Summary: Currently OpKind is stored as an object field called op_ for each IR
node, and one usage of op_ is to avoid dynamic_cast in NodeCast when we
need to downcast a base-node pointer into a concrete sub-node pointer.
As a result, we need to construct and pass in an op when downcasting
nodes, and this becomes quite anonnying when we start to implement the
trie-based IR node reusing. More importantly, the op for each subclass
should be unique for that subclass and thus making it a const static field
is a more logical design.
In this PR, we still keep the object-level op_ for easier XLA adoption. As
furture work, we can come back to remove op_, make the op() method
virtual, and get rid of OpKind in all the node constructors.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76711
Approved by: https://github.com/wconstab, https://github.com/JackCaoG
Summary: TrieCache provides a way to look up an IR node before we
actually create it. If the lookup hits in TrieCache, we reuse the
existing node and move the current pointer in TrieCache to point to that
node; if the lookup misses, we create a new node and insert it into TrieCache.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76542
Approved by: https://github.com/wconstab, https://github.com/JackCaoG
Next stage of breaking up https://github.com/pytorch/pytorch/pull/74710
IR builder class introduced to decouple the explicit usage of `TsNode` in core lazy tensors.
Requires https://github.com/pytorch/pytorch/pull/75324 to be merged in first.
**Background**
- there are ~ 5 special ops used in lazy core but defined as :public {Backend}Node. (DeviceData, Expand, Scalar...)
- we currently require all nodes derive from {Backend}Node, so that backends can make this assumption safely
- it is hard to have shared 'IR classes' in core/ because they depend on 'Node'
**Motivation**
1. avoid copy-paste of "special" node classes for each backend
2. in general decouple and remove all dependencies that LTC has on the TS backend
**Summary of changes**
- new 'IRBuilder' interface that knows how to make 5 special ops
- move 'special' node classes to `ts_backend/`
- implement TSIRBuilder that makes the special TS Nodes
- new backend interface API to get the IRBuilder
- update core code to call the builder
CC: @wconstab @JackCaoG @henrytwo
Partially Fixes#74628
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75433
Approved by: https://github.com/wconstab
- Allow registering custom decompositions
- Add easier API for invoking decompositions
- Shorten API names (no users yet)
I am doing these as one pr because they are fairly short/simple and because github first does not support ghstack yet.
cc @Chillee @zou3519
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76252
Approved by: https://github.com/davidberard98
This **roughly** corresponds to Goal 3.2 in https://docs.google.com/document/d/1iiLNwR5ohAsw_ymfnOpDsyF6L9RTUaHMpD8YLw-jxEw/edit#
Namely:
It adds the following:
* SymbolicIntNode interface
* LazySymbolicIntNode implementation
* Lazy `narrow_copy` implementation
* Need add support for SymInt in codegen
* Test (below)
```cpp
TEST(LazyDynamicOpsTest, NarrowCopy) {
auto x = torch::rand({5, 10, 10}).to(kLazy);
const size_t Y_DIM = 3;
const size_t X_DIM_INDEX = 2;
auto y = torch::rand({Y_DIM}).to(kLazy);
auto ly = torch::lazy::TryGetLtcTensor(y);
auto dim_node = MakeNode<SizeNode>(ly->GetIrValue(), 0);
auto lmn = new torch::lazy::SymbolicIntNode(dim_node);
auto z = x.narrow_copy(X_DIM_INDEX, 0, lmn->toSymInt());
AllClose(z.cpu(), x.cpu().narrow_copy(X_DIM_INDEX, 0, Y_DIM));
}
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75759
Approved by: https://github.com/wconstab
Fixes#75177
Couldn't find any utility method to get directory name in pytorch repo, hence creating a function for that.
Let me know if a new function is not needed.
I also referred [this](https://github.com/pytorch/pytorch/blob/master/c10/test/util/tempfile_test.cpp#L15) for directory check.
