When caching is enabled, an internal model fails with
```
assert_size_stride(bmm_9, (17, s0, 512), (54784, 512, 1))
AssertionError: expected size 17==17, stride 57344==54784 at dim=0
```
looking at this model, the exact problem is when the cache is hit on the forward graph, the generated code for backward fails since the strides of the outputs of forward, passed to backward as inputs, are not what we expected.
This PR changes the evaluation logic so that we defer evaluation of output stride exprs to load path as opposed to eagerly doing it on save path.
I have not been able to come up with a unit test repro for this problem.
Differential Revision: [D58796503](https://our.internmc.facebook.com/intern/diff/D58796503)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128997
Approved by: https://github.com/ezyang
I found it helpful to be able to see, given some inductor output code, which AOT graph it came from. When you have large models with multiple graphs floating around this can be difficult, so I added the aot_config.aot_id to the printed inductor output.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118647
Approved by: https://github.com/ezyang
`python benchmarks/dynamo/microbenchmarks/dynamo_microbenchmarks.py`
- Before: `symbolic_convert_overhead_stress_test: 10.7s`
- After: `symbolic_convert_overhead_stress_test: 8.6s`
`tx.step()` is a small part of that benchmark, so likely the speedup in that isolated function is larger than the top line.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121790
Approved by: https://github.com/oulgen
Context: view fake-ification should handle closed-over state in ViewFuncs for use in view replay by:
* fake-ifying tensors
* symbolicizing SymInts
This avoids invalid specialization during view replay. However, the symbols / tensors created as intermediates in the view chain should not stick around or be guarded on. This PR introduces an `EphemeralSource` intended to be used as a source for this purpose. It has the following properties:
* Considered first to be simplified out in symbol simplification logic
* Errors if guarded on
Differential Revision: [D54561597](https://our.internmc.facebook.com/intern/diff/D54561597)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120948
Approved by: https://github.com/ezyang
# Note: Returning Fake Tensors on First AOT Autograd Call
#
# Inductor will optimize strides of outputs when it deems it profitable.
# For instance, converting to channels last. When we split the graph here
# into multiple inductor compilations, we need to make sure that the
# output strides of one compilation is appropriately passed to the subsequent
# compilations. However, the mapping from inductor output to dynamo output
# is non-trivial due to aot_autograd's deduping, de-aliasing, mutation, re-writing,
# subclass handling, etc. In order to replay all this logic we set a flag such that
# the first invocation of inductor in aot_autograd will return Fake Tensors with
# appropriate strides. Then, all of aot autograd's runtime logic is replayed.
# This gives us the appropriately strided outputs here which will reflect runtime strides.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120523
Approved by: https://github.com/yf225, https://github.com/bdhirsh
This adds support for backwards hooks that are *both*:
1) Interior to the graph; and
2) Dynamically generated (e.g. lambdas)
We do this by creating a BackwardState object that is used to register the hooks in the forward, then populated by dynamo *after* the forwards runs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120382
Approved by: https://github.com/xmfan
Reduces backend=eager compile time from 33 to 19 seconds for `MobileBertForQuestionAnswering`. This also helps an internal model where guards.add function is taking 124 seconds.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120520
Approved by: https://github.com/mlazos
There is no need to make a `frame_summary_stack` copy in case it's not modified. Proposed change uses copy-on-write functional approach that is easy to understand and is more efficient in case `self.loc_in_frame` is `None`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119115
Approved by: https://github.com/Skylion007
UPDATE - I changed the PR because from discussion with @jansel it was clear that someone else was holding on to a reference to f_locals. This PR now solves that problem first. I removed the eval_frame.c part because it was failing tests that use `exec` or `eval` with weird error like `no no locals found when storing 'math'`. I would debug that in a separate PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118447
Approved by: https://github.com/Skylion007, https://github.com/jansel
ghstack dependencies: #118975, #118420
Make variables in dict lazy and remove DICT_KEYS guard.
We build the keys of a dict depth-first and we rely on the guards of
each element in the dict to create the correct guards. This allows us to
remove the rather buggy DICT_KEYS guard and make the guard lazy.
The guards are not completely lazy yet, as we instantiate them in
`_HashableTracker._eq_impl` but it should be possible to make them
truly lazy.
Also, adding new types to the supported types within keys should be less
error prone.
