File size: 1,862 Bytes
49670c7 | 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 | # MIT License
#
# Copyright (c) 2026 audio-embeddings contributors
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from __future__ import annotations
import torch
import torch.nn as nn
import torch.nn.functional as F
class MixedPrecisionRMSNorm(nn.RMSNorm):
"""RMSNorm that keeps its master weight while matching activation dtype.
PyTorch's fused RMSNorm requires the input and weight to have the same dtype.
Mixed-precision training keeps parameters in FP32, so use a differentiable
low-precision view of the weight for the operation itself.
"""
def forward(self, input: torch.Tensor) -> torch.Tensor:
weight = self.weight
if weight is not None and weight.dtype != input.dtype:
weight = weight.to(dtype=input.dtype)
return F.rms_norm(input, self.normalized_shape, weight, self.eps)
|