Fixes https://github.com/pytorch/pytorch/issues/125720
I was earlier worried that DELETE_* or STORE_* on referent values should result in a graph break, because they could invalidate the weak ref. But then @zou3519 pointed out that weakref invalidation will happen EVENTUALLY, CPython provides no guarantees when the weakref will be invalidated (even when the user calls del x and x is the last reference).
So any code that relies on del x to invalidate the weakref of x right away is BAD code. CPython provide no guarantees. Therefore we can (ab)use this nuance, and can just ignore DELETE_* or STORE_* on the referent objects.
The only corner case is when Dynamo is reconstructing the weakref object. Dynamo will have a hard time being correct here, so just SKIP_FRAME on such a case. This is rare.
Cpython notes
1) https://docs.python.org/3/library/weakref.html
2) https://docs.python.org/3/reference/datamodel.html#index-2
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128533
Approved by: https://github.com/jansel
FIXES#113263. Same idea as in https://github.com/pytorch/pytorch/pull/113417, but we need a more intrusive C API to silently nop default saved tensor hooks, in order to support user-code that use torch.autograd.disable_saved_tensors_hooks (see test_unpack_hooks_can_be_disabled). We mock the output of get_hooks while leaving push/pop untouched.
For compiled autograd, we're firing pack hooks once and unpack hooks twice right now, I'll look into this separately from this issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123196
Approved by: https://github.com/soulitzer
Today inlining builtin nn modules is not compatible with parameter freezing. Freezing parameters and then constant folding them through the graph relies on the assumption that they will not be inputs and will be static across calls to the same graph. When inlining builtin nn modules this assumption is broken and we reuse the same graph for different instances of the same nn module. There are three options 1) abandon constant folding, 2) create a dispatcher layer (like cudagraphs) which will dispatch to the correct constant-folded graph for each distinct set of parameters or 3) recompile
This PR implements 3 by introducing guards on the parameter pointers. This was due to freezing being relatively rare and performance sensistive. 2 Had many more unknowns and 1 is not a viable option due to the drop in performance.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128355
Approved by: https://github.com/anijain2305
This PR requires a little justification, but let's start with what it does first:
1. When you have a 0d CPU scalar int64/float64 tensor input to a graph, we will preallocate a backed SymInt/SymFloat corresponding to what you would get if you call item() on this tensor. This means you can freely change your input to be a Python int/float or a Tensor with an item() call and end up with exactly the same level of expressivity (specifically, you can guard on the internal SymInt/SymFloat no matter what). By default, the source of the backed SymInt/SymFloat is `L['tensor'].item()`, but if you have promoted a float input into a Tensor, we will cancel out `torch.as_tensor(L['float']).item()` into just `L['float']`.
2. We switch wrap_symfloat to use this, instead of hand crafting the new SymNodeVariable. Everything works out, except that we carefully pass the item() result to tracked fakes (and not the fake Tensor argument)
OK, so why do this at all? There is some marginal benefit where now some item() calls on scalar inputs can be guarded on, but IMO this is a pretty marginal benefit, and if it was the only reason, I wouldn't do this. The real reason for this is that I need to be able to propagate fake tensors through the graphs that are produced by Dynamo, and if I am doing the old custom wrap_symfloat logic, there's no way I can do this, because ordinarily an item() call will cause an unbacked SymInt when I reallocate.
The other obvious way to solve the problem above is to make a HOP alternative that item() that "bakes in" the backed SymInt its supposed to return. But this strategy seems more parsimonious, and it does have the marginal benefit I mentioned above. The main downside is that what I have to do next, is make it so that when I run tensor computation, I also apply the equivalent operations to the SymInt/SymFloat as well. That's next PR.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126245
Approved by: https://github.com/eellison
ghstack dependencies: #126637
The big idea is that floats are treated as Tensors on input/output to the FX graph, but on the inside, we immediately call item() on the synthetic Tensor and record regular float operations on it. Canonicalization to Tensor operations will happen in a standalone FX pass. This behavior is controlled by `specialize_float` config variable when set to False.