This is marginally less efficient when we graph break, but in turn we
should graph break much less. It also makes the dicts code easier to maintain
(removes `is_hashable_python_var`).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117625
Approved by: https://github.com/jansel, https://github.com/peterbell10, https://github.com/anijain2305
ghstack dependencies: #117982, #118098, #117983
Summary:
The primary problem we are setting out to solve here is fake tensor freshness. Before this PR, fake tensors after dynamo represented fake tensors *at the end* of trace, so subsequent retraces like aot_autograd would start off with fake tensors in the wrong (end result) state, rather than their expected fresh state. The solution here is to start a fresh fake mode, and re-fakify the tensors. The nuance comes from ensuring that symbols are uniformly created for the symbolic sizes and strides of the tensor.
This PR is the result of *a lot* of back and forth with ezyang and eellison. Initially, the first pass at this was not super different from what we have in the PR - the broad strokes were the same:
1) We cache source->symbol in shape_env
2) We pass policy objects around, stored at dynamo fakificaiton time, and reused for later fakification
3) We create a new fake mode for backends
(from https://github.com/pytorch/pytorch/pull/113605/files)
This is ugly, and has some layering violations. We detoured our decision making through a few other alternatives. Immutable/mutable fake tensor mode was the most interesting alternative, https://github.com/pytorch/pytorch/pull/113653, and was struck down on concerns of complexity in fake mode combined with it not covering all edge cases. We also detoured on what to do about tensor memoization returning back potentially different tensors than requested, and if that was an anti pattern (it is) we want to hack in with the symbol cache (we don't).
We went back to the drawing board here, but with a few concessions:
1) the cache for source->symbol must live outside of shape_env, for both lifecycle, and layering reasons
2) A good amount of work needs to be done to pipe policy around fake_mode and meta_utils correctly, to cover all the cases (ezyang did this)
cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng
imported-using-ghimport
Test Plan: Imported from OSS
Reviewed By: huydhn, Chillee
Differential Revision: D51566250
Pulled By: voznesenskym
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114526
Approved by: https://github.com/Chillee, https://github.com/huydhn
The primary problem we are setting out to solve here is fake tensor freshness. Before this PR, fake tensors after dynamo represented fake tensors *at the end* of trace, so subsequent retraces like aot_autograd would start off with fake tensors in the wrong (end result) state, rather than their expected fresh state. The solution here is to start a fresh fake mode, and re-fakify the tensors. The nuance comes from ensuring that symbols are uniformly created for the symbolic sizes and strides of the tensor.
This PR is the result of *a lot* of back and forth with @ezyang and @eellison. Initially, the first pass at this was not super different from what we have in the PR - the broad strokes were the same:
1) We cache source->symbol in shape_env
2) We pass policy objects around, stored at dynamo fakificaiton time, and reused for later fakification
3) We create a new fake mode for backends
(from https://github.com/pytorch/pytorch/pull/113605/files)
This is ugly, and has some layering violations. We detoured our decision making through a few other alternatives. Immutable/mutable fake tensor mode was the most interesting alternative, https://github.com/pytorch/pytorch/pull/113653, and was struck down on concerns of complexity in fake mode combined with it not covering all edge cases. We also detoured on what to do about tensor memoization returning back potentially different tensors than requested, and if that was an anti pattern (it is) we want to hack in with the symbol cache (we don't).
We went back to the drawing board here, but with a few concessions:
1) the cache for source->symbol must live outside of shape_env, for both lifecycle, and layering reasons
2) A good amount of work needs to be done to pipe policy around fake_mode and meta_utils correctly, to cover all the cases (@ezyang did this)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113926
Approved by: https://github.com/ezyang, https://github.com/eellison
They are used in many contexts that don't actually check if the returned
type is `None`. I have also created `try_get()` for the cases where we
do actually want an Optional type returned.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113535
Approved by: https://github.com/ezyang
ghstack dependencies: #113412
The motivation for removing this is already present in the pre-PR comments. Copying it
~~~
# NB - SuperSource is a weird one.
# it is our only source with 2 bases, so we use the objec
# as the base, rather than the type, since an invocation
# like super(Foo, foo) is represented here, the source object base is more spiritually
# aligned with the instance, rather than the type.
# This whole construction is questionable tho, and we should probably find a way to
# avoid this exception to our otherwise nice source parentage invariant.
~~~
Instead of using super(a, b), we can use `type(b).__mro__[index]`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110475
Approved by: https://github.com/jansel
Summary:
The basic concept behind this diff is to modify Dynamo's tracing behavior when it encounters a KeyedJaggedTensor that is synced (aka has `_length_per_key` and `_offset_per_key` populated). These fields are lists of integers; ordinarily, Dynamo will optimistically try to specialize on integers, however, for KJTs, we know that these integers will definitely vary from run-to-run. Furthermore, ordinarily, we would also specialize these integers if they are 0/1, but we will frequently expect features in KJTs to be 0/1.
