Add top-level `layer_norm` wrapper with docs
Browse files
build/torch-universal/triton_layer_norm/__init__.py
CHANGED
@@ -1,5 +1,114 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
from . import layers
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
|
5 |
-
__all__ = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Triton layer normalization kernels.
|
2 |
+
|
3 |
+
This kernel implements layers normalization using Triton. This kernel is from
|
4 |
+
the `flash-attention <https://github.com/Dao-AILab/flash-attention>`_ project.
|
5 |
+
"""
|
6 |
+
|
7 |
+
from typing import Optional
|
8 |
+
|
9 |
+
import torch
|
10 |
|
11 |
from . import layers
|
12 |
+
from .layer_norm import layer_norm_fn, layer_norm_linear_fn, rms_norm_fn
|
13 |
+
|
14 |
+
|
15 |
+
def layer_norm(
|
16 |
+
x: torch.Tensor,
|
17 |
+
weight: torch.Tensor,
|
18 |
+
bias: torch.Tensor,
|
19 |
+
residual: Optional[torch.Tensor] = None,
|
20 |
+
x1: Optional[torch.Tensor] = None,
|
21 |
+
weight1: Optional[torch.Tensor] = None,
|
22 |
+
bias1: Optional[torch.Tensor] = None,
|
23 |
+
eps: float = 1e-6,
|
24 |
+
dropout_p: float = 0.0,
|
25 |
+
rowscale=None,
|
26 |
+
prenorm: bool = False,
|
27 |
+
residual_in_fp32: bool = False,
|
28 |
+
is_rms_norm: bool = False,
|
29 |
+
return_dropout_mask: bool = False,
|
30 |
+
out: Optional[torch.Tensor] = None,
|
31 |
+
residual_out: Optional[torch.Tensor] = None,
|
32 |
+
):
|
33 |
+
"""
|
34 |
+
Apply layer normalization to the input tensor with Triton acceleration.
|
35 |
+
|
36 |
+
Args:
|
37 |
+
x (`torch.Tensor`):
|
38 |
+
Input tensor to normalize.
|
39 |
+
weight (`torch.Tensor`):
|
40 |
+
Scale parameter for normalization.
|
41 |
+
bias (`torch.Tensor`):
|
42 |
+
Shift parameter for normalization.
|
43 |
+
residual (`torch.Tensor`, *optional*):
|
44 |
+
Optional residual tensor to add to the input before normalization.
|
45 |
+
x1 (`torch.Tensor`, *optional*):
|
46 |
+
Optional second input tensor to combine with `x`. When provided, the function
|
47 |
+
first adds `x1` to `x` and then applies normalization.
|
48 |
+
weight1 (`torch.Tensor`, *optional*):
|
49 |
+
Scale parameter for the second normalization.
|
50 |
+
bias1 (`torch.Tensor`, *optional*):
|
51 |
+
Shift parameter for the second normalization.
|
52 |
+
eps (`float`, *optional*, defaults to 1e-6):
|
53 |
+
Small constant added for numerical stability in normalization.
|
54 |
+
dropout_p (`float`, *optional*, defaults to 0.0):
|
55 |
+
Dropout probability. If greater than 0, applies dropout to the input before
|
56 |
+
normalization and residual addition.
|
57 |
+
rowscale (`torch.Tensor`, *optional*):
|
58 |
+
Optional scaling factor applied to each row of the input tensor.
|
59 |
+
Not compatible with the use of `x1`.
|
60 |
+
prenorm (`bool`, *optional*, defaults to False):
|
61 |
+
If True, returns both the normalized output and the unnormalized input+residual.
|
62 |
+
residual_in_fp32 (`bool`, *optional*, defaults to False):
|
63 |
+
If True, performs the residual connection in FP32 precision.
|
64 |
+
is_rms_norm (`bool`, *optional*, defaults to False):
|
65 |
+
If True, uses RMS normalization instead of layer normalization.
|
66 |
+
return_dropout_mask (`bool`, *optional*, defaults to False):
|
67 |
+
If True, returns the dropout mask used for the computation.
|
68 |
+
out (`torch.Tensor`, *optional*):
|
69 |
+
Output tensor for the normalized result. If `None`, a new tensor is allocated.
|
70 |
+
residual_out (`torch.Tensor`, *optional*):
|
71 |
+
Output tensor for the residual result when using prenorm. If `None`, a new tensor
|
72 |
+
is allocated when needed.
|
73 |
+
|
74 |
+
Returns:
|
75 |
+
`torch.Tensor` or tuple of `torch.Tensor`:
|
76 |
+
- The normalized input.
|
77 |
+
- The second normalization of the input if `weight1` is provided.
