id int64 2.74B 3.05B | title stringlengths 1 255 | user stringlengths 2 26 | state stringclasses 2
values | labels listlengths 0 24 | comments int64 0 206 | author_association stringclasses 4
values | body stringlengths 7 62.5k ⌀ | is_title bool 1
class |
|---|---|---|---|---|---|---|---|---|
2,941,421,749 | Demote logger of runtime_asserts_frozen to be fired only on debug mode | tugsbayasgalan | closed | [
"Merged",
"ciflow/trunk",
"release notes: fx",
"fx",
"ciflow/inductor"
] | 8 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #149833
* __->__ #149832
cc @ezyang @SherlockNoMad @EikanWang @jgong5 @wenzhe-nrv
Differential Revision: [D71702305](https://our.internmc.facebook.com/intern/diff/D71702305) | true |
2,941,420,130 | Only print dde partial fx graph for export | bobrenjc93 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"fx",
"module: dynamo",
"ciflow/inductor",
"release notes: export"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149831
Lazos correctly pointed out this doesn't make sense for compile since
we graph break in compile. This results in tons of unwanted user log
spew. We do want this in export though since it's drastiaclly reduced
the support load ... | true |
2,941,295,476 | cd: Add script for generating binary build matrix | seemethere | open | [
"topic: not user facing"
] | 2 | MEMBER | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #150713
* __->__ #149830
This script currently exists in
.github/scripts/generate_binary_build_matrix.py however I think there's
a lot of legacy cruft associated with that script so I'm going to
attempt to do a complete refactor starting wit... | true |
2,941,266,551 | TF32 acceleration on top of oneDNN is available for Intel GPUs. The current Torch version does not have Intel GPU Support | justinchuby | closed | [
"module: cpu",
"triaged",
"module: python frontend"
] | 7 | COLLABORATOR | The warning message
> /opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/backends/mkldnn/__init__.py:78: UserWarning: TF32 acceleration on top of oneDNN is available for Intel GPUs. The current Torch version does not have Intel GPU Support. (Triggered internally at /var/lib/jenkins/workspace/aten/src/ATen/Conte... | true |
2,941,256,497 | bump XNNPACK dependency to fix GCC 14 build on aarch64-linux | prusnak | open | [
"module: build",
"triaged",
"actionable",
"module: xnnpack",
"module: arm"
] | 3 | NONE | ### 🐛 Describe the bug
bundled version of XNNPACK cannot be built on aarch64-linux with GCC14 because of this issue https://github.com/google/XNNPACK/issues/7726
the issue has been fixed in XNNPACK in the meanwhile: https://github.com/google/XNNPACK/commit/3bc2a32a44db62434248197bceefa37f4f05153e
suggestion: bump t... | true |
2,941,248,356 | Why has my linear regression always been NaN? | bbhxwl | closed | [] | 1 | NONE | I use chatgpt to learn linear regression, but I don't understand why it can't predict?
Where is the mistake?
```
import torch
import torch.nn as nn
import torch.optim as optim
# 1. 数据准备：构造老人年龄（特征）和花费金额（目标）的数据
# 注意：数据形状必须是二维张量，每一行代表一个样本
ages = torch.tensor([[65], [70], [75], [80], [85], [90], [95], [100]], dtype=torch... | true |
2,941,049,317 | How to handle dynamic output size with torch.onnx.export (through dynamo) for Resize | FabianSchuetze | closed | [
"module: onnx",
"triaged",
"oncall: pt2"
] | 9 | CONTRIBUTOR | ### 🐛 Describe the bug
I would like to export with torch.onnx.export (through dynamo) some code that contains a resize operation. The output width and height is dynamic. An example model is as follows:
```
import torch
class Model(torch.nn.Module):
def __init__(self):
super().__init__()
def forwar... | true |
2,941,033,781 | [AOTInductor] Free folded constants that's managed by AOTInductor | muchulee8 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149825
internally.
Summary:
This diff allows freeing the usage of folded constants that's created by
AOTInductor through CUDACachingAllocator instead of the constant blob
from cudaMalloc directly.
Test Plan:
LD_LIBRARY_PATH=/data/u... | true |
2,941,032,971 | flex_attention raises error at compile | dslisleedh | closed | [
"triaged",
"oncall: pt2",
"module: higher order operators",
"module: flex attention"
] | 3 | NONE | ### 🐛 Describe the bug
I'm trying to accelerate WindowAttention with flex_attention.
However, when the window size equals 8, it raises an error when compiling.
Please refer to this [code](https://github.com/dslisleedh/ESC/blob/main/scripts/compare_attn.py)
```bash
python compare_attn.py --h 64 --w 64 --window_size ... | true |
2,940,868,279 | Please support python 3.13! | wuhuang2 | closed | [
"needs reproduction",
"module: binaries",
"module: windows",
"triaged"
] | 4 | NONE | ### 🚀 The feature, motivation and pitch
I'm using python3.13 to develop a project, and I need to use the Whisper library, which depends on the Pytorch library, but I encountered the problem of "unable to find the applicable version" when pip, and after checking the Internet, I found that the Pytorch library does not ... | true |
2,940,844,170 | `INTERNAL ASSERT FAILED` when using `torch.max` with mixed device tensors | default1360 | closed | [
"module: cuda",
"triaged"
] | 0 | NONE | ### 🐛 Describe the bug
Code:
```
import torch
x = torch.ones(10)