The fix is to detect KJTs and treat these integers as *unbacked integers*. This is NOT a universally sound optimization: when treating these integers as unbacked, we never report them as equal to zero or one. In return, we always generate graphs that generalize no matter the length of values on features. This is enough to trace through APS sparse arch, torchrec_dlrm and some small split-cat examples.
The special integer behavior is triggered by a dynamically scoped `force_unspec_int_unbacked_size_like` variable on TracingContext, which we trigger when we wrap a KJT. There probably are other ways to do this, but this was simple and worked.
Test Plan:
```
buck2 test mode/dev-nosan //pytorch/benchmark/fb/test_gpu:run_test_gpu
```
from aakhundov
1. first build feed_lower_benchmark:
```
buck2 build --show-output mode/opt -c python.package_style=inplace -c fbcode.enable_gpu_sections=true -c fbcode.platform=platform010 -c fbcode.split-dwarf=true hpc/new/models/feed/benchmark:feed_lower_benchmark
```
2. then run the lowering of the model with it:
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCHINDUCTOR_UNIQUE_KERNEL_NAMES=1 TORCH_LOGS="output_code,graph_code" TORCH_COMPILE_DEBUG=1 ../buck-out/v2/gen/fbcode/79c6b019ee0f9469/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --load=manifold://ig_inference_model/tree/user/facebook/fblearner/predictor/960999465/60/gpu_lowering/input.predictor --skip-trt --skip-ait --sync-mode=0 --enable-aot-inductor --lower-presets="ig_stories" --gpu-trace
```
cf https://docs.google.com/document/d/1yD30xYrdmM8r2HTdmXnZTg0-MHVexfVrAa0294m1AUE/edit?pli=1#heading=h.qiv3fp7e6zg0
From torchrec: https://www.internalfb.com/intern/wiki/Torchrec/Development/Testing_production_models/
From ge0405
baseline (without your diff): f477293168
your diff: f477292363
```
buck2 test //caffe2/test/dynamo:test_dynamo_torchrec
buck2 run 'fbcode//mode/opt' fbcode//pytorch/benchmark/fb/test_gpu:run_test_gpu -- 'pytorch.benchmark.fb.test_gpu.test_gpu.TestBenchmarkFbGpu.test_train_blue_reels_vdd_v3_inductor_speedup'
```
Differential Revision: D49236757
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109216
Approved by: https://github.com/voznesenskym
Summary:
The basic concept behind this diff is to modify Dynamo's tracing behavior when it encounters a KeyedJaggedTensor that is synced (aka has `_length_per_key` and `_offset_per_key` populated). These fields are lists of integers; ordinarily, Dynamo will optimistically try to specialize on integers, however, for KJTs, we know that these integers will definitely vary from run-to-run. Furthermore, ordinarily, we would also specialize these integers if they are 0/1, but we will frequently expect features in KJTs to be 0/1.
The fix is to detect KJTs and treat these integers as *unbacked integers*. This is NOT a universally sound optimization: when treating these integers as unbacked, we never report them as equal to zero or one. In return, we always generate graphs that generalize no matter the length of values on features. This is enough to trace through APS sparse arch, torchrec_dlrm and some small split-cat examples.
The special integer behavior is triggered by a dynamically scoped `force_unspec_int_unbacked_size_like` variable on TracingContext, which we trigger when we wrap a KJT. There probably are other ways to do this, but this was simple and worked.
