Summary:
`np.asscalar` was deprecated and removed in a recent Numpy. It used to be implemented the following way, and the recommended alternative is to call `item()` directly:
```python
def asscalar(a):
return a.item()
```
This fixes all of the references.
Test Plan: visual inspection and automated tests
Differential Revision: D54697760
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121545
Approved by: https://github.com/malfet
This patch addresses the major limitations in our previous [PR #115981](https://github.com/pytorch/pytorch/pull/115981) through the new dedicated repository [AOTriton](https://github.com/ROCm/aotriton)
- [x] Only supports MI200 series GPU (i.e., `gcnArchName == gfx90a:sramecc+:xnack-`).
* MI300X is supported. More architectures will be added once Triton support them.
- [x] Only supports power of two sequence lengths.
* Now it support arbitrary sequence length
- [ ] No support for varlen APIs.
* varlen API will be supported in the next release of AOTriton
- [x] Only support head dimension 16,32,64,128.
* Now it support arbitrary head dimension <= 256
- [x] Performance is still being optimized.
* Kernel is selected according to autotune information from Triton.
Other improvements from AOTriton include
* Allow more flexible Tensor storage layout
* More flexible API
This is a more extensive fix to #112997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121561
Approved by: https://github.com/malfet, https://github.com/atalman
Summary: The current C shim layer manually implements a C interface for a handful of ops. Obviously that's not scalable if we want to extend it to cover all aten ops. This new torchgen script automatically generates C shim interfaces for CPU and CUDA backends. The interface follows the same parameter passing rules as the current C shim layer, such as
* Use plain C data types to pass parameters
* Use AtenTensorHandle to pass at::Tensor
* Use pointer type to pass optional parameter
* Use pointer+length to pass list
* Use device_type+device_index to pass device
* When a parameter is a pointer of pointer, e.g. AtenTensorHandle**, the script generates either a list of optional values or an optional list of values
https://gist.github.com/desertfire/83701532b126c6d34dae6ba68a1b074a is an example of the generated torch/csrc/inductor/aoti_torch/generated/c_shim_cuda.cpp file. The current version doesn't generate C shim wrappers for all aten ops, and probably generates more wrappers than needed on the other hand, but it should serve as a good basis.
This PR by itself won't change AOTI codegen and thus won't introduce any FC breakage. The actual wrapper codegen changes will come in another PR with some version control flag to avoid FC breakage.
Differential Revision: [D54258087](https://our.internmc.facebook.com/intern/diff/D54258087)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120513
Approved by: https://github.com/jansel
Summary:
`nullptr` is typesafe. `0` and `NULL` are not. In the future, only `nullptr` will be allowed.
This diff helps us embrace the future _now_ in service of enabling `-Wzero-as-null-pointer-constant`.
Test Plan: Sandcastle
Reviewed By: meyering
Differential Revision: D54163060
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120740
Approved by: https://github.com/Skylion007
Fixes#115331.
This PR increases the number of valid GPU devices to 512 (from 64) in order to future-proof PyTorch for providers that offer [single nodes with a large device count](https://www.tensorwave.com/). Until now, `DeviceIndex` was an `int8_t`, thus multiple changes were necessary:
- `DeviceIndex` changed to `int16_t`. Updated consumers that assume it to be an `int8_t`.
- Updated bounds checking for `torch.device()` in the Python frontend. Right now, we allow funny things like `torch.device('cpu', 200).index == -56`, which is undefined behavior. I inserted some checks to only allow values between 0 and `c10::Device::MAX_NUM_DEVICES - 1`.
- Updated the `ArgumentInfo` struct as it hardcodes the device index as 8 bit field [^1]. Might be a breaking change, not sure if users rely on this.
- Introduced `c10::Device::MAX_NUM_DEVICES` as a replacement for the old `C10_COMPILE_TIME_MAX_GPUS`
[^1]: This field was unsigned, so I guess this has also been undef behavior the whole time? Our default device index is -1, so this always wrapped around to 255 when written to the `ArgumentInfo` struct. When I switched the `DeviceIndex` to `int16_t`, it actually stayed 255 after unpacking from `ArgumentInfo` again, as the `DeviceIndex` was now wide enough that it didn't wrap back to -1.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119639
Approved by: https://github.com/cyyever, https://github.com/albanD, https://github.com/huydhn
Fixes#115331.
