That fixes `index_put(..., accumulate=True)` for all dtypes
int64 operation is not really atomic, but eventually consistent from the `index_put_accumulate` kernel point of view: i.e. by the end of the operation results in the global memory are indeed accumulation of the operands at given indices
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158179
Approved by: https://github.com/dcci, https://github.com/Skylion007
ghstack dependencies: #158064, #158178
Move `MetalShaderLibrary::bind_tensors` private method to OperatorUtils.h and extract `iter_tensor_offset` method, that returns an offset from the start of the storage associated with given tensor inside the iterator
Migrated `index`, `index_put[_accumulate][_serial]` to the new paradigm that does not require additional tensor for indices nor special handling for 32 vs 64-bit offset, which resulted in almost 2x perf gain for 2000x2000 tensor, see results below before
```
[------------------------------------------------------------ -----------------------------------------------------------]
| 11x50x50 | 11x100x100 | 11x500x500 | 11x1000x1000 | 11x2000x2000
1 threads: ----------------------------------------------------------------------------------------------------------------
__getitem__ (torch.int8, torch.int64) | 383.5 | 379.8 | 470.9 | 1232.9 | 4410.3
__getitem__ (torch.float16, torch.int64) | 379.6 | 354.5 | 533.2 | 1290.3 | 4442.2
__getitem__ (torch.float32, torch.int64) | 360.8 | 338.6 | 478.6 | 1348.9 | 4870.4
Times are in microseconds (us).
```
and after
```
[------------------------------------------------------------ -----------------------------------------------------------]
| 11x50x50 | 11x100x100 | 11x500x500 | 11x1000x1000 | 11x2000x2000
1 threads: ----------------------------------------------------------------------------------------------------------------
__getitem__ (torch.int8, torch.int64) | 349.8 | 330.5 | 432.6 | 764.5 | 1961.2
__getitem__ (torch.float16, torch.int64) | 342.5 | 330.7 | 434.7 | 741.0 | 1969.4
__getitem__ (torch.float32, torch.int64) | 332.2 | 326.1 | 445.4 | 751.3 | 1972.6
Times are in microseconds (us).
```
While migrating also fixed index_put_accumulate for boolean types, by using compare_and_exchange trick over uint
Fixes https://github.com/pytorch/pytorch/issues/153560
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158064
Approved by: https://github.com/dcci
Note on backward precision over fp16:
A float16 number has 10 bits of mantissa, 5 bits of exponent, and 1 bit for the sign. If the sign bit is positive, then with a mantissa $m$ and exponent $e$ represented in base 10, the number that the float16 format represents is $(1 + m / 1024) \exp2(e)$. ([source](https://en.wikipedia.org/wiki/Half-precision_floating-point_format))
Consider adding two numbers $a$ and $b$ which have arbitrary mantissas, and say their exponents are $e_a = 1$ (so $2 \le a \lt 4$) and $e_b=-3$ (so $0.175 \le b \lt 0.25$). Assume that the result has the same exponent as $a$. Since the exponents differ by 4, we'll effectively need to truncate the 4 rightmost bits of $b$'s mantissa, which would introduce a maximum error on the order of $(2^4 / 1024) \exp2(-3) \approx 0.002$.
The error is nearly the same if $e_b = -2$ (so $0.25 \le b \lt 0.5$), where the 3 rightmost bits are truncated, giving a maximum error on the order of $(2^3 / 1024) \exp2(-2) \approx 0.002$. Same for $e_b=-1$.
So if we're adding up nine different numbers that all have exponents -3, -2, or -1, and they sum to a number with exponent 1, then we would expect a maximum error of several times greater than 0.002. In my comments above, summing those particular nine numbers in different ways gave results that ranged between 3.1816 and 3.1758, a difference of $0.0058 \approx 2.9 * 0.002$.
