# Motivation
According to [[1/2] Intel GPU Runtime Upstreaming for Stream](https://github.com/pytorch/pytorch/pull/117611), as mentioned in [[RFC] Intel GPU Runtime Upstreaming](https://github.com/pytorch/pytorch/issues/114842), the second PR covers the changes under `python frontend`.
# Design
Currently, it primarily offers stream-related APIs, including
- `torch.xpu.StreamContext`
- `torch.xpu.current_stream`
- `torch.xpu.set_stream`
- `torch.xpu.synchronize`
- `torch._C._xpu_getCurrentRawStream`
# Additional Context
We will implement functions like `torch.xpu.Stream.wait_event`, `torch.xpu.Stream.wait_stream`, and `torch.xpu.Stream.record_event` in the next PR related with `Event`.
The differences with CUDA:
no default and external stream in XPU and lack of below APIs:
- `torch.cuda.ExternalStream`
- `torch.cuda.default_stream`
- `toch.cuda.is_current_stream_capturing`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117619
Approved by: https://github.com/EikanWang, https://github.com/jgong5, https://github.com/gujinghui, https://github.com/albanD
ghstack dependencies: #117611
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
This PR intends to extend `cuda_lazy_init` to `device_lazy_init` which is a device-agnostic API that can support any backend. And change `maybe_initialize_cuda` to `maybe_initialize_device` to support lazy initialization for CUDA while maintaining scalability.
# Design
We maintain a flag for each backend to manage the lazy initialization state separately.
# Additional Context
No need more UTs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118846
Approved by: https://github.com/malfet
# 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
Summary: Now we can allocate an AOTIModelContainerRunner object instead of relying on torch.utils.cpp_extension.load_inline. Also renamed AOTInductorModelRunner to AOTIRunnerUtil in this PR.
Test Plan: CI
Reviewed By: khabinov
Differential Revision: D52339116
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116269
Approved by: https://github.com/khabinov
## 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
## 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 PR introduces a native version of c10d_functional ops. The main goal is to add collective support in AOTInductor and allow collective ops to work in multi-threaded native runtimes.
The native version also incorporated API improvements we wished to implement in Python c10d_functional:
- Removed `ranks` and `group_size` from collective op signatures which were proven to be redundant.
- Use tensor storage as opposed to `void*` to resolve in-flight work.
The native process group registration/resolution mechansim is only used for native c10d_functional in the PR. It will become the single source of truth in upcoming PRs.
The upcoming PRs will implement Inductor/AOTInductor support for c10d_functional, after which native c10d_functional will replace Python c10d_functional.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110570
Approved by: https://github.com/wanchaol
Removes the existing integration code & build of nvfuser in TorchScript.
Note that I intentionally left the part where we wipe out `third_party/nvfuser` repo. I'll do that in a separate PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111093
Approved by: https://github.com/albanD
This reverts commit 314a502eb04c6382e2cc9af0573533efba54109d.
Changes since original PR:
Reland 1
* rename torch.distributed.hooks to torch.distributed._hooks
Reland 2
* make _hooks importable even if !distributed.is_available()
* handle cuda driver exit intermittent failure caused by new cuda api usage in callback caller (see prev PR in stack)
(original PR https://github.com/pytorch/pytorch/pull/108815 desc copied below)
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/111072
Approved by: https://github.com/malfet
ghstack dependencies: #111061
`libshm.so` depends on the torch library exclusively for `at::RefcountedMapAllocator`,
so it makes sense to move it to c10 along with the other memory allocators.
This means `libshm.so` only depends on `c10` and we don't need to relink
`libshm.so` for every ATen change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109881
Approved by: https://github.com/albanD
Summary: Introduce a utility class AOTIModelRunner to take care of running an AOTInductor compiled model. It does things like dlopen a model, initialize the model container, setup inputs and outputs, and destroy the model container.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110891
Approved by: https://github.com/chenyang78
ghstack dependencies: #110652
This reverts commit ff0358b0384d6a3a5b8ceeae625c93221612ba8e.
(original PR https://github.com/pytorch/pytorch/pull/108815 desc copied below)
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/110907
Approved by: https://github.com/fduwjj
`libshm.so` depends on the torch library exclusively for `at::RefcountedMapAllocator`,
so it makes sense to move it to c10 along with the other memory allocators.
This means `libshm.so` only depends on `c10` and we don't need to relink
`libshm.so` for every ATen change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109881
Approved by: https://github.com/albanD
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
