Prior to this patch, Dynamo conveniently modelled torch profiler context
objects (e.g., `torch.profiler.profile`) as `NullContextVariable`
because `torch.compile` ignore the effect of these profiler contexts.
However, the semantics of these profiler contexts diverges from
`contextlib.nullcontext` in the `__enter__` function, where the former
returns `self` and the latter returns `None`. This causes subtle error
as observed in #125021.
This patch adds back a `ProfilerContextVariable`, which addresses the
aforementioned semantic discrepency.
Fixes#125021.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145537
Approved by: https://github.com/zou3519, https://github.com/williamwen42
Prior to this patch, the `test_cuda_event_created_outside_of_graph`
is flaky in CI, and that's because we have read and write to the same
`foo` tensor buffer from 2 different streams. This patch eliminates that
by adding a synchronization to wait till read finishes before starting
the write.
Fixes#133837, #133828.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145658
Approved by: https://github.com/yifuwang
Fixes#130559
* Intro
This PR adds support for `@contextmanager` in Dynamo. We chose to limit the
scope of this work to only `@contextmanager` and plan to handle generators fully
in #141055 (still in draft).
* Motivation
Dynamo lacks support for generator functions. When it encounters one, it traces
it as if it were a regular function. This is problematic because it can lead to
incorrect behavior. To illustrate, consider the test case below:
```python
import torch
import contextlib
@contextlib.contextmanager
def set_default_dtype(dtype):
old_dtype = torch.get_default_dtype()
try:
torch.set_default_dtype(dtype)
yield
finally:
torch.set_default_dtype(old_dtype)
@torch.compile(backend="eager", fullgraph=True)
def fn():
with set_default_dtype(torch.float64):
x = torch.tensor([3.0, 3.0 + 5.0j])
return x
```
Before this work, Dynamo would not stop at the `yield`, and the graph produced
would contain both calls to `set_default_dtype` executed one after the other.
This is incorrect because the context manager should execute code before and
after the `yield`.
* List of changes
`YIELD_VALUE` now raises an exception (`YieldValueOp`) to signal that control
flow must be suspended and returned to the caller. Additionally, `RETURN_VALUE`
behaves differently in a generator function. Unlike regular functions, where
`RETURN_VALUE` indicates the final result, in generators it signifies that the
generator is exhausted and implicitly raises `StopIteration`.
A new `VariableTracker` named `FunctionDecoratedByContextlibContextManagerVariable`
was introduced to handle `@contextmanager`. This variable tracker acts not just
as a wrapper for the original function but also maintains an internal `tx`
(InstructionTranslator) object to suspend and return control flow to the parent
tracer when a `yield` is encountered.
* Corner cases
Returning a context manager from a compiled function is not supported. This
would require PyTorch to synchronize the generator state between Dynamo and the
interpreter. Any attempt to return it will result in an `IncorrectUsage`
exception.
Graph breaks require special handling as well. In the event of a graph break,
the frame associated with the context manager is skipped, and the context
manager runs in eager mode.
* This PR is breaking my code
There is a configuration flag (`enable_trace_contextlib`) that can be set to
`False` to disable tracing context managers. If this still causes crashes,
please revert this PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136033
Approved by: https://github.com/zou3519
`torch.cuda.Event` objects are different from `torch.cuda.Stream` in that events are not pooled, meaning we can't look up a previously created CUDA event object by ID. This prevents CUDA event object created outside of the Dynamo graph from being used within the graph (since Dynamo needs a way to emit a `call_function` line in the graph that does the retrieval of the event object for downstream op use). This PR adds a simple object pool within Dynamo utility, to support looking up CUDA event object by ID from within the Dynamo graph.
After this PR, if a user creates a CUDA event object outside of the graph and use that event within the graph, the behavior will exactly match eager.
Test commands:
- `pytest -rA test/dynamo/test_ctx_manager.py::CtxManagerTests::test_cuda_event_created_outside_of_graph`
- `pytest -rA test/dynamo/test_ctx_manager.py::CtxManagerTests::test_cuda_event_across_graph_break`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133635
Approved by: https://github.com/yifuwang
ghstack dependencies: #133532, #133531, #133636
Fixes#128059
I'm not sure if this is the right way, since Inductor doesn't always respect the device id set by users, so probably we should just wrap it as null context manager and print a warning. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @jansel @anijain2305 @mlazos @williamwen42
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133385
Approved by: https://github.com/jansel
Fix https://github.com/pytorch/pytorch/issues/124900.
