Expose a set of observability hooks into C10D such that our users can
detect collectives failure both faster and more easily.
The design is similar to NCCL desync debug that it minimized the
overhead by doing most of the work out of the main thread.
This PR introduces a new module torch.distributed.hooks that exposes the following set of methods:
register_collective_start_hook
register_collective_end_hook
register_process_group_hook
The process group hook exposes PG creation on the member ranks and call them inline from the
the PG creation code. This is fine since this happens during initialization and a limited number of times.
The collective start/end hooks are fired from a single background thread. It reads
events from a C++ queue and dispatches over.
Queue notification is oddly done using a pipe, this is needed so python can abort the thread on shutdown
and have it as background thread. This is not possible with more reasonable choices like a condvar.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108815
Approved by: https://github.com/wconstab, https://github.com/fduwjj
We want users to be able to define custom ops in C++ but put the
abstract impl in Python (since it is easier to write them in Python and
the abstract impl better models device semantics and data-dependent
operators).
`m.impl_abstract_pystub(opname, python_module, context)` declares the
abstract_impl of the operator to exist in the given python module.
When the abstract_impl needs to be accessed (either via FakeTensor or
Meta), and it does not exist, the PyTorch Dispatcher will yell
with a descriptive error message.
Some details:
- We construct a new global AbstractImplPyStub mapping in
Dispatcher.cpp. Read/write to this map is protected by the Dispatcher
lock.
- We add a new Meta Tensor fallback kernel. The fallback errors out if there is
no meta kernel, but also offers a nicer error message if we see that there is
a pystub.
- We create a `torch._utils_internal.throw_abstract_impl_not_imported_error`
helper function to throw errors. This way, we can throw different error
messages in OSS PyTorch vs internal PyTorch. To invoke this from C++, we
added a PyInterpreter::throw_abstract_impl_not_imported_error.
Differential Revision: [D49464753](https://our.internmc.facebook.com/intern/diff/D49464753/)
Differential Revision: [D49464753](https://our.internmc.facebook.com/intern/diff/D49464753)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109529
Approved by: https://github.com/ezyang, https://github.com/bdhirsh
Summary:
This PR adds a limited C shim layer for libtorch. The ultimate goal is to ban any direct reference to aten/c10 data structures or functions, to avoid ABI breakage by providing stable C interfaces.
To make the review and landing easier, we broke the changes into several steps. In this PR (a combination of https://github.com/pytorch/pytorch/pull/109022 and https://github.com/pytorch/pytorch/pull/109351), we add C interfaces for certain libtorch functions and modify the wrapper codegen to generate calls to those interfaces. There are a few other items to be addressed in future PRs:
* The AOTInductor runtime interface still takes lists of aten tensors as input and output
* The interaction with ProxyExecutor (general fallback support) needs to move away from aten tensor
* Remove all references to aten/c10 headers in the AOTInductor-generated code
Differential Revision: D49302669
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109391
Approved by: https://github.com/chenyang78
# 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
**Background**: "TorchDynamo Cache Lookup" events appear in traces to indicate a dynamo cache lookup; it's useful to check when cache lookups are taking a long time. To add a profiler event, one can use the `torch.profiler.record_function` context manager, or the C++ equivalent. Previously, the python version was used; first, when the profiler was enabled, callbacks for record_function_enter and record_function_exit were registered; then those would be called before and after every cache lookup.
**This PR**: Instead of calling the python bindings for `torch.profiler.record_function`, directly call the C++ implementation. This simplifies a lot of the code for binding C/C++. It also improves performance; previously there was a lot of overhead in the "TorchDynamo Cache Lookup" event, making the event artificially take a long time. After this change the events now appear shorter, because there's less overhead in starting/stopping the event: in other words, the profiler no longer distorts the results as much.
**Performance results**:
I ran using the script below on a cpu-only 1.6GHz machine. I report the median time (from 100 measurements) of a "TorchDynamo Cache Lookup" event before and after this PR. I think it is reasonable to consider the difference to be due to a reduction in overhead.
