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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346
|
# mypy: ignore-errors
import functools
import unittest
import torch
from functorch.experimental.control_flow import map
from torch.nn.attention.flex_attention import _create_empty_block_mask, flex_attention
from torch.testing import make_tensor
from torch.testing._internal.common_device_type import onlyCUDA
from torch.testing._internal.common_dtype import all_types_and, custom_types
from torch.testing._internal.opinfo.core import DecorateInfo, OpInfo, SampleInput
from torch._higher_order_ops.invoke_subgraph import mark_compile_region
def sample_inputs_map(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(
[make_arg(2, 2, 2, low=0.1, high=2), make_arg(2, 2, 2, low=0.1, high=2)],
args=(make_arg(1, low=0.1, high=2), make_arg(1, low=0.1, high=2)),
)
def inner_f(x, y0, y1):
return [x[0].cos().add_(1.0) * y0, (x[1] + y1.sin()).cos_().view(x[1].size())]
def simple_map(xs, y0, y1):
def f(x, y0, y1):
return inner_f(x, y0, y1)
return map(f, xs, y0, y1)
def nested_map(xs, y0, y1):
def f1(xx, y0, y1):
def f2(x, y0, y1):
return inner_f(x, y0, y1)
return map(f2, xx, y0, y1)
return map(f1, xs, y0, y1)
def triple_nested_map(xs, y0, y1):
def f0(xs, y0, y1):
def f1(xx, y0, y1):
def f2(x, y0, y1):
return inner_f(x, y0, y1)
return map(f2, xx, y0, y1)
return map(f1, xs, y0, y1)
return map(f0, xs, y0, y1)
# Please consult with torch.export team before
# adding new entry to this list.
hop_that_doesnt_have_opinfo_test_allowlist = [
"custom_function_call",
"autograd_function_apply",
"run_and_save_rng_state",
"run_with_rng_state",
"out_dtype",
"trace_wrapped",
"map", # T183144629
"map_impl",
"with_effects",
"strict_mode",
"_export_tracepoint",
"call_torchbind",
"triton_kernel_wrapper_mutation",
"triton_kernel_wrapper_functional",
"hints_wrapper",
]
torch.library.define(
"testlib::mutating_custom_op",
"(Tensor(a!) x, Tensor(b!) z) -> (Tensor, Tensor, Tensor)",
tags=torch.Tag.pt2_compliant_tag,
)
@torch.library.impl("testlib::mutating_custom_op", "cpu")
def foo_impl_cpu(x, z):
x.add_(5)
z.add_(5)
return x, z, x + z
@torch.library.impl("testlib::mutating_custom_op", "cuda")
def foo_impl_cuda(x, z):
x.add_(5)
z.add_(5)
return x, z, x + z
@torch.library.register_fake("testlib::mutating_custom_op")
def foo_impl_abstract(x, z):
return x, z, x + z
def sample_inputs_cond(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(make_arg(2, 2, 2, low=0.1, high=2))
def simple_cond(x):
return torch.cond(x.sum() > 2, lambda x: (x.cos(),), lambda x: (x.sin(),), [x])
def sample_inputs_invoke_subgraph(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(make_arg(2, 2, 2, low=0.1, high=2))
@mark_compile_region
def fn_for_invoke_subgraph(x):
return torch.sin(x)
def simple_invoke_subgraph(x):
return fn_for_invoke_subgraph(x)
def sample_inputs_auto_functionalize(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=False
)
yield SampleInput(
make_arg(2, 2, 2, low=0.1, high=2), make_arg(2, 2, 2, low=0.1, high=2)
)
def simple_auto_functionalize(x, z):
return torch.ops.testlib.mutating_custom_op(x, z)
def sample_inputs_flex_attention(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
def score_mod(score, b, h, m, n):
return score + h
q, k, v = (make_arg(2, 2, 128, 8, low=0.1, high=2) for _ in range(3))
block_mask = _create_empty_block_mask(q, k)
yield SampleInput(q, k, v, score_mod, block_mask)
def sample_inputs_while_loop(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=False
)
yield SampleInput(
torch.tensor(3),
make_arg(2, 3, 4, low=0.1, high=2),
)
def simple_while_loop(iter_t, x):
def cond_fn(iter_t, x):
return iter_t > 0
def body_fn(iter_t, x):
return iter_t - 1, x.cos()
return torch._higher_order_ops.while_loop(cond_fn, body_fn, (iter_t, x))
def sample_inputs_scan(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(
make_arg(2, 2, low=0.1, high=2),
make_arg(2, 2, 2, low=0.1, high=2),
)
def simple_scan(init, xs):
def combine_fn(carry, x):
result = carry @ x + x
return result, carry.clone()
return torch._higher_order_ops.scan(combine_fn, init, xs)
hop_db = [
OpInfo(
name="scan",
variant_test_name="simple",
op=simple_scan,
sample_inputs_func=sample_inputs_scan,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=False,
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="invoke_subgraph",
variant_test_name="simple",
op=simple_invoke_subgraph,
sample_inputs_func=sample_inputs_invoke_subgraph,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=True,
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="map",
variant_test_name="simple",
op=simple_map,
sample_inputs_func=sample_inputs_map,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
),
OpInfo(
name="map",
variant_test_name="nested",
op=nested_map,
sample_inputs_func=sample_inputs_map,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
),
OpInfo(
name="map",
variant_test_name="triple_nested",
op=triple_nested_map,
sample_inputs_func=sample_inputs_map,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
),
OpInfo(
name="cond",
variant_test_name="simple",
op=simple_cond,
sample_inputs_func=sample_inputs_cond,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=True,
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="while_loop",
variant_test_name="simple",
op=simple_while_loop,
sample_inputs_func=sample_inputs_while_loop,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=False,
),
OpInfo(
name="auto_functionalize",
variant_test_name="simple",
op=simple_auto_functionalize,
sample_inputs_func=sample_inputs_auto_functionalize,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=False,
),
OpInfo(
name="flex_attention",
variant_test_name="simple",
op=flex_attention,
sample_inputs_func=sample_inputs_flex_attention,
dtypes=custom_types(torch.float16, torch.float32),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
skips=(
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_aot_export"),
DecorateInfo(
unittest.expectedFailure, "TestHOP", "test_pre_dispatch_export"
),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_serialize_export"),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_retrace_export"),
),
decorators=[onlyCUDA],
),
OpInfo(
name="flex_attention_backward",
variant_test_name="simple",
op=flex_attention,
sample_inputs_func=sample_inputs_flex_attention,
dtypes=custom_types(torch.float16, torch.float32),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
skips=(
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_aot_export"),
DecorateInfo(
unittest.expectedFailure, "TestHOP", "test_pre_dispatch_export"
),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_serialize_export"),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_retrace_export"),
),
decorators=[onlyCUDA],
),
]
|