This PR increases the number of valid GPU devices to 512 (from 64) in order to future-proof PyTorch for providers that offer [single nodes with a large device count](https://www.tensorwave.com/). Until now, `DeviceIndex` was an `int8_t`, thus multiple changes were necessary:
- `DeviceIndex` changed to `int16_t`. Updated consumers that assume it to be an `int8_t`.
- Updated bounds checking for `torch.device()` in the Python frontend. Right now, we allow funny things like `torch.device('cpu', 200).index == -56`, which is undefined behavior. I inserted some checks to only allow values between 0 and `c10::Device::MAX_NUM_DEVICES - 1`.
- Updated the `ArgumentInfo` struct as it hardcodes the device index as 8 bit field [^1]. Might be a breaking change, not sure if users rely on this.
- Introduced `c10::Device::MAX_NUM_DEVICES` as a replacement for the old `C10_COMPILE_TIME_MAX_GPUS`
[^1]: This field was unsigned, so I guess this has also been undef behavior the whole time? Our default device index is -1, so this always wrapped around to 255 when written to the `ArgumentInfo` struct. When I switched the `DeviceIndex` to `int16_t`, it actually stayed 255 after unpacking from `ArgumentInfo` again, as the `DeviceIndex` was now wide enough that it didn't wrap back to -1.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119639
Approved by: https://github.com/cyyever, https://github.com/albanD
Replaces `view_func()` closures with a reified `ViewFunc` data structure. Codegen generates a `ViewFunc` subclass for each view op (e.g. `NarrowViewFunc`) containing state needed to reconstruct the view. The `ViewFunc` API allows for querying and hot-swapping any `SymInt`s or `Tensors` in the state through `get_symints()` / `get_tensors()` / `clone_and_set()`, which will be essential for fake-ification later on.
```cpp
/// Base class for view functions, providing reapplication of a view on a new base.
/// Each view op should get a codegenerated subclass of this class containing
/// any state needed to reconstruct the view. The class also provides convenience
/// accessors for saved SymInts / tensor state. This is useful for e.g. fake-ification,
/// where we want to use symbolic values or fake tensors instead.
struct TORCH_API ViewFunc {
virtual ~ViewFunc() {}
/// Returns any SymInts in the saved state.
virtual std::vector<c10::SymInt> get_symints() const { return {}; }
/// Returns the number of SymInts in the saved state.
virtual size_t num_symints() const { return 0; }
/// Returns any tensors in the saved state.
virtual std::vector<at::Tensor> get_tensors() const { return {}; }
/// Returns the number of tensors in the saved state.
virtual size_t num_tensors() const { return 0; }
/// Reapplies the view on the given base using the saved state.
virtual at::Tensor operator()(const at::Tensor&) const = 0;
/// Returns a clone of this ViewFunc, optionally with the specified saved state.
virtual std::unique_ptr<ViewFunc> clone_and_set(
std::optional<std::vector<c10::SymInt>> = c10::nullopt,
std::optional<std::vector<at::Tensor>> = c10::nullopt) const = 0;
protected:
/// Sets the values of any SymInts in the saved state. The input vector size must
/// match the number of SymInts in the saved state (i.e. the size of the list
/// returned by get_symints()).
virtual void set_symints(std::vector<c10::SymInt>) {}
/// Sets the values of any Tensors in the saved state. The input vector size must
/// match the number of Tensors in the saved state (i.e. the size of the list
/// returned by get_tensors()).
virtual void set_tensors(std::vector<at::Tensor>) {}
};
```
New codegen files:
* `torch/csrc/autograd/generated/ViewFunc.h`
* `torch/csrc/autograd/generated/ViewFuncs.cpp`
The templates for these also contains impls for `ChainedViewFunc` and `ErroringViewFunc` which are used in a few places within autograd.
