Fix two more leaks of the same variety as #136507 (see that PR desc and attached gdoc for debug details).
This time, also add a test-time check that helped to discover new leaks and ensure we won't accidently regress.
Adds `check_tensor_leak` util which internally asserts no tensors are being kept alive by other objects involved in py ref cycles.
Uses objgraph for a nice debug utility when a leak is found.
Credit to @H-Huang for pointing out objdump and helping debug the 'param_group["intermediates"]` leak.
I manually confirmed that all 3 of the leaks identified/fixed so far are caught by the unit test and checker.
Sample output, if I re-introduce a leak by commenting out `del param_group["intermediates"]` in _backward.py,
and run `python test/distributed/pipelining/test_schedule_multiproc.py -k test_schedule_with_native_zero_bubble`:
```
warnings.warn(
/data/users/whc/pytorch/torch/testing/_internal/common_utils.py:5341: UserWarning: 34 tensors were found in the garbage. Did you introduce a reference cycle?
warnings.warn(
/data/users/whc/pytorch/torch/testing/_internal/common_utils.py:5347: UserWarning: Dumping first 1 objgraphs of leaked tensors rendered to png
Graph written to /tmp/objgraph-ztz642h3.dot (19 nodes)
Graph viewer (xdot) not found, generating a png instead
Image generated as /tmp/objgraph-ztz642h3.png
```
rendering of ` /tmp/objgraph-ztz642h3.png`:
<img width="1671" alt="image" src="https://github.com/user-attachments/assets/9098ff29-224c-4533-935b-83c210ac2e22">
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136584
Approved by: https://github.com/kwen2501, https://github.com/H-Huang
ghstack dependencies: #136507
Co-authored-by: Howard Huang <howardhuang@fb.com>
TLDR; found forward activation tensors were being kept alive "forever"
(or until GC ran), and tracked it down to a cycle involving
`stage_backward.<locals>.extract_tensors_with_grads`.
The reference cycle in question is below. (constructed using gc.get_referrers after doing a gc.collect in gc debug mode)
tensor is kept alive by
`[(<class 'cell'>, '0x7f7360234400')]`
tuple of cell objects
`(<cell at 0x7f73602343d0: function object at 0x7f734fff0ee0>, <cell at 0x7f7360234400: list object at 0x7f734e4d9a80>, <cell at 0x7f73602a4190: list object at 0x7f734eff8b00>)`
is kept alive by
`[(<class 'function'>, '0x7f734fff0ee0')]`
`<function stage_backward.<locals>.extract_tensors_with_grads at 0x7f734fff0ee0>`
is kept alive by
`[(<class 'cell'>, '0x7f73602343d0')]`
Put into more plain terms,
```
def stage_backward(...):
...
stage_output_tensors = []
# a cell object will exist that contains the variables defined in stage_backward and used by
# both stage_backward and nested functions
# in this case, the cell object contains 'stage_output_tensors' but
# this function object will hold a reference to a 'cell' that contains any vars from
# the parent scope not explicitly passed into the function as args.
def extract_tensors_with_grads(...):
...
# extract_tensors_with_grads refers to stage_output_tensors, so stage_output_tensors
# is in the cell
stage_output_tensors.append(output_val)
...
# but extract_tensors_with_grads ALSO refers to itself (extract_tensors_with_grads),
# so `extract_tensors_with_grads` will be in the cell
extract_tensors_with_grads(...)
```
More debug details:
https://docs.google.com/document/d/1QPH1Lz0tnieIFPM2tyHrjVB-bjlnHuDgjx1p2am3cmE/edit?usp=sharing
In pdb:
```
gc.collect()
g = gc.garbage
g[-1]
[rank0]:(Pdb) [rank0]:<function
stage_backward.<locals>.extract_tensors_with_grads at 0x7fee5c3392d0>
g[-2]
[rank0]:(Pdb) [rank0]:(<cell at 0x7fee7abbcf40: function object at
0x7fee5c3392d0>, <cell at 0x7fee7abbcf70: list object at
0x7fee7ab68940>, <cell at 0x7fee5c3210c0: list object at 0x7fee5e1
d6340>)
g[-3]
[rank0]:(Pdb) [rank0]:[tensor([[[-4.1127e-06, -3.3826e-06, 2.6226e-06,
..., 6.4969e-06,
[rank0]: -4.4405e-06, -4.7684e-06],
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136507
Approved by: https://github.com/awgu, https://github.com/kwen2501
Moved all the backward functions (`stage_backward_input`, `stage_backward_weight`, `stage_backward`) under the same `backward_maybe_with_nosync` function which controls the logic of the data parallel wrappers.
FSDP was not working with zero bubble PP because there will be twice as many "backward" calls and we update the weight gradients after `autograd.grad` is called. As a result, we need to manually call the FSDP `post_backward_hook()` after the weights have the correct gradients.
Fixes the tests:
`python test/distributed/_composable/test_composability/test_pp_composability.py ComposabilityTest.test_manual_with_data_parallel_dp_type_FSDP_ScheduleClass0_use_new_runtime_False`
`python test/distributed/_composable/test_composability/test_pp_composability.py ComposabilityTest.test_manual_with_data_parallel_dp_type_DDP_ScheduleClass0_use_new_runtime_False`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134052
Approved by: https://github.com/kwen2501
Add `stage_backward_input` and `stage_backward_weight` functions to perform the weight updates for inputs and weights independently.
We still support `self.dw_builder` argument for a custom backward, but it has become optional. It takes a separate code path and cannot be used in conjuction with the native zero backward.
Added tests:
`python test/distributed/pipelining/test_schedule_multiproc.py -k test_schedule_with_native_zero_bubble`
`python test/distributed/pipelining/test_backward.py`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132691
Approved by: https://github.com/wconstab