<details>
<summary>Benchmarking script</summary>
```python
def fn(x, y):
return (x * y).relu()
a, b = [torch.rand((4, 4), requires_grad=True) for _ in range(2)]
opt_fn = torch.compile(fn)
opt_fn(a, b)
opt_fn(a, b)
with torch.profiler.profile() as prof:
opt_fn(a, b)
```
</details>
Median before PR: 198-228 us (median of 100, measured 5 times)
Median after PR: 27us
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108436
Approved by: https://github.com/anijain2305, https://github.com/jansel
# 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
We have a plethora of error types for various errors raised from c10d. These include `RuntimeError`, `TimeoutError`, `SocketError`, `DistBackendError` etc.
This results in messy code during error handling somewhat like this:
```
if "NCCL" in exception_str:
...
if "Timed out initializing process group in store based barrier on rank" in exception_str:
...
if "The client socket has timed out after" in exception_str:
...
if "Broken pipe" in exception_str:
...
if "Connection reset by peer" in exception_str:
...
```
To address this issue, in this PR I've ensured added these error types:
1. **DistError** - the base type of all distributed errors
2. **DistBackendError** - this already existed and referred to PG backend errors
3. **DistStoreError** - for errors originating from the store
4. **DistNetworkError** - for general network errors coming from the socket library
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108191
Approved by: https://github.com/H-Huang
We have a plethora of error types for various errors raised from c10d. These include `RuntimeError`, `TimeoutError`, `SocketError`, `DistBackendError` etc.
This results in messy code during error handling somewhat like this:
```
if "NCCL" in exception_str:
...
if "Timed out initializing process group in store based barrier on rank" in exception_str:
...
if "The client socket has timed out after" in exception_str:
...
if "Broken pipe" in exception_str:
...
if "Connection reset by peer" in exception_str:
...
```
To address this issue, in this PR I've ensured added these error types:
1. **DistError** - the base type of all distributed errors
2. **DistBackendError** - this already existed and referred to PG backend errors
3. **DistStoreError** - for errors originating from the store
4. **DistNetworkError** - for general network errors coming from the socket library
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107651
Approved by: https://github.com/H-Huang
Feature RFC: https://github.com/pytorch/rfcs/pull/56.
The flash attention CPU kernel is added, for forward path FP32. Blocking is applied on dimensions of query length and kv length and the fusion of gemm + softmax update + gemm is done at once for each block. Parallelization is on the dimensions of batch size, head number and query length. In addition, the causal attention mask is supported. As the attention is masked for the unseen tokens, early termination is applied and we only calculate the blocks in the lower triangular part.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103826
Approved by: https://github.com/drisspg, https://github.com/jgong5
ghstack dependencies: #104583, #104584
Fixes #ISSUE_NUMBER
as the title, add context support for custom device and testcase.
And in the future, we may want to refactor these hooks for different device to unify the APIs, would you agree my
idea？ @albanD
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105056
Approved by: https://github.com/albanD
This branch:
1) converts the autograd tape into an FX graph
2) caches that conversion using a "shadow" graph
3) compiles and runs the generated FX graph instead of the normal autograd
What works currently:
1) Caching, capture, and initial integration
2) Backwards hooks
3) Inlining AotAutograd generated subgraphs
4) torch.compiling the generated FX graph
5) Auto-detecting dynamic shapes based on changes
Future work
1) Larger scale testing
1) Boxed calling convention, so memory can be freed incrementally
1) Support hooks on SavedTensor
1) Additional testing by running eager autograd tests under compiled_autograd.enable()
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103822
Approved by: https://github.com/ezyang, https://github.com/albanD
**Summary**
- Update the quantization document that default qconfig with oneDNN backend is recommended to be used on CPUs with Vector Neural Network Instruction support.
- Add the warning message when user uses default qconfig with oneDNN backend on CPU without Vector Neural Network Instruction support.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103653
Approved by: https://github.com/jgong5, https://github.com/malfet
At high current implementation of constrains functions (constrain_as_**) will raise exception for the following code snippets:
```
def f(x):
a = x.item()
constrain_as_size(a, 4, 7)
return torch.empty((a, 4))
inp = torch.tensor([5])
ep = torch._export.export(f, (inp,))
```
The reason is because current constrain logic is:
1) Purely python so it won't survive AOT export (the full node is gone after AOT export since AOT export only maintains aten level op).
2) Utilize side effect to add range constraints for traced symbol's shape env ([code](9591e52880/torch/fx/experimental/symbolic_shapes.py (L370-L372))).
