Manual changes for moving dynamo to core #86621

jansel · 2022-10-10T18:33:46Z

This is the subset of the changes in #86461 not auto-generated by copy_to_core.sh.

pytorch-bot · 2022-10-10T18:33:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/86621

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit our office hours

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 1bf2406:
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

jansel · 2022-10-10T18:34:38Z

torch/csrc/dynamo/guards.cpp

@@ -0,0 +1,378 @@
+#define PY_SSIZE_T_CLEAN


This file is mostly copied from https://github.com/pytorch/torchdynamo/blob/main/torchdynamo/_guards.cpp

jansel · 2022-10-10T18:35:01Z

torch/csrc/dynamo/eval_frame.c

@@ -0,0 +1,578 @@
+#define PY_SSIZE_T_CLEAN


This file is mostly copied from https://github.com/pytorch/torchdynamo/blob/main/torchdynamo/_eval_frame.c

malfet

Can you please use stacks and say land just dyanmo cpp module as is? (Which sounds totally fine to me)

jansel · 2022-10-10T20:06:04Z

Can you please use stacks and say land just dyanmo cpp module as is? (Which sounds totally fine to me)

Isn't this PR almost entirely that? I am thinking we just land this, then rebase #86461.

albanD

guards.cpp not done yet.

setup.py

albanD · 2022-10-10T19:47:04Z

torch/csrc/dynamo/eval_frame.c

+#include <pystate.h>
+
+// see https://bugs.python.org/issue35886
+#if PY_VERSION_HEX >= 0x03080000


Should this be located in torch/csrc/utils/python_compat.h with all other similar changes?

We may want to keep this one in this file because those python headers don't import cleanly in C++.

albanD · 2022-10-10T19:47:30Z

torch/csrc/dynamo/eval_frame.c

+#define true 1
+
+#ifdef _WIN32
+#define unlikely(x) (x)


Use C10_UNLIKELY for this.

This is C code not C++.

Ho, c10/macros/Macros.h is not C compliant?

albanD · 2022-10-10T19:47:54Z

torch/csrc/dynamo/eval_frame.c

+#undef Py_BUILD_CORE
+#endif
+
+#define bool char


C doesn't have a bool type.

From Jason: this has to be plain C and so we redefine these.
Adding a comment would be nice.

Err? C99 supports bool, one just need to include <stdbool.h>

albanD · 2022-10-10T19:48:51Z

torch/csrc/dynamo/eval_frame.c

+#define unlikely(x) __builtin_expect((x), 0)
+#endif
+
+#define NULL_CHECK(val)                                         \


Should we have a TODO to migrate these to use TORCH_CHECK and TORCH_INTERNAL_ASSERT_DEBUG_ONLY?

Also we have throw python_error(); that should be used here.

C code, not C++ -- so can't trow.

We can migrate this in another PR. For now this is just a copy (with a minor tweak to the glue code in the last function).

albanD · 2022-10-10T21:23:29Z

torch/csrc/dynamo/eval_frame.c

+inline static PyObject* eval_frame_callback_get(void) {
+  void* result = PyThread_tss_get(&eval_frame_callback_key);
+  if (unlikely(result == NULL)) {
+    Py_RETURN_NONE;


This one returns a new reference while the one below returns a borrowed reference.
The callers, AFAICT expect a borrowed one.

albanD · 2022-10-10T21:27:55Z

torch/csrc/dynamo/eval_frame.c

+  //  - Python callable(): enables TorchDynamo
+  PyObject* old_callback = eval_frame_callback_get();
+
+  // owned by caller


This incref/decref is not necessary since we're strealing the reference here.
And the ref from the tss can directly be returned.

albanD · 2022-10-10T21:34:42Z

torch/csrc/dynamo/eval_frame.c

+    _methods};
+
+PyObject* torch_c_dynamo_eval_frame_init(void) {
+  extra_index = _PyEval_RequestCodeExtraIndex(ignored);


Also shouldn't we pass in free as the release functions to properly cleanup cache on code objects?
That won't work with the custom SKIP_CODE I guess, but a simple custom function should do it.

albanD · 2022-10-10T21:39:04Z

torch/csrc/dynamo/eval_frame.c

+#endif
+
+// Flag to just run a frame normally
+#define SKIP_CODE ((void*)0x1)


nit: could we use a long lived cache entry instead of a fake pointer?

albanD · 2022-10-10T21:45:12Z

torch/csrc/dynamo/eval_frame.c

+  Py_RETURN_NONE;
+}
+
+static PyMethodDef _methods[] = {


All of these should have their content wrapped into HANDLE_TH_ERRORS / END_HANDLE_TH_ERRORS{_RET}
to ensure proper propagation of error and warnings.