# If CUDA is available, use a CUDA tensor for the output.
if torch.cuda.is_available():
out_values = torch.empty(10, device="cuda")
out_indices = torch.empty(10, dtype=torch.long, device="cpu")
torch.max(x, 0, out=(out_values, out_indices))
... | true |
2,940,838,975 | `Aborted` error when using `torch.cuda.memory.caching_allocator_delete` | default1360 | open | [
"module: cuda",
"triaged",
"module: CUDACachingAllocator"
] | 2 | NONE | ### 🐛 Describe the bug
Code:
```
import torch
from torch.cuda.memory import caching_allocator_delete
torch.cuda.empty_cache()
dev_props = torch.cuda.get_device_properties(0)
total_memory = dev_props.total_memory
allocation = int(total_memory * 0.5)
tmp_tensor = torch.empty(allocation, dtype=torch.int8, device='cuda')... | true |
2,940,837,662 | `Segmentation fault` when using `torch.sparse.mm` with `torch.sparse_csr_tensor` | default1360 | open | [
"module: sparse",
"module: crash",
"triaged"
] | 3 | NONE | ### 🐛 Describe the bug
Code:
```
import torch
m, n, p = 7, 8, 9
nnz = 20
crow_indices = torch.tensor([0, nnz], dtype=torch.int64)
col_indices = torch.arange(nnz, dtype=torch.int32)
values = torch.randn(nnz)
S = torch.sparse_csr_tensor(crow_indices, col_indices, values, size=(m, n))
D = torch.randn(n, p)
result = torc... | true |
2,940,834,456 | `free(): invalid next size` error when using `torch.linalg.ldl_solve` | default1360 | closed | [
"module: crash",
"triaged",
"module: linear algebra"
] | 3 | NONE | ### 🐛 Describe the bug
Code:
```python
import torch
LD = torch.tensor([[1.0, 2.0, 3.0],
[2.0, 5.0, 6.0],
[3.0, 6.0, 9.0]], dtype=torch.float32)
pivots = torch.tensor([0, 1, 2], dtype=torch.int32)
B = torch.tensor([[1.0], [2.0], [3.0]], dtype=torch.float32)
torch.linalg.ldl_sol... | true |
2,940,825,135 | Fix `torch.cuda.MemPool()` internal assertion failure when changing devices | fzyzcjy | open | [
"triaged",
"open source"
] | 2 | CONTRIBUTOR | Fix https://github.com/pytorch/pytorch/issues/149802
This is just a prototype, and I would like to hear feedbacks, e.g. is this direction OK? or shall we let MemPool to support multi devices?
After feedbacks I will refine the PR, by e.g. making code better, adding tests, etc.
| true |
2,940,816,310 | [executorch hash update] update the pinned executorch hash | pytorchupdatebot | closed | [
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor"
] | 39 | COLLABORATOR | This PR is auto-generated nightly by [this action](https://github.com/pytorch/pytorch/blob/main/.github/workflows/nightly.yml).
Update the pinned executorch hash. | true |
2,940,815,107 | [MPS] Add support for `chebyshev_polynomial_t` in eager. | dcci | closed | [
"Merged",
"release notes: mps",
"ciflow/mps"
] | 4 | MEMBER | null | true |
2,940,791,142 | comparison operators only accept scalars as the 2nd argument but not as a 1st argument | ev-br | open | [
"triaged",
"module: python frontend"
] | 1 | COLLABORATOR | ### 🐛 Describe the bug
Comparison 'ufuncs' seem to be missing a `Number, Tensor` overload:
```
In [31]: x = torch.as_tensor([1.0])
In [32]: torch.less_equal(x, 1.0)
Out[32]: tensor([True])
In [33]: torch.less_equal(1.0, x)
---------------------------------------------------------------------------
TypeError ... | true |
2,940,786,761 | [inductor] [bug fix] Enable type promotions in slice_scatter in inductor | golkir | open | [
"triaged",
"open source",
"ciflow/trunk",
"topic: not user facing",
"module: inductor"
] | 9 | CONTRIBUTOR | Fixes #147842. Specifically, enables type promotions when calling `torch.slice_scatter` with tensors of different `dtype` thereby enforcing uniform behaviour in eager mode and inductor compilation mode.
To test:
`pytest -s -v test/inductor/test_torchinductor.py -k test_slice_scatter_types_promotion`
cc @voznesens... | true |
2,940,780,031 | torch.tril introduces NaNs on MPS when matrix contained Infs (when diagonal is negative) | twoertwein | closed | [
"triaged",
"module: NaNs and Infs",
"module: correctness (silent)",
"module: mps"
] | 1 | CONTRIBUTOR | ### 🐛 Describe the bug
```python
# bug
(Pdb) torch.tril(torch.full((3, 3), float("inf"), device="mps"), diagonal=-1)
tensor([[nan, nan, nan],
[inf, nan, nan],
[inf, inf, nan]], device='mps:0')
# working examples
# works with non-infs
(Pdb) torch.tril(torch.full((3, 3), 1.0, device="mps"), diagonal=-1... | true |
2,940,765,851 | [AOTInductor] Refine error message for dlopen in AOTInductor | muchulee8 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149812
Summary:
Refine the error message if dlopen failed in AOTInductor.
The original error message was ominous, modified to recommend user to
rebuild AOTInductor if needed, otherwise it's fine.
Test Plan:
None. Error message chang... | true |
2,940,648,934 | Rename README.txt to README.md | Jzhyang1 | closed | [
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 4 | CONTRIBUTOR | I am 99% sure this is meant to be a .md file rather than a .txt file
Fixes an issue with viewing the README on github, idk what else this accomplishes but it's been bothering me
| true |
2,940,576,242 | [AOTInductor] Bug fix for freeing buffers when freeing multiple times | muchulee8 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149810
Summary:
We might free the active buffer if we free the buffer twice.
