Also partially fixes#122109
This PR:
- We add a C++ flag (only_lift_cpu_tensors) to toggle the
torch.tensor(1, device='cuda') ctor strategy.
When false (default), it does the current PyTorch behavior
of unconditionally constructing a concrete CUDA tensor then calling
lift_fresh on it. When true, we instead construct a concrete CPU
tensor, call lift_fresh, and then call Tensor.to(device) (under any ambient
modes).
- FakeTensorMode flips this flag depending on if CUDA is available or
not. We don't unconditionally set the flag to True because that is
likely BC-breaking.
Test Plan:
- existing tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124413
Approved by: https://github.com/eellison
It's annoying grepping for `__call__` call-sites so they're now all explicit now. I'd do this to MetaConverter too but that one is way more public, a lot more sites.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122270
Approved by: https://github.com/eellison
ghstack dependencies: #122044
Fixes https://github.com/pytorch/pytorch/issues/121085
This PR pretty involved so pay attention to this description. At a high
level, the refactor is intended to be mechanical: anywhere in
MetaConverter where previously we took a Tensor as argument, we now take
a MetaTensorDesc, which contains all of the information that we would
have queried off of the Tensor, but placed into a separate data
structure which we can serialize or use to recreate a fake tensor in
a separate fake tensor mode in exact fidelity to the original.
However, this transformation is not always entirely mechanical. Here
is what you need to pay attention to:
- The memo table from real Tensor -> meta/fake Tensor is now broken
into two memo tables: real Tensor -> stable int id -> meta/fake
Tensor. The stable int id is needed so that when we do serialization,
we know when tensors/storages alias each other and can ensure we preserve
this aliasing upon deserialization.
The way I have implemented changes the weak reference behavior.
Previously, when either the real Tensor OR the meta/fake Tensor went
dead, we would remove the entry from the memo table. Now, this only
removes entries from one of the two memo tables. This semantically
makes sense, because the user may have held on to the stable int id
out of band, and may expect a real Tensor to continue to be numbered
consistently / expect to be able to lookup a meta/fake tensor from
this id. If this is unacceptable, it may be possible to rejigger
the memo tables so that we have real Tensor -> stable int id
and real Tensor -> meta/fake Tensor, but TBH I find the new
implementation a lot simpler, and arranging the memo tables in this
way means that I have to muck around with the real tensor to save
to the memo table; in the current implementation, I never pass the
Tensor to meta_tensor function AT ALL, which means it is impossible
to accidentally depend on it.
- When I fill in the fields of MetaTensorDesc in describe_tensor, I need
to be careful not to poke fields when they are not valid. Previously,
preconditions were implicitly checked via the conditional structure
("is this sparse? is this nested?") that is tested before we start
reading attributes. This structure has to be replicated in
describe_tensor, and I have almost assuredly gotten it wrong on my
first try (I'll be grinding through it on CI; a careful audit will
help too, by auditing that I've tested all the same conditionals that
the original access was guarded by.)
- I originally submitted https://github.com/pytorch/pytorch/pull/121821
for the symbolic shapes change, but it turned out the way I did it
there didn't actually work so well for this PR. I ended up just
inlining the symbolic shapes allocation logic into MetaConverter
(look for calls to maybe_specialize_sym_int_with_hint), maybe there
is a better way to structure it, but what I really want is to
just read sizes/strides/offset directly off of MetaTensorDesc; I
don't want another intermediate data structure.
- Some fields aren't serializable. These are documented as "NOT
serializable". ctx/type should morally be serializable and I just
need to setup a contract with subclasses to let them be serialized.
The fake_mode is used solely to test if we are refakefying with
a pre-existing ShapeEnv and we want to reuse the SymInt
directly--serializing this case is hopeless but I am kind of hoping
after this refactor we do not need this at all. view_func is not
serializable because it's a bound C implemented method. Joel has
promised me that this is not too difficult to actually expose as a
true data structure, but this is the edgiest of edge cases and there
is no reason to deal with it right now.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122044
Approved by: https://github.com/eellison
Fixes https://github.com/pytorch/pytorch/issues/121085
This PR pretty involved so pay attention to this description. At a high
level, the refactor is intended to be mechanical: anywhere in
MetaConverter where previously we took a Tensor as argument, we now take
a MetaTensorDesc, which contains all of the information that we would
have queried off of the Tensor, but placed into a separate data
structure which we can serialize or use to recreate a fake tensor in
a separate fake tensor mode in exact fidelity to the original.
