## Background
This PR adds `torch.utils.serialization.config.load.calculate_storage_offsets`. This option relies on the previous PR in this stack, where storage order was changed to non lexicographical. A `.format_version` entry was added to the zipfile and `calculate_storage_offsets` will only work on checkpoints with `.format_version`.
When this is turned on, for `torch.load(mmap=True)`, offsets of each storage record (other than the 0th storage will be calculated instead of relying on `miniz` APIs to determine this).
The existing APIs will issue multiple random reads (reading the end of central directory record, then reading the zipfile header for the record) to determine the storage offset where the record starts. This can greatly degrade `torch.load(mmap=True)` performance for non-filesystem cases.
6aaae9d78f/caffe2/serialize/inline_container.cc (L589-L605)
## How does this work
The format for the checkpoint is as such
```
archive_name/
|_ data.pkl
|_.format_version
|_byteorder
|_data/
|_ 0
|_ 1
|_ 2
|_ ...
|_
```
Each `data/i` record represents a storage, where storages are written in the order that the Pickler encounters them.
For each storage, our `persistent_load` logic saves the following metadata to the pickle file `dtype, numel, key, location` where `numel` is the number of bytes in the storage.
Note that we always use `miniz` writer in the zip64 mode per [here](7796e308d0/caffe2/serialize/inline_container.cc (L701)) A zipfile record written by miniz looks as such
```
---------------- ----------------- ------------------- ---------------- --------- ------------------------------
| 30 byte header | n byte filename | zip64_extra_data | m byte padding | storage | 16 or 24 byte local dir footer |
---------------- ----------------- ------------------- ---------------- --------- ------------------------------
```
- The header size (30) is given by [`MZ_ZIP_LOCAL_DIR_HEADER_SIZE`](https://github.com/pytorch/pytorch/blob/main/third_party/miniz-3.0.2/miniz.c?fbclid=IwZXh0bgNhZW0CMTEAAR2O8Vysd--UoSCxW70gabXIS1dbz733oHwuUQ5_Ff1hY2WU6PL2i6CSH4A_aem_J9oaU2HpDeWtJKOU9EnVqw#L3290)
- filename will be `"{archive_name}/{filepath}"`
- `zip64_extra_data` is determined by [`mz_zip_writer_create_zip64_extra_data`](7796e308d0/third_party/miniz-3.0.2/miniz.c (L6202)). Note that [we only create zip64_extra_data if storage_size >= 0xFFFFFFFF or the offset of the start of the header >= 0xFFFFFFFF](7796e308d0/third_party/miniz-3.0.2/miniz.c (L6519-L6524))
- `m` is determined by [`getPadding`](7796e308d0/caffe2/serialize/inline_container.cc (L254)), which accounts for filename, zip64_extra_data to determine `m` such that the start of `storage` is aligned to 64 bytes. The `m` bytes will always start with `F B padding_size" as the first 4 bytes
- The local dir footer size is determined based on [this snippet ](7796e308d0/third_party/miniz-3.0.2/miniz.c (L6610-L6632)): if the buffer size is 0 it is skipped. If the zip64_extra_data was created, it is 24, otherwise it is 16.
When `torch.utils.serialization.config.load.calculate_storage_offsets` is set we do the following
- We keep track of where the "cursor" is in the file using `current_offset`, after each persistent_load call, it will be at the offset where the header for the next record starts
- for the 0th storage, "data/0", we use the regular get_record_offset to determine the start of the storage
- for any other storage, (where the storages will be in order encountered by the unpickler, 0, 1, 2, 3, ...) we use `get_record_offset_no_read`, which re-uses the `getPadding` logic to determine the offset of the storage
- Note that `load_tensor` will only ever be called again with the same key if the storage's `._data_ptr()` is 0 [[pointer1](https://github.com/pytorch/pytorch/blob/main/torch/serialization.py#L1917-L1918)][[pointer2](https://github.com/pytorch/pytorch/blob/main/torch/serialization.py#L1936-L1937)], so we cache the offsets for this edge case
- After each storage, if the storage is non-zero, we account for the local dir footer based on the logic described above
## Testing strategy
The agreed upon testing strategy was as follows:
- Add debug code gated by an environment flag `TORCH_SERIALIZATION_DEBUG` that will run this offset calculation logic and verify it against getRecordOffset for each storage (when mmap=False)
- This flag is set throughout CI, which means that every time `torch.load` is called, the offset calculation logic is implicitly being tested.