The generated graph looks like this for the test `test_unspec_float_output`:
```
def forward(self, L_x_: "f32[3]", L_y_: "f32[]"):
l_x_ = L_x_
l_y_ = L_y_
# File: /data/users/ezyang/a/pytorch/test/dynamo/test_unspec.py:511 in f, code: return x + 1, y * 2
add: "f32[3]" = l_x_ + 1; l_x_ = None
item: "Sym(zf0)" = l_y_.item(); l_y_ = None
mul: "Sym(2*zf0)" = item * 2; item = None
scalar_tensor: "f32[]" = torch.scalar_tensor(mul); mul = None
return (add, scalar_tensor)
```
The ingredients:
* **torch/_dynamo/variables/builder.py** When `specialize_float` is False, we wrap float literals with `wrap_symfloat`. This is an unholy mashup of `wrap_symint` and `wrap_unspecialized_primitive`. The overall strategy is that we first generate a tensor argument (because that's what we want to show up into the FX graph), but then immediately call item() on the tensor argument to get a SymNodeVariable, which we will do the rest of the tracing with. Importantly, this SymNodeVariable is backed with the source of the original float: this means we can guard on the resulting value (something we could NOT do with UnspecializedPythonVariable). This has to be done manually, because if you literally call item() on the tensor, you will end up with an unbacked float. There is a bit of copy paste from wrap_symint and wrap_unspecialized_primitive which we can try to factor out, but this really is its own thing and you should review every line of code in the function.
* **torch/fx/experimental/symbolic_shapes.py** We now can generate guards on float inputs, and these guards are handled inside of ShapeEnv. So we need to be able to allocate (backed!) float symbols, and produce guards for them. Fairly straightforward generalization.
* **torch/_dynamo/codegen.py** I also need to maintain the invariant that there are no float outputs to the FX graph. I chose to do this at codegen time. When we detect a SymNodeVariable on the return stack for a float, we on the fly convert it (via `as_tensor`) to a TensorVariable, which is the true output. We then special case the output bytecode to call item() on it again. The tensor conversion is memoized on SymNodeVariable since we typically run the code generation process twice.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125325
Approved by: https://github.com/lezcano, https://github.com/jansel
- `FakeContext` hides all fields other than ctx.saved_tensors, this dynamo errors when the autograd.Function.backward uses other attrs on ctx and it also doesn't allow fallback to eager.
- If we remove it, we still can't fallback to eager: node variables are already freed (ctx.saved_tensors throws)
- However, we can fallback to "pseudo-eager" by using a duck-typed ctx and routing the ctx.saved_tensors to lifted tensors
- Dynamo tries to inline external_utils.call_backward, treats BackwardCFunction as a AutogradFunctionContextVariable (only used up until we create the fake context: FakeBackwardCFunction)
- we call_function backward from the forward class AutogradFunctionVariable, and we still pass in the fake context as a UserDefinedObjectVariable (can later use AutogradFunctionContextVariable + HOO graph speculate)
Fixes#125489#124827
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125661
Approved by: https://github.com/jansel
While there are some similarities, they are also quite different (one
handles Numpy numbers while the other handles ints. I am also going to
add a wrap_symfloat soon which will do even more different behavior.
So split these out for clarity.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125483
Approved by: https://github.com/lezcano
ghstack dependencies: #125395, #125419
We guard on key order
1) When a key is a non-constant object
2) When we actually need key order - like .values, .items etc
For dicts/OrderedDicts that do not require key order guarding, we just rely on usual `GuardManger + DictGetItemGuardAccessor`. This is faster than going through the `list(d.keys())` based design for OrderedDicts.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124779
Approved by: https://github.com/jansel
Update ruff to 0.4.1 .
This version fixes a lot false negatives/false positives, is 20-40% faster, and has various other bug fixes.
Below is a before and after table showing the execution time of ruff lint and ruff format in milliseconds courtesy of https://astral.sh/blog/ruff-v0.4.0
| Repository | Linter (v0.3) | Linter (v0.4) | Formatter (v0.3) | Formatter (v0.4) |
|----------------------------------------------------|---------------|---------------|------------------|------------------|
| [pytorch/pytorch](https://github.com/pytorch/pytorch) | 328.7 | 251.8 | 351.1 | 274.9 |
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124549
Approved by: https://github.com/ezyang
Closes#114966
Frozen field assignment in `__init__` in Python 3.8-3.9:
f5bd65ed37/Lib/dataclasses.py (L402-L411)
```python
import builtins
BUILTINS = builtins
def _field_assign(frozen, name, value, self_name):