Also I am using TORCH_CHECK to show the error. This is highly verbose with the entire stack visible. Is there any alternative for the same so that it is easier to read? This could happen a frequently, so small and concise error would be more helpful here.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75681
Approved by: https://github.com/albanD
Summary:
The `Buf` uses `std::vector<ExprHandle>` to represent its strides. The `ExprHandle` could be an immediate value or a mathematical expression with variables involved both for the static shape and dynamic shape. So it is hard to directly deduce the channels-last contiguous layout based on the numerical calculation. Hence, the utility functions of this PR are based on the pattern match to check whether the `Buf` is channels-last contiguous.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75938
Reviewed By: cpuhrsch
Differential Revision: D35724091
Pulled By: ZolotukhinM
fbshipit-source-id: f79ae21749d0aad8601f0434b52df88602ff09bf
(cherry picked from commit 3712bbbe4bea57c5c1abe1eafde4b8778e13e0c4)
Summary:
Next stage of breaking up https://github.com/pytorch/pytorch/pull/74710
Move shape cache implementation to the backend interface. Also, clean up some of the hashing logic in the base node class.
CC: wconstab JackCaoG henrytwo
Partially Fixes https://github.com/pytorch/pytorch/issues/74628
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75324
Reviewed By: anjali411
Differential Revision: D35730823
Pulled By: wconstab
fbshipit-source-id: cf6fa326319b9324e5f422a78817b6fb5bf7e9b8
(cherry picked from commit faec5043df56639e2fd23de2d91ae796e4f3df70)
Moves jit shape function registration to python. Like jit decompositions, a script must be run after adding new definitions which serializes them in a c++ file.
This was a request so that torch-mlir could define functions in python and upstream their shape functions. cc @silvasean @makslevental
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75546
Approved by: https://github.com/davidberard98
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75807
There is a tension in RecordFunction between two use cases:
1) In the normal eager path we don't run any callbacks, so we need to bail out of the profiling path as soon as possible to minimize eager overhead.
2) When profiling we want to determine which callbacks to run as efficiently as possible to minimize instrumentation overhead.
The confounding factor in all of this is sampling callbacks because they change which callbacks will run on each call, even in steady state operation. This has traditionally been handled with a two stage procedure: first we flip a coin to determine if a sampled callback *might* run. If false (which it usually is), do nothing. This solves (1). If true, check to see if we need to build the full callback set or if it was a false positive. This procedure has two negative effects:
* It forces us to rebuild the set of callbacks to run on every step when profiling
* It leaks the sampling abstraction, requiring other parts of the code to bump certain values and forces RecordFunction to lazily initialize.
This change introduces a multi-level cache which can (in the common case) quickly determine which callbacks *will* run, rather than if callbacks *might* run. This means that rather than call `shouldRunRecordFunction`, we can simply get the callbacks for an invocation and check if they are empty. (And completely removes the pre-sampling heuristic.) Another major benefit of the new cache structure is that it allows thread-safe registration and unregistration of global callbacks.
It's worth briefly discussing how this maintains eager performance. In the standard eager case (only sampling callbacks registered) the cache first checks that the global callbacks haven't changed (atomic read), decrements a counter to see if a sampling callback fired, and then returns the active callbacks which is simply a SmallVector of pointer pairs and a couple POD values (scope, needs inputs/outputs/ids). The biggest cost according to perf is the SmallVector logic; we could consider adopting a hard limit on active callbacks; more than half a dozen callbacks *running* in a single step would be quite a lot. But the total cost relative to `PYTORCH_DISABLE_PER_OP_PROFILING` is only ~10ns, so debatable if it's worth it to switch to `std::array`.
The primary change is in `record_function.cpp`, which has a more detailed description of the new cache structure. `record_function.h` has some minor changes to align with the new calling convention and the remaining files are simply changes to the call sites.
Future work:
* RecordFunction no longer needs to be lazily initialized.
* We can deprecate the disable/reenable APIs, since we can not safely add and remove global callbacks.
Test Plan:
I tested eager mode performance using the overhead benchmark and found that the non-profiled path was unaffected. However the no-op observer dropped from 0.41us to 0.37us (0.25us if no observers are active) which is about 1/3rd reduction in the cost of the callback selection machinery.
I also added several C++ unit tests, as the core RecordFunction machinery (especially sampling) was largely untested.
Reviewed By: swolchok, davidberard98
Differential Revision: D35276158
fbshipit-source-id: 35135f444724fba4eb97c0ae7f3f710f0f9016fd
(cherry picked from commit 9e359b87422c18f2a195185f32e7e85c82f956fd)
Summary:
## Original commit message:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73368
debug_pkl file inside of pytorch's .pt file consists of a list of SourceRanges. Each SourceRange points to a Source which is a stack track, filename, and start, end numbers. Those are emitted in debug_pkl file as strings.