This only exists in C++

We could move these bindings out of the c file though?

albanD · 2022-10-10T23:39:35Z

torch/csrc/dynamo/guards.cpp

+
+namespace {
+
+struct LocalState {


Why not use at::ThreadLocalState ?

This isn't actually thread local, it is a snapshot of thread local data.

Which is exactly what this class is for!
It holds the thread local data so that we can move it aorund.

It just a cache since reading thread local data is slow.

Sure, that still sounds like the same class. You just have less fields here.
We can keep this for perf for sure.

albanD · 2022-10-10T23:42:11Z

torch/csrc/dynamo/guards.cpp

+      bool dynamic_shapes)
+      : pytype(pt),
+        dispatch_key_(state.apply(v.key_set()).raw_repr()),
+        dtype_(v.dtype().toScalarType()),


v.scalar_type()

in the check functions as well

albanD · 2022-10-10T23:45:41Z

torch/csrc/dynamo/guards.cpp

+  at::ScalarType dtype_;
+  bool requires_grad_;
+  bool dynamic_shapes_;
+  std::vector<int64_t> sizes_;


size and stride without storage_offset is usually wrong.
Is there a good reason not to have it here?

I'd be fine adding it. This check isn't for dynamo, it is for the backends.

With this, we assume backends produce offset independent code. This is true for inductor, but could be false for other backends.

Ok. A comment would be great.

albanD · 2022-10-10T23:47:01Z

torch/csrc/dynamo/guards.cpp

+
+typedef struct {
+  PyObject_HEAD;
+  ChecksList* checks;


unique_ptr ?
That would remove the need for a custom dealloc

albanD · 2022-10-10T23:48:18Z

torch/csrc/dynamo/guards.cpp

+} TensorGuards;
+
+static void TensorGuards_dealloc(TensorGuards* self) {
+  if (self->checks != NULL) {


torch/csrc/dynamo/guards.cpp

albanD · 2022-10-10T23:54:53Z

torch/csrc/dynamo/guards.cpp

+  for (auto i : c10::irange(len)) {
+    PyObject* item = PyTuple_GET_ITEM(args, i);
+    if (!THPVariable_CheckExact(item) && !THPVariable_Check(item)) {
+      PyErr_SetString(PyExc_TypeError, "expected Tensor()");


Should we cleanup the partially filed checks when an error happens?

albanD · 2022-10-11T00:04:13Z

torch/csrc/dynamo/guards.cpp

+  Py_RETURN_TRUE;
+}
+
+static PyMethodDef TensorGuards_methods[] = {


nit: these bindings would be a lot simpler via pybind.

pybind is crazy slow.

albanD · 2022-10-11T00:06:04Z

torch/csrc/dynamo/guards.cpp

+    PyObject* item = PyTuple_GET_ITEM(args, i);
+    if (Py_TYPE(item) != checks[i].pytype) {
+      std::stringstream fail_reason;
+      PyObject* type_str = PyObject_Str(PyObject_Type(item));


PyObject_Type returns a new reference. A decreft or wrap in a THPObjectPtr

albanD

Stamping, comments will be addressed in a follow up by @voznesenskym

jansel · 2022-10-11T16:34:39Z

@pytorchbot merge -l

pytorchmergebot · 2022-10-11T16:38:18Z

Merge started

The -l land checks flag is deprecated and no longer needed. Instead we now automatically add the ciflow\trunk label to your PR once it's approved

Your change will be merged once all checks on your PR pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

pytorchmergebot · 2022-10-11T17:03:54Z

Merge failed

Reason: 2 additional jobs have failed, first few of them are: trunk ,trunk / libtorch-linux-bionic-cuda11.6-py3.7-gcc7 / build

