danieldk HF staff commited on
Commit
5600c5f
·
1 Parent(s): 0b7452b

Fix a couple of bugs and add tests from vLLM

Browse files
ext-torch/__init__.py CHANGED
@@ -6,36 +6,42 @@ except ImportError as e:
6
  # Fallback for local development.
7
  try:
8
  import _activation
 
9
  ops = torch.ops._activition
10
  except ImportError:
11
  raise e
12
-
13
-
14
- def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
15
- ops.silu_and_mul(out, x)
16
-
17
-
18
- def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
19
- ops.gelu_and_mul(out, x)
20
-
21
-
22
- def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
23
- ops.gelu_tanh_and_mul(out, x)
24
-
25
-
26
- def fatrelu_and_mul(out: torch.Tensor,
27
- x: torch.Tensor,
28
- threshold: float = 0.0) -> None:
29
- ops.fatrelu_and_mul(out, x, threshold)
30
-
31
-
32
- def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
33
- ops.gelu_fast(out, x)
34
-
35
-
36
- def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
37
- ops.gelu_new(out, x)
38
-
39
-
40
- def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
 
 
 
 
41
  ops.gelu_quick(out, x)
 
 
6
  # Fallback for local development.
7
  try:
8
  import _activation
9
+
10
  ops = torch.ops._activition
11
  except ImportError:
12
  raise e
13
+
14
+
15
+ def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
16
+ ops.silu_and_mul(out, x)
17
+ return out
18
+
19
+
20
+ def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
21
+ ops.gelu_and_mul(out, x)
22
+ return out
23
+
24
+
25
+ def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
26
+ ops.gelu_tanh_and_mul(out, x)
27
+ return out
28
+
29
+
30
+ def fatrelu_and_mul(out: torch.Tensor, x: torch.Tensor, threshold: float = 0.0) -> None:
31
+ ops.fatrelu_and_mul(out, x, threshold)
32
+ return out
33
+
34
+
35
+ def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
36
+ ops.gelu_fast(out, x)
37
+ return out
38
+
39
+
40
+ def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
41
+ ops.gelu_new(out, x)
42
+ return out
43
+
44
+
45
+ def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
46
  ops.gelu_quick(out, x)
47
+ return out
ext-torch/torch_binding.cpp CHANGED
@@ -28,6 +28,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
28
  // Approximate GELU implementation.
29
  ops.def("gelu_fast(Tensor! out, Tensor input) -> ()");
30
  ops.impl("gelu_fast", torch::kCUDA, &gelu_fast);
 
 
 
 
31
  }
32
 
33
  REGISTER_EXTENSION(TORCH_EXTENSION_NAME)
 
28
  // Approximate GELU implementation.
29
  ops.def("gelu_fast(Tensor! out, Tensor input) -> ()");
30
  ops.impl("gelu_fast", torch::kCUDA, &gelu_fast);
31
+
32
+ // Quick GELU implementation.
33
+ ops.def("gelu_quick(Tensor! out, Tensor input) -> ()");
34
+ ops.impl("gelu_quick", torch::kCUDA, &gelu_quick);
35
  }
36
 