That's within the acceptable bounds, and we can safely just increase the error tolerance used in test_output_grad_match for the case of max_pool3d_backward with float16.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157498
Approved by: https://github.com/malfet
They might have been slow on CUDA-11.3, but this version of CUDA is long gone. More fundamental underlying issue were linear complexity of the recursive polynomial definitions for higher order polynomials, for example see this loop from implementation of Chebyshev polynomial of the first kind
7081b8233a/aten/src/ATen/native/Math.h (L2969-L2973)
which were tested by `test_compare_cpu` using following values (as sample index 16)
7081b8233a/torch/testing/_internal/opinfo/core.py (L2079)
Luckily chebyshev polynomials for absolute values higher than 1 pretty quickly reach infinity, see below
```
python3 -c "import torch;print(torch.special.chebyshev_polynomial_v(torch.nextafter(torch.tensor(1.0), torch.tensor(2.0)), torch.tensor(1e6)))"
tensor(nan)
```
Which is not the case for Laguerre polynomials, but it's probably fine to just limit it to 1e7
Before
```
$ PYTORCH_TEST_WITH_SLOW=1 python test_ops.py -k chebyshev_polynomial_
ssssssss..ssssss..ssssss..ssssssssssssssssssssss..ssssss/home/ubuntu/py3.10-nightly/lib/python3.10/site-packages/torch/backends/cuda/__init__.py:131: UserWarning: This API is going to be deprecated, please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:78.)
return torch._C._get_cublas_allow_tf32()
....ssssssssssss..ssssss..ssssss............ssssssssssssssssssssssssssssssssssss..ssssssssssssss..ssssss..ssssssssssssssssssssssssssssss..ssssss....ssssssssssss..ssssss..ssssss............ssssssssssssssssssssssssssssssssssss..ssssss..ssssssssssssss..ssssss..ssssss..ssssssssssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssssssssssss
----------------------------------------------------------------------
Ran 432 tests in 8.575s
OK (skipped=344)
```
After
```
$ PYTORCH_TEST_WITH_SLOW=1 python test_ops.py -k chebyshev_polynomial_
ssssssss........................ssssssssssssssss......../home/ubuntu/pytorch/torch/backends/cuda/__init__.py:131: UserWarning: This API is going to be deprecated, please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /home/ubuntu/pytorch/aten/src/ATen/Context.cpp:78.)
return torch._C._get_cublas_allow_tf32()
........................................................................................xxxxxxxx................ssssssssssssssssssssssss........................................................................................................ssssssss........................ssssssss........................................................................................ssssssss
----------------------------------------------------------------------
Ran 432 tests in 45.580s
OK (skipped=72, expected failures=8)
```
Fixes https://github.com/pytorch/pytorch/issues/79528
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157464
Approved by: https://github.com/Skylion007, https://github.com/dcci
ghstack dependencies: #157488
They were slow on CUDA-11.3, which has long been gone, let's see if they work now
Before
```
$ python test_ops.py -k chebyshev_polynomial_
ssssssss..ssssss..ssssss..ssssssssssssssssssssss..ssssss/home/ubuntu/py3.10-nightly/lib/python3.10/site-packages/torch/backends/cuda/__init__.py:131: UserWarning: This API is going to be deprecated, please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:78.)
return torch._C._get_cublas_allow_tf32()
....ssssssssssss..ssssss..ssssss............ssssssssssssssssssssssssssssssssssss..ssssssssssssss..ssssss..ssssssssssssssssssssssssssssss..ssssss....ssssssssssss..ssssss..ssssss............ssssssssssssssssssssssssssssssssssss..ssssss..ssssssssssssss..ssssss..ssssss..ssssssssssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssss..ssssssssssssss
----------------------------------------------------------------------
Ran 432 tests in 8.575s
OK (skipped=344)
```
After
```
$ python test_ops.py -k chebyshev_polynomial_
ssssssss........................ssssssssssssssss......../home/ubuntu/py3.10-nightly/lib/python3.10/site-packages/torch/backends/cuda/__init__.py:131: UserWarning: This API is going to be deprecated, please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:78.)