Details for Dev Infra team

Raised by workflow job

jansel · 2022-10-11T22:59:45Z

@pytorchbot merge

pytorchmergebot · 2022-10-11T23:01:13Z

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

@mruberry

commit f925b26 Author: Ivan Yashchuk <[email protected]> Date: Thu Oct 13 21:45:09 2022 +0300 Allow skipping view with skip_ops commit ddb769e Author: Ivan Yashchuk <[email protected]> Date: Thu Oct 13 21:38:04 2022 +0300 Add varargs support for view commit a9cdefa Author: Ivan Yashchuk <[email protected]> Date: Wed Oct 12 18:46:46 2022 +0300 Use ops.view name commit 986d76b Author: Ivan Yashchuk <[email protected]> Date: Wed Oct 12 18:27:37 2022 +0300 Fix duplicate commit 1c9c9c6 Author: Ivan Yashchuk <[email protected]> Date: Wed Oct 12 16:49:52 2022 +0300 Add print for ViewOpRecord commit a67e6c2 Merge: b07eeb0 2344135 Author: Ivan Yashchuk <[email protected]> Date: Wed Oct 12 16:43:53 2022 +0300 Merge remote-tracking branch 'upstream/viable/strict' into nvprims-view commit 2344135 Author: Khushi <[email protected]> Date: Wed Oct 12 07:00:40 2022 +0000 [primTorch] special: entr, expit (pytorch#86592) Add _refs for `entr` & `expit`. cc @mruberry @kshitij12345! Pull Request resolved: pytorch#86592 Approved by: https://github.com/mruberry commit a47f93b Author: Sherlock Huang <[email protected]> Date: Wed Oct 12 02:26:02 2022 +0000 Add type and shape annotation for gm.print_readable() (pytorch#86562) For ``` def f(a, b): dim0 = a.shape[0] + b.shape[0] dim1 = a.shape[1] + b.shape[1] d = a.new_empty(dim0, dim1) return d fx_g = make_fx(f, tracing_mode="symbolic")(torch.randn(5, 3), torch.randn(4, 3)) fx_g.print_readable() ``` Tracing with 'real' and 'fake' mode yields ``` class f(torch.nn.Module): def forward(self, a_1: Tensor<f32>[5, 3], b_1: Tensor<f32>[4, 3]): # No stacktrace found for following nodes new_empty: Tensor<f32>[9, 6] = torch.ops.aten.new_empty.default(a_1, [9, 6], dtype = torch.float32, layout = torch.strided, device = device(type='cpu'), pin_memory = False); a_1 = None return new_empty ``` Tracing with 'symbolic' mode yields ``` def forward(self, a_1: Tensor<f32>[t0.size(0), t0.size(1)], b_1: Tensor<f32>[t1.size(0), t0.size(1)]): # No stacktrace found for following nodes sym_size: Symint(t0.size(0)) = torch.ops.aten.sym_size(a_1, 0) sym_size_1: Symint(t1.size(0)) = torch.ops.aten.sym_size(b_1, 0) add: Symint(t0.size(0) + t1.size(0)) = sym_size + sym_size_1; sym_size = sym_size_1 = None sym_size_2: Symint(t0.size(1)) = torch.ops.aten.sym_size(a_1, 1) sym_size_3: Symint(t0.size(1)) = torch.ops.aten.sym_size(b_1, 1); b_1 = None add_1: Symint(2*t0.size(1)) = sym_size_2 + sym_size_3; sym_size_2 = sym_size_3 = None new_empty: Tensor<f32>[t0.size(0) + t1.size(0), 2*t0.size(1)] = torch.ops.aten.new_empty.default(a_1, [add, add_1], dtype = torch.float32, layout = torch.strided, device = device(type='cpu'), pin_memory = False); a_1 = add = add_1 = None return new_empty ``` Pull Request resolved: pytorch#86562 Approved by: https://github.com/Chillee commit e0d6898 Author: PyTorch MergeBot <[email protected]> Date: Wed Oct 12 04:12:43 2022 +0000 Revert "Backport currently dont work with some models if: (pytorch#86510)" This reverts commit 4bfb734. Reverted pytorch#86510 on behalf of https://github.com/facebook-github-bot due to Diff reverted internally commit 25725fd Author: Eddie Yan <[email protected]> Date: Wed Oct 12 03:44:21 2022 +0000 (Re-open) Adds cudaMallocAsync as an alternative backend for the CUDA allocator (pytorch#82682) Rebased version of @mcarilli 's cudaMallocAsync pytorch#65365 for continued testing Pull Request resolved: pytorch#82682 Approved by: https://github.com/ngimel commit a216f47 Author: Nikita Shulga <[email protected]> Date: Wed Oct 12 01:45:21 2022 +0000 Add testing on A10G GPU to periodic workflow (pytorch#85524) This enables testing on lots of modern CUDA features on sm_86 capable GPU While migrating to that platform, discovered that `functorch` tests for `nn.functional.conv.transpose3d` produce garbage on sm_80+ as well as 2 `nvfuser` tests unexpectedly pass and one unexpectedly fails. TODO: - Investigate unexpected success for `test_vmapvjp_linalg_householder_product_cuda_float32` and add `functorch` shard Pull Request resolved: pytorch#85524 Approved by: https://github.com/ngimel commit c4f0b93 Author: Elias Ellison <[email protected]> Date: Tue Oct 11 01:24:48 2022 +0000 Disable autocast in aot autograd (pytorch#86515) Fix for pytorch/torchdynamo#1368 From comment: > When we invoke a Composite Implicit autograd operator that has an autocast rule, such as Einsum, autocast is disabled