Reland of #160532
Summary:
To support exporting a cuda model on a CPU-only machine under fake tensor mode.
User commonly need to move sample inputs to the cuda device with .to("cuda:0") or .to("cuda") call.
This diff supports this.
I expect the following pattern to work
```
with FakeTensorMode(allow_non_fake_inputs=True):
cuda_module = module.to("cuda:0")
cuda_sample_inputs = tuple([x.to("cuda:0") for x in sample_inputs])
with torch.no_grad():
ep = torch.export.export(cuda_module, cuda_sample_inputs)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/163016
Approved by: https://github.com/huydhn
Summary:
To support exporting a cuda model on a CPU-only machine under fake tensor mode.
User commonly need to move sample inputs to the cuda device with .to("cuda:0") or .to("cuda") call.
This diff supports this.
I expect the following pattern to work
```
with FakeTensorMode(allow_non_fake_inputs=True):
cuda_module = module.to("cuda:0")
cuda_sample_inputs = tuple([x.to("cuda:0") for x in sample_inputs])
with torch.no_grad():
ep = torch.export.export(cuda_module, cuda_sample_inputs)
```
Test Plan:
CI
Rollback Plan:
Differential Revision: D80181887
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160532
Approved by: https://github.com/henryoier, https://github.com/ezyang
Summary:
Add _package_executorch_files to archive apis. Allow us to package a PTE file into the archive.
I don't think there's a use-case to have more than one PTE file at the moment, but left it as `EXECUTORCH_FILES` just in case.
Test Plan:
Tested in D81992612
Rollback Plan:
Differential Revision: D81977483
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162520
Approved by: https://github.com/angelayi
`vmap(F.embedding)(DTensor, DTensor)` was failing because F.embedding's
batching rule generates a new tensor via at::arange, at::arange
generates a regular tensor, and DTensor rightfully errors on mixed
DTensor-regular Tensor operations.
This PR fixes the problem by activating DTensor implicit replication on
just the at::arange and the subsequent add operation.
In order to accomplish this I move the DTensor implicit replication flag
to C++ (most batching rules are in C++).
Test Plan:
- new test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162117
Approved by: https://github.com/bdhirsh
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
`get_remote_tensor `: return a symmetric tensor given a peer rank.
The difference between `get_buffer` API and `get_remote_tensor` API:
- the former accepts an offset, whereas the latter doesn't
- the latter returns a symmetric tensor at `hdl.offset` on `peer`.
As a refactorization, this PR also moves the implementation of `get_buffer` and `get_signal_pad` to the `SymmetricMemory` level as their code is common to all backends.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161533
Approved by: https://github.com/ngimel
ghstack dependencies: #161470, #161471, #161532
(Porting most of #161008)
Hooking SymmetricMemory Allocator to MemPool so that user can create symmetric tensors with regular `torch.zeros`, `torch.arange` etc factories. Also so that our ops can have functional variants that create `out` tensors on symmetric memory.
To end users, this PR supports a python UI as follows:
```
allocator = symm_mem.get_mempool_allocator(device)
mempool = torch.cuda.MemPool(allocator)
with torch.cuda.use_mem_pool(mempool):
tensor = torch.arange(numel, dtype=dtype, device=device)
```
Added tests for both use cases above.
Differential Revision: [](https://our.internmc.facebook.com/intern/diff/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161471
Approved by: https://github.com/ngimel
ghstack dependencies: #161470
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
`get_remote_tensor `: return a symmetric tensor given a peer rank.
The difference between `get_buffer` API and `get_remote_tensor` API:
- the former accepts an offset, whereas the latter doesn't
- the latter returns a symmetric tensor at `hdl.offset` on `peer`.
As a refactorization, this PR also moves the implementation of `get_buffer` and `get_signal_pad` to the `SymmetricMemory` level as their code is common to all backends.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161533
Approved by: https://github.com/ngimel
ghstack dependencies: #161470, #161471, #161532
(Porting most of #161008)
Hooking SymmetricMemory Allocator to MemPool so that user can create symmetric tensors with regular `torch.zeros`, `torch.arange` etc factories. Also so that our ops can have functional variants that create `out` tensors on symmetric memory.
To end users, this PR supports a python UI as follows:
```
allocator = symm_mem.get_mempool_allocator(device)
mempool = torch.cuda.MemPool(allocator)
with torch.cuda.use_mem_pool(mempool):
tensor = torch.arange(numel, dtype=dtype, device=device)
```
Added tests for both use cases above.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161471
Approved by: https://github.com/ngimel
ghstack dependencies: #161470
(Porting most of #161008)
Hooking SymmetricMemory Allocator to MemPool so that user can create symmetric tensors with regular `torch.zeros`, `torch.arange` etc factories. Also so that our ops can have functional variants that create `out` tensors on symmetric memory.
