Fixes#115607
We were missing guards when the grads were set to `None`. So if we compiled the optimizer with grads set to their proper value, and then with the grads set to `None` we'd continuously run the `None` version because all of the guards would pass and it would be ordered before the correct version in the cache.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121291
Approved by: https://github.com/Skylion007, https://github.com/anijain2305
This enables the new way of writing guards for dicts. Before we were
doing things like
```
L['self'].param_groups[0][___dict_keys_getitem(L['self'].param_groups[0], 0)][3] is L['self'].param_groups[0]['params'][3]
```
without knowing whether `L['self'].param_groups[0][___dict_keys_getitem(L['self'].param_groups[0], 0)]` was a list.
On a different note, I'll probably write a pass to recover the previous
way to place guards on dicts via something like `DICT_KEYS` as an
optimisation, as it seems relevant for optimisers.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117983
Approved by: https://github.com/mlazos
ghstack dependencies: #117982, #118098
The original motivation for MYPYINDUCTOR was a faster type checking configuration that only checked a subset of files. With the removal of `follow_imports = ignore`, we are now able to use dmypy to do fast incremental typechecking, eliminating the need for this.
Perhaps erroneously, when I tee'ed up this PR I elected to delete the `follow_imports = skip` designations in the mypy-inductor.ini. This lead to a number of extra type error suppressions that I manually edited. You will need to review.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118432
Approved by: https://github.com/Skylion007
ghstack dependencies: #118414, #118418
We split install_global_once into two APIs:
- `install_global_by_id(prefix, value) -> name`: installs a global if it hasn't
been installed yet
- `install_global(prefix, value) -> name`: always installs the global (and
generates a unique name for it)
Then, we refactor most callsites of `install_global_unsafe` to one of
the previous. Some callsites cannot be refactored because we create the
global name first, do a lot of stuff with it, and then install it.
This fixes more test flakiness.
Test Plan:
- Existing tests; I can't reliably repro the flakiness
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118100
Approved by: https://github.com/ezyang, https://github.com/mlazos
1. Removes calls to `replace_all` and `clone` and makes VTs mutable.
2. Properly handles Tuple Iterator mutation. Previously TupleIterator variables would only be properly reconstructed if they were advanced at least once in a frame. On calls to `next`, the source information would be lost (due to constructing a new iterator without using builder), which would ensure that during codegen the variable would be reconstructed from scratch. Now that VTs are mutated, the source is never lost, so we need to properly track mutation and handle it by replaying calls to `next` at the end of the modified bytecode.
3. Added test for checking iadd side effects, this was missing in our unit test coverage.
4. Fixed two incorrect sources, DelayGraphBreakVariable, and UserMethodVariable both relied on setting the source to AttrSource(parent, name) at the callsite of `var_getattr`.
5. Fixed a bug in inplace adding for lists, it would set the resulting VariableTracker's source to `None` which would utilize a different reconstruct path in codegen. Now this is handled explicitly by reconstructing vars when allow_cache=`False`, so that during side effect replay, the mutated var is correctly updated.
In subsequent PRs:
* Refactoring side effect tracking to be significantly simpler (I think we only need an `is_modified` flag)
* Refactor `next_variables` iterator to match the signature of `next`
* Remove all references to `options` in the code
* Refactor VTs representing mutable collections to implement their own mutation update handling
* Remove clone and/or make it specific to lists for creating slices
* Add mutation tracking/replay for sets
* Add mutation tracking/replay for iter.py
* Removing setting source in builder (it's set at the top level after a var is returned)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113725
Approved by: https://github.com/jansel
Marks all params/optimizer state as static addresses and a finalizer which cleans up the graph attributes when the optimizer goes out of scope.
**Note: this does not mark grads as static because this will increase memory usage significantly
There are two cases:
1. The upstream graph is cudagraphed - this case will work fine OOTB
2. The upstream graph is not cudagraphed - in this case, there will be a lot of copies introduced from the upstream (to copy the grads) into cudagraphed-owned memory, unless the user explicitly marks the grads as static. If the user does this, this will also require not deallocating the grads in zero_grad() (either the mod or optimizer version) by setting them to zero vs None. There is a PR (https://github.com/pytorch/pytorch/pull/107853) in flight to throw an error if zero_grad attempts to set static grads to None.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107504
Approved by: https://github.com/eellison
Manually generate guards for optimizer rather than use variable builder, which can be slow with lots of params.
This is the reason for ~10s compile slowdown
Redisable `_init_group`. This is important, because if for any reason a frame which calls `_init_group` is run in the python interpreter, we will trace it, which we don't want to do. We only want to call it when it is accessed via the fast path implemented with the optimizer variable during symbolic interpretation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103121
Approved by: https://github.com/jansel
On calls to `_init_group` rather than tracing through it, extract python values from the arguments, and call the initialization. This avoids having to trace this function which is very slow with large parameters, and also avoids graph breaking on it. This is sound in this case because the state is only initialized once in the eager case. Guards on the state and params are generated explicitly rather than via tracing the initialization.
Caveats:
`_init_group` also gathers various state tensors into lists via mutating list arguments to pass to the functional optimizer implementation. These state tensors exist on the optimizer itself, but we don't know exactly how the gathering is done and which tensors correspond to which attributes of the optimizer module (each optimizer has different states). To rectify this, we keep weak_ptrs to all of the tensors collected in the lists in globals (similar to how parameter keys are stored for dictionaries). These pointers are guaranteed to be alive as long as the optimizer object is alive if the internal state is not interfered with and they are guarded with weakref guards
Pull Request resolved: https://github.com/pytorch/pytorch/pull/102640
Approved by: https://github.com/jansel