However, this transformation is not always entirely mechanical. Here
is what you need to pay attention to:
- The memo table from real Tensor -> meta/fake Tensor is now broken
into two memo tables: real Tensor -> stable int id -> meta/fake
Tensor. The stable int id is needed so that when we do serialization,
we know when tensors/storages alias each other and can ensure we preserve
this aliasing upon deserialization.
The way I have implemented changes the weak reference behavior.
Previously, when either the real Tensor OR the meta/fake Tensor went
dead, we would remove the entry from the memo table. Now, this only
removes entries from one of the two memo tables. This semantically
makes sense, because the user may have held on to the stable int id
out of band, and may expect a real Tensor to continue to be numbered
consistently / expect to be able to lookup a meta/fake tensor from
this id. If this is unacceptable, it may be possible to rejigger
the memo tables so that we have real Tensor -> stable int id
and real Tensor -> meta/fake Tensor, but TBH I find the new
implementation a lot simpler, and arranging the memo tables in this
way means that I have to muck around with the real tensor to save
to the memo table; in the current implementation, I never pass the
Tensor to meta_tensor function AT ALL, which means it is impossible
to accidentally depend on it.
- When I fill in the fields of MetaTensorDesc in describe_tensor, I need
to be careful not to poke fields when they are not valid. Previously,
preconditions were implicitly checked via the conditional structure
("is this sparse? is this nested?") that is tested before we start
reading attributes. This structure has to be replicated in
describe_tensor, and I have almost assuredly gotten it wrong on my
first try (I'll be grinding through it on CI; a careful audit will
help too, by auditing that I've tested all the same conditionals that
the original access was guarded by.)
- I originally submitted https://github.com/pytorch/pytorch/pull/121821
for the symbolic shapes change, but it turned out the way I did it
there didn't actually work so well for this PR. I ended up just
inlining the symbolic shapes allocation logic into MetaConverter
(look for calls to maybe_specialize_sym_int_with_hint), maybe there
is a better way to structure it, but what I really want is to
just read sizes/strides/offset directly off of MetaTensorDesc; I
don't want another intermediate data structure.
- Some fields aren't serializable. These are documented as "NOT
serializable". ctx/type should morally be serializable and I just
need to setup a contract with subclasses to let them be serialized.
The fake_mode is used solely to test if we are refakefying with
a pre-existing ShapeEnv and we want to reuse the SymInt
directly--serializing this case is hopeless but I am kind of hoping
after this refactor we do not need this at all. view_func is not
serializable because it's a bound C implemented method. Joel has
promised me that this is not too difficult to actually expose as a
true data structure, but this is the edgiest of edge cases and there
is no reason to deal with it right now.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122044
Approved by: https://github.com/eellison
ghstack dependencies: #122018
Summary: an fbcode test exposed a shortcoming where we serve a FakeTensor from the cache with the wrong inference_mode. Take the current mode into account in the cache key so we only serve entries from the same mode we're in currently
Test Plan: New unit test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119963
Approved by: https://github.com/eellison
Partially fixes https://github.com/pytorch/pytorch/issues/105077
Repro:
```python
import tempfile
import torch
from torch._subclasses import fake_tensor
class TheModelClass(torch.nn.Module):
def __init__(self):
super(TheModelClass, self).__init__()
self.fc1 = torch.nn.Linear(5, 10)
def forward(self, x):
return self.fc1(x)
with tempfile.NamedTemporaryFile() as state_dict_file:
# Create state_dict to be loaded later
model = TheModelClass()
torch.save(model.state_dict(), state_dict_file.name)
fake_mode = fake_tensor.FakeTensorMode()
with fake_mode:
# This is where the bug is triggered
state_dict = torch.load(state_dict_file.name)
```
Error:
```bash
Traceback (most recent call last):
File "issue_gh_torch_105077.py", line 22, in <module>