during its invocation. When we trace out the operators in an implicit op, re-applying on autocast rules on those operators might yield divergence from what was executed at runtime. This pass checks for divergence. If divergence is found, we will disable autocast. We would like to avoid disabling autocast if possible because accessing TLS is slow. Concretely, the problem found was when invoked `sum` in `einsum`: As seen by the following divergence: ``` >>> with torch.cuda.amp.autocast(enabled=True): ... print(torch.ops.aten.sum.dim_IntList(torch.rand([2, 2, 2], device="cuda", dtype=torch.half), [1, 2]).dtype) ... torch.float32 >>> print(torch.ops.aten.sum.dim_IntList(torch.rand([2, 2, 2], device="cuda", dtype=torch.half), [1, 2]).dtype) torch.float16 ``` Edit: we've decided to accept the overhead of universally disabling autocast instead Pull Request resolved: pytorch#86515 Approved by: https://github.com/bdhirsh, https://github.com/Chillee commit d598290 Author: Christian Puhrsch <[email protected]> Date: Wed Oct 12 01:27:57 2022 +0000 Basic SDP benchmark harness (pytorch#86729) Basic benchmark for reference and discussion. Pull Request resolved: pytorch#86729 Approved by: https://github.com/drisspg commit 4bfb734 Author: Han Qi (qihqi) <[email protected]> Date: Wed Oct 12 00:39:25 2022 +0000 Backport currently dont work with some models if: (pytorch#86510) Backport currently dont work with some models if: * model is originally exported with interface call enabled (backport would disable it) * model is flatbuffer (flatbuffer support is soft enabled via link time registry), so we manually trigger it Fixes #ISSUE_NUMBER Pull Request resolved: pytorch#86510 Approved by: https://github.com/cccclai commit ce48df9 Author: Bin Bao <[email protected]> Date: Tue Oct 11 20:31:12 2022 +0000 Re-enable torchdynamo unit tests (pytorch#86658) Pull Request resolved: pytorch#86658 Approved by: https://github.com/jansel commit 692b525 Author: Nikita Shulga <[email protected]> Date: Wed Oct 12 00:32:53 2022 +0000 [MPS] Extend unary ops to int64 (pytorch#86615) Most of them are already supported for `int64` except for: - rounding operations (`floor`, `ceil` and `round`), which are no-ops for integral types anyway - sign operation, when it can be emulated by clamping it tensor to [-1, 1] range Test new types by test MPS Fixes pytorch#86319 Pull Request resolved: pytorch#86615 Approved by: https://github.com/DenisVieriu97, https://github.com/huydhn commit f912b58 Author: PyTorch MergeBot <[email protected]> Date: Tue Oct 11 23:53:12 2022 +0000 Revert "Enable max.unary_out (pytorch#85926)" This reverts commit 16a0fa1. Reverted pytorch#85926 on behalf of https://github.com/osalpekar due to The internal diff for this commit shows a number of pytorch quantization test failures. Here is a sample output: AssertionError: Tensor-likes are not close! Mismatched elements: 319 / 320 (99.7%). Greatest absolute difference: 0.056652069091796875 at index (0, 0, 4, 5) (up to 1e-05 allowed). Link to the diff: [D40232598](https://www.internalfb.com/diff/D40232598). Link to the Sandcastle job that is failing: https://www.internalfb.com/intern/sandcastle/job/18014399302908587/ commit 2aa981a Author: PyTorch MergeBot <[email protected]> Date: Tue Oct 11 23:39:50 2022 +0000 Revert "Reland 2 of Merge more symbolic meta kernels and symint changes from branch (pytorch#86334) (pytorch#86488)" This reverts commit 978b46d. Reverted pytorch#86488 on behalf of https://github.com/osalpekar due to Broke executorch builds internally with the following message: RuntimeError: Missing out variant for functional op: aten::split.Tensor(Tensor(a -> *) self, SymInt split_size, int dim=0) -> Tensor(a)[] . Make sure you have loaded your custom_ops_generated_lib commit 9eb4f9d Author: Nikita Shulga <[email protected]> Date: Tue Oct 11 19:49:23 2022 +0000 Tweak test tolerances to be compatible with A10G (pytorch#86538) Pull Request resolved: pytorch#86538 Approved by: https://github.com/ngimel commit 7fa601b Author: Nikita Shulga <[email protected]> Date: Tue Oct 11 23:27:30 2022 +0000 Skip chalf.mean in test_reductions_large_half_tensors (pytorch#86747) As `mean_reduce` is not implemented for complex half Fixes pytorch#86743 and unblock A10G testing Pull Request resolved: pytorch#86747 Approved by: https://github.com/ngimel commit 811b8e0 Author: PyTorch MergeBot <[email protected]> Date: Tue Oct 11 23:12:40 2022 +0000 Revert "min/max support for SymInt/Floats, finish as_strided/scatter/squeeze() backward symint support (pytorch#86643)" This reverts commit 86f914e. Reverted pytorch#86643 on