Since many SourceRange shares the same source, the string for trace can be deduped.
The newer format saves a set of unique traces in a tuple, then each SourceRange will save the offset of it's trace w.r.t. position in that tuple. (i.e. manually applying dictionary compression).
The above helps with smaller file size. On loading, if we copy each trace to Source as string the runtime memory would still blowup.
To mitigate this, we use SourceView directly instead of source which will take the reference of string inside of Deserializer and make that into string_view. This is safe because Deserializer is hold by Unpickler by shared_ptr, and Unpickler is also hold by shared_ptr by another Source object. That Source object will be alive during the model construction.
Test Plan:
## Original Test plan
unit test
Took original file (312271638_930.predictor.disagg.local); loaded with `torch.jit.load` save again with `torch.jit.save`. Unzip both, look at contents:
```
[qihan@devvm5585.vll0 ~]$ du archive -h
4.0K archive/xl_model_weights
3.7M archive/extra
8.0K archive/code/__torch__/caffe2/torch/fb/model_transform/splitting
8.0K archive/code/__torch__/caffe2/torch/fb/model_transform
8.0K archive/code/__torch__/caffe2/torch/fb
8.0K archive/code/__torch__/caffe2/torch
8.0K archive/code/__torch__/caffe2
20M archive/code/__torch__/torch/fx/graph_module
20M archive/code/__torch__/torch/fx
8.0K archive/code/__torch__/torch/classes
20M archive/code/__torch__/torch
20M archive/code/__torch__
20M archive/code
2.7M archive/constants
35M archive
[qihan@devvm5585.vll0 ~]$ du resaved -h
4.0K resaved/extra
8.0K resaved/code/__torch__/caffe2/torch/fb/model_transform/splitting
8.0K resaved/code/__torch__/caffe2/torch/fb/model_transform
8.0K resaved/code/__torch__/caffe2/torch/fb
8.0K resaved/code/__torch__/caffe2/torch
8.0K resaved/code/__torch__/caffe2
1.3M resaved/code/__torch__/torch/fx/graph_module
1.3M resaved/code/__torch__/torch/fx
8.0K resaved/code/__torch__/torch/classes
1.4M resaved/code/__torch__/torch
1.4M resaved/code/__torch__
1.4M resaved/code
2.7M resaved/constants
13M resaved
[qihan@devvm5585.vll0 ~]$
```
## Additional test:
`buck test mode/dev-tsan //caffe2/benchmarks/static_runtime:static_runtime_cpptest -- --exact 'caffe2/benchmarks/static_runtime:static_runtime_cpptest - StaticRuntime.to'` passes
test jest.fbios.startup_cold_start.local.simulator f333356873 -
Differential Revision: D35196883
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74869
Approved by: https://github.com/gmagogsfm
Summary:
Tensor.is_alias_of relies on Storage to perform. However, LTCTensorImpl was
not implemented with that in mind. This commit adds a fake storage to LazyTensor
as a marker to mark LazyTensors that point to the same storage. The reason
why it's not done at LTCTensorImpl is that LazyTensor maintains the view ops/alias
logic in LazyTensor class instead of relying on TensorImpl to do the check.
Test Plan:
./build/bin/test_lazy --gtest_filter=LazyOpsTest.IsAliasOf
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75246
Approved by: https://github.com/bdhirsh
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75230
Op diagonal is a view op which we can't code-gen yet. Therefore, support
it by making hand-written IR construction and lowering.
Test Plan: ./build/bin/test_lazy --gtest_filter=LazyOpsTest.TestDiagonal*
Reviewed By: wconstab
Differential Revision: D35378316
Pulled By: alanwaketan
fbshipit-source-id: 7958d00107aef20ac37aabcf2868346240977530
(cherry picked from commit 84155528fce484627c9688cfd92fd4aeb68219e5)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75201
In this diff:
1. Bump supported version to 9, which will serve as a placeholder for upcoming version bump to v9 for flatbuffer format migration.
2. Implements backport from v9 flatbuffer file to v8 pickle file.
ghstack-source-id: 153225189
(Note: this ignores all push blocking failures!)