Test Plan:
```
LD_LIBRARY_PATH=/data/users/$USER/pytorch/build/lib
/home/$USER/local/pytorch/build/bin/test_aoti_inference
```
Reviewers:
Subsc... | true |
2,940,562,575 | LoadHIP.cmake should find_package(composable_kernel) | trixirt | open | [
"module: build",
"module: rocm",
"triaged"
] | 4 | NONE | ### 🐛 Describe the bug
When building on Fedora, there is this build error
aten/src/ATen/native/hip/ck_types.h:19:10: fatal error: 'ck/ck.hpp' file not found
19 | #include <ck/ck.hpp>
| ^~~~~~~~~~~
1 error generated when compiling for host.
The ck/ck.hpp header is part of the composable_kernel pack... | true |
2,940,534,154 | Fix #149806 : Fix path lookup in _preload_cuda_deps | Divain | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"bug"
] | 9 | CONTRIBUTOR | @pytorchbot label "bug"
Fixes #149806
| true |
2,940,513,319 | Fix #149806 : Fix path lookup in _preload_cuda_deps | Divain | closed | [
"open source",
"bug"
] | 6 | CONTRIBUTOR | Fixes #149806
| true |
2,940,510,876 | _preload_cuda_deps cannot find libraries located in path/lib_folder | Divain | closed | [
"module: binaries",
"module: cuda",
"triaged"
] | 1 | CONTRIBUTOR | ### 🐛 Describe the bug
Hi,
I'm facing an issue when loading torch with CUDA from a PEX file. The function [_preload_cuda_deps](https://github.com/pytorch/pytorch/blob/2b848ab192e51498fb626355aedfd210df7da27e/torch/__init__.py#L282) seems to have a bug that prevents it from locating CUDA dependencies when they are pl... | true |
2,940,464,271 | Wrong location of rocm_version.h for Fedora and OpenSUSE | trixirt | closed | [
"module: build",
"module: rocm",
"triaged"
] | 2 | NONE | ### 🐛 Describe the bug
LoadHIP.cmake makes the assumption that ROCm is installed only from AMD to /opt/rocm
For several linux distributions including Fedora and OpenSUSE, this is /usr
This can be worked around sometimes if the user knows to set ROCM_PATH.
For some header files, it can not.
For rocm_version.h set, no... | true |
2,940,451,901 | torch.linalg.norm RuntimeError with torch.func.grad( vmap( hessian(.) ) ) | BurgerAndreas | open | [
"triaged",
"module: linear algebra",
"module: vmap",
"module: functorch"
] | 0 | NONE | ### 🐛 Describe the bug
`torch.linalg.norm(a)` causes `RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation`, but `torch.sum(a**2).sqrt()` works fine
```python
import torch
import numpy as np
def forward(samples):
# this causes
# RuntimeError: one of t... | true |
2,940,426,743 | checking out NCCL when it is not used | trixirt | open | [
"module: build",
"triaged"
] | 0 | NONE | ### 🐛 Describe the bug
NCCL is conditionally used with the USE_NCCL environmental variable
But it is unconditionally git cloned here
https://github.com/pytorch/pytorch/blob/main/tools/build_pytorch_libs.py#L122
This causes a problem for packaging pytorch v2.7.0 on Fedora.
packaging requires network isolation, so the... | true |
2,940,289,317 | (With PR) `torch.cuda.MemPool()` internal assertion failure when changing devices | fzyzcjy | open | [
"module: cuda",
"triaged"
] | 1 | CONTRIBUTOR | ### Potential cause analysis
Quickly glanced at the code, quick thoughts:
* When creating `MemPool` on device 0, it creates a MemPool on device 0, let's say it has mempool_id 111
* When first call to `use_mem_pool`, it tells C++ to find mempool with id 111 on device 1 (!), but that does not exist, so C++ side creates... | true |
2,940,287,457 | `INTERNAL ASSERT FAILED` in `torch.func.vmap` and `torch.scatter_add` | vwrewsge | closed | [
"triaged",
"module: vmap",
"oncall: pt2",
"module: functorch",
"module: pt2-dispatcher"
] | 4 | NONE | ### 🐛 Describe the bug
Code:
```
import torch
def buggy_vmap_fn(input_tensor, index_tensor, src_tensor):
return torch.func.vmap(lambda t: torch.scatter_add(t, 0, index_tensor, src_tensor))(input_tensor)
input_tensor = torch.randn(3)
index_tensor = torch.tensor([0, 1, 2])
src_tensor = torch.tensor([1.0, 2.0, 3.0]... | true |
2,940,240,700 | `Segmentation fault` in `torch.jit.jit_module_from_flatbuffer` | vwrewsge | open | [
"oncall: jit"
] | 0 | NONE | ### 🐛 Describe the bug
Code:
```
import torch
from torch import nn
simple_model = nn.Sequential(
nn.Linear(10, 20),
nn.BatchNorm2d(5),
nn.ReLU()
)
scripted_model = torch.jit.script(simple_model)
torch.jit.save_jit_module_to_flatbuffer(scripted_model, 'model.ff')
loaded_model = torch.jit.jit_module_from_... | true |
2,940,226,203 | bug in pytorch/torch/nn/parameter: | said-ml | closed | [] | 1 | NONE | ### 🐛 Describe the bug
```python
class UninitializedBuffer(UninitializedTensorMixin, torch.Tensor):
r"""A buffer that is not initialized.
Uninitialized Buffer is a a special case of :class:`torch.Tensor`
where the shape of the data is still unknown.