When we reconstruct `ContextWrappingVariables`s, we only reconstruct the context class, not the object. Normally, contexts are active (via `with ctx:`) and we initialize the context object in the resume function. But for the case of inactive contexts (contexts declared ahead of time before the `with` block), we do not reconstruct them properly in the optimized bytecode or resume function. So this PR adds initialization for inactive contexts in the resume function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125203
Approved by: https://github.com/jansel
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
Before the pr, we have a graph break for:
```python
def f():
if torch.cuda.current_stream() is not None:
return torch.randn(2, 2)
torch.compile(f, backend="eager", fullgraph=True)()
```
This pr supports comparson ops of StreamVariable and ConstantVariable by returning a constant.
It's safe to return a constant in this case becuase the StreamVariable is guarded by ID_MATCH when created.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119199
Approved by: https://github.com/yifuwang, https://github.com/anijain2305, https://github.com/jansel
This is a placeholder implementation for reconstructing streams via global storage to unblock FSDP, pending proper stream support design
This PR does a few things:
1) fixes registration for devices with indices. We were only supporting "cuda", we now support "cuda:k" interfaces where k is # of gpu
2) Changes the stream objects in dynamo to take devices as device types, instead of strings, and updates the string based device APIs to gracefully take device types.
3) Introduces a reconstruct-by-global (using existing cleanup hook structures) to streams as a placeholder impl for now
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117386
Approved by: https://github.com/jansel
Note about the Updates:
This PR:
1. skips more flash attention related UTs on MI200
2. Fix additional ATen compiling errors after hipification
3. Fix the author "root" of a specific commit
4. Includes the patch from Nikita in favor of block level static initialization.
CAVEAT: This revised PR has a commit that modifies the CI to force its running on MI200 nodes. That specific commit must be reverted before merge.
Original PR (https://github.com/pytorch/pytorch/pull/114309) Note:
This pull requests add initial Flash Attention support for AMD/ROCM platform. It added a specialized Triton repository/branch as a compile-time dependency for Flash Attention math library on AMD/ROCM. This triton submodule is not used at runtime and will not be shipped to the final pytorch package. We have the plan to release this specialized Triton as a separate project.
Know limitations:
- Only supports MI200 series GPU (i.e., `gcnArchName == gfx90a:sramecc+:xnack-`.
- Only supports power of two sequence lengths.
- No support for varlen APIs.
- Only support head dimension 16,32,64,128.
- Performance is still being optimized.
Fixes#112997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115981
Approved by: https://github.com/malfet
This PR implements 2 things:
1. support the device agnostic stream and runtime APIs captured by the dynamo.
2. support the stream methods(include the event) captured by the dynamo.
Here are details for 1st.
Previously the stream captured in dynamo was tightly bind to CUDA. Here we implement a global singleton container named `StreamMethodContainer` for different backends to register their associated stream methods to dynamo. When import the backend’s product, the stream operations can be registered directly by calling
```
device_stream_method = {'current_stream': method_1,
'create_stream_context': method_2,
'set_stream': method_3,
'set_stream_by_id': method_4}
torch._dynamo.stream.register_stream_method(device_name, device_stream_method)
```
Stream methods need to be passed in this API according to the precise semantics represented by the dict key in `device_stream_method`. After register, these methods can be used by dynamo to capture the stream operations in users’ script, for example, get the current stream or set the specific stream. Additionally, the wrapped stream variable and the stream context variable are changed to be the device-agnostic, the proxy functions of these variables are assigned by the associated methods in the container. All of this are illustrated in the below. Below is a illustration.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108312
Approved by: https://github.com/jansel, https://github.com/jgong5
# 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
# 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
```python
def wrapper_fn(x):
with torch.autograd.graph.disable_saved_tensors_hooks("ERROR"):
y = x + 1
print("HI")
return y + 2
x = torch.randn(())
a = wrapper_fn(x)
opt = torch.compile(wrapper_fn, backend='eager', fullgraph=False)
e = opt(x)
```
Without the fix fails with,
```
Traceback (most recent call last):
File "/home/kshiteej/Pytorch/pytorch_functorch/test/test_trace_grad.py", line 182, in <module>
e = opt(x)
File "/home/kshiteej/Pytorch/pytorch_functorch/torch/_dynamo/eval_frame.py", line 333, in _fn
return fn(*args, **kwargs)
File "/home/kshiteej/Pytorch/pytorch_functorch/test/test_trace_grad.py", line 165, in wrapper_fn
def wrapper_fn(x):
AttributeError: module 'torch.autograd.graph' has no attribute 'disable_saved_tensors_hook'
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106875
Approved by: https://github.com/zou3519