behalf of https://github.com/osalpekar due to Need to revert this to cleanly revert pytorch#86488. This should be safe to re-land later commit f1fdb6e Author: Jason Ansel <[email protected]> Date: Tue Oct 11 23:01:21 2022 +0000 Manual changes for moving dynamo to core (pytorch#86621) This is the subset of the changes in pytorch#86461 not auto-generated by `copy_to_core.sh`. Pull Request resolved: pytorch#86621 Approved by: https://github.com/albanD commit 09364f4 Author: Nikita Shulga <[email protected]> Date: Tue Oct 11 22:39:58 2022 +0000 Compile C10 with `Wshadow` (pytorch#86666) This should prevent further regressions like pytorch#86646 Update `fmt` to `7.1.0` to fix variable shadowing in that library Pull Request resolved: pytorch#86666 Approved by: https://github.com/seemethere commit 0337f0a Author: Zain Rizvi <[email protected]> Date: Tue Oct 11 21:56:01 2022 +0000 Add error checking to flaky test bot platform parser (pytorch#86632) If an invalid platform is specified when disabling a test with flaky test bot, the CI crashes, skipping all tests that come after it. This turns it into a console message instead. Not erroring out here since it'll affect random PRs. Actual error message should go into the bot that parses the original issue so that it can respond on that issue directly Pull Request resolved: pytorch#86632 Approved by: https://github.com/huydhn commit 42bd275 Author: Partho <[email protected]> Date: Tue Oct 11 21:41:48 2022 +0000 [doc] LR scheduler example fix (pytorch#86629) Fixes issue pytorch#86208 As suggested in the issue, updated the LR scheduler example to use a regular nn.Module like the other examples on the same page. Pull Request resolved: pytorch#86629 Approved by: https://github.com/soulitzer commit 32152ce Author: jimku9 <[email protected]> Date: Tue Oct 11 21:21:53 2022 +0000 Add original sources/references to Wishart.py in distributions (pytorch#86543) @fritzo As discussed, add original sources/references to Wishart.py in distributions and corrected typos in the error messages. Pull Request resolved: pytorch#86543 Approved by: https://github.com/fritzo commit 50af1ac Author: Sherlock Huang <[email protected]> Date: Tue Oct 11 17:56:59 2022 +0000 Mark aten ops as canonical (pytorch#86215) This is the first batch of canonical aten ops. 87 in total. More to come in the future PRs. native_dropout abs add.Tensor add.Scalar arange.start_step bitwise_not bmm cat clamp constant_pad_nd convolution convolution_backward div.Tensor div.Scalar embedding_dense_backward erf exp expand fill.Scalar grid_sampler_2d native_group_norm native_group_norm_backward native_layer_norm native_layer_norm_backward log _log_softmax max.dim amax mean.dim min.dim amin mm mul.Tensor mul.Scalar native_batch_norm permute scalar_tensor reciprocal neg repeat relu gelu rsqrt sigmoid slice.Tensor slice_scatter _softmax squeeze.dim sum.dim_IntList sqrt tanh unsqueeze var.dim where.self clone sub.Tensor sub.Scalar addmm _to_copy view scatter_add bitwise_and.Tensor bitwise_or.Tensor eq.Scalar ge.Scalar le.Scalar gt.Scalar lt.Scalar index_select nonzero gather maximum minimum pow.Tensor_Scalar hardtanh leaky_relu _adaptive_avg_pool2d _adaptive_avg_pool2d_backward avg_pool2d avg_pool2d_backward max_pool2d_with_indices max_pool2d_with_indices_backward upsample_bilinear2d.vec upsample_bilinear2d_backward.vec upsample_nearest2d.vec upsample_nearest2d_backward.vec col2im Pull Request resolved: pytorch#86215 Approved by: https://github.com/suo, https://github.com/anjali411 commit 8db3025 Author: Jeff Daily <[email protected]> Date: Tue Oct 11 20:55:58 2022 +0000 [ROCm] set nvfuser default to disabled, keep CI (pytorch#86369) Bug fix. nvfuser is functional for ROCm on gfx906, but some tests are failing for other gfx targets. Disable nvfuser until all features are verified. Users may still opt-in by setting the known env var PYTORCH_JIT_ENABLE_NVFUSER=1. This PR sets this env var for the github actions workflow for ROCm since all current CI hosts are gfx906. Pull Request resolved: pytorch#86369 Approved by: https://github.com/huydhn commit 5ffe24f Author: Stephen Jia <[email protected]> Date: Tue Oct 11 20:16:56 2022 +0000 [vulkan][ez] fix always printing out a warning when retrieving the global context (pytorch#86697) Summary: D40151818 (pytorch@82ed5ca) replaces the `TORCH_CHECK` with a `TORCH_WARN` but since it does not check if the context is valid the message gets printed every time. This diff fixes that. Test Plan: Referring to [Pytorch Vulkan Testing Procedures](https://fb.quip.com/fZALAc9zhlcU) On Mac: 1. `vulkan_api_test` on Mac 2. model