=PJRRTp83M0jIEBA_iWzfWwHZelSbL3TBrSDqlO3SQtk,49919
|
|
41
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_bf16.h,sha256=fucv1_06JHVm82T0TmvERBbmtZTDQK6WJi_58oGQOXg,40634
|
|
42
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_bfloat16.h,sha256=cFJlQEELGau_9geACeuiiFHyuAWCD6-VuSqcTnqajX0,9484
|
|
43
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_common.h,sha256=dzkuIzuklqTRaNJjKLqfFEm6Fh4tK_FkTjYHFsZkmCI,1370
|
|
44
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_complex.h,sha256=SEygl8X_MCXDVXxNIBm5Ds0eWwa-ojVXUUW48SIgsX8,5855
|
|
45
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_cooperative_groups.h,sha256=SvrkniHiDGt-ztZRBvbkyajfUxTbGQzpZC1gnd4T-i8,31624
|
|
46
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_fp16.h,sha256=86Nw97iaiC4QV5xBv8d3Bwc4FioMh5DQuCHj3sh_Yrw,57854
|
|
47
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_gl_interop.h,sha256=9vxiV6rYRMGx12TPnrAVRvrfLyoRp74XRgKSPBPa2hk,3860
|
|
48
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_math_constants.h,sha256=u1fIaf-AiWF70ZA1zxVkUIbRqoJLu5lrfYbgt_usySk,5890
|
|
49
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_runtime.h,sha256=ZvDsQ0AiZnJ178NuAsA7AuHrySXbN3aFs5Z9m2tsIDg,13954
|
|
50
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_runtime_pt_api.h,sha256=fc4mtHBkWmiSRh8m-dxIxvu9zsweLTwEgohkntYcgJw,9997
|
|
51
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_unsafe_atomics.h,sha256=w9nJ1S32GRl_ejDiGacteM6Zf84iovIifAzWX8Bze0Q,24202
|
|
52
|
+
triton/backends/amd/include/hip/amd_detail/amd_hip_vector_types.h,sha256=qPdmRJnzlgtjVshkafoHxdHoMLkoYS9U-ZD-TjLznr0,57088
|
|
53
|
+
triton/backends/amd/include/hip/amd_detail/amd_math_functions.h,sha256=46wiaEMStCczEsHtccgHlATfw_0O5j6Z8rlFkC7bmUA,3171
|
|
54
|
+
triton/backends/amd/include/hip/amd_detail/amd_surface_functions.h,sha256=rsQuylNqmNhLb7PZjBz7WbruD_6YIXtOptY2BNJDxVU,11062
|
|
55
|
+
triton/backends/amd/include/hip/amd_detail/amd_warp_functions.h,sha256=p8DdtuxqlgGHzKdVPMHDnZOD8zA5f6GjLHYMr0_FKjQ,18966
|
|
56
|
+
triton/backends/amd/include/hip/amd_detail/concepts.hpp,sha256=7EOkpr2w2-jclUQ115yxtFCkBWJ7btUzhBOe-mR0N0M,1252
|
|
57
|
+
triton/backends/amd/include/hip/amd_detail/device_library_decls.h,sha256=4clSpgf898UVjfZFVnDkcYi75A27crPsuFtLcs1s4KU,7457
|
|
58
|
+
triton/backends/amd/include/hip/amd_detail/functional_grid_launch.hpp,sha256=u7hRB9kQXX575a5C7cV3gKow55DSBUCwO0dTjIswlag,8129
|
|
59
|
+
triton/backends/amd/include/hip/amd_detail/grid_launch.h,sha256=tNS7CQw9gy-z930CElH3n6c5iMvpsQ_WFZK024mNzEo,1830
|
|
60
|
+