Test Plan:
```
buck2 test mode/dev-nosan //pytorch/benchmark/fb/test_gpu:run_test_gpu
```
from aakhundov
1. first build feed_lower_benchmark:
```
buck2 build --show-output mode/opt -c python.package_style=inplace -c fbcode.enable_gpu_sections=true -c fbcode.platform=platform010 -c fbcode.split-dwarf=true hpc/new/models/feed/benchmark:feed_lower_benchmark
```
2. then run the lowering of the model with it:
```
TORCHINDUCTOR_MAX_AUTOTUNE=1 TORCHINDUCTOR_UNIQUE_KERNEL_NAMES=1 TORCH_LOGS="output_code,graph_code" TORCH_COMPILE_DEBUG=1 ../buck-out/v2/gen/fbcode/79c6b019ee0f9469/hpc/new/models/feed/benchmark/__feed_lower_benchmark__/feed_lower_benchmark.par --load=manifold://ig_inference_model/tree/user/facebook/fblearner/predictor/960999465/60/gpu_lowering/input.predictor --skip-trt --skip-ait --sync-mode=0 --enable-aot-inductor --lower-presets="ig_stories" --gpu-trace
```
cf https://docs.google.com/document/d/1yD30xYrdmM8r2HTdmXnZTg0-MHVexfVrAa0294m1AUE/edit?pli=1#heading=h.qiv3fp7e6zg0
From torchrec: https://www.internalfb.com/intern/wiki/Torchrec/Development/Testing_production_models/
From ge0405
baseline (without your diff): f477293168
your diff: f477292363
Differential Revision: D49019987
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108960
Approved by: https://github.com/voznesenskym
Now looks like:
```
[2023-09-08 06:04:48,532] [0/0] torch._dynamo.symbolic_convert: [DEBUG] TRACE STORE_ATTR foo [ConstantVariable(int), NNModule
Variable()]
[2023-09-08 06:04:48,532] [0/0] torch._dynamo.convert_frame: [INFO]
Restarting analysis due to _dynamo/variables/nn_module.py
:138 in convert_to_unspecialized
[2023-09-08 06:04:48,533] [0/0_1] torch._dynamo.symbolic_convert: [INFO] Step 1: torchdynamo start tracing f /data/users/ezyang/c/pytorch/a.py:6
[2023-09-08 06:04:48,533] [0/0_1] torch._dynamo.symbolic_convert.__trace_source: [DEBUG] TRACE starts_line f /data/users/ezyang/c/pytorch/a.py:6
```
I'm happy to bikeshed the exact formatting of the attempt number if you
want.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108864
Approved by: https://github.com/mlazos, https://github.com/voznesenskym
All log messages that occur while running Dynamo compilation now have `[X/Y]` added to the beginning of their message. X represents the frame being compiled, while Y says which compilation of the frame. For example, if you are debugging a frame that is repeatedly recompiling, you can look for N/0, N/1, N/2, etc. for the same N. Here is what the logs look like as you transition from one frame to another:
<img width="1372" alt="image" src="https://github.com/pytorch/pytorch/assets/13564/4897e368-1e50-4807-b342-54e911bcf087">
To accurately get this prefix added to all messages, I had to expand the scope of the `tracing` context manager. Its scope now coincides with `log_compilation_event`. To do this, I had to populate fake mode lazily in the TracingContext, since it isn't created until later, inside the OutputGraph.
This subsumes the previous X.Y logging that was solely for dynamic shapes.
Unfortunately I had to reindent some stuff. Review the diff with whitespace off.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107530
Approved by: https://github.com/anijain2305
ghstack dependencies: #107505, #107516
The new guard printout looks like this:
```
[DEBUG] GUARDS:
[DEBUG] ___check_type_id(L['name'], 7605632) # if name == "special_attr": # test/dynamo/test_misc.py:1155 in __getattribute__
[DEBUG] L['name'] == '_backward_pre_hooks' # if name == "special_attr": # test/dynamo/test_misc.py:1155 in __getattribute__
[DEBUG] ___check_obj_id(L['self'], 139746432564960) # return super().__getattribute__(name) # test/dynamo/test_misc.py:1157 in __getattribute__
[DEBUG] ___check_obj_id(L['__class__'], 1451499216) # return super().__getattribute__(name) # test/dynamo/test_misc.py:1157 in __getattribute__
[DEBUG] ___is_grad_enabled() # _dynamo/output_graph.py:346 in init_ambient_guards
[DEBUG] not ___are_deterministic_algorithms_enabled() # _dynamo/output_graph.py:342 in init_ambient_guards
[DEBUG] ___is_torch_function_enabled() # _dynamo/output_graph.py:350 in init_ambient_guards
[DEBUG] utils_device.CURRENT_DEVICE == None # _dynamo/output_graph.py:348 in init_ambient_guards
```
Along with the guards, we also print what line of user code caused the guard to be added, or what line of Dynamo internal code added the guard (if there is no user stack trace, which is typically the case for ambient guards.)