Example codegen for `slice.Tensor`:
```cpp
// torch/csrc/autograd/generated/ViewFuncs.h
#define SLICE_TENSOR_VIEW_FUNC_AVAILABLE
struct SliceTensorViewFunc : public torch::autograd::ViewFunc {
SliceTensorViewFunc(int64_t dim, c10::optional<c10::SymInt> start, c10::optional<c10::SymInt> end, c10::SymInt step) : dim(dim), start(start), end(end), step(step)
{};
virtual ~SliceTensorViewFunc() override {};
virtual std::vector<c10::SymInt> get_symints() const override;
virtual size_t num_symints() const override;
virtual std::vector<at::Tensor> get_tensors() const override;
virtual size_t num_tensors() const override;
virtual at::Tensor operator()(const at::Tensor&) const override;
virtual std::unique_ptr<ViewFunc> clone_and_set(
std::optional<std::vector<c10::SymInt>> = c10::nullopt,
std::optional<std::vector<at::Tensor>> = c10::nullopt) const override;
protected:
virtual void set_symints(std::vector<c10::SymInt>) override;
virtual void set_tensors(std::vector<at::Tensor>) override;
private:
int64_t dim;
c10::optional<c10::SymInt> start;
c10::optional<c10::SymInt> end;
c10::SymInt step;
};
...
// torch/csrc/autograd/generated/ViewFuncs.cpp
std::vector<c10::SymInt> SliceTensorViewFunc::get_symints() const {
::std::vector<c10::SymInt> symints;
symints.reserve((start.has_value() ? 1 : 0) + (end.has_value() ? 1 : 0) + 1);
if(start.has_value()) symints.insert(symints.end(), *(start));
if(end.has_value()) symints.insert(symints.end(), *(end));
symints.push_back(step);
return symints;
}
size_t SliceTensorViewFunc::num_symints() const {
return static_cast<size_t>((start.has_value() ? 1 : 0) + (end.has_value() ? 1 : 0) + 1);
}
void SliceTensorViewFunc::set_symints(std::vector<c10::SymInt> symints) {
TORCH_INTERNAL_ASSERT(symints.size() == num_symints());
auto i = 0;
if(start.has_value()) start = symints[i];
i += (start.has_value() ? 1 : 0);
if(end.has_value()) end = symints[i];
i += (end.has_value() ? 1 : 0);
step = symints[i];
}
std::vector<at::Tensor> SliceTensorViewFunc::get_tensors() const {
::std::vector<at::Tensor> tensors;
return tensors;
}
size_t SliceTensorViewFunc::num_tensors() const {
return static_cast<size_t>(0);
}
void SliceTensorViewFunc::set_tensors(std::vector<at::Tensor> tensors) {
TORCH_INTERNAL_ASSERT(tensors.size() == num_tensors());
}
at::Tensor SliceTensorViewFunc::operator()(const at::Tensor& input_base) const {
return at::_ops::slice_Tensor::call(input_base, dim, start, end, step);
}
std::unique_ptr<ViewFunc> SliceTensorViewFunc::clone_and_set(
std::optional<std::vector<c10::SymInt>> symints,
std::optional<std::vector<at::Tensor>> tensors) const {
auto output = std::make_unique<SliceTensorViewFunc>(dim, start, end, step);
if (symints.has_value()) {
output->set_symints(std::move(*(symints)));
}
if (tensors.has_value()) {
output->set_tensors(std::move(*(tensors)));
}
return output;
}
```
The `_view_func()` / `_view_func_unsafe()` methods now accept two additional (optional) args for `symint_visitor_fn` / `tensor_visitor_fn`. If these are defined, they are expected to be python callables that operate on a single SymInt / tensor and return a new one. This allows for the hot-swapping needed during fake-ification.
For testing, there are extensive pre-existing tests, and I added a test to ensure that hot-swapping functions correctly.
```sh
python test/test_autograd.py -k test_view_func_replay
python test/test_ops.py -k test_view_replay
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118404
Approved by: https://github.com/ezyang
Summary:
In March 2019 D14468816 introduced some infra to mark tests as flaky
while still running them. In July 2019 D15797371 removed the last use of this
feature. Remove the related code as well.