triton/backends/amd/include/hip/amd_detail/grid_launch.hpp,sha256=EuAlM3olyrArebqwW5eSxo4gfjvWCGOAGAuLLmFttgw,1370
|
|
61
|
+
triton/backends/amd/include/hip/amd_detail/grid_launch_GGL.hpp,sha256=KpQAuyy1Dyt45WcPaR_x-Ex-onPGEHA01DBbla7TT-k,1219
|
|
62
|
+
triton/backends/amd/include/hip/amd_detail/helpers.hpp,sha256=hi2pW1mXQnbIwvmwWt_nG6A38sqLOd-QP5S9sETTs60,5707
|
|
63
|
+
triton/backends/amd/include/hip/amd_detail/hip_api_trace.hpp,sha256=d01j4SFQP_6ALwUHByxznZV8SrQHbuujRYon8rxFw-I,94612
|
|
64
|
+
triton/backends/amd/include/hip/amd_detail/hip_assert.h,sha256=fNsG23KISuY-k5JFoX-5hZ7qGQScisXuHcdEwYlXOqw,3978
|
|
65
|
+
triton/backends/amd/include/hip/amd_detail/hip_cooperative_groups_helper.h,sha256=tQ_XIvGKhvrj1h7gY-IVLmKvIPhsQa0YsBflxdhUHP8,7957
|
|
66
|
+
triton/backends/amd/include/hip/amd_detail/hip_fp16_gcc.h,sha256=BtFsKmTptN4TOHocEicfNbBl2JCdZWKm_bd5mc5OzYY,6660
|
|
67
|
+
triton/backends/amd/include/hip/amd_detail/hip_fp16_math_fwd.h,sha256=63tKWMPdW56qWlH_HbCaF_isVXufm514ol_SxL4YjTQ,5134
|
|
68
|
+
triton/backends/amd/include/hip/amd_detail/hip_ldg.h,sha256=KAEZb9H4z4DDrkaloMOeWzahiDfI2V6c68vWT3jb5fU,3652
|
|
69
|
+
triton/backends/amd/include/hip/amd_detail/hip_prof_str.h,sha256=s1T2IrCwYzZQOuCs5ppuegFQbjXSF2JA1eUSCmZg9AA,621355
|
|
70
|
+
triton/backends/amd/include/hip/amd_detail/hip_runtime_prof.h,sha256=6GVfh1la0wtBVwdKX5y0C32dPD9shJp1o8wZdHsjZHA,2715
|
|
71
|
+
triton/backends/amd/include/hip/amd_detail/host_defines.h,sha256=h_ZpFE4Clm2iyRyJevDb57Y-gC-6RVPjhnZ5rzPxiUo,7038
|
|
72
|
+
triton/backends/amd/include/hip/amd_detail/hsa_helpers.hpp,sha256=Os-sJQOFI_0Abh8Ql05s0Rtfruk4NsSMfg7BtugxMgg,3232
|
|
73
|
+
triton/backends/amd/include/hip/amd_detail/macro_based_grid_launch.hpp,sha256=6ocsArNa9_R6D6XCuNy8Zq23KG-j2uYsjqNCtnMrJws,67925
|
|
74
|
+
triton/backends/amd/include/hip/amd_detail/math_fwd.h,sha256=nup5YhceJnngoLJCESI8qX08dNpbZci0i78WKu-wfdI,17000
|
|
75
|
+
triton/backends/amd/include/hip/amd_detail/ockl_image.h,sha256=LzRPGMb515_iIAIIcbb2uQB-bTvT4xOjY51VdARD7lc,10538
|
|
76
|
+
triton/backends/amd/include/hip/amd_detail/program_state.hpp,sha256=8QE9OmB8OKTy7rBr3EYEizJI2s-_1tgXpgU7zCA2Ky0,3154
|
|
77
|
+
triton/backends/amd/include/hip/amd_detail/texture_fetch_functions.h,sha256=Ex1lF2gBWJxtC3yP9pXRSFywMp3gbEmyl0Sw8iL91yM,17787
|
|
78
|
+
triton/backends/amd/include/hip/amd_detail/texture_indirect_functions.h,sha256=KkW5o5gMpoVMTRwzfXHA7-kZ9ynI8OaIw6jJ1EB1s98,18447
|
|
79
|
+
triton/backends/amd/include/hsa/Brig.h,sha256=5H-btCHq40qgjjpwVAoRWf3E0ccf-J6UCPEcKx_hGKw,32705
|
|
80
|
+
triton/backends/amd/include/hsa/amd_hsa_common.h,sha256=q_zN0eq-dwR7FnQ84PcpV3yZyvjHsouIAjJgKltGoX8,3912