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107505
Approved by: https://github.com/mlazos, https://github.com/voznesenskym, https://github.com/anijain2305
It looks like this:
```
[DEBUG] GUARD: ___check_type_id(L['z'][L["MyEnum"].BAR], 7640416) and L['z'][L["MyEnum"].BAR] == 10
[DEBUG] Stack:
[DEBUG] File "/data/users/ezyang/b/pytorch/test/dynamo/test_misc.py", line 6657, in <module>
[DEBUG] run_tests()
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/test_case.py", line 38, in run_tests
[DEBUG] run_tests()
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/testing/_internal/common_utils.py", line 985, in run_tests
[DEBUG] unittest.main(argv=argv)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/main.py", line 101, in __init__
[DEBUG] self.runTests()
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/main.py", line 271, in runTests
[DEBUG] self.result = testRunner.run(self.test)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/runner.py", line 184, in run
[DEBUG] test(result)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/suite.py", line 84, in __call__
[DEBUG] return self.run(*args, **kwds)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/suite.py", line 122, in run
[DEBUG] test(result)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/suite.py", line 84, in __call__
[DEBUG] return self.run(*args, **kwds)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/suite.py", line 122, in run
[DEBUG] test(result)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/case.py", line 650, in __call__
[DEBUG] return self.run(*args, **kwds)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/testing/_internal/common_utils.py", line 2521, in run
[DEBUG] self._run_with_retry(
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/testing/_internal/common_utils.py", line 2450, in _run_with_retry
[DEBUG] super_run(result=result)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/case.py", line 591, in run
[DEBUG] self._callTestMethod(testMethod)
[DEBUG] File "/home/ezyang/local/b/pytorch-env/lib/python3.10/unittest/case.py", line 549, in _callTestMethod
[DEBUG] method()
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/testing/_internal/common_utils.py", line 2377, in wrapper
[DEBUG] method(*args, **kwargs)
[DEBUG] File "/data/users/ezyang/b/pytorch/test/dynamo/test_misc.py", line 2529, in test_enum_as_dict_key_with_overloaded_str
[DEBUG] res = opt_fn(x)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/eval_frame.py", line 333, in _fn
[DEBUG] return fn(*args, **kwargs)
[DEBUG] File "/data/users/ezyang/b/pytorch/test/dynamo/test_misc.py", line 2519, in fn
[DEBUG] torch._dynamo.graph_break()
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/eval_frame.py", line 493, in catch_errors
[DEBUG] return callback(frame, cache_size, hooks, frame_state)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/convert_frame.py", line 637, in _convert_frame
[DEBUG] result = inner_convert(frame, cache_size, hooks, frame_state)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/convert_frame.py", line 133, in _fn
[DEBUG] return fn(*args, **kwargs)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/convert_frame.py", line 371, in _convert_frame_assert
[DEBUG] return _compile(
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/convert_frame.py", line 567, in _compile
[DEBUG] guarded_code = compile_inner(code, one_graph, hooks, transform)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/utils.py", line 181, in time_wrapper
[DEBUG] r = func(*args, **kwargs)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/convert_frame.py", line 466, in compile_inner
[DEBUG] out_code = transform_code_object(code, transform)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/bytecode_transformation.py", line 1028, in transform_code_object
[DEBUG] transformations(instructions, code_options)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/convert_frame.py", line 416, in transform
[DEBUG] tracer = InstructionTranslator(
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/symbolic_convert.py", line 2018, in __init__
[DEBUG] self.symbolic_locals = collections.OrderedDict(
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/symbolic_convert.py", line 2021, in <genexpr>
[DEBUG] VariableBuilder(
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 211, in __call__
[DEBUG] vt = self._wrap(value).clone(**self.options())
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 404, in _wrap
[DEBUG] result = {
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 405, in <dictcomp>
[DEBUG] k: VariableBuilder(
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 211, in __call__
[DEBUG] vt = self._wrap(value).clone(**self.options())
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 354, in _wrap
[DEBUG] return type_dispatch(self, value)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 837, in wrap_literal
[DEBUG] return self.wrap_unspecialized_primitive(value)
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 1073, in wrap_unspecialized_primitive
[DEBUG] guards=self.make_guards(GuardBuilder.CONSTANT_MATCH),