state_dict = torch.load(state_dict_file.name)
File "/opt/pytorch/torch/serialization.py", line 1014, in load
return _load(opened_zipfile,
File "/opt/pytorch/torch/serialization.py", line 1422, in _load
result = unpickler.load()
File "/opt/pytorch/torch/_utils.py", line 205, in _rebuild_tensor_v2
tensor = _rebuild_tensor(storage, storage_offset, size, stride)
File "/opt/pytorch/torch/_utils.py", line 184, in _rebuild_tensor
return t.set_(storage._untyped_storage, storage_offset, size, stride)
File "/opt/pytorch/torch/utils/_stats.py", line 20, in wrapper
return fn(*args, **kwargs)
File "/opt/pytorch/torch/_subclasses/fake_tensor.py", line 1288, in __torch_dispatch__
return self.dispatch(func, types, args, kwargs)
File "/opt/pytorch/torch/_subclasses/fake_tensor.py", line 1468, in dispatch
self.invalidate_written_to_constants(func, flat_arg_fake_tensors, args, kwargs)
File "/opt/pytorch/torch/_subclasses/fake_tensor.py", line 1733, in invalidate_written_to_constants
_, new_kwargs = normalize_function(
File "/opt/pytorch/torch/fx/operator_schemas.py", line 297, in normalize_function
torch_op_schemas = get_signature_for_torch_op(target)
File "/opt/pytorch/torch/fx/operator_schemas.py", line 167, in get_signature_for_torch_op
signatures = [_torchscript_schema_to_signature(schema) for schema in schemas]
File "/opt/pytorch/torch/fx/operator_schemas.py", line 167, in <listcomp>
signatures = [_torchscript_schema_to_signature(schema) for schema in schemas]
File "/opt/pytorch/torch/fx/operator_schemas.py", line 70, in _torchscript_schema_to_signature
arg_type = _torchscript_type_to_python_type(arg.type)
File "/opt/pytorch/torch/fx/operator_schemas.py", line 64, in _torchscript_type_to_python_type
return eval(ts_type.annotation_str, _type_eval_globals)
File "<string>", line 1, in <module>
NameError: name 'Storage' is not defined
```
This PR adds the ability to create fake tensors during `torch.load` by wrapping the `torch.tensor.set_` call around a `torch.utils._mode_utils.no_dispatch()` to skip fake mode dispatcher for it and thus create a real tensor. It later calls `fake_mode.from_tensor(t)` to finally create the fake tensor.
Co-authored-by: Edward Z. Yang <ezyang@mit.edu>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108186
Approved by: https://github.com/ezyang
Partially fixes https://github.com/pytorch/pytorch/issues/105077
Repro:
```python
import tempfile
import torch
from torch._subclasses import fake_tensor
class TheModelClass(torch.nn.Module):
def __init__(self):
super(TheModelClass, self).__init__()
self.fc1 = torch.nn.Linear(5, 10)
def forward(self, x):
return self.fc1(x)
with tempfile.NamedTemporaryFile() as state_dict_file:
# Create state_dict to be loaded later
model = TheModelClass()
torch.save(model.state_dict(), state_dict_file.name)
fake_mode = fake_tensor.FakeTensorMode()
with fake_mode:
# This is where the bug is triggered
state_dict = torch.load(state_dict_file.name)
```
Error:
```bash
Traceback (most recent call last):
File "issue_gh_torch_105077.py", line 22, in <module>
state_dict = torch.load(state_dict_file.name)
File "/opt/pytorch/torch/serialization.py", line 1014, in load
return _load(opened_zipfile,
File "/opt/pytorch/torch/serialization.py", line 1422, in _load
result = unpickler.load()
File "/opt/pytorch/torch/_utils.py", line 205, in _rebuild_tensor_v2
tensor = _rebuild_tensor(storage, storage_offset, size, stride)
File "/opt/pytorch/torch/_utils.py", line 184, in _rebuild_tensor
return t.set_(storage._untyped_storage, storage_offset, size, stride)
File "/opt/pytorch/torch/utils/_stats.py", line 20, in wrapper
return fn(*args, **kwargs)
File "/opt/pytorch/torch/_subclasses/fake_tensor.py", line 1288, in __torch_dispatch__
return self.dispatch(func, types, args, kwargs)
File "/opt/pytorch/torch/_subclasses/fake_tensor.py", line 1468, in dispatch
self.invalidate_written_to_constants(func, flat_arg_fake_tensors, args, kwargs)
File "/opt/pytorch/torch/_subclasses/fake_tensor.py", line 1733, in invalidate_written_to_constants
_, new_kwargs = normalize_function(
File "/opt/pytorch/torch/fx/operator_schemas.py", line 297, in normalize_function
torch_op_schemas = get_signature_for_torch_op(target)
File "/opt/pytorch/torch/fx/operator_schemas.py", line 167, in get_signature_for_torch_op
signatures = [_torchscript_schema_to_signature(schema) for schema in schemas]
File "/opt/pytorch/torch/fx/operator_schemas.py", line 167, in <listcomp>
