Commit Graph

7 Commits

Author SHA1 Message Date
Edward Z. Yang
f641c55c9b Make torch._dynamo.mark_static work inside graph (#118962)
I livecoded the entire PR authoring process, you can watch it at https://youtu.be/06HuwNR9-uI

Signed-off-by: Edward Z. Yang <ezyang@meta.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/118962
Approved by: https://github.com/yanboliang
2024-02-02 20:01:27 +00:00
Edward Z. Yang
68f9c28e00 Don't make default arguments dynamic (#118772)
Noticed this while working on
https://github.com/pytorch/pytorch/issues/114590

Signed-off-by: Edward Z. Yang <ezyang@meta.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/118772
Approved by: https://github.com/anijain2305
2024-02-01 18:11:57 +00:00
Jez Ng
5da9abfec2 [dynamo] Enable typechecking for comptime.py (#112999)
I made `comptime` a callable instance instead of a function because mypy
doesn't allow creating extra attributes on a plain function.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/112999
Approved by: https://github.com/ezyang
ghstack dependencies: #112130, #112970, #112971, #112972, #112973, #112974, #112975
2023-11-08 21:17:45 +00:00
Mark Saroufim
95fced4483 Pretty dataclass dynamo explain (#102869)
Also thinking out loud: maybe we only print graph break reasons? And for the rest we have a verbose print which prints everything?

TODO: some tests are failing based on what they expect a guard string to look like, easy to fix i'll do it early next week

# After

```
(sourcetorch) ubuntu@ip-172-31-1-136:~/test$ python pretty.py
BREAK
Graph Count: 2
Graph Break Count: 1
Op Count: 2
Break Reasons:
  Break Reason 1:
    Reason: call_function BuiltinVariable(print) [ConstantVariable(str)] {}
    User Stack:
      <FrameSummary file /home/ubuntu/test/pretty.py, line 6 in fn>
Ops per Graph:
  Ops 1:
    <built-in function add>
  Ops 2:
    <built-in function add>
Out Guards:
  Guard 1:
    Name: ''
    Source: global
    Create Function: GRAD_MODE
    Guard Types: ['GRAD_MODE']
    Code List: ['___is_grad_enabled()']
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 2:
    Name: ''
    Source: global
    Create Function: DEFAULT_DEVICE
    Guard Types: ['DEFAULT_DEVICE']
    Code List: ['utils_device.CURRENT_DEVICE == None']
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 3:
    Name: "G['print']"
    Source: global
    Create Function: BUILTIN_MATCH
    Guard Types: None
    Code List: None
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 4:
    Name: ''
    Source: global
    Create Function: DETERMINISTIC_ALGORITHMS
    Guard Types: ['DETERMINISTIC_ALGORITHMS']
    Code List: ['not ___are_deterministic_algorithms_enabled()']
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 5:
    Name: "L['x']"
    Source: local
    Create Function: TENSOR_MATCH
    Guard Types: None
    Code List: None
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 6:
    Name: ''
    Source: global
    Create Function: GRAD_MODE
    Guard Types: ['GRAD_MODE']
    Code List: ['___is_grad_enabled()']
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 7:
    Name: ''
    Source: global
    Create Function: DEFAULT_DEVICE
    Guard Types: ['DEFAULT_DEVICE']
    Code List: ['utils_device.CURRENT_DEVICE == None']
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 8:
    Name: ''
    Source: global
    Create Function: DETERMINISTIC_ALGORITHMS
    Guard Types: ['DETERMINISTIC_ALGORITHMS']
    Code List: ['not ___are_deterministic_algorithms_enabled()']
    Object Weakref: None
    Guarded Class Weakref: None
  Guard 9:
    Name: "L['x']"
    Source: local
    Create Function: TENSOR_MATCH
    Guard Types: None
    Code List: None
    Object Weakref: None
    Guarded Class Weakref: None
Compile Times: TorchDynamo compilation metrics:
Function                        Runtimes (s)
------------------------------  --------------
_compile                        0.0164, 0.0035
OutputGraph.call_user_compiler  0.0000, 0.0000
```

