The cpu inductor processes .to(torch.uint8) incorrectly, leading to numerical inconsistencies. The convert_float_to_int8 function may return incorrect results for negative inputs, such as -2.xx, when the data type is uint8_t, producing 0 instead of 255. This issue stems from the clamping logic; we should avoid converting min_val to uint8_t too early
Fixes https://github.com/pytorch/pytorch/issues/156788
@leslie-fang-intel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157267
Approved by: https://github.com/leslie-fang-intel
Summary: We found there's a special case in recent APS model where the input tensor has smaller size compared to the split size. It will be automatically truncated in split.Tensor thus we add extra condition check for split_with_sizes when do the normalization.
Test Plan:
### unit
```
buck2 test 'fbcode//mode/dev-nosan' fbcode//caffe2/test/inductor:split_cat_fx_aten_passes -- test_split_aten_normalization
```
Buck UI: https://www.internalfb.com/buck2/2ecd1ef8-8efe-4245-b4c8-282c23645b3c
Test UI: https://www.internalfb.com/intern/testinfra/testrun/7599824648585787
Network: Up: 3.9GiB Down: 9.2GiB (reSessionID-1396c91e-0dd2-457b-a49b-a6ab1f2a7d8f)
Loading targets. Remaining 0/5344 99617 dirs read, 1074949 targets declared
Analyzing targets. Remaining 0/123279 4988547 actions, 5966764 artifacts declared
Executing actions. Remaining 0/728058 209:52:59.9s exec time total
Command: test. Finished 12466 local, 209448 remote, 1226 cache (1% hit) 42:10.5s exec time cached (0%)
Time elapsed: 26:07.6s
Tests finished: Pass 2. Fail 0. Fatal 0. Skip 0. Build failure 0
### E2E
before fix:
aps-afoc_apop_pt2_v0-db2fe0449a
after fix:
aps-afoc_apop_pt2_v0-755ad0cdc6
Rollback Plan:
Differential Revision: D77961394
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157857
Approved by: https://github.com/anijain2305
Summary: When `compile_standalone` is True, we set `package_cpp_only` to True as well. We raise an error if `package_cpp_only` is explicitly set to False in config.
Test Plan:
```
buck2 run mode/dev-nosan fbcode//caffe2/test/inductor:test_aot_inductor -- -r TestAOTInductorConfig
```
Rollback Plan:
Differential Revision: D77889754
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157731
Approved by: https://github.com/desertfire
This implements a new `wait_stream` API in Work that matches how `wait` works for ProcessGroupNCCL for CPU based backends such as Gloo.
The idea is to support Gloo communication overlap in FSDPv2/HSDP with minimal changes to FSDP.
There was a previous attempt to make FSDPv2 use Work.wait but given the extensive stream semantics used it doesn't play nicely. https://github.com/pytorch/pytorch/pull/148780
This uses a "Baton" CUDA kernel which spinlocks on a pinned CPU tensor waiting for it to be set.
Test plan:
```
pytest test/distributed/test_c10d_gloo.py -v -k wait_stream
pytest test/distributed/test_c10d_nccl.py -v -k wait_stream
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156883
Approved by: https://github.com/kwen2501, https://github.com/fduwjj
# Feature
If a Triton kernel has a complicated indexing expression, Inductor may decide to precompute it on the host and pass it to the kernel as an argument. This happens in situations like broadcasts with dynamic shapes.
This PR adds support for this feature to Inductor's FX IR backend.
We generate FX IR for precomputed size args in 3 steps:
1. In `PythonWrapperCodegen`, this PR refactors the relevant code to use a `SymbolicCallArgLine` instead of raw Python strings. This stores a (symbol, expr) pair. (Prior to this PR, it was (str, expr), but changing this to a symbol makes it easier to do substitutions later on.)
2. In `WrapperFxCodegen`, keep a dict of {symbol: expr} arg defs which gets updated whenever we see a `SymbolicCallArgLine`.
3. When the FX backend sees a `KernelCallLine`, it uses this dict to replace symbolic call args with their definitions.
In the longer run, it might be desirable to emit FX nodes defining these symbolic call args. That way, we could reuse the size computation when the same kernel is called multiple times. However, I wasn't sure if there was an existing way to generate FX nodes from a sympy expression, and implementing that seemed like overkill for the present purposes.