comparison binary on Mac On Android: 1. `vulkan_api_test` on Android 2. benchmark binary on Android Reviewed By: salilsdesai Differential Revision: D40266820 Pull Request resolved: pytorch#86697 Approved by: https://github.com/kirklandsign commit f32aeea Author: Han Qi (qihqi) <[email protected]> Date: Tue Oct 11 20:07:58 2022 +0000 Set interface_call to true be default (pytorch#86668) Summary: ASR models need it Test Plan: existing unit tests Reviewed By: cccclai Differential Revision: D40251788 Pull Request resolved: pytorch#86668 Approved by: https://github.com/cccclai commit 7f02f2a Author: Huy Do <[email protected]> Date: Tue Oct 11 19:34:44 2022 +0000 [Experimentation] Add TSAN build and test (pytorch#85313) Some parts of the PR are adopted from the previously abandoned pytorch#36694. This PR is the first part to setup TSAN jobs in the CI. The data race warnings from TSAN will need to be reviewed later in a separate PR. Pull Request resolved: pytorch#85313 Approved by: https://github.com/osalpekar commit 9256204 Author: 胡玮文 <[email protected]> Date: Tue Oct 11 19:03:43 2022 +0000 Optimize __dlpack_device__ performance (pytorch#86665) This can be critical when processing a large number of tensors ```bash python -m timeit --setup 'import torch; t = torch.empty(1000, device="cuda")' 't.__dlpack_device__()' ``` based on 1.12.1: before: 100000 loops, best of 5: 2.32 usec per loop after: 500000 loops, best of 5: 844 nsec per loop Pull Request resolved: pytorch#86665 Approved by: https://github.com/SunDoge, https://github.com/soulitzer commit c12f829 Author: Jerry Zhang <[email protected]> Date: Tue Oct 11 18:49:09 2022 +0000 [nn] Add remove_duplicate flag to named_buffers (#674) (pytorch#85903) Summary: X-link: meta-pytorch/torchrec#674 Pull Request resolved: pytorch#84984 this is to allow named_buffers to return the same buffer objects with different names multiple times, needed by internal use cases ghstack-source-id: 168589597 Test Plan: python test/test_nn.py -k test_buffers_and_named_buffers Imported from OSS Reviewed By: albanD Differential Revision: D39493161 Pull Request resolved: pytorch#85903 Approved by: https://github.com/albanD commit 693250a Author: David <[email protected]> Date: Tue Oct 11 18:05:53 2022 +0000 Docs: fx.Node docs incorrectly state that the self argument is included in args for module calls (pytorch#86685) It seems like the [torch.fx.Node docs](https://pytorch.org/docs/stable/fx.html#torch.fx.Node) are incorrect regarding the inclusion of the self argument for module call nodes. While the docs state that self (the module) is included in `args`, it is in fact not, as demonstrated by this code: ```python import torch from torch import fx, nn class Net(nn.Module): def __init__(self): super().__init__() self.submod = nn.Linear(10, 10) def forward(self, x): x = x.flatten() return self.submod(x) graph_module = fx.symbolic_trace(Net()) print(graph_module.graph) # doesn't show self for the submodule call submod_node = list(graph_module.graph.nodes)[2] print(submod_node.op) # call_module print(submod_node.args) # (flatten,) => would need to have len 2 if self was included flatten_node = list(graph_module.graph.nodes)[1] print(flatten_node.op) # call_method print(flatten_node.args) # (x,) => here self is included (and docs are correct) ``` Since [torch.fx.Interpreter also uses `args` as if self was is not included](https://github.com/pytorch/pytorch/blob/2fe580859012d2d24a54e452195ccbc7f3191036/torch/fx/interpreter.py#L288), I assume the docs are incorrect. Pull Request resolved: pytorch#86685 Approved by: https://github.com/soulitzer commit 160118d Author: Fang Wang <[email protected]> Date: Tue Oct 11 17:52:18 2022 +0000 Add test case for matrix multiply-add with large inputs (pytorch#85550) Summary: - Added test case for addmm, baddbmm and linear with large inputs - Testing with torch types: float32, float16, bfloat16 Test Plan: Run unit tests with: `buck2 run mode/opt //caffe2/test:linalg_re_cuda` ``` ... test_addmm_baddbmm_large_input_1_10000_10000_10000_cpu_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_1_10000_10000_10000_cpu_float16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_1_10000_10000_10000_cpu_float32 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_1_10000_1000_10000_cpu_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_1_10000_1000_10000_cpu_float16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_1_10000_1000_10000_cpu_float32 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_2_1000_1000_1000_cpu_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_2_1000_1000_1000_cpu_float16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_2_1000_1000_1000_cpu_float32 