File size: 10,301 Bytes
c4cbbbc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
import ctypes
import glob
import os
import sys
import sysconfig
import importlib


def is_windows():
    return sys.platform.startswith("win")


module_name = ".Release._compiled_module" if is_windows() else "._compiled_module"

_pybind_module = importlib.import_module(module_name, package=__name__)

symbols_to_import = [
    "backend_version",
    "backend_version_string",
    "get_last_error_string",
    "destroy_handle",
    "norm_forward_phase",
    "reduction_mode",
    "behavior_note",
    "knob_type",
    "create_handle",
    "create_kernel_cache",
    "create_device_properties",
    "get_stream",
    "numerical_note",
    "set_stream",
    "build_plan_policy",
    "data_type",
    "tensor_reordering",
    "heur_mode",
    "pygraph",
    "tensor",
    "knob",
    "cudnnGraphNotSupportedError",
    "diagonal_alignment",
    "attention_implementation",
]

for symbol_name in symbols_to_import:
    globals()[symbol_name] = getattr(_pybind_module, symbol_name)

from .datatypes import _library_type, _is_torch_tensor

__version__ = "1.19.0"


def _tensor(
    self,
    dim,
    stride,
    data_type=data_type.NOT_SET,
    is_virtual=False,
    is_pass_by_value=False,
    ragged_offset=None,
    reordering_type=tensor_reordering.NONE,
    name="",
    uid=-1,
):
    """
    Create a tensor.

    Args:
        dim (List[int]): The dimensions of the tensor.
        stride (List[int]): The strides of the tensor.
        data_type (cudnn.data_type): The data type of the tensor.
        is_virtual (bool): Flag indicating if the tensor is virtual.
        is_pass_by_value (bool): Flag indicating if the tensor is passed by value.
        ragged_offset (cudnn_tensor): The ragged offset tensor.
        reordering_type (cudnn.tensor_reordering): The reordering type of the tensor.
        name (str): The name of the tensor.

    Returns:
        cudnn_tensor: The created tensor.
    """
    return self._make_tensor(
        dim=dim,
        stride=stride,
        data_type=_library_type(data_type),
        is_virtual=is_virtual,
        is_pass_by_value=is_pass_by_value,
        ragged_offset=ragged_offset,
        reordering_type=reordering_type,
        name=name,
        uid=uid,
    )


def _set_data_type(
    self,
    data_type=data_type.NOT_SET,
):
    return self._set_data_type(_library_type(data_type))


_pybind_module.tensor.set_data_type = _set_data_type
pygraph.tensor = _tensor


def _library_device_pointer(input_tensor):
    # either pass in pointers directly
    if type(input_tensor) is int:
        return input_tensor
    # directly extract data pointer for torch tensors
    elif _is_torch_tensor(input_tensor):
        return input_tensor.data_ptr()
    # fall back to dlpack support by library
    else:
        return _pybind_module._get_data_ptr(input_tensor)


def _execute(
    self,
    tensor_to_device_buffer,
    workspace,
    handle=None,
    override_uids=None,
    override_shapes=None,
    override_strides=None,
):
    """
    Execute a cudnn graph.

    Args:
        tensor_to_device_buffer (dict(cudnn_tensor, Union[torch.Tensor, int, __dlpack__])): The dimensions of the tensor.
        workspace (Union[torch.Tensor, int, __dlpack__]): The name of the tensor.
        handle: cudnn_handle created with cudnn.create_handle()
    Returns:
        None
    """
    uid_to_tensor_pointer = {
        x if type(x) is int else x.get_uid(): _library_device_pointer(pointer) for x, pointer in tensor_to_device_buffer.items() if x is not None
    }

    workspace_pointer = _library_device_pointer(workspace)
    self._execute(uid_to_tensor_pointer, workspace_pointer, handle)


def _execute_plan_at_index(
    self,
    tensor_to_device_buffer,
    workspace,
    index,
    handle=None,
    override_uids=None,
    override_shapes=None,
    override_strides=None,
):
    """
    Execute a cudnn graph.

    Args:
        tensor_to_device_buffer (dict(cudnn_tensor, Union[torch.Tensor, int, __dlpack__])): The dimensions of the tensor.
        workspace (Union[torch.Tensor, int, __dlpack__]): The name of the tensor.
        index(int): Location of execution plan to use.
        handle: cudnn_handle created with cudnn.create_handle()
    Returns:
        None
    """
    uid_to_tensor_pointer = {
        x if type(x) is int else x.get_uid(): _library_device_pointer(pointer) for x, pointer in tensor_to_device_buffer.items() if x is not None
    }

    workspace_pointer = _library_device_pointer(workspace)
    self._execute_plan_at_index(
        uid_to_tensor_pointer,
        workspace_pointer,
        index,
        handle,
        override_uids,
        override_shapes,
        override_strides,
    )


pygraph.execute = _execute
pygraph.execute_plan_at_index = _execute_plan_at_index


def load_cudnn():
    # First look at python site packages
    lib_path = glob.glob(os.path.join(sysconfig.get_path("purelib"), "nvidia/cudnn/bin/cudnn64_9.dll"))