Unlike a :class:`torch.Tensor`, uninitialized para... | true |
2,940,178,868 | Build Extension Failed with setuptools==77.0.3 | AlongWY | open | [
"module: cpp-extensions",
"triaged"
] | 1 | NONE | ### 🐛 Describe the bug
When build deepspeed wheels with setuptools==77.0.3, the CUDAExtension throw the error info:
```
File "/opt/python/cp39-cp39/lib/python3.9/site-packages/setuptools/_distutils/command/sdist.py", line 245, in add_defaults
self._add_defaults_ext()
File "/opt/python... | true |
2,940,124,047 | avoid allocation when tensor_new from storage | ppwwyyxx | closed | [
"open source",
"Merged",
"topic: not user facing"
] | 4 | COLLABORATOR | null | true |
2,940,014,555 | Cannot compile SGlang with Torch 2.7 or Torch 2.8 and CUDA 12.8 (sm_120). | shahizat | open | [
"module: build",
"triaged"
] | 0 | NONE | ### 🐛 Describe the bug
Greetings to all,
I want to build SGlang from source on a machine with Nvidia RTX 5090. Using torch 2.6, the build succeeds, but torch 2.6 does not work with Triton version 3.3 and CUDA 12.8 with sm_120 support. Errors appear with versions 2.7 and the latest 2.8.
Error logs related to torch... | true |
2,939,989,403 | Remove outdated instructions from CI scripts | cyyever | closed | [
"open source",
"Merged",
"release notes: releng"
] | 3 | COLLABORATOR | Some instructions about Python 3.8 and CUDA 11.3 are removed. | true |
2,939,953,493 | [MPS/inductor] Add support for modified_scaled_bessel_k{0,1} | dcci | closed | [
"Merged",
"topic: not user facing",
"module: mps",
"ciflow/mps",
"module: inductor"
] | 3 | MEMBER | cc @kulinseth @albanD @malfet @DenisVieriu97 @jhavukainen @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov | true |
2,939,950,965 | [CUDA]][SymmetricMemory] Interpret empty string as `std::nullopt` in `rendezvous` | eqy | closed | [
"oncall: distributed",
"module: cuda",
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: bug fixes",
"topic: not user facing"
] | 13 | COLLABORATOR | this is a "temporary" fix as current internal API requires strings at some interfaces instead of `std::optional` and empty strings are presumably used in-lieu of `nullopt`.
e.g.,
https://github.com/pytorch/pytorch/blob/9d02b3993f7dae7fa3379d5190ac88291ecd4dce/torch/csrc/distributed/c10d/intra_node_comm.cu#L49
this... | true |
2,939,936,349 | [dynamo] Always trace into tensor subclass `__torch_function__` | StrongerXi | closed | [
"Merged",
"Reverted",
"ciflow/trunk",
"module: inductor",
"module: dynamo",
"ciflow/inductor",
"release notes: dynamo",
"ci-no-td"
] | 22 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149792
* #149484
* #149483
* #149482
This patch effectively ignores traceable_tensor_subclasses, allowing
Dynamo to always try tracing into the `__torch_function__` of tensor
subclass. This helps us with 2 things:
1. allowing users t... | true |
2,939,936,277 | [dynamo] Fix handling of setattr with some tensor attributes | StrongerXi | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #149792
* #149484
* #149483
* #149482
* __->__ #149791
* #149481
We weren't handling `setattr(tensor_obj, "real", 42)` correctly, because
the attribute is a `GetSetDescriptorType` that has special setter logic.
See added test and comments fo... | true |
2,939,920,336 | [inductor] Add the largest matmul tile size to default tuning set | bertmaher | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | While we probably don't want to expand the set of default matmul tunings too much, this is the largest tile size usable by H100 and A100, and is usually the top performing tile size for large matmuls. E.g. on H100 adding this tile size improves perf of multiplying 8192-square matrices from 600->700 tflops. (cuBLAS 12... | true |
2,939,898,879 | flex_attention create_block_mask() + inductor: integer division or modulo by zero | rmmr | closed | [
"triaged",
"oncall: pt2",
"module: higher order operators",
"module: pt2-dispatcher",
"module: flex attention"
] | 1 | NONE | ### 🐛 Describe the bug
This occurs very randomly! However i managed to reproduce. Please run this snippet multiple times, if no error happens the first time. In case of a notebook restart the kernel before rerunning.
```python
import torch
import torch._inductor.utils
from torch.nn.attention.flex_attention import (
... | true |
2,939,829,977 | `torch.vmap` does not work with tensor subclasses as expected | hchau630 | open | [
"triaged",
"tensor subclass",
"module: functorch"
] | 4 | NONE | ### 🐛 Describe the bug
vmapped functions do not seem to call `__torch_function__` for tensor subclasses. To illustrate this, I combined the [tutorial](https://pytorch.org/docs/stable/notes/extending.html#subclassing-torch-tensor) for tensor subclassing and the [example](https://pytorch.org/docs/stable/generated/torch... | true |
2,939,827,292 | Slower Mixed Precision Performance with pytorch installed via pip vs. conda | stas-sl | closed | [
"needs reproduction"
] | 2 | NONE | I’ve observed a significant performance difference when running mixed precision workloads with pytorch installed via pip compared to the same version installed via conda. The pip installation is substantially slower in mixed precision (float16) on a GTX 1080 GPU (yeah, it's old). This issue does not appear in full prec... | true |
2,939,818,329 | CudaGraphs Failing on Blackwell | drisspg | closed | [
"module: cuda",
"triaged",
"module: cuda graphs",
"Blackwell"
] | 2 | CONTRIBUTOR | # Summary
Run repro:
```py
import torch
def func(a):
return torch.softmax(a, dim=-1, dtype=torch.float32)
a = torch.randn(4, 16, dtype=torch.float16, device="cuda")
g = torch.cuda.CUDAGraph()
torch.cuda.synchronize()
with torch.cuda.graph(g):
out = func(a)
torch.cuda.synchronize()
g.replay()
torch.cuda.... | true |
2,939,802,895 | [DCP] Cache save plan metadata to reduce the collective overhead | saumishr | closed | [
"oncall: distributed",
"fb-exported",
"Merged",
"ciflow/trunk",
"release notes: distributed (checkpoint)",
"oncall: distributed checkpointing"
] | 9 | CONTRIBUTOR | Summary:
Cache save plan metadata to reduce the collective overhead.