37
  REGISTER_EXTENSION(TORCH_EXTENSION_NAME)
tests/__init__.py ADDED
File without changes
tests/kernels/__init__.py ADDED
File without changes
tests/kernels/allclose_default.py ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ # Reference default values of atol and rtol are from
4
+ # https://github.com/pytorch/pytorch/blob/6d96beb6bec24d73ee3f080bac54d2104068f675/test/test_transformers.py#L67
5
+ default_atol = {torch.float16: 1e-3, torch.bfloat16: 1e-3, torch.float: 1e-5}
6
+ default_rtol = {torch.float16: 1e-3, torch.bfloat16: 1.6e-2, torch.float: 1.3e-6}
7
+
8
+
9
+ def get_default_atol(output) -> float:
10
+ return default_atol[output.dtype]
11
+
12
+
13
+ def get_default_rtol(output) -> float:
14
+ return default_rtol[output.dtype]
tests/kernels/test_activation.py ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import random
3
+ from typing import Type
4
+
5
+ import activation
6
+ import pytest
7
+ import torch
8
+ import torch.nn.functional as F
9
+
10
+ from .utils import opcheck
11
+ from .allclose_default import get_default_atol, get_default_rtol
12
+
13
+ DTYPES = [torch.half, torch.bfloat16, torch.float]
14
+ NUM_TOKENS = [7, 83, 2048] # Arbitrary values for testing
15
+ D = [512, 13824] # Arbitrary values for testing
16
+ SEEDS = [0]
17
+ CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)]
18
+
19
+
20
+ def gelu_fast(x: torch.Tensor) -> torch.Tensor:
21
+ return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 * (1.0 + 0.044715 * x * x)))
22
+
23
+
24
+ def gelu_new(x: torch.Tensor) -> torch.Tensor:
25
+ c = math.sqrt(2.0 / math.pi)
26
+ return 0.5 * x * (1.0 + torch.tanh(c * (x + 0.044715 * torch.pow(x, 3.0))))
27
+
28
+
29
+ def gelu_quick(x: torch.Tensor) -> torch.Tensor:
30
+ return x * torch.sigmoid(1.702 * x)
31
+
32
+
33
+ def fatrelu_and_mul(x: torch.Tensor, threshold: float) -> torch.Tensor:
34
+ d = x.shape[-1] // 2
35
+ x1 = x[..., :d]
36
+ x2 = x[..., d:]
37
+ x1 = F.threshold(x1, threshold, 0.0)
38
+ return x1 * x2
39
+
40
+
41
+ def silu_and_mul(x: torch.Tensor) -> torch.Tensor:
42
+ d = x.shape[-1] // 2
43
+ return F.silu(x[..., :d]) * x[..., d:]
44
+
45
+
46
+ def gelu_and_mul(x: torch.Tensor, approximate: str) -> torch.Tensor:
47
+ d = x.shape[-1] // 2
48
+ return F.gelu(x[..., :d], approximate=approximate) * x[..., d:]
49
+
50
+
51
+ @pytest.mark.parametrize("activation_name", ["silu", "gelu", "gelu_tanh", "fatrelu"])
52
+ @pytest.mark.parametrize("num_tokens", NUM_TOKENS)
53
+ @pytest.mark.parametrize("d", D)
54
+ @pytest.mark.parametrize("dtype", DTYPES)
55
+ @pytest.mark.parametrize("seed", SEEDS)
56
+ @pytest.mark.parametrize("device", CUDA_DEVICES)
57
+ @torch.inference_mode()
58
+ def test_act_and_mul(
59
+ activation_name: str,
60
+ num_tokens: int,
61
+ d: int,
62
+ dtype: torch.dtype,
63
+ seed: int,
64
+ device: str,
65
+ ) -> None:
66
+ random.seed(seed)
67
+ torch.manual_seed(seed)
68
+ torch.set_default_device(device)
69
+ x = torch.randn(num_tokens, 2 * d, dtype=dtype)
70
+ if activation_name == "silu":
71
+ torch_fn = silu_and_mul
72
+ fn = activation.silu_and_mul
73
+ op = activation.ops.silu_and_mul
74
+ elif activation_name == "gelu":
75
+ torch_fn = lambda x: gelu_and_mul(x, "none")
76
+ fn = activation.gelu_and_mul
77
+ op = activation.ops.gelu_and_mul
78
+ elif activation_name == "gelu_tanh":
79
+ torch_fn = lambda x: gelu_and_mul(x, "tanh")
80
+ fn = activation.gelu_tanh_and_mul
81
+ op = activation.ops.gelu_tanh_and_mul
82
+ elif activation_name == "fatrelu":
83
+ threshold = random.uniform(0, 1)
84
+ torch_fn = lambda x: fatrelu_and_mul(x, threshold)
85
+ fn = lambda out, x: activation.fatrelu_and_mul(out, x, threshold)
86
+ op = activation.ops.fatrelu_and_mul
87
+
88
+ out_shape = x.shape[:-1] + (x.shape[-1] // 2,)
89
+ out = torch.empty(out_shape, dtype=x.dtype, device=x.device)