Test Plan: ci
Reviewed By: mlogachev
Differential Revision: D50601204
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112007
Approved by: https://github.com/malfet
Replaces `view_func()` closures with a reified `ViewFunc` data structure. Codegen generates a `ViewFunc` subclass for each view op (e.g. `NarrowViewFunc`) containing state needed to reconstruct the view. The `ViewFunc` API allows for querying and hot-swapping any `SymInt`s or `Tensors` in the state through `get_symints()` / `get_tensors()` / `clone_and_set()`, which will be essential for fake-ification later on.
```cpp
/// Base class for view functions, providing reapplication of a view on a new base.
/// Each view op should get a codegenerated subclass of this class containing
/// any state needed to reconstruct the view. The class also provides convenience
/// accessors for saved SymInts / tensor state. This is useful for e.g. fake-ification,
/// where we want to use symbolic values or fake tensors instead.
struct TORCH_API ViewFunc {
virtual ~ViewFunc() {}
/// Returns any SymInts in the saved state.
virtual std::vector<c10::SymInt> get_symints() const { return {}; }
/// Returns the number of SymInts in the saved state.
virtual size_t num_symints() const { return 0; }
/// Returns any tensors in the saved state.
virtual std::vector<at::Tensor> get_tensors() const { return {}; }
/// Returns the number of tensors in the saved state.
virtual size_t num_tensors() const { return 0; }
/// Reapplies the view on the given base using the saved state.
virtual at::Tensor operator()(const at::Tensor&) const = 0;
/// Returns a clone of this ViewFunc, optionally with the specified saved state.
virtual std::unique_ptr<ViewFunc> clone_and_set(
std::optional<std::vector<c10::SymInt>> = c10::nullopt,
std::optional<std::vector<at::Tensor>> = c10::nullopt) const = 0;
protected:
/// Sets the values of any SymInts in the saved state. The input vector size must
/// match the number of SymInts in the saved state (i.e. the size of the list
/// returned by get_symints()).
virtual void set_symints(std::vector<c10::SymInt>) {}
/// Sets the values of any Tensors in the saved state. The input vector size must
/// match the number of Tensors in the saved state (i.e. the size of the list
/// returned by get_tensors()).
virtual void set_tensors(std::vector<at::Tensor>) {}
};
```
New codegen files:
* `torch/csrc/autograd/generated/ViewFunc.h`
* `torch/csrc/autograd/generated/ViewFuncs.cpp`
The templates for these also contains impls for `ChainedViewFunc` and `ErroringViewFunc` which are used in a few places within autograd.
Example codegen for `slice.Tensor`:
```cpp
// torch/csrc/autograd/generated/ViewFuncs.h
#define SLICE_TENSOR_VIEW_FUNC_AVAILABLE
struct SliceTensorViewFunc : public torch::autograd::ViewFunc {
SliceTensorViewFunc(int64_t dim, c10::optional<c10::SymInt> start, c10::optional<c10::SymInt> end, c10::SymInt step) : dim(dim), start(start), end(end), step(step)
{};
virtual ~SliceTensorViewFunc() override {};
virtual std::vector<c10::SymInt> get_symints() const override;
virtual size_t num_symints() const override;
virtual std::vector<at::Tensor> get_tensors() const override;
virtual size_t num_tensors() const override;
virtual at::Tensor operator()(const at::Tensor&) const override;
virtual std::unique_ptr<ViewFunc> clone_and_set(
std::optional<std::vector<c10::SymInt>> = c10::nullopt,
std::optional<std::vector<at::Tensor>> = c10::nullopt) const override;
protected:
virtual void set_symints(std::vector<c10::SymInt>) override;
virtual void set_tensors(std::vector<at::Tensor>) override;
private:
int64_t dim;
c10::optional<c10::SymInt> start;
c10::optional<c10::SymInt> end;
c10::SymInt step;
};
...