Global plan dedupe and metadata creation are the main overheads on Rank 0. This change saves all this cost for the subsequent saves if the plans do not change. A quick experiment with the 256 rank job, Global step overhead drops by ~99%, from 90s+... | true |
2,939,791,586 | [ca] torch._dynamo.disable the checkpoint unpack hook | xmfan | open | [
"module: inductor",
"ciflow/inductor"
] | 2 | MEMBER | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149784
* #149773
* #149897
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov | true |
2,939,787,810 | [MPS] Add support for scaled_modified_bessel_k1 to eager. | dcci | closed | [
"Merged",
"topic: improvements",
"module: mps",
"release notes: mps",
"ciflow/mps"
] | 4 | MEMBER | Another day another op
cc @kulinseth @albanD @malfet @DenisVieriu97 @jhavukainen | true |
2,939,779,645 | [graph partition] support splitting on custom ops | BoyuanFeng | closed | [
"Merged",
"ciflow/trunk",
"module: inductor",
"ciflow/inductor",
"release notes: inductor"
] | 7 | CONTRIBUTOR | This PR adds support for graph partition on custom ops. Land after #149458.
### API
This PR provides a new API to set torch._C.Tag.cudagraph_unsafe tag for custom ops. This tag will be used for graph partition.
Example usage:
```python
@torch.library.custom_op(
"mylib::mysin",
mutates_args=["out_li... | true |
2,939,777,155 | ProcessGroupGloo: support ReduceOp::AVG | d4l3k | closed | [
"oncall: distributed",
"Merged",
"ciflow/trunk",
"release notes: distributed (c10d)"
] | 3 | MEMBER | This adds AVG support to ProcessGroupGloo to better support FSDP on CPU. I expect there will be more issues but this is easy enough to support in a naive fashion.
This applies to both reduce and allreduce.
This is a simple SUM + division and may not be the most numerically stable but that's expected. FSDP for lo... | true |
2,939,773,393 | Remove aten.elu core ATen decomp because it is now core ATen | swolchok | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 5 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149780
Per @larryliu0820.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov | true |
2,939,757,143 | Removed ROCM ifdef that governs thread count + smem parallel reduction. | 5had3z | closed | [
"module: rocm",
"triaged",
"open source",
"Merged",
"release notes: cuda"
] | 10 | CONTRIBUTOR | #149548 Fixed the arbitrarily missing parallelism for NLL, but they also added an arbritrary #ifdef ROCM guard around this fix to prevent its use on CUDA gpus. There is also a problem with the way the kernel does the reduction from the intermediate shared memory, using only thread 0 walking linearly. This has been chan... | true |
2,939,755,118 | ci/docker: use NCCL 2.26.2-1 | d4l3k | closed | [
"Merged",
"topic: not user facing"
] | 6 | MEMBER | Related to #149153
This updates some build scripts to hopefully fix the nightly builds which are somehow building against nccl 2.25.1 and using 2.26.2 from pip.
Test plan:
After merging rerun nightly linux jobs and validate that nccl version matches | true |
2,939,753,347 | [WIP] stop writing Max(*, 1) for strides | pianpwk | closed | [
"release notes: fx",
"fx",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | This should help us move away from size-oblivious. Looking at what the code was before sym_max(*, 1) was introduced, I think this is appropriate (https://github.com/pytorch/pytorch/pull/94400 in `_prims_common/__init__.py`) | true |
2,939,737,320 | [WIP] guard or false | pianpwk | closed | [
"release notes: fx",
"fx",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Fixes #ISSUE_NUMBER
cc @ezyang @SherlockNoMad @EikanWang @jgong5 @wenzhe-nrv | true |
2,939,737,224 | Use schema as source of truth + support ones_like/empty_like | pytorchbot | closed | [
"open source",
"ciflow/inductor"
] | 1 | COLLABORATOR | NOTE: THIS MUST BE LANDED ONLY AFTER https://github.com/pytorch/pytorch/pull/149644 IS LANDED IN THE RELEASE, otherwise tests will fail
This change does 2 important things:
(a) Instead of relying on IValue type as source of truth, we use the schema as the source of truth, which is important as IValue types are over... | true |
2,939,712,627 | bound_sympy() produces incorrect result for mod | pianpwk | open | [
"triaged",
"oncall: pt2",
"module: dynamic shapes",
"export-triage-review",
"oncall: export"
] | 2 | CONTRIBUTOR | ### 🐛 Describe the bug
`bound_sympy(s0 - (s0 % 8))` produces an incorrect range of [-5, inf], when the correct answer is [0, inf] (s0 has a bound of [2, inf].
My guess is this happens because each term is evaluated individually, with s0 resolving to [2, inf], and -(s0 % 8) resolving to [-7, 0], combining for a range... | true |
2,939,703,212 | [ca][aot cache] disable caching on joint graphs when CA is enabled | xmfan | open | [
"module: dynamo",
"ciflow/inductor"
] | 2 | MEMBER | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #149784
* __->__ #149773
* #149897
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,939,686,949 | [inductor] fix combo_kernel logging #2 | YUNQIUGUO | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 5 | CONTRIBUTOR | Summary:
fix another combo kernel logging error:
File "/home/guorachel/local/fbsource/buck-out/v2/gen/fbcode/4bcbfa3ef39dbd6f/caffe2/test/inductor/__combo_kernels__/combo_kernels#link-tree/torch/_inductor/scheduler.py", line 2036, in _init
self.create_combo_kernel_nodes(num_ck_nodes=None)
File "/home/guorachel... | true |
2,939,678,315 | How to remove the “internal api” notice? | justinchuby | closed | [
"module: docs",
"triaged"
] | 4 | COLLABORATOR | ### 📚 The doc issue
What is the option that will remove this notice?