|
78 |
+
- The residual tensor if `prenorm` is set.
|
79 |
+
- The dropout mask if `return_dropout_mask` is set.
|
80 |
+
- The dropout mask for `x1` if `x1` is provided and `return_dropout_mask` is set.
|
81 |
+
"""
|
82 |
+
return layer_norm_fn(
|
83 |
+
x,
|
84 |
+
weight,
|
85 |
+
bias,
|
86 |
+
residual,
|
87 |
+
x1,
|
88 |
+
weight1,
|
89 |
+
bias1,
|
90 |
+
eps,
|
91 |
+
dropout_p,
|
92 |
+
rowscale,
|
93 |
+
prenorm,
|
94 |
+
residual_in_fp32,
|
95 |
+
is_rms_norm,
|
96 |
+
return_dropout_mask,
|
97 |
+
out=out,
|
98 |
+
residual_out=residual_out,
|
99 |
+
)
|
100 |
+
|
101 |
+
|
102 |
+
__kernel_metadata__ = {
|
103 |
+
"license": "bsd-3-clause",
|
104 |
+
}
|
105 |
+
|
106 |
|
107 |
+
__all__ = [
|
108 |
+
"__kernel_metadata__",
|
109 |
+
"layers",
|
110 |
+
"layer_norm",
|
111 |
+
"layer_norm_fn",
|
112 |
+
"layer_norm_linear_fn",
|
113 |
+
"rms_norm_fn",
|
114 |
+
]
|
torch-ext/triton_layer_norm/__init__.py
CHANGED
@@ -1,5 +1,114 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
from . import layers
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
|
5 |
-
__all__ = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Triton layer normalization kernels
|
2 |
+
|
3 |
+
This kernel implements layers normalization using Triton. This kernel is from
|
4 |
+
the `flash-attention <https://github.com/Dao-AILab/flash-attention>`_ project.
|
5 |
+
"""
|
6 |
+
|
7 |
+
from typing import Optional
|
8 |
+
|
9 |
+
import torch
|
10 |
|
11 |
from . import layers
|
12 |
+
from .layer_norm import layer_norm_fn, layer_norm_linear_fn, rms_norm_fn
|
13 |
+
|
14 |
+
|
15 |
+
def layer_norm(
|
16 |
+
x: torch.Tensor,
|
17 |
+
weight: torch.Tensor,
|
18 |
+
bias: torch.Tensor,
|
19 |
+
residual: Optional[torch.Tensor] = None,
|
20 |
+
x1: Optional[torch.Tensor] = None,
|
21 |
+
weight1: Optional[torch.Tensor] = None,
|
22 |
+
bias1: Optional[torch.Tensor] = None,
|
23 |
+
eps: float = 1e-6,
|
24 |
+
dropout_p: float = 0.0,
|
25 |
+
rowscale=None,
|
26 |
+
prenorm: bool = False,
|
27 |
+
residual_in_fp32: bool = False,
|
28 |
+
is_rms_norm: bool = False,
|
29 |
+
return_dropout_mask: bool = False,
|
30 |
+
out: Optional[torch.Tensor] = None,
|
31 |
+
residual_out: Optional[torch.Tensor] = None,
|
32 |
+
):
|
33 |
+
"""
|
34 |
+
Apply layer normalization to the input tensor with Triton acceleration.
|
35 |
+
|
36 |
+
Args:
|
37 |
+
x (`torch.Tensor`):
|
38 |
+
Input tensor to normalize.
|
39 |
+
weight (`torch.Tensor`):
|
40 |
+
Scale parameter for normalization.
|
41 |
+
bias (`torch.Tensor`):
|
42 |
+
Shift parameter for normalization.
|
43 |
+
residual (`torch.Tensor`, *optional*):
|
44 |
+
Optional residual tensor to add to the input before normalization.
|
45 |
+
x1 (`torch.Tensor`, *optional*):
|
46 |
+
Optional second input tensor to combine with `x`. When provided, the function
|
47 |
+
first adds `x1` to `x` and then applies normalization.
|
48 |
+
weight1 (`torch.Tensor`, *optional*):
|
49 |
+
Scale parameter for the second normalization.
|
50 |
+
bias1 (`torch.Tensor`, *optional*):
|
51 |
+
Shift parameter for the second normalization.
|
52 |
+
eps (`float`, *optional*, defaults to 1e-6):
|
53 |
+
Small constant added for numerical stability in normalization.
|
54 |
+
dropout_p (`float`, *optional*, defaults to 0.0):
|
55 |
+
Dropout probability. If greater than 0, applies dropout to the input before
|
56 |
+
normalization and residual addition.
|
57 |
+
rowscale (`torch.Tensor`, *optional*):
|
58 |
+
Optional scaling factor applied to each row of the input tensor.