// torch/csrc/autograd/generated/ViewFuncs.cpp
std::vector<c10::SymInt> SliceTensorViewFunc::get_symints() const {
::std::vector<c10::SymInt> symints;
symints.reserve((start.has_value() ? 1 : 0) + (end.has_value() ? 1 : 0) + 1);
if(start.has_value()) symints.insert(symints.end(), *(start));
if(end.has_value()) symints.insert(symints.end(), *(end));
symints.push_back(step);
return symints;
}
size_t SliceTensorViewFunc::num_symints() const {
return static_cast<size_t>((start.has_value() ? 1 : 0) + (end.has_value() ? 1 : 0) + 1);
}
void SliceTensorViewFunc::set_symints(std::vector<c10::SymInt> symints) {
TORCH_INTERNAL_ASSERT(symints.size() == num_symints());
auto i = 0;
if(start.has_value()) start = symints[i];
i += (start.has_value() ? 1 : 0);
if(end.has_value()) end = symints[i];
i += (end.has_value() ? 1 : 0);
step = symints[i];
}
std::vector<at::Tensor> SliceTensorViewFunc::get_tensors() const {
::std::vector<at::Tensor> tensors;
return tensors;
}
size_t SliceTensorViewFunc::num_tensors() const {
return static_cast<size_t>(0);
}
void SliceTensorViewFunc::set_tensors(std::vector<at::Tensor> tensors) {
TORCH_INTERNAL_ASSERT(tensors.size() == num_tensors());
}
at::Tensor SliceTensorViewFunc::operator()(const at::Tensor& input_base) const {
return at::_ops::slice_Tensor::call(input_base, dim, start, end, step);
}
std::unique_ptr<ViewFunc> SliceTensorViewFunc::clone_and_set(
std::optional<std::vector<c10::SymInt>> symints,
std::optional<std::vector<at::Tensor>> tensors) const {
auto output = std::make_unique<SliceTensorViewFunc>(dim, start, end, step);
if (symints.has_value()) {
output->set_symints(std::move(*(symints)));
}
if (tensors.has_value()) {
output->set_tensors(std::move(*(tensors)));
}
return output;
}
```
The `_view_func()` / `_view_func_unsafe()` methods now accept two additional (optional) args for `symint_visitor_fn` / `tensor_visitor_fn`. If these are defined, they are expected to be python callables that operate on a single SymInt / tensor and return a new one. This allows for the hot-swapping needed during fake-ification.
For testing, there are extensive pre-existing tests, and I added a test to ensure that hot-swapping functions correctly.
```sh
python test/test_autograd.py -k test_view_func_replay
python test/test_ops.py -k test_view_replay
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118404
Approved by: https://github.com/ezyang
# Motivation
According to [[1/4] Intel GPU Runtime Upstreaming for Device](https://github.com/pytorch/pytorch/pull/116019), As mentioned in [[RFC] Intel GPU Runtime Upstreaming](https://github.com/pytorch/pytorch/issues/114842), this third PR covers the changes under `libtorch_python`.
# Design
This PR primarily offers device-related APIs in python frontend, including
- `torch.xpu.is_available`
- `torch.xpu.device_count`
- `torch.xpu.current_device`
- `torch.xpu.set_device`
- `torch.xpu.device`
- `torch.xpu.device_of`
- `torch.xpu.get_device_name`
- `torch.xpu.get_device_capability`
- `torch.xpu.get_device_properties`
- ====================
- `torch.xpu._DeviceGuard`
- `torch.xpu._is_compiled`
- `torch.xpu._get_device`
# Additional Context
We will implement the support of lazy initialization in the next PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116850
Approved by: https://github.com/EikanWang, https://github.com/jgong5, https://github.com/gujinghui, https://github.com/malfet
This is a lot of files changed! Don't panic! Here's how it works:
* Previously, we set `follow_imports = silent` for our mypy.ini configuration. Per https://mypy.readthedocs.io/en/stable/running_mypy.html#follow-imports, what this does is whenever we have an import to a module which is not listed as a file to be typechecked in mypy, we typecheck it as normal but suppress all errors that occurred in that file.
* When mypy is run inside lintrunner, the list of files is precisely the files covered by the glob in lintrunner.toml, but with files in excludes excluded.
* The top-level directive `# mypy: ignore-errors` instructs mypy to typecheck the file as normal, but ignore all errors.
* Therefore, it should be equivalent to set `follow_imports = normal`, if we put `# mypy: ignore-errors` on all files that were previously excluded from the file list.