> This page describes an internal API which is not intended to be used outside of the PyTorch codebase and can be modified or removed without notice.
We would like to remove it for https://pytorch.org/docs/stable/onnx_dynamo.html and a few onnx p... | true |
2,939,663,451 | Include other accelerators in capturable docstr for optimizers | janeyx99 | closed | [
"Merged",
"ciflow/trunk",
"topic: docs",
"release notes: optim"
] | 5 | CONTRIBUTOR | Fixes #149722
@ILCSFNO is this better?
| true |
2,939,656,581 | fix inductor logging for torch._scaled_mm | vkuzo | open | [
"ciflow/trunk",
"topic: bug fixes",
"module: inductor",
"module: dynamo",
"ciflow/inductor",
"release notes: inductor"
] | 4 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149769
Summary:
https://github.com/pytorch/pytorch/pull/148800 made inductor logs
throw an exception of the model contains `torch._scaled_mm` because
it split a string by underscore and making some assumptions about the resulting
li... | true |
2,939,568,144 | [FSDP2][DTensor] numeric bug for DTensor + python float in gradient clipping | weifengpy | open | [
"oncall: distributed",
"triaged",
"module: fsdp"
] | 6 | CONTRIBUTOR | ### 🐛 Describe the bug
```
python3.12/site-packages/torch/nn/utils/clip_grad.py", line 155, in _clip_grads_with_norm_
[rank1]: clip_coef = max_norm / (total_norm + 1e-6)
```
for DTensor + 1e-6, each rank adds 1e-6 to local tensor, instead of adding once
this was reported by others as well
### Versions
na
cc ... | true |
2,939,550,367 | `flex_attention` slower than manual attention implementation | abdulfatir | open | [
"triaged",
"oncall: pt2",
"module: higher order operators",
"module: pt2-dispatcher",
"module: flex attention"
] | 6 | NONE | ### 🐛 Describe the bug
I'm not sure if this is a bug or if I am using `flex_attention` incorrectly. Also, not completely sure if flex attention is designed for such masked language modeling-style use cases. That said, I found `flex_attention` to be slower than a (compiled) manual torch implementation of attention for... | true |
2,939,531,591 | Move emulate_precision_casts to be controlled by a JK. | c00w | closed | [
"fb-exported",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Summary: This allows us to selectively turn this on for internal use cases.
Test Plan: Relying primarily on existing models not breaking at test time.
Reviewed By: Yuzhen11
Differential Revision: D71647650
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv... | true |
2,939,504,820 | Parallelize sort | pytorchbot | closed | [
"open source"
] | 1 | COLLABORATOR | PR #142391 erroneously used `USE_OMP` instead of `USE_OPENMP`.
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 | true |
2,939,492,678 | [DTensor] Error on illegal view op during sharding prop | wconstab | closed | [
"oncall: distributed",
"Merged",
"ciflow/trunk",
"ciflow/inductor",
"release notes: distributed (dtensor)"
] | 21 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149764
* #152045
Adds explicit error checking during sharding propagation for view ops
rather than relying on runtime errors during local op execution.
Before:
An error is thrown by aten.view op called by DTensor dispatch, beca... | true |
2,939,479,668 | [scan] Support None return in combine_fn | angelayi | open | [
"module: dynamo",
"ciflow/inductor"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149763
Wondering if we could support the case if the user doesn't want to put anything in the accumulated tensor:
```python
def add(x, y):
return x + y[0], None
def f(init, x):
return ... | true |
2,939,457,110 | terminate called after throwing an instance of 'c10::Error' | guarin | open | [
"oncall: distributed",
"triaged",
"module: ddp"
] | 0 | NONE | Hi, I am running into the stacktrace shown below and am not sure where the error is coming from.
Setup is a 4xRTX4090 node with DDP training. The error happens randomly, in this case after 20 epochs.
Library versions:
```
Platform: Linux-6.8.0-52-generic-x86_64-with-glibc2.39
CUDA Version: 12.4
CUDA Driver Version: 5... | true |
2,939,442,183 | [Inductor] Introducing Subgraph as a Choice | PaulZhang12 | closed | [
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149761
Introduce autotuning on a subgraph as a choice in Inductor.
Working with decomposing mm -> bmm + sum, with repro https://pastebin.com/UZq3VtyK
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSu... | true |
2,939,441,365 | AssertionError: Unexpected key _export_root.mods.embedding_model.blocks.0 | ivyw-ts | closed | [
"module: onnx",
"triaged"
] | 1 | NONE | ### 🐛 Describe the bug
This is a followup/continuation of <a href="https://github.com/pytorch/pytorch/issues/149533">bug report #149533</a>.
We ran into this error when trying to convert the <a href="https://huggingface.co/speechbrain/lang-id-voxlingua107-ecapa">VoxLingua107 ECAPA-TDNN Spoken Language Identification ... | true |
2,939,360,023 | [BE]: Update cudnn frontend submodule to 1.11.0 | Skylion007 | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"release notes: cudnn"
] | 9 | COLLABORATOR | Update CUDNN frontend submodule to 11.1.0. Adds some new features like score_mod from flex_attention and adds a lot of bugfixes and new feature knobs. | true |
2,939,332,557 | [dynamo][hooks] config to wrap the top frame in a wrapper | anijain2305 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149758
* #149712
This should be done by default but there are too many issues. This PR is a
workaround.
https://github.com/pytorch/pytorch/issues/117584
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @z... | true |
2,939,325,223 | [Profiler] Give non-zero default values to start events | mcalman | open | [
"triaged",
"open source",
"Merged",
"Reverted",
"topic: not user facing",
"ci-no-td"
] | 20 | CONTRIBUTOR | The intent of the existing code is to
> // Assign system TIDs to start events based on the system TID of the next
// observed event with the same Python TID.