To end users, this PR supports a python UI as follows:
```
allocator = symm_mem.get_mempool_allocator(device)
mempool = torch.cuda.MemPool(allocator)
with torch.cuda.use_mem_pool(mempool):
tensor = torch.arange(numel, dtype=dtype, device=device)
```
Added tests for both use cases above.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161471
Approved by: https://github.com/ngimel
ghstack dependencies: #161470
Summary:
We split the refactoring in two parts for forward compatibility concerns
First, we land the deserialization (loading part)
Then, we land the serialization (saving part)
Save weights and constants as individual files in PT2 archive. Each weight/constant will be saved as raw bytes, unless it is a custom object (TorchBind object) or a non-fake tensor subclass, for these two special cases we still save them using pickle.
The metadata of saved tensors along with the file name will be saved as `PayloadMeta`.
The mapping from FQN to `PayloadMeta` will be saved as `PayloadConfig` under `WEIGHTS_CONFIG_FORMAT` and `CONTANTS_CONFIG_FORMAT`
This changes the serialization in python side when calling `torch.export.save()`.
For deserialization in python `torch.export.load()`, we make it BC-safe by allowing loading legacy format weights/constants.
For deserialization in C++ `torch/nativert/ModelRunner.cpp`, we make this a BC breaking change as currently the OSS ModelRunner API is not being used.
The file structure
```
├── archive_format
├── archive_version
├── byteorder
├── .data
│ ├── serialization_id
│ └── version
├── data
│ ├── sample_inputs
│ │ └── model.pt
│ ├── constants
│ │ ├── tensor_0
│ │ ├── tensor_1
│ │ └── model_constants_config.json
│ └── weights
│ ├── weight_0
│ ├── weight_1
│ ├── weight_2
│ ├── weight_3
│ └── model_weights_config.json
└── models
└── model.json
```
Test Plan:
CI
Rollback Plan:
Differential Revision: D80035490
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160394
Approved by: https://github.com/SherlockNoMad
This is far simpler than #155164 since we never destroy the cudaGraphExec_t.
The request comes from TRT-LLM specifically. The motivation is that some power users would like to mutate specific kernel parameters via APIs like `cudaGraphExec*SetParams` after a cuda graph has been instantiated. For example, a common request has been to be able to change the sequence length of attention kernels, after having captured a graph for the largest possible sequence length. It turns out that the host overhead you eliminate via cuda graphs in LLM inference ends up causing an increase in computation time when you size your kernels to the maximum possible sequence length (which I believe is done in both TRT-LLM and vLLM). Attention is the most problematic kernel because its computation time is quadratic in the sequence length, rather than linear.
This can work if your attention kernel can work for arbitrary shapes (this is not the case for all attention implementations! Many of them specialize with templates), and you have a persistent kernel that allocates only as many blocks as you have SM's (so you don't have to figure out how many blocks to allocate for a specific sequence length). Using a conditional SWITCH node is a better generic approach to this problem, but that requires more infrastructure work.
Note that this requires knowledge of the exact location of the value in your kernel's parameter buffer to mutate. It won't work with arbitrary stream capture code whose kernels you don't know before hand. So I expect this code path to be rarely used.
Testing:
```
pytest -s -k raw_graph_exec test/test_cuda.py
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161294
Approved by: https://github.com/ngimel, https://github.com/BoyuanFeng, https://github.com/eellison, https://github.com/eqy
expose the pointer so that we can create the `ncclConfig_t` object from pytorch and use it elsewhere. this is useful to control the nccl communicator parameters for multiple nccl communicators.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/161136
Approved by: https://github.com/kwen2501
Using good old IOKit to get `gpu-core-count` property from device implementing `AGXAccelerator` service
Expose this one as `torch.backend.mps.get_core_count()` and make it accessible via `MpsInterface` to the inductor
Test Plan: Run `python3 -c "import torch;print(torch.backends.mps.get_name(), torch.backends.mps.get_core_count())"` and compare it to `system_profiler SPDisplaysDataType|head -n10`
```
% python3 -c "import torch;print(torch.backends.mps.get_name(), torch.backends.mps.get_core_count())"
Apple M1 Pro 16
% system_profiler SPDisplaysDataType|head -n10
Graphics/Displays:
Apple M1 Pro:
Chipset Model: Apple M1 Pro
Type: GPU
Bus: Built-In
Total Number of Cores: 16
Vendor: Apple (0x106b)
Metal Support: Metal 3
```
This would significantly improve occupancy for torch.compile generated kernels
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160414
Approved by: https://github.com/dcci
Adds `OperatorEntry::getComputedKernelForDispatchKey` which returns the KernelFunction corresponding to `OperatorEntry.dispatchTable_[dispatch_ix]` for a given dispatch key
- Specifically it returns a `SafeKernelFunction` that holds a `KernelToken`. This `KernelToken` is registered to the `KernelFunction` in `OperatorEntry.kernels_` and will be invalidated when the `KernelFunction` is destructed (i.e. when the `AnnotatedKernel` that holds this `KernelFunction` is removed from `kernels_`, which happens when the corresponding impl is deregistered).
- `SafeKernelFunction` can be called via `callBoxed`, the validity of the token will be checked before this happens
- `SafeKernelFunction` is pybinded and `getComputedKernelForDispatchKey` is exposed to the frontend ia `torch.library.get_kernel`
Related to https://github.com/pytorch/pytorch/issues/155330
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158393
Approved by: https://github.com/albanD