|
59 |
+
Not compatible with the use of `x1`.
|
60 |
+
prenorm (`bool`, *optional*, defaults to False):
|
61 |
+
If True, returns both the normalized output and the unnormalized input+residual.
|
62 |
+
residual_in_fp32 (`bool`, *optional*, defaults to False):
|
63 |
+
If True, performs the residual connection in FP32 precision.
|
64 |
+
is_rms_norm (`bool`, *optional*, defaults to False):
|
65 |
+
If True, uses RMS normalization instead of layer normalization.
|
66 |
+
return_dropout_mask (`bool`, *optional*, defaults to False):
|
67 |
+
If True, returns the dropout mask used for the computation.
|
68 |
+
out (`torch.Tensor`, *optional*):
|
69 |
+
Output tensor for the normalized result. If `None`, a new tensor is allocated.
|
70 |
+
residual_out (`torch.Tensor`, *optional*):
|
71 |
+
Output tensor for the residual result when using prenorm. If `None`, a new tensor
|
72 |
+
is allocated when needed.
|
73 |
+
|
74 |
+
Returns:
|
75 |
+
`torch.Tensor` or tuple of `torch.Tensor`:
|
76 |
+
- The normalized input.
|
77 |
+
- The second normalization of the input if `weight1` is provided.
|
78 |
+
- The residual tensor if `prenorm` is set.
|
79 |
+
- The dropout mask if `return_dropout_mask` is set.
|
80 |
+
- The dropout mask for `x1` if `x1` is provided and `return_dropout_mask` is set.
|
81 |
+
"""
|
82 |
+
return layer_norm_fn(
|
83 |
+
x,
|
84 |
+
weight,
|
85 |
+
bias,
|
86 |
+
residual,
|
87 |
+
x1,
|
88 |
+
weight1,
|
89 |
+
bias1,
|
90 |
+
eps,
|
91 |
+
dropout_p,
|
92 |
+
rowscale,
|
93 |
+
prenorm,
|
94 |
+
residual_in_fp32,
|
95 |
+
is_rms_norm,
|
96 |
+
return_dropout_mask,
|
97 |
+
out=out,
|
98 |
+
residual_out=residual_out,
|
99 |
+
)
|
100 |
+
|
101 |
+
|
102 |
+
__kernel_metadata__ = {
|
103 |
+
"license": "bsd-3-clause",
|
104 |
+
}
|
105 |
+
|
106 |
|
107 |
+
__all__ = [
|
108 |
+
"__kernel_metadata__",
|
109 |
+
"layers",
|
110 |
+
"layer_norm",
|
111 |
+
"layer_norm_fn",
|
112 |
+
"layer_norm_linear_fn",
|
113 |
+
"rms_norm_fn",
|
114 |
+
]
|
torch-ext/triton_layer_norm/layers.py
CHANGED
@@ -5,10 +5,32 @@ from .layer_norm import rms_norm_fn
|
|
5 |
|
6 |
|
7 |
class LlamaRMSNorm(nn.Module):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
weight: torch.Tensor
|
9 |
variance_epsilon: float
|
10 |
|
11 |
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
return rms_norm_fn(
|
13 |
hidden_states,
|
14 |
self.weight,
|
|
|
5 |
|
6 |
|
7 |
class LlamaRMSNorm(nn.Module):
|
8 |
+
"""
|
9 |
+
RMS Layer Norm for Llama models.
|
10 |
+
|
11 |
+
Triton-optimized RMS layer norm. The interface is compatible with `LLamaRMSNorm` in
|
12 |
+
`transformers`.
|
13 |
+
|
14 |
+
Attributes:
|
15 |
+
weight (`torch.Tensor`): The learnable scaling parameter.
|
16 |
+
variance_epsilon (`float`): The epsilon value for numerical stability.
|
17 |
+
"""
|
18 |
weight: torch.Tensor
|
19 |
variance_epsilon: float
|
20 |
|
21 |
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
22 |
+
"""
|
23 |
+
Apply RMS normalization to the input hidden states.
|
24 |
+
|
25 |
+
Args:
|
26 |
+
hidden_states (`torch.Tensor`):
|
27 |
+
Input tensor of shape `(batch_size, sequence_length, hidden_size)` or any shape
|
28 |
+
where the last dimension is the feature dimension to be normalized.
|
29 |
+
|
30 |
+
Returns:
|
31 |
+
`torch.Tensor`:
|
32 |
+
The normalized tensor with the same shape as the input `hidden_states`.
|
33 |
+
"""
|
34 |
return rms_norm_fn(
|
35 |
hidden_states,
|
36 |
self.weight,
|