3) If runtime assertion is turned on (by default). [`_AddRuntimeAssertionsForConstraintsPass`](9591e52880/torch/_export/passes/add_runtime_assertions_for_constraints_pass.py (L98-L100)) will try to append assertion node based on range constrains extracted from shape env of symbol during another interpretation round.
4). However, since 1), in the round of AOT export, range constraints logic won't run for symbols generated during this round. And later there is no range constrains information available for assertion round and caused issue.
5) As a result of above, it will failure at `torch.empty((a, 4))` (there is no constrains for `a` that it must be positive).
The fix here is just to implement range constrain logic as a native aten op (CPU implementation as no-op) to make it be able to survive AOT export.
**NOTE:**
[Logic](2d745b95d7/torch/fx/experimental/symbolic_shapes.py (L350-L365C15)) within [`constrain_range`](2d745b95d7/torch/fx/experimental/symbolic_shapes.py (LL313C74-L313C74)) is split out as `constrain_range_int` to capture case when non `SymInt` is passed in and reused in the new `_constrain_range`. The reason is when non `SymInt` is provided:
* If it directly calls `sym_constrain_range`, the C++ version will be called which will be no-op.
* So in this case it calls `constrain_range_int` instead to be able to capture issue like user provides a input whose tensor's shape could be out of range during exporting, like the following for above code example:
```
...
inp = torch.tensor([10])
ep = torch._export.export(f, (inp,)) # immediately raise error
```
Differential Revision: [D46734204](https://our.internmc.facebook.com/intern/diff/D46734204)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103346
Approved by: https://github.com/tugsbayasgalan
Now, when you do an inplace mutation and the view is naughty, you get this message:
```
RuntimeError: A view was created in no_grad mode and is being modified inplace with grad mode enabled. Given that this use case is ambiguous and error-prone, it is forbidden. You can clarify your code by moving both the view and the inplace either both inside the no_grad block (if you don't want the inplace to be tracked) or both outside (if you want the inplace to be tracked). To find out where this view was allocated, run your entire forward region under anomaly mode (torch.autograd.detect_anomaly(check_nan=False)).
```
When you run under anomaly mode, you get:
```
RuntimeError: A view was created in no_grad mode and is being modified inplace with grad mode enabled. Given that this use case is ambiguous and error-prone, it is forbidden. You can clarify your code by moving both the view and the inplace either both inside the no_grad block (if you don't want the inplace to be tracked) or both outside (if you want the inplace to be tracked). This view was allocated at:
File "/data/users/ezyang/c/pytorch/test/test_autograd.py", line 4299, in arglebargle
File "/data/users/ezyang/c/pytorch/test/test_autograd.py", line 4306, in test_anomaly_gives_view_stack
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/case.py", line 549, in _callTestMethod
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/case.py", line 591, in run
File "/data/users/ezyang/c/pytorch/torch/testing/_internal/common_utils.py", line 2266, in _run_with_retry
File "/data/users/ezyang/c/pytorch/torch/testing/_internal/common_utils.py", line 2337, in run
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/case.py", line 650, in __call__
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/suite.py", line 122, in run
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/suite.py", line 84, in __call__
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/suite.py", line 122, in run
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/suite.py", line 84, in __call__
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/runner.py", line 184, in run
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/main.py", line 271, in runTests
File "/home/ezyang/local/c/pytorch-env/lib/python3.10/unittest/main.py", line 101, in __init__
File "/data/users/ezyang/c/pytorch/torch/testing/_internal/common_utils.py", line 894, in run_tests
File "/data/users/ezyang/c/pytorch/test/test_autograd.py", line 11209, in <module>
```
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103185
Approved by: https://github.com/zdevito
In this stack of PRs we adding caching to output tensors for cudagraph trees after we've done initial recording. On initial recording we do not cache tensor outputs because this prevents memory from being reclaimed. On subsequent exeuctions we do cache them to avoid overhead. However, because there is an extra reference around, this caused divergent recording & execution behavior in both autocast caching and autograd gradient stealing. Divergent recording & execution would keep on re-recording and eventually stabilize, but it's not what you want to see happen.
This pr makes the autocast cache and buffer stealing aware of the cudagraph static output tensors.
I will add this to the other cudagraph impl in another pr.
Not sure if this should be in autograd or in autocast since it affects both.. Or somewhere else
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99368
Approved by: https://github.com/albanD, https://github.com/ezyang