# If we're a frozen class, then assign to our fields in __init__
# via object.__setattr__. Otherwise, just use a simple
# assignment.
#
# self_name is what "self" is called in this function: don't
# hard-code "self", since that might be a field name.
if frozen:
return f'BUILTINS.object.__setattr__({self_name},{name!r},{value})'
return f'{self_name}.{name}={value}'
```
Frozen field assignment in `__init__` in Python 3.10+:
812245ecce/Lib/dataclasses.py (L436-L445)
```python
__dataclass_builtins_object__ = object
def _field_assign(frozen, name, value, self_name):
# If we're a frozen class, then assign to our fields in __init__
# via object.__setattr__. Otherwise, just use a simple
# assignment.
#
# self_name is what "self" is called in this function: don't
# hard-code "self", since that might be a field name.
if frozen:
return f'__dataclass_builtins_object__.__setattr__({self_name},{name!r},{value})'
return f'{self_name}.{name}={value}'
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124393
Approved by: https://github.com/jansel
I'm going to setup some extra behavior when we set example value, so
I need a convenient place to interpose. I cannot easily do it on
meta itself because its a generic dict with no interposition point.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124176
Approved by: https://github.com/oulgen
ghstack dependencies: #124105, #124059
### Context
In today's Dynamo, we lift all tensors encountered during tracing to be individual graph inputs, even when they were in a container.
And [Dynamo generates](fdc281f258/torch/_dynamo/codegen.py (L371)) the runtime function's signature using the graph's graphargs.
This means that the generated function will have each grapharg as an argument, which is problematic if we want to free the inputs in inductor codegen. See [python function arguments are kept alive for the duration of the function call](https://github.com/pytorch/pytorch/pull/83137#issuecomment-1211320670).
```python
# original code
def forward(inputs):
a, b, c, d, e = inputs
inputs.clear()
out = a
out += b
del b # frees memory
out += c
del c # frees memory
out += d
del d # frees memory
out += e
del e # frees memory
return out
# compiled code:
def forward(a, b, c, d, e):
# b, c, d, e can't be freed before end of function
```
This isn't a concern when compiling forward because a, b, c, d, e are all from user code, and should be kept alive. But when compiling backwards, a, b, c, d, e may be intermediate results i.e. activations, that we DO want to clear ASAP to remain on par with eager peak memory.
### Solution
We have encountered similar memory problems in AOTAutograd before, where we adopted the boxed calling convention (wrapping to-be-freed objects in a list), adding list clearing to inductor codegen, and being careful about holding references to elements in the input list. We need to do something similar, but for inputs from the user program (compiled autograd fx graph in this case).
This PR support lists as graphargs/placeholder nodes. When tracing a list of tensors, we create a node for it, and pre-emptively initialize variable trackers for its elements before they are used in the user program. Subsequent uses of those variables will find hits in the lookup table `input_source_to_var`.
With the inputs as a list in the graph args, our compiled code can free inputs just like in the eager case.
```python
def forward(inputs):
# a, b, c, d, e can be freed within the function now
```
Currently, AOT/Inductor flattens list input via [flatten_graph_inputs wrapper](597f479643/torch/_inductor/compile_fx.py (L1454-L1478)), which is why this PR's CI can be green. Additional changes are needed to its runtime wrapper, done in the next PR. The next step is to ensure that we are careful in forwarding the list to inductor codegen without holding additional references.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122353
Approved by: https://github.com/jansel
ghstack dependencies: #123630, #123674
FSDP2 creates CUDA streams outside of compile region in its 1st iteration eager run, and then torch.compile will attempt to record method calls on these streams (e.g. `stream.record_event()`) in >1st iteration compiled run.
Before this PR, stream proxy is None which causes "None doesn't have attribute record_event" error when we try to call `record_event()` on it. After this PR, stream proxy has the correct value which makes calling methods on it possible.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123487
Approved by: https://github.com/jansel
I am ok if people don't want this PR to be merged.
For optimizers, we know that the state dict and param_group have same parameters. So, I think its ok to skip TENSOR_MUST_ALIAS guards.
Similarly for state tensors, all of them are different. Therefore, we can skip the tensor aliasing guards.
With this PR, these are the numbers for Megatron which has 394 parameters
<img width="290" alt="image" src="https://github.com/pytorch/pytorch/assets/13822661/0ce75dc6-4299-46bb-bf3c-7989ebc7cfc4">
C++ numbers jump a lot because of 2 reasons
1) We are now not doing INCREF/DECREF for a large number of tensors.