signatures = [_torchscript_schema_to_signature(schema) for schema in schemas]
File "/opt/pytorch/torch/fx/operator_schemas.py", line 70, in _torchscript_schema_to_signature
arg_type = _torchscript_type_to_python_type(arg.type)
File "/opt/pytorch/torch/fx/operator_schemas.py", line 64, in _torchscript_type_to_python_type
return eval(ts_type.annotation_str, _type_eval_globals)
File "<string>", line 1, in <module>
NameError: name 'Storage' is not defined
```
This PR adds the ability to create fake tensors during `torch.load` by wrapping the `torch.tensor.set_` call around a `torch.utils._mode_utils.no_dispatch()` to skip fake mode dispatcher for it and thus create a real tensor. It later calls `fake_mode.from_tensor(t)` to finally create the fake tensor.
Co-authored-by: Edward Z. Yang <ezyang@mit.edu>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108186
Approved by: https://github.com/ezyang
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
Subsumes half of https://github.com/pytorch/pytorch/pull/113605
We support fakeifying an already fake tensor, which will give you a new fake tensor mirroring the same structure as the original fake tensor, which is what is needed by https://github.com/pytorch/pytorch/issues/113643 . However, when this refakeification happens, we will naively reallocate all new sizes for all of the fake tensor. This is the right thing to do if you are re-fakeifying on a fresh ShapeEnv (because you're reparametrizing the sizes or something), but if you have two fake tensor modes which are sharing a shape environment, you would actually rather just reuse the original sizes/strides/offset from the original fake tensor. This ends up being pretty simple. I recommend viewing with whitespace diff turned off.
There's some fuzz around jagged tensor handling; that code is probably not quite right, but I fixed it for this particular case in the most straightforward way.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113651
Approved by: https://github.com/albanD, https://github.com/eellison, https://github.com/bdhirsh
aten.softmax will generate a different decomposition for fp16/bf16 and fp32 because when invoked in lower precision it will upcast the inputs to fp32 and then downcast after. This has been causing us to miss bf16 patterns. For example, Camembert improves 20% with this PR (as do I'm sure many other models).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109142
Approved by: https://github.com/yanboliang
ghstack dependencies: #109663, #108894, #108917
aten.softmax will generate a different decomposition for fp16/bf16 and fp32 because when invoked in lower precision it will upcast the inputs to fp32 and then downcast after. This has been causing us to miss bf16 patterns. For example, Camembert improves 20% with this PR (as do I'm sure many other models).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109142
Approved by: https://github.com/yanboliang
ghstack dependencies: #108894, #108917
# Summary
## PR Dependencies
I don't use ghstack :( this is a PR where it would have been helpful. That beings said I am going to peel off some PRs to make reviewing this easier:
- [x] Separate build flags for Flash and MemEff: #107985
### Description
This pull request updates the version of _scaled_dot_product_flash_attention from version 1 to version 2. The changes are based on the flash attention code originally authored by @tridao
### Changes Made
The majority of the changes in this pull request involve:
- Copying over the flash_attention sources.
- Updating header files.
- Removing padding and slicing code from within the flash_attention kernel and relocating it to the composite implicit region of the SDPA. This was need to make the kernel functional and appease autograd.
- Introducing a simple kernel generator to generate different instantiations of the forward and backward flash templates.
- Adding conditional compilation (ifdef) to prevent building when nvcc is invoked with gencode < sm80.
- Introducing a separate dependent option for mem_eff_attention, as flash_attention v2 lacks support for Windows and cannot be built for sm50 generation codes.
- Modifying build.sh to reduce parallelization on sm86 runners and to lower the maximum parallelization on the manywheel builds. This adjustment was made to address out-of-memory issues during the compilation of FlashAttentionV2 sources.