## Before

```
('Dynamo produced 2 graphs with 1 graph break and 2 ops', [{Guard(name='print', source=<GuardSource.GLOBAL: 1>, create_fn=<function GuardBuilder.BUILTIN_MATCH at 0x7f92ea5009d0>, is_volatile=False, guard_types=None, code_list=None, obj_weakref=None, guarded_class_weakref=None), Guard(name='x', source=<GuardSource.LOCAL: 0>, create_fn=<function GuardBuilder.TENSOR_MATCH at 0x7f92ea501000>, is_volatile=False, guard_types=['TENSOR_MATCH'], code_list=None, obj_weakref=<weakref at 0x7f9224d28f40; dead>, guarded_class_weakref=<weakref at 0x7f92d81734c0; to 'torch._C._TensorMeta' at 0x540b610 (Tensor)>)}, {Guard(name='x', source=<GuardSource.LOCAL: 0>, create_fn=<function GuardBuilder.TENSOR_MATCH at 0x7f92ea501000>, is_volatile=False, guard_types=['TENSOR_MATCH'], code_list=None, obj_weakref=<weakref at 0x7f9224d5e700; dead>, guarded_class_weakref=<weakref at 0x7f92d81734c0; to 'torch._C._TensorMeta' at 0x540b610 (Tensor)>)}], [GraphModule(), GraphModule()], [[<built-in function add>], [<built-in function add>]], [GraphCompileReason(reason='call_function BuiltinVariable(print) [ConstantVariable(str)] {}', user_stack=[<FrameSummary file <ipython-input-1-9e2ddb639697>, line 6 in fn>]), GraphCompileReason(reason='return_value', user_stack=[<FrameSummary file <ipython-input-1-9e2ddb639697>, line 8 in <graph break in fn>>])], 'Dynamo produced 2 graphs with 1 graph break and 2 ops\n Break reasons: \n\n1. call_function BuiltinVariable(print) [ConstantVariable(str)] {}\n  File "<ipython-input-1-9e2ddb639697>", line 6, in fn\n    print("BREAK")\n \n2. return_value\n  File "<ipython-input-1-9e2ddb639697>", line 8, in <graph break in fn>\n    return x\n \nTorchDynamo compilation metrics:\nFunction                        Runtimes (s)\n------------------------------  --------------\n_compile                        0.0418, 0.0084\nOutputGraph.call_user_compiler  0.0001, 0.0001')

```

## Program

```python
import torch
import torch._dynamo

def fn(x):
    x = x + 1
    print("BREAK")
    x = x + 1
    return x

out = torch._dynamo.explain(fn, torch.randn(10))
print(out)

```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/102869
Approved by: https://github.com/voznesenskym
2023-06-07 22:38:57 +00:00
Edward Z. Yang
b740d3b014 Add comptime.breakpoint (#102758)
This sets a pdb breakpoint to fire whenever we *compile* this
Python code in Dynamo.

Signed-off-by: Edward Z. Yang <ezyang@meta.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/102758
Approved by: https://github.com/zou3519, https://github.com/voznesenskym
2023-06-02 17:44:16 +00:00
Edward Z. Yang
7dd95ad7f3 Add a convenience shortcut for accessing size on ComptimeVar (#95404)
Signed-off-by: Edward Z. Yang <ezyang@meta.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/95404
Approved by: https://github.com/voznesenskym
2023-02-27 02:02:02 +00:00
Edward Z. Yang
dfe916ca88 Dynamo comptime, with public ComptimeContext API (#90983)
This PR adds `@comptime`, a decorator that causes a given function to be executed at compile time when Dynamo is symbolically evaluating their program. To query the Dynamo state, we offer a public ComptimeContext API which provides a limited set of APIs for querying Dynamo's internal state. We intend for users to use this API and plan to keep it stable. Here are some things you can do with it:

* You want to breakpoint Dynamo compilation when it starts processing a particular line of user code: give comptime a function that calls breakpoint
* You want to manually induce a graph break for testing purposes; give comptime a function that calls unimplemented
* You want to perform a debug print, but you don't want to induce a graph break; give comptime a function that prints.
* You can print what the symbolic locals at a given point in time are.
* You can print out the partial graph the Dynamo had traced at this point.
* (My original motivating use case.) You want to add some facts to the shape env, so that a guard evaluation on an unbacked SymInt doesn't error with data-dependent. Even if you don't know what the final user API for this should be, with comptime you can hack out something quick and dirty. (This is not in this PR, as it depends on some other in flight PRs.)

Check out the tests to see examples of comptime in action.

In short, comptime is a very powerful debugging tool that lets you drop into Dynamo from user code, without having to manually jerry-rig pdb inside Dynamo to trigger after N calls.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Pull Request resolved: https://github.com/pytorch/pytorch/pull/90983
Approved by: https://github.com/jansel
2022-12-19 11:06:01 +00:00