    if lib_path:
        assert len(lib_path) == 1, f"Found {len(lib_path)} libcudnn.dll.x in nvidia-cudnn-cuXX."
        lib = ctypes.windll.LoadLibrary(lib_path[0])
    else:  # Fallback
        lib = ctypes.windll.LoadLibrary("cudnn64_9.dll")

    handle = ctypes.cast(lib._handle, ctypes.c_void_p).value
    _pybind_module._set_dlhandle_cudnn(handle)


def _dlopen_cudnn():
    # First look at python site packages
    lib_path = glob.glob(os.path.join(sysconfig.get_path("purelib"), "nvidia/cudnn/lib/libcudnn.so.*[0-9]"))

    if not lib_path:
        lib_path = glob.glob(os.path.join(sysconfig.get_path("purelib"), "nvidia/cudnn_jit/lib/libcudnn.so.*[0-9]"))

    if lib_path:
        assert len(lib_path) == 1, f"Found {len(lib_path)} libcudnn.so.x in nvidia-cudnn-cuXX."
        lib = ctypes.CDLL(lib_path[0])
    else:  # Fallback
        try:
            lib = ctypes.CDLL("libcudnn.so.9")
        except Exception:
            try:
                lib = ctypes.CDLL("libcudnn.so")
            except Exception:
                lib = None

    if lib is not None:
        handle = ctypes.cast(lib._handle, ctypes.c_void_p).value
        _pybind_module._set_dlhandle_cudnn(handle)


if is_windows():
    load_cudnn()
else:
    _dlopen_cudnn()

from .graph import graph, jit, graph_cache
from .wrapper import Graph

from typing import Any


def __getattr__(name: str) -> Any:
    if name == "NSA":
        try:
            from .native_sparse_attention import NSA as _NSA

            return _NSA
        except Exception as e:
            raise ImportError(f"NSA requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}") from e

    elif name == "GemmSwigluSm100":
        try:
            from .gemm_swiglu import GemmSwigluSm100 as _GemmSwigluSm100

            return _GemmSwigluSm100
        except Exception as e:
            raise ImportError(f"GemmSwigluSm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}") from e

    elif name == "gemm_swiglu_wrapper_sm100":
        try:
            from .gemm_swiglu import (
                gemm_swiglu_wrapper_sm100 as _gemm_swiglu_wrapper_sm100,
            )

            return _gemm_swiglu_wrapper_sm100
        except Exception as e:
            raise ImportError(
                f"gemm_swiglu_wrapper_sm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}"
            ) from e

    elif name == "GemmAmaxSm100":
        try:
            from .gemm_amax import GemmAmaxSm100 as _GemmAmaxSm100

            return _GemmAmaxSm100
        except Exception as e:
            raise ImportError(f"GemmAmaxSm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}") from e

    elif name == "gemm_amax_wrapper_sm100":
        try:
            from .gemm_amax import (
                gemm_amax_wrapper_sm100 as _gemm_amax_wrapper_sm100,
            )

            return _gemm_amax_wrapper_sm100
        except Exception as e:
            raise ImportError(f"gemm_amax_wrapper_sm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}") from e

    # Grouped GEMM module
    elif name == "grouped_gemm":
        try:
            from . import grouped_gemm as _grouped_gemm

            return _grouped_gemm
        except Exception as e:
            raise ImportError(f"grouped_gemm requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}") from e

    elif name == "GroupedGemmSwigluSm100":
        try:
            from .grouped_gemm import GroupedGemmSwigluSm100 as _GroupedGemmSwigluSm100

            return _GroupedGemmSwigluSm100
        except Exception as e:
            raise ImportError(f"GroupedGemmSwigluSm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}") from e

    elif name == "grouped_gemm_swiglu_wrapper_sm100":
        try:
            from .grouped_gemm import (
                grouped_gemm_swiglu_wrapper_sm100 as _grouped_gemm_swiglu_wrapper_sm100,
            )

            return _grouped_gemm_swiglu_wrapper_sm100
        except Exception as e:
            raise ImportError(
                f"grouped_gemm_swiglu_wrapper_sm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}"
            ) from e

    elif name == "GroupedGemmDswigluSm100":
        try:
            from .grouped_gemm import GroupedGemmDswigluSm100 as _GroupedGemmDswigluSm100

            return _GroupedGemmDswigluSm100
        except Exception as e:
            raise ImportError(f"GroupedGemmDswigluSm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}") from e

    elif name == "grouped_gemm_dswiglu_wrapper_sm100":
        try:
            from .grouped_gemm import (
                grouped_gemm_dswiglu_wrapper_sm100 as _grouped_gemm_dswiglu_wrapper_sm100,
            )

            return _grouped_gemm_dswiglu_wrapper_sm100
        except Exception as e:
            raise ImportError(
                f"grouped_gemm_dswiglu_wrapper_sm100 requires optional dependencies. Install with 'pip install nvidia-cudnn-frontend[cutedsl]': {e}"
            ) from e

    else:
        raise AttributeError(name)