- Adding/Updating tests.
### Notes for Reviewers
This is not a fun review, and I apologize in advance.
Most of the files-changed are in the flash_attn/ folder. The only files of interest here IMO:
- aten/src/ATen/native/transformers/cuda/flash_attn/flash_api.cpp
- aten/src/ATen/native/transformers/cuda/flash_attn/kernels/generate_kernels.py ( this has been incorporated upstream to flash-attention github)
There are a number of files all related to avoiding OOMs in CI/CD. These are typically shell scripts.
### Follow up items
- Include the updates from e07aa036db and 9e5e8bc91e | https://github.com/pytorch/pytorch/issues/108108
### Work Items
- [x] I don't think Windows will be supported for 3.1.0 - Need to update cmakee
- [x] Let multi_query/attention pass through and test | UPDATE: I have the fast path implemented here: https://github.com/pytorch/pytorch/pull/106730 but since this will require changes to semantics of math to call repeat_interleave, I think this should be done as a followup.
- [x] Had to drop cutlass back to 3.0.0 to get it to compile. Need to figure out how to upgrade to 3.1.0 and later. Spoke with Tri and he is going to be taking a look. Note: compiling with clang currently errors for the cute headers.
- [x] Update test exercise above codepath
- [x] Still need to disable on seq_len % 128 != 0 for backward( Tri beat me to it a4f148b6ab)
- [x] Add determinism warning to BWD, Tri got to this one as well: 1c41d2b
- [x] Update dispatcher to universally prefer FlashV2
- [x] Update tests to exercise new head_dims
- [x] Move the head_dim padding from kernel to top level composite implicit function in order to make it purely functional
- [x] Create template generator script
- [x] Initial cmake support for building kernels/ folder
- [x] Replay CudaGraph changes
### Results
#### Forward only
The TFlops are reported here are on a100 that is underclocked.
#### Forward+Backward
Ran a sweep and for large compute bound sizes we do see a ~2x performance increase for forw+back.
<img width="1684" alt="Screenshot 2023-07-20 at 3 47 47 PM" src="https://github.com/pytorch/pytorch/assets/32754868/fdd26e07-0077-4878-a417-f3a418b6fb3b">
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105602
Approved by: https://github.com/huydhn, https://github.com/cpuhrsch
- Update cross-ref FakeMode test to use ShapeEnv. Dynamic ops can now
return an unbacked SymInt. We always accept this as equal to whatever
the real value was.
- Relax test so it works on all classes, not just unittest.TestCase
- Properly wrap the original method, so things like
pytree.mark.parametrize are carried over
- Support dynamic shapes by default for make_fx `tracing_mode="fake"` without symbolifying everything else
Fixes https://github.com/pytorch/pytorch/issues/108927
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108929
Approved by: https://github.com/zou3519
# Summary
## PR Dependencies
I don't use ghstack :( this is a PR where it would have been helpful. That beings said I am going to peel off some PRs to make reviewing this easier:
- [x] Separate build flags for Flash and MemEff: #107985
### Description
This pull request updates the version of _scaled_dot_product_flash_attention from version 1 to version 2. The changes are based on the flash attention code originally authored by @tridao
### Changes Made
The majority of the changes in this pull request involve:
- Copying over the flash_attention sources.
- Updating header files.
- Removing padding and slicing code from within the flash_attention kernel and relocating it to the composite implicit region of the SDPA. This was need to make the kernel functional and appease autograd.
- Introducing a simple kernel generator to generate different instantiations of the forward and backward flash templates.
- Adding conditional compilation (ifdef) to prevent building when nvcc is invoked with gencode < sm80.
- Introducing a separate dependent option for mem_eff_attention, as flash_attention v2 lacks support for Windows and cannot be built for sm50 generation codes.
- Modifying build.sh to reduce parallelization on sm86 runners and to lower the maximum parallelization on the manywheel builds. This adjustment was made to address out-of-memory issues during the compilation of FlashAttentionV2 sources.
- Adding/Updating tests.
### Notes for Reviewers
This is not a fun review, and I apologize in advance.
Most of the files-changed are in the flash_attn/ folder. The only files of interest here IMO:
- aten/src/ATen/native/transformers/cuda/flash_attn/flash_api.cpp
- aten/src/ATen/native/transformers/cuda/flash_attn/kernels/generate_kernels.py ( this has been incorporated upstream to flash-attention github)
There are a number of files all related to avoiding OOMs in CI/CD. These are typically shell scripts.