* Having done this, we can remove the exclude list from .lintrunner.toml, since excluding a file from typechecking is baked into the files themselves.
* torch/_dynamo and torch/_inductor were previously in the exclude list, because they were covered by MYPYINDUCTOR. It is not OK to mark these as `# mypy: ignore-errors` as this will impede typechecking on the alternate configuration. So they are temporarily being checked twice, but I am suppressing the errors in these files as the configurations are not quite the same. I plan to unify the configurations so this is only a temporary state.
* There were some straggler type errors after these changes somehow, so I fixed them as needed. There weren't that many.
In the future, to start type checking a file, just remove the ignore-errors directive from the top of the file.
The codemod was done with this script authored by GPT-4:
```
import glob
exclude_patterns = [
...
]
for pattern in exclude_patterns:
for filepath in glob.glob(pattern, recursive=True):
if filepath.endswith('.py'):
with open(filepath, 'r+') as f:
content = f.read()
f.seek(0, 0)
f.write('# mypy: ignore-errors\n\n' + content)
```
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118414
Approved by: https://github.com/thiagocrepaldi, https://github.com/albanD
Summary:
Our upcoming compiler upgrade will require us not to have shadowed variables. Such variables have a _high_ bug rate and reduce readability, so we would like to avoid them even if the compiler was not forcing us to do so.
This codemod attempts to fix an instance of a shadowed variable. Please review with care: if it's failed the result will be a silent bug.
**What's a shadowed variable?**
Shadowed variables are variables in an inner scope with the same name as another variable in an outer scope. Having the same name for both variables might be semantically correct, but it can make the code confusing to read! It can also hide subtle bugs.
This diff fixes such an issue by renaming the variable.
- If you approve of this diff, please use the "Accept & Ship" button :-)
Test Plan: Sandcastle
Reviewed By: igorsugak
Differential Revision: D52582853
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117996
Approved by: https://github.com/PaliC, https://github.com/kit1980, https://github.com/malfet
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
<!--
copilot:summary
-->
### <samp>🤖 Generated by Copilot at 27084ed</samp>
This pull request simplifies and cleans up the code that uses the cuDNN library for convolution, batch normalization, CTC loss, and quantized operations. It removes the unnecessary checks and conditions for older cuDNN versions and the experimental cuDNN v8 API, and ~~replaces them with the stable `cudnn_frontend` API that requires cuDNN v8 or higher. It also adds the dependency and configuration for the `cudnn_frontend` library in the cmake and bazel files.~~ Correction: The v7 API will still be available with this PR, and can still be used, without any changes to the defaults. This change simply always _builds_ the v8 API, and removes the case where _only_ the v7 API is built.
This is a re-land of https://github.com/pytorch/pytorch/pull/91527
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95722
Approved by: https://github.com/malfet, https://github.com/atalman
## Summary
This PR added 3 intra-node GPU allreduce algorithms to PyTorch:
- One-shot allreduce (inspired by FasterTransformer): all ranks simultaneously read and accumulate data from other ranks.
- Two-shot allreduce (inspired by FasterTransformer): all ranks simultanesouly read and accumulate `1 / world_size` data from other ranks. Then all ranks read accumulated data from other ranks. (effectively one-shot reduce-scatter + one-shot all-gather).
- Hybrid cube mesh allreduce (original): a one-shot allreduce variant that avoids transmission over PCIe on HCM topology.
## Micro Benchmarks
## Details
The intra-node algos are organized behind `c10d::IntraNodeComm`, which is responsible for:
- Managing handshaking and cuda IPC handle exchange among ranks.
- Querying NVLink connection and detecting topology.
- Performing algo selection based on available info.
- Launching the selected allreduce kernel.
`c10d::IntraNodeComm` is integrated into `c10d::ProcessGroupNCCL` as follows:
- When the `ENABLE_INTRA_NODE_COMM` environment variable is set, `c10d::ProcessGroupNCCL` initialize a `c10d::IntraNodeComm` for its ranks.
- If the setup is not suitable for intra-node comm (e.g. not all ranks are from the same node), the rendezvous logic guarantees all participants fall back consistently.