|
|
81
|
+
triton/backends/amd/include/hsa/amd_hsa_elf.h,sha256=r3xymEjYeTIBCPvlKBDJxKyI1Dfg6KDXc5VqO9Uy1iM,16352
|
|
82
|
+
triton/backends/amd/include/hsa/amd_hsa_kernel_code.h,sha256=C55F8a480QsW16-iwN9TIT3cKnGh6GoeoEaEv3aVh4g,12659
|
|
83
|
+
triton/backends/amd/include/hsa/amd_hsa_queue.h,sha256=ZJ-k5wY30heLmQnGB0VUz36XCiVHRmspg5FRNMGIk_U,4766
|
|
84
|
+
triton/backends/amd/include/hsa/amd_hsa_signal.h,sha256=FDegZnWQC04GtnqHjXOBsB-AoVSaqdhNY6Mwbua5FGA,2947
|
|
85
|
+
triton/backends/amd/include/hsa/hsa.h,sha256=Jft1K5uFAcasOD9IYW6wD5GsGQcPQTrmbpjie-0Wh00,190916
|
|
86
|
+
triton/backends/amd/include/hsa/hsa_amd_tool.h,sha256=pyZSyIVl-UA5AOhte78jvn4V3hCd0dxJAIv7KeADsPs,2843
|
|
87
|
+
triton/backends/amd/include/hsa/hsa_api_trace.h,sha256=2iuwHcpyW9wvr-WPKCgatQzYBaA8rTa3w1BRMXBGcSI,28982
|
|
88
|
+
triton/backends/amd/include/hsa/hsa_ext_amd.h,sha256=V7bs8agiXxNEpJd9yoLQjUQ_vJeOagdfdrQ3Pgb4ioA,115908
|
|
89
|
+
triton/backends/amd/include/hsa/hsa_ext_finalize.h,sha256=sv0AZbDM-B1wIdQ3cHTMlpUtNacQN2PkOgX90IZol_o,20227
|
|
90
|
+
triton/backends/amd/include/hsa/hsa_ext_image.h,sha256=t5YJm_aw9EePCeFL1hoIfQ8ubIjBte-ptfReq6Ts-8Y,54232
|
|
91
|
+
triton/backends/amd/include/hsa/hsa_ven_amd_aqlprofile.h,sha256=9uev2nT29MCdu7-HMkg9iItHop6QMOBMQL5DAFnftSg,19777
|
|
92
|
+
triton/backends/amd/include/hsa/hsa_ven_amd_loader.h,sha256=c6cxPAzAox7u6IbFzEkQZfCuRl-Kr39WhY2_w23X1R4,26146
|
|
93
|
+
triton/backends/amd/include/roctracer/hip_ostream_ops.h,sha256=WNXFZxawBXHmFGMDFIOZqXkCw6VzyDexwGPkGJre4w0,184840
|
|
94
|
+
triton/backends/amd/include/roctracer/hsa_ostream_ops.h,sha256=AYwF-IT9Dhl2FX-GuvCJZX6fSmHK0xkKLORx9QxuSK8,57857
|
|
95
|
+
triton/backends/amd/include/roctracer/hsa_prof_str.h,sha256=ctT-KKsIGayp7RUGUsFNR-dE65VydyXla_Qgvf-efTU,122884
|
|
96
|
+
triton/backends/amd/include/roctracer/roctracer.h,sha256=B8sHz2DMNprP7EqNWIGwVLY1KQMpxmhfVy4UoR8dzzY,23849
|
|
97
|
+
triton/backends/amd/include/roctracer/roctracer_ext.h,sha256=vLaZ8peAxSy0cwrdEalKnUApkKspfa04iw1Mr_Zcio0,2940
|
|
98
|
+
triton/backends/amd/include/roctracer/roctracer_hcc.h,sha256=NlF3R8JQ9oX9lGpm0b2n-EWJ0r3y9sP9wbwnoucaCuY,1303
|
|
99
|
+
triton/backends/amd/include/roctracer/roctracer_hip.h,sha256=RCzYuNw1vLR7xK4rb06TtM9TU546UYKHJ83IMHmZEm8,1432
|
|
100
|
+
triton/backends/amd/include/roctracer/roctracer_hsa.h,sha256=M8APM64XNAWSslxQisM-pcmKoUQaUdTMaKvSACyt0Ag,4108
|
|
101
|
+
triton/backends/amd/include/roctracer/roctracer_plugin.h,sha256=8GGE1zDbdPCVJtbmwOCYq7X0mwFjfWRtzDYKLD4cKys,4786
|
|
102
|
+
triton/backends/amd/include/roctracer/roctracer_roctx.h,sha256=gBjBk5vb0l3PbBSQ7V9iFtaM_RzkIDJEW1A_PXBihBM,2014