However, if there are start events that don't share the same Python TID as later observed events, then they are left with the default initialization of... | true |
2,939,265,807 | Removing doc references to PRE_CXX11_ABI. | AlannaBurke | closed | [
"module: docs",
"Merged",
"topic: not user facing"
] | 5 | CONTRIBUTOR | Fixes #149550
cc @svekars @sekyondaMeta | true |
2,939,217,204 | [sigmoid] Fix scalar resolution for Scalar_mode aten ops. | zhxchen17 | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 5 | CONTRIBUTOR | Summary: For Scalar variant resolution, we didn't handle a corner case of "Tensor_mode" variant (from aten::div). Adding the missing case to the graph pass.
Test Plan: buck test mode/opt caffe2/test:test_export -- -r test_operator_aten_tensor_mode_variant_cpp_runtime
Differential Revision: D71638433
| true |
2,939,217,042 | [sigmoid] Support _operator.neg/truediv | zhxchen17 | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 5 | CONTRIBUTOR | Summary: adding operator.truediv and operator.neg support to the runtime
Test Plan: buck run mode/opt caffe2/test:test_export -- -r test_sym_float_operators_cpp_runtime_nonstrict
Differential Revision: D71637267
| true |
2,939,215,094 | Stash tensors for reduce_scatter_v and all_gather_v | kwen2501 | closed | [
"oncall: distributed",
"release notes: distributed (c10d)"
] | 2 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149753
* #148590
https://github.com/pytorch/pytorch/pull/148590 removed record_stream. Since previous AVOID_RECORD flag does not cover reduce_scatter_v and all_gather_v which are in coalescing form, these two ops were missed. Causin... | true |
2,939,187,660 | [MPS][BE] Move `polar`/`complex` to stubs | malfet | closed | [
"Merged",
"topic: not user facing",
"release notes: mps",
"ciflow/mps"
] | 4 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #149730
* __->__ #149752
* #149729
* #149728
* #149727
No need to have in-place MPS kernel, as it just copy-n-paste of code
from TensorFactories.cpp into Binarykernel.mm | true |
2,939,186,088 | [AOTAutogradCache] Allow Custom Autograd functions behind a flag | jamesjwu | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 7 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149751
This adds a new env var and flag,
autograd_cache_allow_custom_autograd_functions, (env var: `TORCHINDUCTOR_AUTOGRAD_CACHE_ALLOW_CUSTOM_AUTOGRAD`) which allows custom autograd functions into AOTAutogradCache.
@hirsheybar... | true |
2,939,124,233 | [FSDP2] warning that reshard_after_forward=1 and True are different | weifengpy | closed | [
"oncall: distributed",
"Merged",
"ciflow/inductor",
"release notes: distributed (fsdp2)"
] | 3 | CONTRIBUTOR | people complains about spending time to debug reshard_after_forward=1. What they actually want is reshard_after_forward=True. 1 and True can be used interchangeably in programming generally, add one-time warning to remind they are different
* reshard_after_forward=1 means resharding parameters to world size 1, by keep... | true |
2,939,099,055 | Support None return type in torchbind and Add more AOTI torchbind e2e tests | yushangdi | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 7 | CONTRIBUTOR | Summary:
- Add more tests for torchbind in aoti
**FallBackKernel**
- In FallbackKernel.find_device, do not check the device of torchbind obj because they don't have a fixed "device"
- If no device found for CallTorchBindObject, use cpu
- handle None output in `export_extern_kernel_node`
Test Plan:
```
buck run //sigm... | true |
2,939,090,050 | Remove `torch.utils.deterministic` from `MOD_SKIPLIST` | guilhermeleobas | open | [
"open source",
"module: dynamo",
"ciflow/inductor"
] | 2 | COLLABORATOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149748
* #149643
Attempt to trace `torch.utils.deterministic` module
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,939,066,739 | Support torchbind in OSS proxy executor | yushangdi | closed | [
"fb-exported",
"Merged",
"Reverted",
"ciflow/trunk",
"module: inductor",
"ciflow/inductor",
"release notes: export",
"ci-no-td"
] | 16 | CONTRIBUTOR | Summary:
Implement torchbind support in OSSProxyExecutor.
Exactly the same as the implementation in FbProxyExecutor.
D69693697 - fbProxyExecutor
D69887230 - fbProxyExecutor but for torchbind method
Other changes:
- When generating the schema of the CallTrochBind HOP, the arg name of the torchbind object... | true |
2,939,058,839 | enabled dynamic rblock scaling on H100 | shunting314 | open | [
"triaged",
"oncall: pt2",
"module: inductor"
] | 0 | CONTRIBUTOR | ### 🐛 Describe the bug
Apply https://github.com/pytorch/pytorch/pull/109275 to h100.
### Error logs
_No response_
### Versions
.
cc @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @aakhun... | true |
2,939,052,831 | [ONNX] Clean up legacy dynamo export code | justinchuby | closed | [
"module: onnx",
"open source",
"Merged",
"ciflow/trunk",
"release notes: onnx",
"topic: bc breaking",
"suppress-bc-linter"
] | 3 | COLLABORATOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149745
Clean up code that is unused and obsolete. The public `torch.onnx.dynamo_export` is kept for now but the legacy implementation is removed.