# Test plan
Added a new CI test exercising this feature.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157758
Approved by: https://github.com/jansel
Summary: These changes in D76442012 got reverted after the PR landed due to aps_models/ads/launchers/pearl/tests/ne/e2e_deterministic_tests:pearl_e2e_ne_tests failing with `Config not loaded due to no timely response from configerator. Likely configerator_proxy or falcon_proxy are not healthy`, but that test failing is definitely transient and unrelated to my changes, so re-creating the diff
Test Plan:
ensure tests pass
Rollback Plan:
Differential Revision: D77871099
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157715
Approved by: https://github.com/meetv18
This is useful for vLLM, which runs AOTAutograd directly on graphs after
they have been split.
I created a new flag for this instead of reusing
`keep_original_node_name` (please let me know if you think I should reuse this).
The reasoning is:
- The names of the placeholder nodes is different from the targets of
the placehoder nodes. The targets are the actual input names.
- Backwards compatibility: this API has been out for ~4 years, it
looks public, and it has extensive public use. For example, this change
would actually be BC-breaking to vLLM (they rely on the subgraph input
names being different at the moment).
Test Plan:
- new tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157733
Approved by: https://github.com/ezyang
Update s390x test marks
test_logs_out from test/dynamo/test_logging.py is updated
and no longer fails on s390x.
test_qengine from test/test_torch.py doesn't work on s390x:
no QEngine is available.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157541
Approved by: https://github.com/huydhn
Target determination sorts the tests in a PR CI run based on heuristics about which tests are more relevant to the PR's changes. This can help provide faster CI signal as well as help alleviate capacity concerns as job durations should decrease due to catching failures earlier.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156545
Approved by: https://github.com/jeffdaily, https://github.com/clee2000
# Motivation
https://github.com/pytorch/pytorch/pull/155451 decoupled `torch._C._storage_Use_Count` from CUDA and introduced a corresponding unit test:
815545f2dd/test/test_torch.py (L257-L262)
However, this test fails when PyTorch is built with debug assertions enabled. @clee2000 disabled this UT in https://github.com/pytorch/pytorch/pull/156731. The root cause is that `_cdata` is obtained from an `intrusive_ptr`, not a `weak_intrusive_ptr`. As a result, calling `c10::weak_intrusive_ptr::use_count` on it triggers the internal assertion:
815545f2dd/c10/util/intrusive_ptr.h (L912-L917)
For example:
```python
a = torch.randn(10, device=device) # refcount=1, weakcount=1
prev_cf = torch._C._storage_Use_Count(a.untyped_storage()._cdata) # violate the assertation
```
This violates the expected invariant inside `weak_intrusive_ptr::use_count`, which assumes the pointer was originally constructed from a valid `weak_intrusive_ptr`. Actually, `storage_impl` is obtained from an `intrusive_ptr`.
815545f2dd/torch/csrc/Module.cpp (L2105-L2109)
# Solution
Use `c10::intrusive_ptr::use_count` instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157694
Approved by: https://github.com/albanD
This PR uses `find_library` to replace `find_path`.
It also searches for NVSHMEM host lib and device lib separately.
Tested against system install location: /usr/local/lib and /usr/local/include.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157695
Approved by: https://github.com/Skylion007
ghstack dependencies: #157513
This PR adds a new config option, `caching_precompile`, and a `DynamoCache`, which loads and saves Dynamo Cache entries automatically. It also hooks up DynamoCache to PrecompileContext, so that we can save multiple cache entries.
When this configuration is turned on, we:
- Automatically create and initialize a CompilePackage on every torch.compile
- Automatically use BundledAutogradcache
- Automatically save the CompilePackage entry to DynamoCache after every compile
You can also use PrecompileContext.serialize() to manually serialize a full object.
I've added unit tests to exhibit this behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155913
Approved by: https://github.com/zhxchen17
This PR uses `find_library` to replace `find_path`.
It also searches for NVSHMEM host lib and device lib separately.
Tested against system install location: /usr/local/lib and /usr/local/include.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157695
Approved by: https://github.com/Skylion007
ghstack dependencies: #157513
Summary: We want to add versioning to DCP to the metadata so that whenever planner logic changes, we can use the version on save to determine how to load the data
Test Plan:
added a test
Rollback Plan:
Differential Revision: D76135887
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155343
Approved by: https://github.com/teja-rao
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
This PR adds a new config option, `caching_precompile`, and a `DynamoCache`, which loads and saves Dynamo Cache entries automatically. It also hooks up DynamoCache to PrecompileContext, so that we can save multiple cache entries.
When this configuration is turned on, we:
- Automatically create and initialize a CompilePackage on every torch.compile
- Automatically use BundledAutogradcache
- Automatically save the CompilePackage entry to DynamoCache after every compile
You can also use PrecompileContext.serialize() to manually serialize a full object.
I've added unit tests to exhibit this behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155913
Approved by: https://github.com/zhxchen17