### Follow up items
- Include the updates from e07aa036db and 9e5e8bc91e | https://github.com/pytorch/pytorch/issues/108108
### Work Items
- [x] I don't think Windows will be supported for 3.1.0 - Need to update cmakee
- [x] Let multi_query/attention pass through and test | UPDATE: I have the fast path implemented here: https://github.com/pytorch/pytorch/pull/106730 but since this will require changes to semantics of math to call repeat_interleave, I think this should be done as a followup.
- [x] Had to drop cutlass back to 3.0.0 to get it to compile. Need to figure out how to upgrade to 3.1.0 and later. Spoke with Tri and he is going to be taking a look. Note: compiling with clang currently errors for the cute headers.
- [x] Update test exercise above codepath
- [x] Still need to disable on seq_len % 128 != 0 for backward( Tri beat me to it a4f148b6ab)
- [x] Add determinism warning to BWD, Tri got to this one as well: 1c41d2b
- [x] Update dispatcher to universally prefer FlashV2
- [x] Update tests to exercise new head_dims
- [x] Move the head_dim padding from kernel to top level composite implicit function in order to make it purely functional
- [x] Create template generator script
- [x] Initial cmake support for building kernels/ folder
- [x] Replay CudaGraph changes
### Results
#### Forward only
The TFlops are reported here are on a100 that is underclocked.
#### Forward+Backward
Ran a sweep and for large compute bound sizes we do see a ~2x performance increase for forw+back.
<img width="1684" alt="Screenshot 2023-07-20 at 3 47 47 PM" src="https://github.com/pytorch/pytorch/assets/32754868/fdd26e07-0077-4878-a417-f3a418b6fb3b">
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105602
Approved by: https://github.com/huydhn, https://github.com/cpuhrsch
This PR adds dedicated FakeTensor testing to operator_compile_check. We
reuse CrossRefFakeMode to do this and improve the error messages on it.
Note that this only really runs detailed tests for operators that do not
have data-dependent output shape. In the future we should add something
like a dynamic CrossRefFakeMode.
Test Plan:
- existing tests (these now have improved error messages).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103595
Approved by: https://github.com/ezyang, https://github.com/soulitzer
FakeTensor has a default device logic that wraps meta tensors to the right device after running meta kernels and throws on multiple devices. This logic was only running on the wrapping from meta kernels -> fake. For out variants, where the output of the meta kernel was already a fake tensor because it was an input, the device logic wasn't running.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/101807
Approved by: https://github.com/ngimel
This PR:
- adds an abstract registration API for CustomOp (CustomOp.impl_abstract)
that is used for both FakeTensor and meta tensors
- deletes CustomOp.impl_meta
The user story behind this API is that it is the one-stop shop for
registering implementations for data-less Tensors, i.e. FakeTensor and
Meta tensor.
The abstract implementation provided by the user:
- gets registered as the FakeTensor implementation AND the meta formula
- can be written like a regular meta formula. If the user decides that
they need something more special (i.e. data-dependent output shape),
then they are able to query a current context object (FakeTensorImplCtx)
that has methods to construct new unbacked symints.
Caveats:
- we really need to make FakeTensor/FakeTensorMode public. Otherwise,
there isn't a way for the user to interactively test that their abstract
implementation is correct without running through large pieces of the
PT2 stack (make_fx or torch.compile).
- We do not memoize the symints produced by
ctx.create_unbacked_symint(). It is possible to do this in the
future, but it is difficult to do soundly and I am not convinced of
the utility outside of the nonzero() usecase mentioned in #95399
Public API:
- More docs will come when we actually expose this API to users by
putting it in a public namespace, unless you folks want it now.
- The APIs mentioned in `__all__` are the ones that are intended to be
public.
Test Plan:
- Updated existing custom_op_db operators
- Added new numpy_nonzero and numpy_nms operations that test operations
that have data-dependendent output shape.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99439
Approved by: https://github.com/ezyang
I got too confused by the FakeTensor printing, so this PR fixes it to
print normally.