Differential Revision: [D67673026](https://our.internmc.facebook.com/intern/diff/D67673026)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143880
Approved by: https://github.com/albanD
ghstack dependencies: #143879
## Background
This PR adds `torch.utils.serialization.config.load.calculate_storage_offsets`. This option relies on the previous PR in this stack, where storage order was changed to non lexicographical. A `.format_version` entry was added to the zipfile and `calculate_storage_offsets` will only work on checkpoints with `.format_version`.
When this is turned on, for `torch.load(mmap=True)`, offsets of each storage record (other than the 0th storage will be calculated instead of relying on `miniz` APIs to determine this).
The existing APIs will issue multiple random reads (reading the end of central directory record, then reading the zipfile header for the record) to determine the storage offset where the record starts. This can greatly degrade `torch.load(mmap=True)` performance for non-filesystem cases.
6aaae9d78f/caffe2/serialize/inline_container.cc (L589-L605)
## Testing strategy
The agreed upon testing strategy was as follows:
- Add debug code gated by an environment flag `TORCH_SERIALIZATION_DEBUG` that will run this offset calculation logic and verify it against getRecordOffset for each storage (when mmap=False)
- This flag is set throughout CI, which means that every time `torch.load` is called, the offset calculation logic is implicitly being tested.
Differential Revision: [D67673026](https://our.internmc.facebook.com/intern/diff/D67673026)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143880
Approved by: https://github.com/albanD
ghstack dependencies: #143879
Add `PyTorchFileWriter.write_record_metadata(record_name, num_bytes)` that
- writes the zipfile header/end of central directory metadata for an entry*
- reserves `num_bytes` in the zipfile for the payload.
*Since the payload is not provided, the CRC32 computation is skipped and 0s are written in the corresponding entry of the zipfile header
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125184
Approved by: https://github.com/albanD
Summary:
Use concurrent multiple readers to access record from different start index. It can provide better performance when the data being accessed is large.
bypass-github-pytorch-ci-checks
Test Plan:
```
buck2 run @//mode/dev //caffe2/caffe2/serialize:inline_container_test
```
Reviewed By: YazhiGao
Differential Revision: D50957607
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112818
Approved by: https://github.com/houseroad, https://github.com/huydhn
Summary:
Zion-4s core has poor perf when it comes to reading the large tensor (e.g. 300G), no matter for manifold downloading or reading from files. In this diff, I changed the getRecord function from single thread to multiple threads by passing multiple readers to getRecord function and access the same record at different chunks with different readers.
We control the number of additional reader with the`sigrid_model_manager_additional_reader` flag. The default value is 0. When `additional_reader=2`, we allocate `2` extra read client threads.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111426
Approved by: https://github.com/jiayisuse
Summary:
serialization_id was added in a previous change to be written as a random GUID associated with each time saving of a module is called, for the purpose of adding tracking for saved artifacts. In order not to disturb existing systems that rely on the serialized bytes to be deterministic for serializing the same module, this change uses the combined hash of uncompressed content and file names instead of GUID for serialization id.
The use of this hashing reuses the same CRC32 that is already calculated for zip writing, so it doesn't incur additional computational overhead.
Data descriptor is one of the file headers inside the zip format https://en.wikipedia.org/wiki/ZIP_(file_format)#Data_descriptor. It contains the CRC32 of the uncompressed data. By inspecting the written data in PyTorchStreamWriter, the CRC32 is found for each written record.
In order to make serialization_id a unique and deterministic id for the
serialized files without computation overhead, the updated `serialization_id` is computed based on all files written, and is composed of:
1) a combined hash of record name hashes
2) a combined crc32 of the record uncompressed data
Example value: "15656915541136177431866432772"
Test Plan: buck2 test @//mode/dev //caffe2/caffe2/serialize:inline_container_test
Differential Revision: D46038973
Pull Request resolved: https://github.com/pytorch/pytorch/pull/101964
Approved by: https://github.com/davidberard98
Summary:
In order to better track models after serialization, this change writes a serialization_id as a UUID to inline container. Having this ID enables traceability of model in saving and loading events.
serialization_id is generated as a new UUID everytime serialization takes place. It can be thought of as a model snapshot identifier at the time of serialization.