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_2_100_100_100_cpu_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_2_100_100_100_cpu_float16 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_2_100_100_100_cpu_float32 (test_linalg_re_cuda.TestLinalgReCudaCPU) ... skipped 'Only runs on cuda' test_addmm_baddbmm_large_input_1_10000_10000_10000_cuda_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_1_10000_10000_10000_cuda_float16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_1_10000_10000_10000_cuda_float32 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_1_10000_1000_10000_cuda_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_1_10000_1000_10000_cuda_float16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_1_10000_1000_10000_cuda_float32 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_2_1000_1000_1000_cuda_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_2_1000_1000_1000_cuda_float16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_2_1000_1000_1000_cuda_float32 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_2_100_100_100_cuda_bfloat16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_2_100_100_100_cuda_float16 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok test_addmm_baddbmm_large_input_2_100_100_100_cuda_float32 (test_linalg_re_cuda.TestLinalgReCudaCUDA) ... ok ---------------------------------------------------------------------- Ran 24 tests in 63.224s OK (skipped=12) ``` Differential Revision: D39718256 Pull Request resolved: pytorch#85550 Approved by: https://github.com/IvanYashchuk, https://github.com/malfet commit 212fa87 Author: vfdev <[email protected]> Date: Tue Oct 11 17:52:16 2022 +0000 Fix torch histogramdd docstring (pytorch#86593) Fixed torch histogramdd docsting with missing common_args Pull Request resolved: pytorch#86593 Approved by: https://github.com/soulitzer commit f26292d Author: Jane Xu <[email protected]> Date: Tue Oct 11 17:42:51 2022 +0000 [BE] Fix python docs typos up till torch.chunk (pytorch#86642) Was doing the Views lab linked https://github.com/pytorch/pytorch/wiki/Tensor-and-Operator-Basics and noticed a few typos, which led to this PR. Test plan: verified in preview Pull Request resolved: pytorch#86642 Approved by: https://github.com/soulitzer commit 86f914e Author: albanD <[email protected]> Date: Tue Oct 11 10:35:18 2022 -0400 min/max support for SymInt/Floats, finish as_strided/scatter/squeeze() backward symint support (pytorch#86643) Pull Request resolved: pytorch#86643 Approved by: https://github.com/anjali411 commit b07eeb0 Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 29 17:01:50 2022 +0300 Use string names for matching view-like functions commit d8c005a Merge: 59cb4be ad87365 Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 29 17:01:03 2022 +0300 Merge remote-tracking branch 'upstream/viable/strict' into nvprims-view commit 59cb4be Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 18:37:59 2022 +0300 lint commit 92edd1a Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 18:15:35 2022 +0300 Add view_copy commit 79c18da Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 18:08:25 2022 +0300 Add _unsafe_view to list commit 254161d Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 18:07:51 2022 +0300 Add _unsafe_view to tests commit 487a7a8 Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 18:00:30 2022 +0300 Use func == torch.ops.aten.view.default commit 24e61bf Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 17:57:48 2022 +0300 Update torch/_prims/nvfuser_prims.py commit abad276 Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 17:53:42 2022 +0300 Modify python frontend according latest changes commit 712447f Merge: a135db1 0c46e3e Author: Ivan Yashchuk <[email protected]> Date: Thu Sep 22 17:22:44 2022 +0300 Merge remote-tracking branch 'upstream/viable/strict' into nvprims-view commit a135db1 Author: Ivan Yashchuk <[email protected]> Date: Wed Sep 7 17:06:30 2022 +0300 Add interception of view for TorchRefsNvfuserCapabilityMode commit f0c039e Author: Ivan Yashchuk <[email protected]> Date: Wed Sep 7 17:06:07 2022 +0300 Add test for view -> nvprims.view lowering commit 246c999 Author: Ivan Yashchuk <[email protected]> Date: Wed Sep 7 16:40:13 2022 +0300 Add tests commit c48ba8e Author: Ivan Yashchuk <[email protected]> Date: Wed Sep 7 16:39:59 2022 +0300 Add nvprims.view commit 3980f32 Author: Ivan Yashchuk <[email protected]> Date: Wed Sep 7 16:39:38 2022 +0300 Add fd.ops.view