Before:
```
with FakeTensorMode():
x = torch.empty(2, 2, device="cpu")
print(x)
# FakeTensor(FakeTensor(..., device='meta', shape=(2, 2)), cpu)
```
After (Tensor printing doesn't print the default device):
```
FakeTensor(..., shape=(2, 2))
```
Test Plan:
- new test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99205
Approved by: https://github.com/eellison
Original Issue from #92670
pytest ./generated/test_XuyangBai_PointDSC.py -k test_004
==> RuntimeError: as_strided_scatter: sizes [4], strides [85], storage offset 256 and itemsize 4 requiring a storage size of 2048 are out of bounds for storage of size 1024
Repro:
```
class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()
def forward(self, x):
x[1].fill_diagonal_(0) # this check size failed
device = torch.device("cpu")
model = Model()
model.to(device)
torch._dynamo.reset()
compiled_model = torch._dynamo.optimize("inductor")(model)
arg = [torch.rand([4, 1, 1])]
compiled_model(*arg)
```
The error was raised at the checking required size in as_strided_scatter.
https://github.com/pytorch/pytorch/blob/master/torch/_prims/__init__.py#L1818
In the case of input is a tensor with storage offset(a view), when compute input's storage length, should also take input's base tensor's size/stride/offset into account instead of compare it with number of element of input.
This diff fix the bug and add test.
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98483
Approved by: https://github.com/ngimel
Summary: This fixes the case when some of the input tensors were
real tensors and fakified in `validate_and_convert_non_fake_tensors`,
but `flat_arg_fake_tensors` would not contain all the inputs
because it was computed before the fakification. We fix this by
recomputing `flat_arg_fake_tensors` after fakification as well.
Test Plan:
python test/dynamo/test_export.py ExportTests.test_mixed_real_and_fake_inputs
Reviewers: Chillee, voznesenskym
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98769
Approved by: https://github.com/voznesenskym
This was leftover for when we had more logic in the FakeTensor and not FakeTensorMode, and wasn't firing correctly. It also makes more sense for it to be in the other validation function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97186
Approved by: https://github.com/bdhirsh
This was leftover for when we had more logic in the FakeTensor and not FakeTensorMode, and wasn't firing correctly. It also makes more sense for it to be in the other validation function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97186
Approved by: https://github.com/bdhirsh
Fix for https://github.com/pytorch/pytorch/issues/95693.
From https://pytorch.org/tutorials/intermediate/memory_format_tutorial.html:
> There are minor difference between the two APIs to and contiguous. We suggest to stick with to when explicitly converting memory format of tensor.
For general cases the two APIs behave the same. However in special cases for a 4D tensor with size NCHW when either: C==1 or H==1 && W==1, only to would generate a proper stride to represent channels last memory format.
We hit this case in convolution_backward in calling `contiguous()`. Even though we were determining that we should run the backward in channels_last forward, as FakeTensor had gathered from the output of [determine_backend_memory_format](https://github.com/pytorch/pytorch/blob/master/torch/_subclasses/fake_tensor.py#L559), we were still outputting a contiguous tensor. That led to the mismatch in strides in the issue.
Should we be calling `to` instead of `contiguous` more liberally throughout the codebase, especially in convolution related code ? Not sure if there are reasons not to do this.
Another fix would be to update `cudnn_conv_suggest_memory_format` so that it would output a contiguous_format in this case.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96791
Approved by: https://github.com/ngimel
Two small changes that I'm bundling together because one of them needs to touch fbcode and I'm not sure how to do stacked diffs + internal changes + land before release cut.
Remove allow_meta from ctor, and allow by default: we should be able to trace through meta with fake tensors, so in some senses it's a bit weird to expose to user to disallow this. However, it's still useful debug wise to error from time to time, so I've added an option to the config that will get back previous behavior.
Remove `throw_on_data_dependent_ops=True`: this was intended as a temporary behavior as we were smoothing things turning on the erroring. There are no uses anywhere of `throw_on_data_dependent_ops=False` I could find.
These are technically backward-incompatble, but fake tensor is new since the last release / in a private namespace, and I don't want to release it with baggage that would be hard to remove later.
