Fixes big endian arch bugs. #26383
Closed
Fixes big endian arch bugs. #26383
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Serialization.cpp fails on big endian machines.
This patch fixes the endian bugs and also makes the pytorch
model files portable across different endian architectures.
x86 generated model file can be read on s390 arch.
First problem, is serialization.cpp forgets to convert "size" value
of the storage elements to the native byte order.
torch.load throws an assertion as a result
(see the first stack trace below).
Second problem is when it reads the model from storage (doRead)
it decodes values to little endian which is the wrong order
on a big endian machine. The decode should be
to THP_nativeByteOrder() instead
(see the model dump below)
loaded_model = torch.load( opt.model_file, map_location=torch.device("cpu"))
File "/usr/local/lib/python3.6/dist-packages/torch/serialization.py", line 422, in load
return _load(f, map_location, pickle_module, **pickle_load_args)
File "/usr/local/lib/python3.6/dist-packages/torch/serialization.py", line 616, in _load
deserialized_objects[key]._set_from_file(f, offset, f_should_read_directly)
RuntimeError: storage has wrong size: expected 2305843009213693952 got 32
(the very long number is actually 32 in the wrong endianness)
x86 (original)
>>> import torch
>>> torch.load('400f2k_best.model', map_location=torch.device("cpu"))
{'epoch': 24, 'model_type': 'emb_aec', 'classifier_model': OrderedDict([('model.0.weight', tensor([[ 2.4608e-01, -1.1174e-01, -1.0854e-01, 4.0124e-01, -1.5261e-02,
-1.2206e-01, 1.3229e-01, -1.2615e-01, -5.2773e-01, 2.6333e-01,
-3.1462e-03, -1.4902e-01, 9.8545e-02, -1.5789e-01, -2.2625e-01,
-1.0776e-01, -9.0895e-02, -3.8530e-01, 9.1152e-01, -3.9720e-01,
-8.5848e-01, -4.7837e-02, -1.5178e-01, 8.5023e-02, 1.5013e-01,
-9.9294e-02, -2.7422e-01, -4.3986e-01, -4.4297e-01, -3.9570e-01,
s390x (wrong endianness)
>>> import torch
>>> torch.load( "400f2k_best.model", map_location=torch.device("cpu"))
{'epoch': 24, 'model_type': 'emb_aec', 'classifier_model': OrderedDict([('model.0.weight', tensor([[ 9.2780e+21, -9.7722e-11, 4.1350e+33, 7.782e+34, 4.2056e-31,
9.0784e+18, 1.1846e-32, 3.3320e-32, -4.8288e-28, -7.2679e+12,
1.5379e-16, -5.2604e+12, -4.7240e+17, 4.6092e-21, -1.8360e-20,
-2.7712e-31, 1.4548e-16, -2.5089e-27, 7.9094e-10, 7.1977e+34,
1.1930e+26, 8.4536e+15, 2.7757e+23, -5.8455e-10, -1.5611e+09,
-1.1311e-23, 6.6451e+19, -2.0970e+20, 3.4878e-19, -1.0857e-12,
7.8098e+22, 5.3998e-35],
pytorchbot
added
the
module: internals
Contributor
Author
|
@geert56 FYI |
Contributor
|
Does this fix an existing functional test on your big endian target, or should we add a dedicated test for it? |
Contributor
Author
@resistor Fixes an existing functional test. I don't think we need to add a test. The fix is operational and working as expected on our system. We also tested in on x86 (little endian) just to make sure we didn't break anything. |
resistor
reviewed
Sep 19, 2019
resistor
approved these changes
Sep 19, 2019
Contributor
resistor
left a comment
resistor
left a comment
There was a problem hiding this comment.
Approved modulo one nit.
detect erroneous usage if it ever arises.
Contributor
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@resistor has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Contributor
|
This pull request has been merged in afa5d08. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Serialization.cpp fails on big endian machines.
This patch fixes the endian bugs and also makes the pytorch
model files portable across different endian architectures.
x86 generated model file can be read on s390 arch.
First problem, is serialization.cpp forgets to convert "size" value
of the storage elements to the native byte order.
torch.load throws an assertion as a result
(see the first stack trace below).
Second problem is when it reads the model from storage (doRead)
it decodes values to little endian which is the wrong order
on a big endian machine. The decode should be
to THP_nativeByteOrder() instead
(see the model dump below)
Model file load on x86 (correct output)
Model file load on s390x (wrong endianness; notice the exponents)