Manual for moving dynamo to core

c68f4dd

jansel requested a review from albanD as a code owner October 10, 2022 18:33

pytorch-bot bot added the release notes: fx release notes category label Oct 10, 2022

jansel commented Oct 10, 2022

View reviewed changes

jansel changed the title ~~Manual for moving dynamo to core~~ Manual changes for moving dynamo to core Oct 10, 2022

jansel added the module: dynamo label Oct 10, 2022

malfet reviewed Oct 10, 2022

View reviewed changes

albanD reviewed Oct 10, 2022

View reviewed changes

albanD added the ciflow/binaries Trigger all binary build and upload jobs on the PR label Oct 10, 2022

jansel and others added 2 commits October 10, 2022 16:47

Skip python 3.11

3244623

Attempt to fix c2039

53180ea

albanD reviewed Oct 11, 2022

View reviewed changes

lints

1bf2406

jansel mentioned this pull request Oct 11, 2022

Move TorchDynamo into PyTorch core #86461

Closed

jansel added the topic: not user facing topic category label Oct 11, 2022

soumith mentioned this pull request Oct 11, 2022

Move functorch tests from functorch/test/* to test/*; delete functorch CI configs #86618

Closed

jansel mentioned this pull request Oct 11, 2022

Migration to PyTorch core pytorch/torchdynamo#1588

Closed

13 tasks

albanD reviewed Oct 11, 2022

View reviewed changes

albanD approved these changes Oct 11, 2022

View reviewed changes

pytorch-bot bot added the ciflow/trunk Trigger trunk jobs on your pull request label Oct 11, 2022

pytorchmergebot added the Merged label Oct 11, 2022

pytorchmergebot closed this in f1fdb6e Oct 11, 2022

jansel deleted the dynamo_changes branch October 11, 2022 23:06

Manual changes for moving dynamo to core #86621

Manual changes for moving dynamo to core #86621

Uh oh!

Conversation

jansel commented Oct 10, 2022

Uh oh!

pytorch-bot bot commented Oct 10, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/86621

✅ No Failures

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

malfet left a comment

Choose a reason for hiding this comment

Uh oh!

jansel commented Oct 10, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

albanD left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jansel Oct 10, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pytorch-bot bot commented Oct 10, 2022 •

edited

Loading

jansel commented Oct 10, 2022 •

edited

Loading

jansel Oct 10, 2022 •

edited

Loading

albanD left a comment •

edited

Loading