Test Plan:
fb:
```
cd ~/fbsource/fbcode/ && buck test -c fbcode.caffe2_enable_flatbuffer=1 caffe2/test/cpp/jit:jit -- LiteInterpreterTest.BackPortByteCodeModelAllVersions
Parsing buck files: finished in 0.7 sec
Downloaded 0/25 artifacts, 0.00 bytes, 100.0% cache miss (for updated rules)
Building: finished in 20.7 sec (100%) 21783/21783 jobs, 5/21783 updated
cd ~/fbsource/fbcode/ && buck test caffe2/test/cpp/jit:jit -- FlatbufferTest.FlatbufferBackPortTest
Parsing buck files: finished in 0.7 sec
Building: finished in 4.5 sec (100%) 12972/53298 jobs, 0/53298 updated
Total time: 5.3 sec
More details at https://www.internalfb.com/intern/buck/build/b658d597-d358-4293-97cb-28e7612b96e8
BUILD SUCCEEDED
Tpx test run coordinator for Facebook. See https://fburl.com/tpx for details.
Running with tpx session id: 35d5542d-6ee3-4c28-be10-1d822c7a6fef
Trace available for this run at /tmp/tpx-20220308-090347.891303-35d5542d-6ee3-4c28-be10-1d822c7a6fef/trace.log
RemoteExecution session id: reSessionID-35d5542d-6ee3-4c28-be10-1d822c7a6fef-tpx
Started reporting to test run: https://www.internalfb.com/intern/testinfra/testrun/8444249379196000
✓ ListingSuccess: caffe2/test/cpp/jit:jit : 490 tests discovered (22.838)
✓ Pass: caffe2/test/cpp/jit:jit - FlatbufferTest.FlatbufferBackPortTest (0.289)
Summary
Pass: 1
ListingSuccess: 1
If you need help understanding your runs, please follow the wiki: https://fburl.com/posting_in_tpx_users
Finished test run: https://www.internalfb.com/intern/testinfra/testrun/8444249379196000
```
Reviewed By: iseeyuan
Differential Revision: D34702597
fbshipit-source-id: 5c203c29d13360d7934ce6e57557739e7038c05e
(cherry picked from commit 6189e08a2bd968fdab636f77cb6bd73d6c36beb2)
Summary:
PyTorch supports registering a custom operator by `TORCH_LIBRARY_FRAGMENT` / `TORCH_LIBRARY_IMPL` and `torch::jit::tensorexpr::getNNCLoweringRegistry` could insert a custom operator. But the te fuser passes conditional check does not support custom operator. The `isSupported` of `tensorexpr_fuser` checks whether the `Node` is `get_tensorexpr_elementwise_set()`, `supported_non_eltwise_set()`, `supported_misc_set` and `supported_reduction_set`. If a custom operator needs to be added to the TE fusion group, the checked will block it.
Taking the RN50 as an example, we can speed up the model by fusing the convolution and consecutive element-wise operator into a custom operator. The framework overhead becomes non-negligible when the computation becomes more efficient, especially for the latency mode and the tiny models. If the TE fuser allows adding the custom operator to the fusion group, then the entire RN50 model could be fused by TE as a single operator/function consisting of "ExternalCalls" and TE-IR. This could significantly reduce framework overhead, which in turn improves RN50 E2E performance. The same goes for other models.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73073
Reviewed By: pbelevich
Differential Revision: D35453165
Pulled By: ZolotukhinM
fbshipit-source-id: a764cf340b0b1e05fe230649cbe44f5786bdd37d
(cherry picked from commit ee95aa4d36714540fbb216a338799e6a6bb966d5)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57775
The minimum supported bytecode version is updated from 3 to 4. We no longer support version 3 bytecode models.
Why?
* There are hacky codes in operator loading, that performs differently on one operator on the global bytecode version 3. Instead operator related metadata should be passed (for example, in #56845). To allow future development, we remove the hacky way first.
* The bytecode version was bumped from 3 to 4 more than half a year ago. Since all the production models are all bumped to version 4, it's not practical to keep and maintain version 3. The risk to deprecate version 3 is low.
Test Plan: Imported from OSS
Reviewed By: raziel
Differential Revision: D28270791
Pulled By: cccclai
fbshipit-source-id: 70b1bd6352fdaae5f8d2173b81578d77018c8e44
(cherry picked from commit 3e930fa381cd01f3705116795c6426df992372fc)