- `c10d::ProcessGroupNCCL::allreduce` consults `c10d::IntraNodeComm` whether to use intra-node allreduce and carries out the communication accordingly.
We currently detect two types of topoloies from the nNVLink connection mesh:
- Fully connected: all GPU pairs has direct NVLink connection (e.g. NVSwitch or fully connected sub-set of hybrid cube mesh)
- `msg <= 256KB`: one-shot allreduce.
- `256KB < msg <= 10MB`: two-shot allreduce.
- `msg > 10MB`: instructs the caller to fallback to NCCL.
- Hybrid cube mesh
- `msg <= 256KB`: one-shot allreduce.
- `msg > 256KB`: instructs the caller to fallback to NCCL.
## Next Steps
- Fine tune algo selection based on GPU model, topology, link speed.
- Potentially optimize the two-shot allreduce impl. Accroding to FasterTransformer, two-shot allreduce is preferred until 50MB. There might be room for improvement, but PyTorch does impose more constraints:
- FasterTransformer uses a single process to drive multiple devices. It can use `cudaDeviceEnablePeerAccess` enable device-level peer access.
- PyTorch uses multiple process to drive multiple devices. With cuda IPC, a device can only share a specific region to other devices. This means extra copies may be unavoidable.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114001
Approved by: https://github.com/yf225
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 https://github.com/pytorch/pytorch/issues/112997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114309
Approved by: https://github.com/jeffdaily, https://github.com/malfet
---------
Co-authored-by: Joseph Groenenboom <joseph.groenenboom@amd.com>
## Summary
This PR added 3 intra-node GPU allreduce algorithms to PyTorch:
- One-shot allreduce (inspired by FasterTransformer): all ranks simultaneously read and accumulate data from other ranks.
- Two-shot allreduce (inspired by FasterTransformer): all ranks simultanesouly read and accumulate `1 / world_size` data from other ranks. Then all ranks read accumulated data from other ranks. (effectively one-shot reduce-scatter + one-shot all-gather).
- Hybrid cube mesh allreduce (original): a one-shot allreduce variant that avoids transmission over PCIe on HCM topology.
## Micro Benchmarks
## Details
The intra-node algos are organized behind `c10d::IntraNodeComm`, which is responsible for:
- Managing handshaking and cuda IPC handle exchange among ranks.
- Querying NVLink connection and detecting topology.
- Performing algo selection based on available info.
- Launching the selected allreduce kernel.
`c10d::IntraNodeComm` is integrated into `c10d::ProcessGroupNCCL` as follows:
- When the `ENABLE_INTRA_NODE_COMM` environment variable is set, `c10d::ProcessGroupNCCL` initialize a `c10d::IntraNodeComm` for its ranks.
- If the setup is not suitable for intra-node comm (e.g. not all ranks are from the same node), the rendezvous logic guarantees all participants fall back consistently.
- `c10d::ProcessGroupNCCL::allreduce` consults `c10d::IntraNodeComm` whether to use intra-node allreduce and carries out the communication accordingly.
We currently detect two types of topoloies from the nNVLink connection mesh:
- Fully connected: all GPU pairs has direct NVLink connection (e.g. NVSwitch or fully connected sub-set of hybrid cube mesh)
- `msg <= 256KB`: one-shot allreduce.
- `256KB < msg <= 10MB`: two-shot allreduce.
- `msg > 10MB`: instructs the caller to fallback to NCCL.
- Hybrid cube mesh
- `msg <= 256KB`: one-shot allreduce.
- `msg > 256KB`: instructs the caller to fallback to NCCL.
## Next Steps
- Fine tune algo selection based on GPU model, topology, link speed.
- Potentially optimize the two-shot allreduce impl. Accroding to FasterTransformer, two-shot allreduce is preferred until 50MB. There might be room for improvement, but PyTorch does impose more constraints:
- FasterTransformer uses a single process to drive multiple devices. It can use `cudaDeviceEnablePeerAccess` enable device-level peer access.