2) For python guards, we can expect higher numbers but that requires some more plumbing because the Python tensor guards are all collapsed into one.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123044
Approved by: https://github.com/jansel, https://github.com/mlazos
Currently, when we create proxies for a list's elements in wrap_fx_proxy_cls, we create them using the same source as the list's e.g. `LocalSource(inputs)` instead of `GetItemSource(LocalSource(inputs), index=i)`. This results in invalid guards when the tensors it contains becomes dynamic, and the guard system thinks the list is a tensor:
```
Malformed guard:
L['sizes'][0] == L['inputs'].size()[0]
Malformed guard:
2 <= L['inputs'].size()[0]
Traceback [...]
AttributeError: 'list' object has no attribute 'size'
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122691
Approved by: https://github.com/jansel, https://github.com/anijain2305
Fixes#118795
This is a graph breaking partial fix for #120914. We still need -actual- module parametrization tracing support, but at least it doesn't blow up hard now.
**Background**: Module parametrization injects a property as the module parameter attribute that calls a `nn.Module` whose forward takes in a module parameter and returns a reparametrized module parameter.
Example:
```
class MyParametrization(nn.Module):
def forward(X):
# This reparametrization just negates the original parameter value
return -X
m = nn.Linear(...)
p = MyParametrization()
register_parametrization(m, "weight", p)
# Accessing the "weight" attribute will invoke p's forward() on m's original weight and return the output as the new weight.
# m.weight here is now an injected property that does the above instead of an actual Parameter.
# This property is defined in torch/nn/utils/parametrize.py.
m.weight
# NB: Parametrization changes the module type (e.g. torch.nn.utils.parametrize.ParametrizedLinear)
print(type(m))
```
**Problem 1**: Dynamo has special tracing rules for things in `torch.nn`. Parametrizing a module changes the type of the module and the parametrized attribute, so now these rules wrongly affect tracing here. To fix this:
* For parametrized modules, call `convert_to_unspecialized()` to restart analysis where Dynamo starts inlining the module.
**Problem 2**: The issue seen in #118795 is that Dynamo will see a dynamically constructed tensor when `m.weight` is called and introduce that to its `tensor_weakref_to_sizes_strides` cache during fake-ification. This tensor is also made to be a graph input, since it's a module parameter. When guards are created for this module parameter input, the logic calls `m.weight` again and tries to look the result up in the cache, but this is a different tensor now, giving the `KeyError` symptom. To fix this:
* Replace Dynamo's `tensor_weakref_to_sizes_strides` cache with a `input_source_to_sizes_strides` cache.
* This cache was originally introduced in #100128.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121041
Approved by: https://github.com/anijain2305
List of changes:
- Replace JVP_NESTING by torch._C._functorch.maybe_current_level()
- Remove all increment nesting functions from wrap_fx_proxy_cls
- fwAD.make_dual receives the dual_level as keyword argument
- Add jvp_increment_nesting, set_fwd_grad_enabled and dual_level context managers to dynamo
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119926
Approved by: https://github.com/zou3519
This PR:
* Introduces an ATen op for creating true jagged views from a dense values buffer
* `_nested_view_from_jagged(values, offsets, lengths, ragged_idx, dummy)`
* This ops is implemented on the Python side using torch.library so we can return a subclass instance
* `jagged_from_list()` now uses this instead of the old autograd.Function `NestedViewFromBuffer`
* The latter op is used for non-contiguous JTs returned via `torch.nested.narrow()`
* `dummy` is an awful hack to ensure that `NestedTensor.__torch_dispatch__()` is invoked for our view
* Introduces an ATen op for accessing the `values` component of an NT via a view
* `_nested_get_values(nt)`
* **Removes** the autograd.Functions `ViewNestedFromBuffer` and `ViewBufferFromNested` in favor of `nested_from_values_offsets()` / `nested_from_values_offsets_lengths()` and `nt.values()`, respectively.
* Changes test code to prefer `as_nested_tensor()` over `jagged_from_list()` directly
* Similarly, avoid `buffer_from_jagged()`, preferring `values()`
* Depends on general subclass view fake-ification on the PT2 side (handled solely in previous PRs in the stack)
With these changes, the semantics of jagged layout NTs are such that they are considered a true view of the underlying `values` buffer. This means views of jagged NTs are views of the underlying buffer as well, simplifying some handling.
Differential Revision: [D54269922](https://our.internmc.facebook.com/intern/diff/D54269922)
Co-authored-by: voznesenskym <voznesenskym@gmail.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113279
Approved by: https://github.com/ezyang