|
|
103
|
+
triton/backends/amd/include/roctracer/roctx.h,sha256=RhJXUXRhSJ5LRE_1gm7E6-bjEMrfcFBLDLuf3UxAIh8,6717
|
|
104
|
+
triton/backends/amd/include/roctracer/ext/prof_protocol.h,sha256=6FAcvVD-dNM7uulFs2B-aTxw5xOAWGy6evdD4yUaebA,3849
|
|
105
|
+
triton/backends/amd/lib/ockl.bc,sha256=wQKCzkKukIHbu0lyjKUYlhndc7S27xto6L54J0Bn-C0,246124
|
|
106
|
+
triton/backends/amd/lib/ocml.bc,sha256=UPNTXW0gCXUNB-c6orSYwb-mz9_mjUc7zny_vfFza44,205964
|
|
107
|
+
triton/backends/nvidia/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
108
|
+
triton/backends/nvidia/compiler.py,sha256=P0AAdEoTbaVlKwf9ORWauHorhtCF8ljVTp_clLZ8dFo,16560
|
|
109
|
+
triton/backends/nvidia/driver.c,sha256=sg-C-zQEW2B9BkTQJ3xKfnRf1rOcC8_eHoQCTuZr_Ow,18872
|
|
110
|
+
triton/backends/nvidia/driver.py,sha256=Pqx9wUluhxTQovBsePBxcPNOJ0zSYuMhGzCDhNCHkEQ,17534
|
|
111
|
+
triton/backends/nvidia/bin/ptxas.exe,sha256=2m-Z7N6rSw_zUngI0Zza4eMzhqglEEQQ6sMkev8cAr8,19580416
|
|
112
|
+
triton/backends/nvidia/include/cuda.h,sha256=6SN2FpBhqz9-wdKnWN4_R3q18-auO67gdcBp-cEiSpE,1072817
|
|
113
|
+
triton/backends/nvidia/lib/libdevice.10.bc,sha256=XC-uN8huaMOjhgWpX1EtfRLV89uYYxC-R_VzBKpype4,473728
|
|
114
|
+
triton/backends/nvidia/lib/x64/cuda.lib,sha256=LU7iq6V6zutLCiMR_ZGaoJAZxz7h7htxjCwh3v_wWWA,155386
|
|
115
|
+
triton/compiler/__init__.py,sha256=kSVpmv2ro25zaF-fVJcpeyxMpRRb5uCiXQo7DqhG1CQ,239
|
|
116
|
+
triton/compiler/code_generator.py,sha256=g80O73MyM9acg7XNfcHgrq6I4bNgEd5_LY7gYp3vPsE,57574
|
|
117
|
+
triton/compiler/compiler.py,sha256=Er1CFZhQc3smV0OCmZRFKYXFGhJKfU5zJLH3oEf0NFI,17580
|
|
118
|
+
triton/compiler/errors.py,sha256=I9Y15pDWcL9heY4SWWdLeMDtW6Iiq2pFXzKfJ6dY_C0,1732
|
|
119
|
+
triton/compiler/make_launcher.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
120
|
+
triton/language/__init__.py,sha256=jGmSXwL_gpfWobg5qwlU124HNi_aLX5cNgx_Y9NiBX0,4852
|
|
121
|
+
triton/language/_utils.py,sha256=bkp98MH2y3mfSI7h1u_T33VPwYqsbnIJkjuwIsNsfE4,646
|
|
122
|
+
triton/language/core.py,sha256=a9A4B5uUBn1ijuqSZFMP7APU1wJ_-DTom51vvDGui9k,93793
|
|
123
|
+
triton/language/math.py,sha256=Lkr348qTen3UxyB-tu4_j368LzCRK1KnIE7qwEC9Kg8,7442
|
|
124
|
+
triton/language/random.py,sha256=s664rmyx6UCFJUo8M2EhNHUsckROwhmWXdf6UuAQp2I,6864
|
|
125
|
+
triton/language/semantic.py,sha256=yJhocGpO3_X4YSk9GTRQMalbhDt7eh9SOzDq02Djg34,79854
|
|
126
|
+
triton/language/standard.py,sha256=NMo6NQOJt81Zxy9s-U9o4xmg5DlKhY04H7WTRKfMBS4,13747
|
|
127
|
+
triton/language/extra/__init__.py,sha256=XRXFvr7416pRsh_Rh-X6qV66SiEyVDVbxp4GSAE1mfc,655
|
|
128
|
+