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 269, in make_guards
[DEBUG] return {source.make_guard(guard) for guard in guards}
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_dynamo/variables/builder.py", line 269, in <setcomp>
[DEBUG] return {source.make_guard(guard) for guard in guards}
[DEBUG] File "/data/users/ezyang/b/pytorch/torch/_guards.py", line 641, in make_guard
[DEBUG] return Guard(self.name(), self.guard_sou
```
One downside is I can't report *why* the guard was added. I'm not entirely sure how to do this; the problem is guards will propagate to a bunch of variables before finally getting included as part of the final set. Maybe a very very verbose version could report stack traces at every handoff point.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107388
Approved by: https://github.com/mlazos
ghstack dependencies: #107438, #107358
I get a 2% inference speedup in HF with this PR. I checked to see if there any models where unfusing was slower than the cublas gelu fusion, and I did not see any, which was surprising to me. Sorry for the cublas-activation api churn 😬
Kicking off another run in cublas 12, it's possible that the results have changed since.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106912
Approved by: https://github.com/jansel
ghstack dependencies: #106911
Previously, you would get an error like
```
Dynamo input and output is a strict subset of traced input/output
```
now you get
```
Cannot export model which references tensors that are neither
buffers/parameters/constants nor are direct inputs. For each tensor, if you'd
like this tensor to be an explicit input, add it as a dummy argument
to the top-level model definition you are exporting; if you would
like its value to be embedded as an exported constant, wrap its access
in a function marked with @assume_constant_result.
G['bulbous_bouffant'], accessed at:
File "test_export.py", line N, in f
return bulbous_bouffant + y
```
This doesn't handle outputs, I'm going to hit that next.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106403
Approved by: https://github.com/tugsbayasgalan
This prefigures a refactor that will move the backward compilation
to entirely ahead of time, so I need to extract these strides some
other way. Straight from the compiler's mouth will do it.
I can't easily get the information via the return result of `fw_compiler` without changing the calling convention, so instead I smuggle it via TracingContext. TracingContext may be None when we are compiling patterns for the joint graph pattern matcher.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105010
Approved by: https://github.com/shunting314
Fixes#95900
Using the following repro as guide:
```python
import torch
import torch._dynamo
from torch._subclasses import fake_tensor
from torch.fx.experimental.symbolic_shapes import ShapeEnv
from torch._dynamo.output_graph import config
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.linear = torch.nn.Linear(2, 2)
self.linear2 = torch.nn.Linear(2, 2)
def forward(self, x):
out = self.linear(x)
out = self.linear2(out)
return out
fake_mode = fake_tensor.FakeTensorMode(allow_non_fake_inputs=False,
allow_fallback_kernels=True,
shape_env=ShapeEnv(
allow_scalar_outputs=config.capture_scalar_outputs,
allow_dynamic_output_shape_ops=config.capture_dynamic_output_shape_ops,
frame_id=0
),
)
# Fakefying input/model before calling torch._dynamo.export
with fake_mode:
fake_x = torch.rand(5, 2, 2)
model = Model()
# Calling torch._dynamo.export without active fake mode
graph_module, guards = torch._dynamo.export(
model,
fake_x,
aten_graph=True,
fake_mode=fake_mode
)
graph_module.print_readable()
graph_module.graph.print_tabular()
```
Summary of changes:
* Plumb fake_mode through torch.export API. When specified, it
replaces the creation of a new FaketendorMode at InstructionTranslator on behalf of OutputGraph
Hacks FakeTensor.__new__ to prevent a
torch.tensor._make_subclass call for inputs that are already fakefied by
user. This probably need to be fixed in a nicer way. Any idea?
* Removed a few asserts that didn't want faked tensors coming
from user script
* Added torch._subclasses.fake_tensor.FakeTensor to type list on a few
asserts check to allow fake inputs
The changes above allowed symbolic tracing with both static and dynamic shapes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100017
Approved by: https://github.com/ezyang
Adds a freezing pass that will constant fold parameters in inductor `config.freezing`. This occurs post functionalization in aot autograd to capture both dispatching and allow passes to occur post functionalization. A few notes:
- There is an option to discard parameters `config.freezing_discard_parameters` which will take the current eager modules and wrap parameters to a Tensor subclass which will error if used.
- I needed to expose flat_params in aot_autograd in order to discard old references when we constant fold away parameters, like with amp. I also exposed `fw_metadata` to avoid constant folding mutated paraemters.