Fix for https://github.com/pytorch/pytorch/issues/92877.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/93993
Approved by: https://github.com/bdhirsh, https://github.com/ezyang
Fixes#90652
Previously, we had assumed that the only way to call `handle_torch_function_no_python_arg_parser` was through the Python key. This is no longer true with FakeTensor. Specifically `_like` functions will call `.device()` on FakeTensors when the args list is being parsed. In order to respect that the mode stack shouldn't run when the python key is off, this just adds that a check that the python key is on/the torch_function equivalent to that function
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91573
Approved by: https://github.com/ezyang
Found this issue from [weekly running 7k github models](https://github.com/pytorch/torchdynamo/issues/1884). This caused regression on pass rate, there are 25 models failed due to this issue.
The reason is argument ```cx``` of ```aten._cudnn_rnn``` can be ```None```, but it doesn't handle well in meta registration, so throws the following error:
```
Traceback (most recent call last):
File "/scratch/ybliang/work/repos/pytorch/torch/_dynamo/utils.py", line 1059, in run_node
return nnmodule(*args, **kwargs)
File "/scratch/ybliang/work/repos/pytorch/torch/nn/modules/module.py", line 1482, in _call_impl
return forward_call(*args, **kwargs)
File "/scratch/ybliang/work/repos/pytorch/torch/nn/modules/rnn.py", line 477, in forward
result = _VF.rnn_tanh(input, hx, self._flat_weights, self.bias, self.num_layers,
File "/scratch/ybliang/work/repos/pytorch/torch/_subclasses/fake_tensor.py", line 916, in __torch_dispatch__
r = func(*args, **kwargs)
File "/scratch/ybliang/work/repos/pytorch/torch/_ops.py", line 284, in __call__
return self._op(*args, **kwargs or {})
File "/scratch/ybliang/work/repos/pytorch/torch/_meta_registrations.py", line 2108, in _cudnn_rnn
cy = cx.new_empty(0 if cx is None else cell_shape)
AttributeError: 'NoneType' object has no attribute 'new_empty'
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91333
Approved by: https://github.com/ezyang
Previously, we planned to lift the parameters and weights while exporting and implement our own transformer to "unlift" the lifted weights and params back to the graph as attributes. But this is bit challenging because:
- We need to maintain correct ordering for weights and parameters that are passed as inputs so that we know how to map them back.
- Some weights are unused in the graph, so our transformer needs to be aware of which weights and parameters are not used in the graph. And we need to distinguish which are real user input and which are parameters.
- There can be more edge cases we haven't seen in other models yet.
I am aware that @Chillee and @bdhirsh mentioned that functionalization won't work with fake-tensor attributes but this is fine for the short term as we don't expect users to be modifying weights and params in inference mode. In fact, we explicitly disable attribute mutation in torchdynamo export mode right now.
Given above condition, it might be ok to just fakify params when we need. I use a flag to guard against this change.
Differential Revision: [D41891201](https://our.internmc.facebook.com/intern/diff/D41891201)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90417
Approved by: https://github.com/eellison
In `FakeTensorMode.__torch_dispatch__`, the output is now always computed by meta kernels in
```python
try:
with in_kernel_invocation_manager(self):
r = func(*args, **kwargs) # <----- "r" can be a real tensor.
except NotImplementedError as not_implemented_error:
# no meta kernel registered, fallback to kernel for the device
if not self.allow_fallback_kernels:
raise not_implemented_error
return run_fallback_kernel(self, func, args, kwargs, not_implemented_error)
return self.wrap_meta_outputs_with_default_device_logic(r, func, args, kwargs)
```
For example, I observed a CPU tensor is generated when executing `aten.addmm` when running `FakeTensorProp`. Therefore, I'd like to allow `FakeTensorMode` to wrap real tensor as `FakeTensor` during the computation. Does this PR look a good direction to fix this problem? If yes, I can go ahead and add some tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88700
Approved by: https://github.com/eellison, https://github.com/ezyang
The logic for determine conv backend and therefore output striding is very complex. It depends on build settings, input striding/contiguity, sizes, etc. Eventually we should port that logic to the meta impl for dynamic shapes but that will require a lot more work and keeping the implementations in sync. See https://github.com/pytorch/torchdynamo/issues/1701
This is a prerequisite to removing the inductor conv stride propagation and more general fake tensor for inductor propagation. In that PR, the meta impls for cpu conv give incorrect striding which led to test failures (https://github.com/pytorch/pytorch/pull/87083).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87305
Approved by: https://github.com/ezyang