Test Plan:
```
buck2 test @//mode/dev //caffe2/caffe2/serialize:inline_container_test
```
Local tests:
```
buck2 run @//mode/opt //scripts/atannous:example_pytorch_package
buck2 run @//mode/opt //scripts/atannous:example_pytorch
buck2 run @//mode/opt //scripts/atannous:example_pytorch_script
```
```
$ unzip -l output.pt
Archive: output.pt
Length Date Time Name
--------- ---------- ----- ----
36 00-00-1980 00:00 output/.data/serialization_id
358 00-00-1980 00:00 output/extra/producer_info.json
58 00-00-1980 00:00 output/data.pkl
261 00-00-1980 00:00 output/code/__torch__.py
326 00-00-1980 00:00 output/code/__torch__.py.debug_pkl
4 00-00-1980 00:00 output/constants.pkl
2 00-00-1980 00:00 output/version
--------- -------
1045 7 files
```
```
unzip -p output.pt "output/.data/serialization_id"
a9f903df-cbf6-40e3-8068-68086167ec60
```
Differential Revision: D45683657
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100994
Approved by: https://github.com/davidberard98
Avoid double exception in destructor if attempting to serialize to
python object that does not have `write` method
Use `Finalizer` class in `PyTorchStreamWriter::writeEndOfFile()` to a
always set `finailized_` property even if excretion occurs. (as there
isn't much one can do at this point)
Add expicit check for the attribue to `_open_zipfile_writer_buffer` and
add unitests
Modernize code a bit by using Python-3 `super()` method
Fixes https://github.com/pytorch/pytorch/issues/87997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88128
Approved by: https://github.com/albanD
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/61805
Similar in spirit to https://github.com/pytorch/pytorch/pull/61371.
While writing two files with the same name is allowed by the ZIP format,
most tools (including our own) handle this poorly. Previously I banned
this within `PackageExporter`, but that doesn't cover other uses of the
zip format like TorchScript.
Given that there are no valid use cases and debugging issues caused by
multiple file writes is fiendishly difficult, banning this behavior enitrely.
Differential Revision:
D29748968
D29748968
Test Plan: Imported from OSS
Reviewed By: Lilyjjo
Pulled By: suo
fbshipit-source-id: 0afee1506c59c0f283ef41e4be562f9c22f21023
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/51754
This API allows you to manage multiple python interpreters in a single
process to deploy PyTorch models packaged with torch.package.
torch/csrc/deploy/deploy.h contains the API definition
torch/csrc/deploy/test_deploy.cpp has some examples.
Notes:
* mutex is added to PyTorchStreamReader to make it safe to use from multiple threads at once.
* USE_DEPLOY is only true for the special libtorch_deployinterpreter.so library, when enabled
we use a hash table to maintain PyObject <> at::Tensor mappping rather than the internal pointer
in Tensor since >1 interpreter may have a reference to the tensor.
* serialization.py has some additional functions for creating pickle objects
but keeping storages in memory for use transfering tensors between interpreters
Test Plan: Imported from OSS
Reviewed By: wconstab
Differential Revision: D26329468
Pulled By: zdevito
fbshipit-source-id: d75f4ebb9a27f1d911179d9996041bcb3ca04a07
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49905
There's size regression in model delivery in D25682312. Only the model version numbers are used. However, the dependency of the entire c10 (128 KB) is pulled in.
This diff is to decouple the version numbers to a separate header file, versions.h. Other targets referring to version numbers only can have deps of ```caffe2:version_headers```.
ghstack-source-id: 119161467
Test Plan: CI
Reviewed By: xcheng16, guangyfb
Differential Revision: D25716601
fbshipit-source-id: 07634bcf46eacfefa4aa75f2e4c9b9ee30c6929d
Summary:
Since caffe2 and torch have been consolidated, CAFFE2_API should be merged with TORCH_API. Addresses a TODO.