Remove public option classes and OnnxRegistry that have been deprecated.
Users:... | true |
2,939,033,198 | [fbcode]Removing `@NoIntBaseDeprecated` annotation in `caffe2.thrift` file (#149742) | Sunnie912 | open | [
"fb-exported",
"Merged",
"Reverted",
"ciflow/trunk",
"topic: not user facing",
"ci-no-td"
] | 19 | CONTRIBUTOR | Summary:
To align with thrift-python, we are adding the int base class for `non-Flag` enums. In order to not break production code, the annotation `python.NoIntBaseClassDeprecated` is added to opt-out some enums
After the related customer code logic changes, we can now safely remove the annotations that were added ea... | true |
2,939,017,311 | dynamic annotations that still allow duck sizing | xmfan | open | [
"triaged",
"oncall: pt2",
"module: dynamic shapes"
] | 2 | MEMBER | ### 🚀 The feature, motivation and pitch
Today if we `mark_dynamic`/`maybe_mark_dynamic`, we will always assign a separate symbol for each of the dims. This is different than automatic dynamic's default behavior which might share symbols between dims, and this difference can incur recompiles and remote cache misses.
... | true |
2,938,982,029 | [fbcode]Removing `@NoIntBaseDeprecated` annotation in `caffe2.thrift` file | williamwen42 | closed | [
"fb-exported"
] | 4 | MEMBER | Summary:
To align with thrift-python, we are adding the int base class for `non-Flag` enums. In order to not break production code, the annotation `python.NoIntBaseClassDeprecated` is added to opt-out some enums
After the related customer code logic changes, we can now safely remove the annotations that were added ear... | true |
2,938,981,469 | Cudagraph fix + comment cleanup | eellison | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149741
Cudagraphs is careful to not allow any memory recorded to escape globally without having a reference to the tensor. This is because we may later reclaim that memory for a cudagraph recording and we need to mark the tensor a... | true |
2,938,955,950 | Torch compile update documentation, listing required dependent packages | atalman | closed | [
"module: docs",
"triaged",
"topic: docs",
"oncall: pt2"
] | 2 | CONTRIBUTOR | ### 🐛 Describe the bug
Please see PR: https://github.com/pytorch/test-infra/pull/6434
On clean Amazon Linux 2023 instance to make torch compile work in addition to the python we need to install following 2 packages:
```
yum groupinstall -y "Development Tools"
yum install -y python-devel
```
Please make sure to doc... | true |
2,938,933,544 | adding logging to capture when a trainer process is sigkilled. | aschhabra | closed | [
"oncall: distributed",
"fb-exported",
"release notes: distributed (torchelastic)"
] | 9 | CONTRIBUTOR | Summary:
It will help us determine when a trainer process is not terminated gracefully due to SIGKILL by torch elastic.
In case of process is terminated with SIGKILL, there is no way to collect logs before termination which causes confusion due to missing logs. Logging when trainer was SIGKILLED when help understand fa... | true |
2,938,929,927 | [ROCm] Extend vectorized elementwise kernel to more heterogenous tensor types. | carlobertolli | closed | [
"module: rocm",
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"release notes: rocm",
"ciflow/rocm"
] | 3 | CONTRIBUTOR | This patch extends the initial support for "vectorized templated" kernels to the following input tensor types: (BFloat16, float)
(float, float16)
(float16, float)
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd | true |
2,938,823,259 | `INTERNAL ASSERT FAILED` in `interpolate` and `torch.import_ir_module` | vwrewsge | open | [
"oncall: quantization",
"module: error checking"
] | 0 | NONE | ### 🐛 Describe the bug
# Bug 1
Code:
```
import torch
from torch.nn.quantized.functional import interpolate
x = torch.rand((1, 1, 4, 4), dtype=torch.float32)
q_x = torch.quantize_per_tensor(x, scale=0.1, zero_point=10, dtype=torch.quint8)
_ = interpolate(q_x, scale_factor=-1.0, mode='nearest')
```
Output:
```
Fi... | true |
2,938,799,179 | `INTERNAL ASSERT FAILED` in `torch.jit.script` | vwrewsge | open | [
"oncall: jit"
] | 0 | NONE | ### 🐛 Describe the bug
Code:
```
import torch
from torch.utils.mobile_optimizer import optimize_for_mobile
class TestModule(torch.nn.Module):
def forward(self, input: torch.Tensor, weight: torch.Tensor) -> torch.Tensor:
return torch.nn.functional.conv2d(input, weight, padding="same")
module = TestModule... | true |
2,938,791,301 | DISABLED test_binary_op_with_scalar_self_support__foreach_pow_is_fastpath_True_cuda_int64 (__main__.TestForeachCUDA) | pytorch-bot[bot] | open | [
"triaged",
"module: flaky-tests",
"skipped",
"module: mta"
] | 5 | NONE | Platforms: linux, slow
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_binary_op_with_scalar_self_support__foreach_pow_is_fastpath_True_cuda_int64&suite=TestForeachCUDA&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/... | true |
2,938,680,560 | [RFC] Multi-backend, multi-device test class instantiation for Inductor | kundaMwiza | open | [
"triaged",
"module: testing",
"oncall: pt2",
"module: inductor"
] | 3 | CONTRIBUTOR | ### 🚀 The feature, motivation and pitch
## Background
Out-of-tree backends (like Graphcore's) are able to utilise the PyTorch test suite to verify their implementation.
For eager tests, to instantiate device-specific test classes, out-of-tree backends can subclass from `DeviceTypeTestBase` and register their device... | true |
2,938,676,439 | [AOTI] Switch AOTI benchmark runner to use run_single_threaded | desertfire | open | [
"module: dynamo",
"ciflow/inductor"
] | 2 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #149733
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.