- Caching parameter transformations/constant folding across different inferences nyi
- Checking version_counter of constant folded params nyi
I'm not really sure what the actual naming should be. In jit there was both "freezing", which was platform agnostic, and "optimize for inference", which made device specific optimizations. We're doing the latter here but maybe freezing is a better name.
Differential Revision: [D46244033](https://our.internmc.facebook.com/intern/diff/D46244033)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100652
Approved by: https://github.com/jansel
Billing of changes:
* Get rid of `print_guards`; instead, you control this with `TORCH_LOGS=torch.fx.experimental.symbolic_shapes`, debug logging toggles stack traces
* Don't incorrectly report the tracing context frame when we're compiling; we just don't have this info anymore! (TODO: use the saved frames instead). This is via a new TracingContext.clear_frame context manager
* Add TracingContext.extract_stack() which gives you the tracing context stack.
* Add ShapeEnvLoggingAdapter to report which ShapeEnv any given operation is from (this is helpful for debugging situations when there are too many ShapeEnvs floating around)
* Tweak create_symbol log message to also report Source
* Add a debug log whenever duck sizing occurs
* Report an excerpt of both the user and system backtrace whenever a guard is added in INFO mode. I found this is a good balance of "where did the guard come from" without full backtrace verbosity.
Example log output with the new output:
```
[2023-04-12 08:25:49,003] torch.fx.experimental.symbolic_shapes: [INFO] 0: create_env
[2023-04-12 08:25:49,021] torch.fx.experimental.symbolic_shapes: [INFO] 0: create_symbol s0 = 32 for L['x'].size()[0]
[2023-04-12 08:25:50,154] torch.fx.experimental.symbolic_shapes: [INFO] 0: evaluate_expr s0 < 128 [guard added] at w.py:11 in forward2 (_dynamo/variables/tensor.py:476 in evaluate_expr)
[2023-04-12 08:25:52,057] torch.fx.experimental.symbolic_shapes: [INFO] 0: evaluate_expr Eq(Mod(s0, 16), 0) [guard added] (_inductor/codegen/triton.py:77 in is_aligned)
```
from running
```
import torch
import torch._dynamo
def f(x, y):
return x + y
def forward(x, y):
return forward2(x, y)
def forward2(x, y):
if x.size(0) < 128:
x = x * 2
else:
x = x * 3
r = f(x, y)
r = r * y
return r
def woof():
fn_compiled = torch.compile(forward, dynamic=True)
x = torch.randn(32, device='cuda')
y = torch.randn(32, device='cuda')
print(fn_compiled(x, y))
woof()
```
(To induce the Triton guard, I synthetically reverted https://github.com/pytorch/pytorch/pull/98471)
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98941
Approved by: https://github.com/wconstab
This replaces fake_mode_from_tensors but it preferentially looks for
fake_mode in TracingContext and also if there is an active fake mode
on the dispatch stack, before groveling in tensors to find it.
This advances PegasusForCausalLM, which was previously failing because
we generated a graph that had a parameter (non-fake) and a SymInt,
and thus previously we failed to detect the correct fake mode.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98321
Approved by: https://github.com/voznesenskym
The purpose of this PR is to remove reliance on argument positions in dedup guards, AND extend the functionality to params.
A version of this PR was stamped prior https://github.com/pytorch/pytorch/pull/95831 - but was kinda gross, because it was based on an underlying PR that did way too much with source names.
This PR leaves most of that alone, in favor of just reusing the same name standardization logic that dynamo module registration does.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96774
Approved by: https://github.com/ezyang
Previously, when starting to trace a function, we would record a frame summary recording the definition loc. This would lead to an unconventional-looking stack trace when used for debugging, e.g., shape guards.
```
File ".../scripts/avik/pt2/example.py", line 407, in forward
def forward(self, x):
...
File ".../transformers/models/bert/modeling_bert.py", line 912, in forward
@add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
...
File ".../transformers/models/bert/modeling_bert.py", line 562, in forward
def forward(
...
File ".../transformers/models/bert/modeling_bert.py", line 484, in forward
def forward(
...
File ".../transformers/models/bert/modeling_bert.py", line 416, in forward
def forward(
...
File ".../transformers/models/bert/modeling_bert.py", line 275, in forward
def forward(
...
File ".../transformers/models/bert/modeling_bert.py", line 351, in forward
attention_scores = attention_scores + attention_mask
```
As noted in https://github.com/pytorch/pytorch/pull/95848#discussion_r1134397096, we would like to change this to record function calls instead, like conventional stack traces do. This diff makes this change. The above stack now looks like the following, which is way more helpful at a glance to understand what's going on.