Manually edited some references of the removed `CAFFE2_API`:
* `CONTRIBUTING.md`
* `caffe2/proto/CMakeLists.txt`
* `cmake/ProtoBuf.cmake`
* `c10/macros/Export.h`
* `torch/csrc/WindowsTorchApiMacro.h`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49496
Reviewed By: malfet, samestep
Differential Revision: D25600726
Pulled By: janeyx99
fbshipit-source-id: 7e068d959e397ac183c097d7e9a9afeca5ddd782
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48863
Support default arguments when invoking a module via PyTorch Lite (`mobile::Module`).
Test Plan:
buck test mode/dbg //caffe2/test/cpp/jit:jit -- LiteInterpreterTest.MethodInvocation
buck test mode/dbg caffe2/test:mobile -- test_method_calls_with_optional_arg
Reviewed By: raziel, iseeyuan
Differential Revision: D25152559
fbshipit-source-id: bbf52f1fbdbfbc6f8fa8b65ab524b1cd4648f9c0
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48802
Current torch::jit::load API only supports unique_ptr ReadAdaptInterface input, but for some cases, torch::jit::load may not be the only consumer of the reader adapter. This diff enables an overload of torch::jit::load to load shared_ptr PyTorchStreamReader.
Reviewed By: malfet, houseroad
Differential Revision: D25241904
fbshipit-source-id: aa403bac9ed820cc0e94342aebfe524a1d5bf913
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/43531
It's useful for building some tooling out of tree to manipulate zip files in PyTorch-y way
Test Plan: contbuild
Reviewed By: houseroad
Differential Revision: D23277361
fbshipit-source-id: e15fad20e792d1e41018d32fd48295cfe74bea8c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/42413
When a default argument is added, it does not break backward compatibility (BC) for full-jit, but does break BC for mobile bytecode. For example, https://github.com/pytorch/pytorch/pull/40737. To make bytecode BC in this case, we
1. Introduce kMinSupportedBytecodeVersion. The loaded model version should be between kMinSupportedBytecodeVersion and kProducedBytecodeVersion.
2. If an operator is updated, and we can handle BC, bump the kProducedBytecodeVersion (for example, from 3 to 4).
3. If model version is at the older version of the operator, add an adapter function at loading. For the added default arg, we push this default arg to stack before calling the actual operator function.
Test Plan: Imported from OSS
Reviewed By: xcheng16
Differential Revision: D22898314
Pulled By: iseeyuan
fbshipit-source-id: 90d339f8e1365f4bb178db8db7c147390173372b
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36439
A proposal of versioning in bytecode, as suggested by dzhulgakov in the internal post: https://fb.workplace.com/groups/pytorch.mobile.work/permalink/590192431851054/
kProducedBytecodeVersion is added. If the model version is not the same as the number in the code, an error will be thrown.
The updated bytecode would look like below. It's a tuple of elements, where the first element is the version number.
```
(3,
('__torch__.m.forward',
(('instructions',
(('STOREN', 1, 2),
('DROPR', 1, 0),
('MOVE', 2, 0),
('OP', 0, 0),
('RET', 0, 0))),
('operators', (('aten::Int', 'Tensor'),)),
('constants', ()),
('types', ()),
('register_size', 2))))
```
Test Plan: Imported from OSS
Differential Revision: D22433532
Pulled By: iseeyuan
fbshipit-source-id: 6d62e4abe679cf91a8e18793268ad8c1d94ce746
Summary:
BC NOTE:
This change makes it so modules saved with torch.jit.save in PyTorch 1.6 can be loaded by previous versions of PyTorch unless they use torch.div or (soon) torch.full. It also lets tensors saved using torch.save be loaded by previous versions. So this is the opposite of BC-breaking, but I'm using that label to highlight this issue since we don't have a "BC-improving" label.
PR NOTE:
When an operator's semantics change in PyTorch we want to do two things:
1) Preserve the semantics of older serialized Torchscript programs that use the operator
2) Ensure the new semantics are respected
Historically, this meant writing a Versioned Symbol that would remap older versions of the operator into current PyTorch code (1), and bumping the produced file format version (2). Unfortunately, bumping the produced file format version is a nuclear option for ensuring semantics are respected, since it also prevents older versions of PyTorch from loading anything (even tensors!) from newer versions.
Dynamic versioning addresses the nuclear consequences of bumping the produced file format version by only bumping it when necessary. That is, when an operator with changed semantics is detected in the serialized Torchscript. This will prevent Torchscript programs that use the changed operator from loading on earlier versions of PyTorch, as desired, but will have no impact on programs that don't use the changed operator.