return torch._C._get_cublas_allow_tf32()
........................................................................................ssssssss................ssssssssssssssssssssssss........................................................................................................ssssssss........................ssssssss........................................................................................ssssssss
----------------------------------------------------------------------
Ran 432 tests in 42.379s
OK (skipped=80)
```
Fixes https://github.com/pytorch/pytorch/issues/79528
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157464
Approved by: https://github.com/Skylion007
Otherwise it turns test into a trivial one(that always succeeds), as following example demonstrates
```python
import torch
from torch.testing._internal.common_utils import serialTest, run_tests, TestCase
class MegaTest(TestCase):
@serialTest
def test_foo(self):
if hasattr(self.test_foo, "pytestmark"):
print("foo has attr and it is", self.test_foo.pytestmark)
print("foo")
@serialTest()
def test_bar(self):
if hasattr(self.test_bar, "pytestmark"):
print("bar has attr and it is", self.test_bar.pytestmark)
print("bar")
if __name__ == "__main__":
run_tests()
```
That will print
```
test_bar (__main__.MegaTest.test_bar) ... bar has attr and it is [Mark(name='serial', args=(), kwargs={})]
bar
ok
test_foo (__main__.MegaTest.test_foo) ... ok
----------------------------------------------------------------------
Ran 2 tests in 0.013s
```
Added assert that arg is boolean in the decorator to prevent such silent skips in the future
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157388
Approved by: https://github.com/clee2000
Fixes#155006
Inductor sometimes codegens triton kernel definitions into a triple-quoted text block. If the text block itself contains triple-quotes, this breaks. Notably, this can happen for user-defined triton kernels, where the user may have added a docstring in their triton kernel.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157322
Approved by: https://github.com/zou3519, https://github.com/drisspg
Fixes#155006
Inductor sometimes codegens triton kernel definitions into a triple-quoted text block. If the text block itself contains triple-quotes, this breaks. Notably, this can happen for user-defined triton kernels, where the user may have added a docstring in their triton kernel.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157322
Approved by: https://github.com/zou3519, https://github.com/drisspg
Based on the [conversation](https://github.com/pytorch/pytorch/issues/121791), we plan to drop the "highest, high, medium" to represent fp32 internal computation data types . Instead, we will directly use the algorithm to represent it.
### Design Choice: Directly use algorithms name like "TF32", "BF16".
#### Pros
- The names are more informative. 'tf32' is more informative than a simple "high".
- Easier to extend new algorithm like `tf32x3`
#### Cons
- "HIGHEST, HIGH, MEDIUM" indicated the relative precision between different algorithms. However, we can have more documents to discuss them.
### We provide a layered structure for backends/operators.
('f32' is short for 'fp32_precision')
### We provide 3 fp32 compute precision can be set:
- **"ieee"**: Not allowed to use any other internal computation data types .
- **"tf32"**: Allowed to use tf32 as internal computation data types.
- **"bf16"**: Allowed to use bf16 as internal computation data types.
- **"none"**: Precision's are not set. Can be override by its father node.
### Overriding Precision Settings
Child node can be override by its father node if it is set to default.
For current default settings:
```
backend = generic, op = all, precision setting = none
backend = cuda, op = all, precision setting = none
backend = cuda, op = conv, precision setting = tf32
backend = cuda, op = rnn, precision setting = tf32
backend = cuda, op = matmul, precision setting = none
backend = matmul, op = all, precision setting = none
backend = matmul, op = conv, precision setting = none
backend = matmul, op = rnn, precision setting = none
backend = matmul, op = matmul, precision setting = none
```
- If the user set `torch.backends.mkldnn.fp32_precision="bf16"`, his child nodes `torch.backends.mkldnn.matmul.fp32_precision` / `torch.backends.mkldnn.conv.fp32_precision` / `torch.backends.mkldnn.rnn.fp32_precision` will also be override to "bf16".
- If the user set `torch.backends.fp32_precision="bf16"`, `torch.backends.mkldnn.fp32_precision` and his child nodes will also we override to "bf16".
### Backward Compatible
Since new API allow user to have more fine-grained control. There will be some conflict. For example, previous `torch.backends.cudnn.allow_tf32` are not enough to represent the status for `torch.backends.cudnn.rnn.fp32_precision="ieee"` and `torch.backends.cudnn.conv.fp32_precision="tf32"`. Therefore, our goal for backward compatible is
- If the user only uses previous APIs, it will work as previous expectations.