```
File ".../scripts/avik/pt2/example.py", line 408, in forward
bert_out = self.bert(**x)
...
File ".../transformers/models/bert/modeling_bert.py", line 1021, in forward
encoder_outputs = self.encoder(
...
File ".../transformers/models/bert/modeling_bert.py", line 610, in forward
layer_outputs = layer_module(
...
File ".../transformers/models/bert/modeling_bert.py", line 496, in forward
self_attention_outputs = self.attention(
...
File ".../transformers/models/bert/modeling_bert.py", line 426, in forward
self_outputs = self.self(
...
File ".../transformers/models/bert/modeling_bert.py", line 351, in forward
attention_scores = attention_scores + attention_mask
```
Differential Revision: [D44101882](https://our.internmc.facebook.com/intern/diff/D44101882/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96882
Approved by: https://github.com/ezyang
I am still reading Dynamo source code...
This is an easy PR to simplify `Source.is_nn_module()` to reuse `GuardSource.is_nn_module()` instead of having the `in (...)` check implemented twice. While simplifying that, I thought I might as well add some type annotations for `Source` methods.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95292
Approved by: https://github.com/ezyang
Whenever you guard on something, you're supposed to tell GuardBuilder about it, so GuardBuilder knows that it has to actually bind it in scope when it creates the guard function. But shape env guards bypass that mechanism completely. Well, now they don't.
For the most part, this didn't matter in practice, because we usually had a `TENSOR_MATCH` guard floating around that made sure that the guard stayed live. But if we ever eliminate those guards (e.g., because we build it into the shape guard directly; something we'll probably want to do when https://github.com/pytorch/pytorch/pull/89707 goes online) then this will indeed matter.
One complication: some of the shape env guards are on globals. You have to make sure to shunt the usage to the correct guard builder in that case. Maybe it would be better if we refactored things so there is only one GuardBuilder. Not sure.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91058
Approved by: https://github.com/voznesenskym
I'm going to need this in the follow up PR. Instead of storing only Source.name() in Symbol, I now store a full on Source. Lots of replumbing reoccurs. In particular:
- Move Source to torch._guards to break cycles
- I have to add TensorPropertySource and NegateSource to handle x.size()[0] and -x codegen that I was doing with string manipulation previously
- I tighten up invariants so that I never pass source=None; instead I pass ConstantSource (these are constant sources right) and test for that rather than source being missing. I think this is more parsimonious
- Some mypy wobbles from new imports
I didn't move LocalSource and friends to torch._guards, but I ended up needing to access them in a few places. The main annoyance with moving these is that then I also need to move the bytecode codegen stuff, and that's not so easy to move without bringing in the kitchen sink.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91057
Approved by: https://github.com/albanD, https://github.com/voznesenskym, https://github.com/zou3519
I'm going to need this in the follow up PR. Instead of storing only Source.name() in Symbol, I now store a full on Source. Lots of replumbing reoccurs. In particular:
- Move Source to torch._guards to break cycles
- I have to add TensorPropertySource and NegateSource to handle x.size()[0] and -x codegen that I was doing with string manipulation previously
- I tighten up invariants so that I never pass source=None; instead I pass ConstantSource (these are constant sources right) and test for that rather than source being missing. I think this is more parsimonious
- Some mypy wobbles from new imports
I didn't move LocalSource and friends to torch._guards, but I ended up needing to access them in a few places. The main annoyance with moving these is that then I also need to move the bytecode codegen stuff, and that's not so easy to move without bringing in the kitchen sink.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91057
Approved by: https://github.com/albanD, https://github.com/voznesenskym
The idea is to make ShapeEnv guards less of a one-off special snowflake, and integrate it more closely with the regular builder infrastructure. But it is not so easy: the shape env code has to live after tensor match code, because we need to know that the values in question are tensors before we start matching on them. So we introduce a new `shape_env_code` field to put the special shape env code, so we can add it to the final constructed code after tensor.
Everything else works the obvious way. There's a new ShapeEnvSource for constructing the singleton SHAPE_ENV guard that drives the shape env guard construction. I added some more docs and also made the printed code for guards include the enclosing lambda for more clarity.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91055
Approved by: https://github.com/albanD, https://github.com/voznesenskym