90
+ out = fn(out, x)
91
+ ref_out = torch_fn(x)
92
+
93
+ # The SiLU, GELU and FatReLU implementations are equivalent to the native
94
+ # PyTorch implementations, so we can do exact comparison.
95
+ torch.testing.assert_close(out, ref_out, atol=0.0, rtol=0.0)
96
+
97
+ d = x.shape[-1] // 2
98
+ output_shape = x.shape[:-1] + (d,)
99
+ out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
100
+ if activation_name == "fatrelu":
101
+ opcheck(op, (out, x, threshold))
102
+ else:
103
+ opcheck(op, (out, x))
104
+
105
+
106
+ @pytest.mark.parametrize(
107
+ "activation_fns",
108
+ [
109
+ (gelu_fast, activation.gelu_fast, activation.ops.gelu_fast),
110
+ (gelu_new, activation.gelu_new, activation.ops.gelu_new),
111
+ (gelu_quick, activation.gelu_quick, activation.ops.gelu_quick),
112
+ ],
113
+ )
114
+ @pytest.mark.parametrize("num_tokens", NUM_TOKENS)
115
+ @pytest.mark.parametrize("d", D)
116
+ @pytest.mark.parametrize("dtype", DTYPES)
117
+ @pytest.mark.parametrize("seed", SEEDS)
118
+ @pytest.mark.parametrize("device", CUDA_DEVICES)
119
+ @torch.inference_mode()
120
+ def test_activation(
121
+ activation_fns,
122
+ num_tokens: int,
123
+ d: int,
124
+ dtype: torch.dtype,
125
+ seed: int,
126
+ device: str,
127
+ ) -> None:
128
+ torch.manual_seed(seed)
129
+ torch.set_default_device(device)
130
+ x = torch.randn(num_tokens, d, dtype=dtype)
131
+ torch_fn, fn, op = activation_fns
132
+ out = fn(torch.empty_like(x), x)
133
+ ref_out = torch_fn(x)
134
+ torch.testing.assert_close(
135
+ out, ref_out, atol=get_default_atol(out), rtol=get_default_rtol(out)
136
+ )
137
+
138
+ out = torch.empty_like(x)
139
+ opcheck(op, (out, x))
tests/kernels/utils.py ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Kernel test utils"""
2
+
3
+ import itertools
4
+ import random
5
+ import unittest
6
+ from numbers import Number
7
+ from typing import Any, Dict, List, NamedTuple, Optional, Sequence, Tuple, Union
8
+
9
+ import pytest
10
+ import torch
11
+ from torch._prims_common import TensorLikeType
12
+
13
+ # For now, disable "test_aot_dispatch_dynamic" since there are some
14
+ # bugs related to this test in PyTorch 2.4.
15
+ DEFAULT_OPCHECK_TEST_UTILS: Tuple[str, ...] = (
16
+ "test_schema",
17
+ "test_autograd_registration",
18
+ "test_faketensor",
19
+ )
20
+
21
+ ALL_OPCHECK_TEST_UTILS: Tuple[str, ...] = (
22
+ "test_schema",
23
+ "test_autograd_registration",
24
+ "test_faketensor",
25
+ "test_aot_dispatch_dynamic",
26
+ )
27
+
28
+
29
+ # Copied/modified from torch._refs.__init__.py
30
+ def fp8_allclose(
31
+ a: TensorLikeType,
32
+ b: TensorLikeType,
33
+ rtol: float = 1e-05,
34
+ atol: float = 1e-08,
35
+ equal_nan: bool = False,
36
+ ) -> bool:
37
+ """
38
+ Reference implementation of torch.allclose
39
+ """
40
+ torch._refs._check_close_args(name="torch.allclose", a=a, b=b, rtol=rtol, atol=atol)
41
+
42
+ return bool(
43
+ torch.all(
44
+ torch.isclose(
45
+ a.double(), b.double(), rtol=rtol, atol=atol, equal_nan=equal_nan
46
+ )
47
+ ).item()
48
+ )
49
+
50
+
51
+ # A special version of op check that has a restricted default set of test_utils
52
+ # and a patched version of allclose that supports fp8 types.
53
+ def opcheck(
54
+ op: Union[
55
+ torch._ops.OpOverload,
56
+ torch._ops.OpOverloadPacket,
57
+ torch._library.custom_ops.CustomOpDef,
58
+ ],
59
+ args: Tuple[Any, ...],
60
+ kwargs: Optional[Dict[str, Any]] = None,
61
+ *,
62
+ test_utils: Union[str, Sequence[str]] = ALL_OPCHECK_TEST_UTILS,
63
+ raise_exception: bool = True,
64
+ cond: bool = True
65
+ ) -> Dict[str, str]:
66
+ with unittest.mock.patch("torch.allclose", new=fp8_allclose):
67
+ return (
68
+ torch.library.opcheck(
69
+ op, args, kwargs, test_utils=test_utils, raise_exception=raise_exception
70
+ )
71
+ if cond
72
+ else {}
73
+ )