- If the user use **new** API to change the status to an **un-representable** status for old API, and try to access the status by **old** API. We will raise Runtime Error and point the document for user.
### Test Plan
```
python test/test_cuda.py -k test_fp32_precision_with_tf32
python test/test_cuda.py -k test_fp32_precision_with_float32_matmul_precision
python test/test_cuda.py -k test_invalid_status_for_legacy_api
python test/test_mkldnn.py -k test_mlkdnn_get_set
python test/test_mkldnn.py -k test_generic_precision
python test/test_mkldnn.py -k test_invalid
python test/test_mkldnn.py -k test_default_use_parent
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125888
Approved by: https://github.com/jgong5, https://github.com/albanD
Co-authored-by: Jiang, Yanbing <yanbing.jiang@intel.com>
Notable new features/optimizations for SDPA operators on AMD systems from AOTriton 0.10b:
* Official support of gfx950/gfx1201
* Experimental support of gfx1101/gfx1151/gfx1150/gfx1200
* Reduce libaotriton.so binary size by over 80%.
+ Without this optimization the binary size of `libaotriton.so` could be
over 100MiB due to 2x more supported architectures compared with 0.9b.
Now it is only about 11MiB.
* Support sliding window attention (SWA) in
`_flash_attention_forward/backward`. Should fix#154582
See https://github.com/ROCm/aotriton/releases/tag/0.10b for full details,
including Known Problems.
Notable changes to SDPA backend:
* `std::optional<int64_t>` `window_size_left/right` are directly passed to
ROCM's SDPA backend, because the default value `-1` is meaningful to
AOTriton's backend and bottom-right aligned causal mask is implemented with
negative `window_size_left/right`
* Some code clean up around `USE_CK_FLASH_ATTENTION`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156499
Approved by: https://github.com/jeffdaily, https://github.com/jithunnair-amd
Usually `world_size=torch.cuda.device_count()` for FSDPTest-based tests
But distributed test class `TestFullyShardAllGatherExtensionsMultiProcess` [forces to use `world_size=2`](0a6e1d6b9b/test/distributed/_composable/fsdp/test_fully_shard_extensions.py (L170)) even for 1 GPU.
Then NCCL fails with errors:
```
HIP_VISIBLE_DEVICES=0 python distributed/_composable/fsdp/test_fully_shard_extensions.py -v -k test_all_gather_extensions_train_parity
...
ncclInvalidUsage: This usually reflects invalid usage of NCCL library.
Duplicate GPU detected : rank 1 and rank 0 both on CUDA device c000
Duplicate GPU detected : rank 0 and rank 1 both on CUDA device c000
```
The test method [has `@skip_if_lt_x_gpu(2)` decorator](0a6e1d6b9b/test/distributed/_composable/fsdp/test_fully_shard_extensions.py (L209)), but test fails during test class initialization before decorator activation
This PR will skip FSDPtest-based tests if `world_size > torch.cuda.device_count()`
```
HIP_VISIBLE_DEVICES=0 python distributed/_composable/fsdp/test_fully_shard_extensions.py -v -k test_all_gather_extensions_train_parity
...
dist init r=0, world=2
dist init r=1, world=2
SKIPPED [15.5507s] (Need at least 2 CUDA devices)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155836
Approved by: https://github.com/jeffdaily
This PR includes the GBID weblink whenever a user encounters a graph break. I also had to include the JSON file in setup.py, so it can be part of the files that are packaged in during CI. It also fixes the issue of the hardcoded error messages stripping away one of the '/' in 'https'.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156033
Approved by: https://github.com/williamwen42
This PR includes the GBID weblink whenever a user encounters a graph break. I also had to include the JSON file in setup.py, so it can be part of the files that are packaged in during CI. It also fixes the issue of the hardcoded error messages stripping away one of the '/' in 'https'.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156033
Approved by: https://github.com/williamwen42
### MOTIVATION
To generalize Distributed test cases for non-CUDA devices
### CHANGES
- test/distributed/optim/test_zero_redundancy_optimizer.py
- test/distributed/test_c10d_logger.py
- test/distributed/test_compute_comm_reordering.py
Replaced hard coded device names with get_devtype from torch.testing._internal.common_fsdp.