- PyTorch uses multiple process to drive multiple devices. With cuda IPC, a device can only share a specific region to other devices. This means extra copies may be unavoidable.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114001
Approved by: https://github.com/yf225
Related to #103973#110532#108404#94891
**Context:**
As commented in 6ae0554d11/cmake/Dependencies.cmake (L1198)
Kernel asserts are enabled by default for CUDA and disabled for ROCm.
However it is somewhat broken, and Kernel assert was still enabled for ROCm.
Disabling kernel assert is also needed for users who do not have PCIe atomics support. These community users have verified that disabling the kernel assert in PyTorch/ROCm platform fixed their pytorch workflow, like torch.sum script, stable-diffusion. (see the related issues)
**Changes:**
This pull request serves the following purposes:
* Refactor and clean up the logic, make it simpler for ROCm to enable and disable Kernel Asserts
* Fix the bug that Kernel Asserts for ROCm was not disabled by default.
Specifically,
- Renamed `TORCH_DISABLE_GPU_ASSERTS` to `C10_USE_ROCM_KERNEL_ASSERT` for the following reasons:
(1) This variable only applies to ROCm.
(2) The new name is more align with #define CUDA_KERNEL_ASSERT function.
(3) With USE_ in front of the name, we can easily control it with environment variable to turn on and off this feature during build (e.g. `USE_ROCM_KERNEL_ASSERT=1 python setup.py develop` will enable kernel assert for ROCm build).
- Get rid of the `ROCM_FORCE_ENABLE_GPU_ASSERTS' to simplify the logic and make it easier to understand and maintain
- Added `#cmakedefine` to carry over the CMake variable to C++
**Tests:**
(1) build with default mode and verify that USE_ROCM_KERNEL_ASSERT is OFF(0), and kernel assert is disabled:
```
python setup.py develop
```
Verify CMakeCache.txt has correct value.
```
/xxxx/pytorch/build$ grep USE_ROCM_KERNEL_ASSERT CMakeCache.txt
USE_ROCM_KERNEL_ASSERT:BOOL=0
```
Tested the following code in ROCm build and CUDA build, and expected the return code differently.
```
subprocess.call([sys.executable, '-c', "import torch;torch._assert_async(torch.tensor(0,device='cuda'));torch.cuda.synchronize()"])
```
This piece of code is adapted from below unit test to get around the limitation that this unit test now was skipped for ROCm. (We will check to enable this unit test in the future)
```
python test/test_cuda_expandable_segments.py -k test_fixed_cuda_assert_async
```
Ran the following script, expecting r ==0 since the CUDA_KERNEL_ASSERT is defined as nothing:
```
>> import sys
>>> import subprocess
>>> r=subprocess.call([sys.executable, '-c', "import torch;torch._assert_async(torch.tensor(0,device='cuda'));torch.cuda.synchronize()"])
>>> r
0
```
(2) Enable the kernel assert by building with USE_ROCM_KERNEL_ASSERT=1, or USE_ROCM_KERNEL_ASSERT=ON
```
USE_ROCM_KERNEL_ASSERT=1 python setup.py develop
```
Verify `USE_ROCM_KERNEL_ASSERT` is `1`
```
/xxxx/pytorch/build$ grep USE_ROCM_KERNEL_ASSERT CMakeCache.txt
USE_ROCM_KERNEL_ASSERT:BOOL=1
```
Run the assert test, and expected return code not equal to 0.
```
>> import sys
>>> import subprocess
>>> r=subprocess.call([sys.executable, '-c', "import torch;torch._assert_async(torch.tensor(0,device='cuda'));torch.cuda.synchronize()"])
>>>/xxxx/pytorch/aten/src/ATen/native/hip/TensorCompare.hip:108: _assert_async_cuda_kernel: Device-side assertion `input[0] != 0' failed.
:0:rocdevice.cpp :2690: 2435301199202 us: [pid:206019 tid:0x7f6cf0a77700] Callback: Queue 0x7f64e8400000 aborting with error : HSA_STATUS_ERROR_EXCEPTION: An HSAIL operation resulted in a hardware exception. code: 0x1016
>>> r
-6
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114660
Approved by: https://github.com/jeffdaily, https://github.com/malfet, https://github.com/jithunnair-amd