triton/language/extra/libdevice.py,sha256=Dki14elRNmQsz-Ytw9CnOaLCCnte4T6cI8bOzWjN63A,6318
|
|
129
|
+
triton/language/extra/cuda/__init__.py,sha256=JqiuryHnWRkfFztXgxbiQ62XA4dEKhsjhIHGobLuzcQ,414
|
|
130
|
+
triton/language/extra/cuda/_experimental_tma.py,sha256=FwtsItBySF70RzS3qMKrlcdxznjFom6JD40QOs_RfNU,3555
|
|
131
|
+
triton/language/extra/cuda/libdevice.py,sha256=crwXcdixYPuvzVOQ0e5styRAwQrUg0RRRlqek7QvXRw,56165
|
|
132
|
+
triton/language/extra/cuda/utils.py,sha256=e1BslV7lZGhi2uVIlo5lI9dcN61HUMIU2asPaRjsyIo,4379
|
|
133
|
+
triton/language/extra/hip/__init__.py,sha256=ieSER4LeX9_0horChGUUVwpuKAprkuka8uGAkEBDyDM,49
|
|
134
|
+
triton/language/extra/hip/libdevice.py,sha256=EVraUfeXzQmN3F5Lleg2mohVcbFWOWlLaAH1nkbqtV4,16841
|
|
135
|
+
triton/runtime/__init__.py,sha256=mKL5cqIBDUw2WO80NRCh4s1G8KYaqgM59TTAbTkPPjQ,621
|
|
136
|
+
triton/runtime/autotuner.py,sha256=BJe69v9MSMSzdkvYSUDrvXrAFeLZ1x6A-7aUmpz2Le0,17271
|
|
137
|
+
triton/runtime/build.py,sha256=Cpi_FTz40f7l0Yqt2xp_-GfXSCMozGJolobmqIsWKUE,4137
|
|
138
|
+
triton/runtime/cache.py,sha256=OQhUkwIW38-kayOL8P6SizWMAYSoVa_TbOYdTUHBkU0,10268
|
|
139
|
+
triton/runtime/driver.py,sha256=VZ-883Xri71R72lHB6usIpLo3gGLbZJkAlLP3ewWSpc,1509
|
|
140
|
+
triton/runtime/errors.py,sha256=oj73dn34qJbLhOjakakAuZPSv-laZyIYylJiJwREA8Y,787
|
|
141
|
+
triton/runtime/interpreter.py,sha256=0SPiXDlM7X7DbCdu2JXoLmxJ8ugCwH_3NPoxuU0tJyg,53201
|
|
142
|
+
triton/runtime/jit.py,sha256=8C8OgvZ0pRRL-8S2PbK9Knp6m6kbGE6O1immZpXVIzA,35303
|
|
143
|
+
triton/tools/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
144
|
+
triton/tools/build_extern.py,sha256=jCr-2hu3nLGBIJhCGUQ1jAyzLttughjkiPGEwRFjLR0,13673
|
|
145
|
+
triton/tools/compile.c,sha256=rjuAQ8b-2DTtbj29SgK1NxJI5BSU2P9ccp9wa5p8Iyc,2090
|
|
146
|
+
triton/tools/compile.h,sha256=n9QKIFZTL4RSsiXtAxBP9XGSnxjyaevQQ9bBpwDsvAg,332
|
|
147
|
+
triton/tools/compile.py,sha256=b3yNnVgoBk8WzOs87JrZPDIyasdSgAslOWmxse1J6yM,6761
|
|
148
|
+
triton/tools/disasm.py,sha256=BBO4bALdLcWgWDLhQdYHLlTx3oo8g_d8maeE_Uu-FmU,5088
|
|
149
|
+
triton/tools/experimental_descriptor.py,sha256=0Wqy96Cc6YLh9o0eTknW-Lfvha6lfRSfe8bswkcPHMs,1260
|
|
150
|
+
triton/tools/link.py,sha256=u7qtfZRLriZkAMEGNvj8YF-k1cthmLL7BwHYqBgT63E,11871
|
|
151
|
+
triton_windows-3.2.0.post11.dist-info/METADATA,sha256=Yf96fuQiVJN0rzLdTEuz0mm4AcD0edZW_D1ngM_wYZE,1514
|
|
152
|
+
triton_windows-3.2.0.post11.dist-info/WHEEL,sha256=fvNoTPbNdRAfsv1hWsmn2TKqULUpllfaSGBE1Ye0N8Q,99
|
|
153
|
+
triton_windows-3.2.0.post11.dist-info/top_level.txt,sha256=iIyUoyO6Ld8jQHWY9plUyco-emBXKTqVmuAXJAGvmEk,211
|
|
154
|
+
triton_windows-3.2.0.post11.dist-info/RECORD,,
|