DistributedTestBase is used instead of MultiProcessTestCase, to make use of helper functions.
- torch/testing/_internal/common_distributed.py
extended common utility functions
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152471
Approved by: https://github.com/d4l3k
Backwards pass simply iterates over all 8 points current point contributed to, and back propagates them with the respective weights
TODO: Benchmark the performance of similar loop for the forward pas (i.e. compiler should be able to do loop unrolling, so no point of unrolling it by hand)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156373
Approved by: https://github.com/dcci
ghstack dependencies: #156375
Notable new features/optimizations for SDPA operators on AMD systems from AOTriton 0.10b:
* Official support of gfx950/gfx1201
* Experimental support of gfx1101/gfx1151/gfx1150/gfx1200
* Reduce libaotriton.so binary size by over 80%.
+ Without this optimization the binary size of `libaotriton.so` could be
over 100MiB due to 2x more supported architectures compared with 0.9b.
Now it is only about 11MiB.
* Support sliding window attention (SWA) in
`_flash_attention_forward/backward`. Should fix#154582
See https://github.com/ROCm/aotriton/releases/tag/0.10b for full details,
including Known Problems.
Notable changes to SDPA backend:
* `std::optional<int64_t>` `window_size_left/right` are directly passed to
ROCM's SDPA backend, because the default value `-1` is meaningful to
AOTriton's backend and bottom-right aligned causal mask is implemented with
negative `window_size_left/right`
* Some code clean up around `USE_CK_FLASH_ATTENTION`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156290
Approved by: https://github.com/jithunnair-amd, https://github.com/jeffdaily
Which was not caught by CI beforehand, as all 3D examples right now are symmetric, so add an uneven shape to `sample_inputs_interpolate`
Though it's indirectly tested by `test_upsample_nearest3d` inductor test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156375
Approved by: https://github.com/atalman
There are a few considerations here:
1. A user might want to modify the cudaGraph_t either during the stream capture or after the stream capture (but before instantiation). This draft implements modification after stream capture only, though support could be added for modification during stream capture by applying
https://github.com/pytorch/pytorch/pull/140979/files#diff-d7302d133bb5e0890fc94de9aeea4d9d442555a3b40772c9db10edb5cf36a35cR391-R404
2. Previously, the cudaGraph_t would be destroyed before the end of capture_end() unless the user had previously called enable_debug_mode(). There is no way to implement this correctly without removing this restriction, or forcing the user to always call enable_debug_mode(). However, enable_debug_mode() is a confusing API (despite being an instance method, it would modify a static global variable; thus, putting one CUDAGraph object into debug mode puts all of them into debug mode, which is not acceptable in my opinion). Therefore, I made enable_debug_mode() into a no-op. This means that the CPU memory usage will increase after this change. I think this is likely to be fine.
3. No python bindings yet. These should be easy to add. It is probably worthwhile to take some time to make sure that the returned cudaGraph_t can be converted into the cuda-python cudaGraph_t in a reasonable, hopefully type-safe, manner (but without making cuda-python a dependency of pytorch), since I imagine most users will use the pip cuda-python package to make modifications.
4. There are two foot guns:
a. The cudaGraph_t returned by raw_cuda_graph() is not owned by the user, so it will be destroyed once the owning CUDAGraph is destroyed (or calls reset()).
b. The following seuquence won't work as intended:
```
g = torch.cuda.CUDAGraph()
with torch.cuda.graph(g):
foo()
g.replay()
raw_graph = g.raw_cuda_graph()
modify(raw_graph)
g.replay()
```
This won't work because the user must call instantiate() again after modifying cudaGraph_t. You could add a "safety" mechanism by traversing the cudaGraph_t to create a hash and seeing if the hash changes between calls to replay(), but this is likely way too expensive.
I think these two foot guns are probably okay given that this a bit of an experts' API.
Fixes#155106
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155164
Approved by: https://github.com/ngimel