Note that this change is only applicable when using torch.jit.save and torch.jit.load. torch.save pickles the given object using pickle (by default), which saves a function's Python directly.
No new tests for this behavior are added since the existing tests for versioned division in test_save_load already validate that models with div are loaded correctly at version 4.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40279
Reviewed By: dzhulgakov
Differential Revision: D22168291
Pulled By: mruberry
fbshipit-source-id: e71d6380e727e25123c7eedf6d80e5d7f1fe9f95
Summary:
BC-breaking NOTE:
In PyTorch 1.6 bool and integral fill values given to torch.full must set the dtype our out keyword arguments. In prior versions of PyTorch these fill values would return float tensors by default, but in PyTorch 1.7 they will return a bool or long tensor, respectively. The documentation for torch.full has been updated to reflect this.
PR NOTE:
This PR causes torch.full to throw a runtime error when it would have inferred a float dtype by being given a boolean or integer value. A versioned symbol for torch.full is added to preserve the behavior of already serialized Torchscript programs. Existing tests for this behavior being deprecated have been updated to reflect it now being unsupported, and a couple new tests have been added to validate the versioned symbol behavior. The documentation of torch.full has also been updated to reflect this change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40364
Differential Revision: D22176640
Pulled By: mruberry
fbshipit-source-id: b20158ebbcb4f6bf269d05a688bcf4f6c853a965
Summary:
Reserves file format version 5 for marking when torch.full(int)->FloatTensor will be deprecated.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40089
Differential Revision: D22066359
Pulled By: mruberry
fbshipit-source-id: 6158e03ca75e3795a2641123ff23d67975170f44
Summary:
**1.6 Deprecation Note**
In PyTorch 1.6 attempting to divide two integer tensors or an integer tensor and an integer scalar will throw a runtime error. This behavior was deprecated with a warning in PyTorch 1.5. In PyTorch 1.7 torch.div and the division operator will always perform true division like Python3 and NumPy.
To divide integer values use either torch.true_divide, for true division, or torch.floor_divide (the // operator) for floor division.
**PR Summary**
This PR updates the warning message when performing integer division to be a runtime error. Because some serialized Torchscript programs may rely on torch.div's historic behavior it also implements a "versioned symbol" for div that lets those models retain their current behavior. Extensive tests of this behavior are the majority of this PR.
Note this change bumps the produced file format version to delineate which programs should have their historic div behavior preserved.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/38620
Differential Revision: D21612598
Pulled By: mruberry
fbshipit-source-id: c9c33591abce2f7e97f67f0f859901f5b03ed47d
Summary:
See https://github.com/pytorch/pytorch/pull/38620 for additional context.
When PyTorch begins producing file format 4 with the updated div behavior it's safe for older PyTorch versions to consume it, since file format 4 only prohibits functionality. Bumping the supported file format version now gives PyTorch users on Master some leeway on updating their services that consume vs. produce PyTorch.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39529
Differential Revision: D21886790
Pulled By: mruberry
fbshipit-source-id: d6098eff06c26f18c3fac5cc85e5db298ba86e27
Summary:
Stacked PRs
* #33474 - [jit] Remove list specializations from pickler
* **#33255 - [jit] Add type tags to lists/dicts in pickle**
This adds a global call to `torch.jit._pickle.restore_type_tags` for
lists and dicts so that we can preserve their types after serialization.
](https://our.intern.facebook.com/intern/diff/20346780/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33255
Pulled By: driazati
Differential Revision: D20346780
fbshipit-source-id: c8534954ef4adb2e3c880401acbee30cd284f3db
Summary:
Stacked PRs
* #33474 - [jit] Remove list specializations from pickler
* **#33255 - [jit] Add type tags to lists/dicts in pickle**
This adds a global call to `torch.jit._pickle.restore_type_tags` for
lists and dicts so that we can preserve their types after serialization.
](https://our.intern.facebook.com/intern/diff/19868637/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33255
Pulled By: driazati
Reviewed By: xman1979, Tianshu-Bao
Differential Revision: D19868637
fbshipit-source-id: 2f1826e6679a786ca209198690269f399a542c04
